feat: Migrate to stable rust

feat: SDKs for Nodejs FFI, Rust, and C (#295 )
* feat: Nodejs FFI package * chore: Update docs for - feat: Nodejs FFI package
2026-03-18 15:42:33 -07:00 · 2026-03-18 15:39:43 -07:00 · 2026-03-17 12:27:48 -07:00 · 2026-03-17 10:40:09 -07:00 · 2026-03-15 08:00:18 -07:00 · 2026-03-15 07:59:38 -07:00
150 changed files with 17725 additions and 4785 deletions
@@ -10,12 +10,7 @@ rustflags = ["-C", "target-feature=-crt-static"]
 [target.aarch64-unknown-linux-musl]
 rustflags = ["-C", "target-feature=-crt-static"]

+# Android/Termux: no hardcoded linker so native Termux builds use the system cc.
+# For CI cross-compilation the linker is set via CARGO_TARGET_AARCH64_LINUX_ANDROID_LINKER env var.
 [target.aarch64-linux-android]
-rustflags = [
-  "-C",
-  "linker=aarch64-linux-android-clang",
-  "-C",
-  "link-args=-rdynamic",
-  "-C",
-  "default-linker-libraries",
-]
+rustflags = ["-C", "link-args=-rdynamic"]
@@ -1,4 +1,4 @@
-name: Lua E2E Tests
+name: e2e Tests

 on:
  push:
@@ -9,11 +9,16 @@ on:
 env:
  CARGO_TERM_COLOR: always
  MACOSX_DEPLOYMENT_TARGET: "13"
+  # Force Node 24 for all JS-based actions to avoid the libuv
+  # process_title assertion crash on Windows (known Node 20 bug).
+  FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true

 jobs:
  lua-tests:
-    name: Lua E2E (${{ matrix.os }})
+    name: e2e (${{ matrix.os }})
    runs-on: ${{ matrix.os }}
+    # e2e tests could be flaky on CI so we do not block release creation if they failed
+    continue-on-error: ${{ github.ref == 'refs/heads/main' && github.event_name == 'push' }}
    strategy:
      fail-fast: false
      matrix:
@@ -23,7 +28,8 @@ jobs:
          - os: windows-latest
            target: x86_64-pc-windows-msvc
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
+      - uses: oven-sh/setup-bun@v2

      - name: Install Zig
        uses: mlugg/setup-zig@v2
@@ -31,17 +37,17 @@ jobs:
          version: 0.15.2

      - name: Install Rust
-        uses: actions-rust-lang/setup-rust-toolchain@v1
+        uses: actions-rust-lang/setup-rust-toolchain@v1.15.4
        with:
          cache: true
-          cache-on-failure: true
-          cache-key: "v1-lua-e2e"
+          cache-on-failure: false
+          cache-key: "v2-lua-e2e"
          rustflags: ""
          target: ${{ matrix.target || '' }}

      - name: Build Rust binary (Windows)
        if: matrix.target
-        run: cargo build --release --target ${{ matrix.target }} -p fff-nvim
+        run: cargo build --release --target ${{ matrix.target }} -p fff-nvim --features zlob

      - name: Copy binary to target/release (Windows)
        if: matrix.target
@@ -68,7 +74,7 @@ jobs:

      - name: Build Rust binary
        if: ${{ !matrix.target }}
-        run: cargo build --release -p fff-nvim
+        run: cargo build --release -p fff-nvim --features zlob

      - name: Install Neovim
        uses: rhysd/action-setup-vim@v1
@@ -82,6 +88,25 @@ jobs:

      - name: Run Lua tests
        shell: bash
-        run: |
-          nvim --headless -u tests/minimal_init.lua \
-            -c "PlenaryBustedFile tests/fff_core_spec.lua" 2>&1
+        run: make test-lua
+
+      - name: Run bun tests
+        shell: bash
+        if: ${{ matrix.os == 'macos-latest' }}
+        run: make test-bun
+
+      - name: Install Node.js
+        if: ${{ matrix.os != 'ubuntu-latest' }}
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+
+      - name: Install node deps
+        if: ${{ matrix.os != 'ubuntu-latest' }}
+        shell: bash
+        run: npm install
+
+      - name: Run node tests
+        if: ${{ matrix.os != 'ubuntu-latest' }}
+        shell: bash
+        run: make test-node
@@ -0,0 +1,50 @@
+name: Lua CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  lua-ls:
+    name: lua-language-server type check
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install Neovim
+        run: |
+          curl -L https://github.com/neovim/neovim/releases/download/v0.11.5/nvim-linux-x86_64.tar.gz -o /opt/nvim.tar.gz
+          mkdir /opt/nvim
+          tar xzf /opt/nvim.tar.gz -C /opt/nvim
+          mv /opt/nvim/nvim-linux-x86_64/* /opt/nvim
+          echo "/opt/nvim/bin" >> $GITHUB_PATH
+
+      - name: Install lua-language-server
+        run: |
+          curl -L "https://github.com/LuaLS/lua-language-server/releases/download/3.17.1/lua-language-server-3.17.1-linux-x64.tar.gz" -o /opt/lls.tar.gz
+          mkdir /opt/lls
+          tar -xzf /opt/lls.tar.gz -C /opt/lls
+          echo "/opt/lls/bin" >> $GITHUB_PATH
+
+      - name: Clone snacks.nvim
+        run: git clone --depth=1 https://github.com/folke/snacks.nvim /opt/snacks.nvim
+
+      - name: Run lua-language-server
+        run: lua-language-server --configpath .luarc.ci.json --check=.
+
+  luacheck:
+    name: luacheck lint
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install luacheck
+        run: |
+          sudo apt-get update -qq
+          sudo apt-get install -y luarocks
+          sudo luarocks install luacheck
+
+      - name: Run luacheck
+        run: luacheck lua/
@@ -13,7 +13,7 @@ jobs:
      id-token: "write"
      contents: "read"
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
      - uses: DeterminateSystems/nix-installer-action@main
      - uses: DeterminateSystems/magic-nix-cache-action@main
      - uses: DeterminateSystems/flake-checker-action@main
@@ -10,7 +10,7 @@ jobs:
    permissions:
      contents: write
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
        with:
          # fetch last 2 commits required for auto force push back
          fetch-depth: 2
@@ -2,9 +2,14 @@ name: Prebuild

 on:
  push:
-    branches: [main, feat/binaries]
+    branches: [main]
+    tags:
+      - "v*"
  pull_request:

+env:
+  FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true
+
 jobs:
  build-nvim:
    name: Build Neovim ${{ matrix.target }}
@@ -36,6 +41,12 @@ jobs:
            artifact_name: target/aarch64-unknown-linux-musl/release/libfff_nvim.so
            ext: so

+          ## Android (Termux)
+          - os: ubuntu-latest
+            target: aarch64-linux-android
+            artifact_name: target/aarch64-linux-android/release/libfff_nvim.so
+            ext: so
+
          ## macOS builds
          - os: macos-latest
            target: x86_64-apple-darwin
@@ -56,7 +67,7 @@ jobs:
            ext: dll

    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
        with:
          persist-credentials: false

@@ -64,7 +75,7 @@ jobs:
        run: rustup target add ${{ matrix.target }}

      - name: Install Zig
-        uses: mlugg/setup-zig@v2
+        uses: goto-bus-stop/setup-zig@v2
        with:
          version: 0.15.2

@@ -73,22 +84,40 @@ jobs:
        run: cargo install cargo-zigbuild

      - name: Build for Linux
-        if: contains(matrix.os, 'ubuntu')
+        if: contains(matrix.os, 'ubuntu') && !contains(matrix.target, 'android')
        run: |
-          cargo zigbuild --release --target ${{ matrix.zigbuild_target || matrix.target }} -p fff-nvim
+          cargo zigbuild --release --target ${{ matrix.zigbuild_target || matrix.target }} -p fff-nvim --features zlob
+          mv "${{ matrix.artifact_name }}" "${{ matrix.target }}.${{ matrix.ext }}"
+
+      - name: Build for Android (Termux)
+        if: contains(matrix.target, 'android')
+        run: |
+          NDK_BIN="$ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/bin"
+
+          # NDK clang for C deps (libgit2, lmdb, blake3) that need Bionic sysroot headers
+          export CC_aarch64_linux_android="$NDK_BIN/aarch64-linux-android24-clang"
+          export CXX_aarch64_linux_android="$NDK_BIN/aarch64-linux-android24-clang++"
+          export AR_aarch64_linux_android="$NDK_BIN/llvm-ar"
+          export CARGO_TARGET_AARCH64_LINUX_ANDROID_LINKER="$NDK_BIN/aarch64-linux-android24-clang"
+
+          cargo build --release --target ${{ matrix.target }} -p fff-nvim --features zlob
          mv "${{ matrix.artifact_name }}" "${{ matrix.target }}.${{ matrix.ext }}"

      - name: Build for macOS
        if: contains(matrix.os, 'macos')
        run: |
-          MACOSX_DEPLOYMENT_TARGET="13" cargo build --release --target ${{ matrix.target }} -p fff-nvim
+          MACOSX_DEPLOYMENT_TARGET="13" cargo build --release --target ${{ matrix.target }} -p fff-nvim --features zlob
          mv "${{ matrix.artifact_name }}" "${{ matrix.target }}.${{ matrix.ext }}"

+      - name: Ad-hoc sign macOS binary
+        if: contains(matrix.os, 'macos')
+        run: codesign --force --sign - "${{ matrix.target }}.${{ matrix.ext }}"
+
      - name: Build for Windows
        if: contains(matrix.os, 'windows')
        shell: bash
        run: |
-          cargo build --release --target ${{ matrix.target }} -p fff-nvim
+          cargo build --release --target ${{ matrix.target }} -p fff-nvim --features zlob
          mv "${{ matrix.artifact_name }}" "${{ matrix.target }}.${{ matrix.ext }}"

      - name: Upload artifacts
@@ -110,26 +139,33 @@ jobs:
            target: x86_64-unknown-linux-gnu
            zigbuild_target: x86_64-unknown-linux-gnu.2.17
            artifact_name: target/x86_64-unknown-linux-gnu/release/libfff_c.so
-            npm_package: fff-bun-linux-x64-gnu
+            npm_package: fff-bin-linux-x64-gnu
            lib_filename: libfff_c.so
            ext: so
          - os: ubuntu-latest
            target: aarch64-unknown-linux-gnu
            zigbuild_target: aarch64-unknown-linux-gnu.2.17
            artifact_name: target/aarch64-unknown-linux-gnu/release/libfff_c.so
-            npm_package: fff-bun-linux-arm64-gnu
+            npm_package: fff-bin-linux-arm64-gnu
            lib_filename: libfff_c.so
            ext: so
          - os: ubuntu-latest
            target: x86_64-unknown-linux-musl
            artifact_name: target/x86_64-unknown-linux-musl/release/libfff_c.so
-            npm_package: fff-bun-linux-x64-musl
+            npm_package: fff-bin-linux-x64-musl
            lib_filename: libfff_c.so
            ext: so
          - os: ubuntu-latest
            target: aarch64-unknown-linux-musl
            artifact_name: target/aarch64-unknown-linux-musl/release/libfff_c.so
-            npm_package: fff-bun-linux-arm64-musl
+            npm_package: fff-bin-linux-arm64-musl
+            lib_filename: libfff_c.so
+            ext: so
+
+          ## Android (Termux)
+          - os: ubuntu-latest
+            target: aarch64-linux-android
+            artifact_name: target/aarch64-linux-android/release/libfff_c.so
            lib_filename: libfff_c.so
            ext: so

@@ -137,13 +173,13 @@ jobs:
          - os: macos-latest
            target: x86_64-apple-darwin
            artifact_name: target/x86_64-apple-darwin/release/libfff_c.dylib
-            npm_package: fff-bun-darwin-x64
+            npm_package: fff-bin-darwin-x64
            lib_filename: libfff_c.dylib
            ext: dylib
          - os: macos-latest
            target: aarch64-apple-darwin
            artifact_name: target/aarch64-apple-darwin/release/libfff_c.dylib
-            npm_package: fff-bun-darwin-arm64
+            npm_package: fff-bin-darwin-arm64
            lib_filename: libfff_c.dylib
            ext: dylib

@@ -151,18 +187,18 @@ jobs:
          - os: windows-latest
            target: x86_64-pc-windows-msvc
            artifact_name: target/x86_64-pc-windows-msvc/release/fff_c.dll
-            npm_package: fff-bun-win32-x64
+            npm_package: fff-bin-win32-x64
            lib_filename: fff_c.dll
            ext: dll
          - os: windows-latest
            target: aarch64-pc-windows-msvc
            artifact_name: target/aarch64-pc-windows-msvc/release/fff_c.dll
-            npm_package: fff-bun-win32-arm64
+            npm_package: fff-bin-win32-arm64
            lib_filename: fff_c.dll
            ext: dll

    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
        with:
          persist-credentials: false

@@ -170,7 +206,7 @@ jobs:
        run: rustup target add ${{ matrix.target }}

      - name: Install Zig
-        uses: mlugg/setup-zig@v2
+        uses: goto-bus-stop/setup-zig@v2
        with:
          version: 0.15.2

@@ -179,25 +215,43 @@ jobs:
        run: cargo install cargo-zigbuild

      - name: Build for Linux
-        if: contains(matrix.os, 'ubuntu')
+        if: contains(matrix.os, 'ubuntu') && !contains(matrix.target, 'android')
        run: |
-          cargo zigbuild --release --target ${{ matrix.zigbuild_target || matrix.target }} -p fff-c
+          cargo zigbuild --release --target ${{ matrix.zigbuild_target || matrix.target }} -p fff-c --features zlob
+          mv "${{ matrix.artifact_name }}" "c-lib-${{ matrix.target }}.${{ matrix.ext }}"
+
+      - name: Build for Android (Termux)
+        if: contains(matrix.target, 'android')
+        run: |
+          NDK_BIN="$ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/bin"
+
+          export CC_aarch64_linux_android="$NDK_BIN/aarch64-linux-android24-clang"
+          export CXX_aarch64_linux_android="$NDK_BIN/aarch64-linux-android24-clang++"
+          export AR_aarch64_linux_android="$NDK_BIN/llvm-ar"
+          export CARGO_TARGET_AARCH64_LINUX_ANDROID_LINKER="$NDK_BIN/aarch64-linux-android24-clang"
+
+          cargo build --release --target ${{ matrix.target }} -p fff-c --features zlob
          mv "${{ matrix.artifact_name }}" "c-lib-${{ matrix.target }}.${{ matrix.ext }}"

      - name: Build for macOS
        if: contains(matrix.os, 'macos')
        run: |
-          MACOSX_DEPLOYMENT_TARGET="13" cargo build --release --target ${{ matrix.target }} -p fff-c
+          MACOSX_DEPLOYMENT_TARGET="13" cargo build --release --target ${{ matrix.target }} -p fff-c --features zlob
          mv "${{ matrix.artifact_name }}" "c-lib-${{ matrix.target }}.${{ matrix.ext }}"

+      - name: Ad-hoc sign macOS binary
+        if: contains(matrix.os, 'macos')
+        run: codesign --force --sign - "c-lib-${{ matrix.target }}.${{ matrix.ext }}"
+
      - name: Build for Windows
        if: contains(matrix.os, 'windows')
        shell: bash
        run: |
-          cargo build --release --target ${{ matrix.target }} -p fff-c
+          cargo build --release --target ${{ matrix.target }} -p fff-c --features zlob
          mv "${{ matrix.artifact_name }}" "c-lib-${{ matrix.target }}.${{ matrix.ext }}"

      - name: Prepare npm package
+        if: "!contains(matrix.target, 'android')"
        shell: bash
        run: |
          # Copy the built binary into the platform npm package directory
@@ -210,19 +264,108 @@ jobs:
          path: c-lib-${{ matrix.target }}.*

      - name: Upload npm package artifact
+        if: "!contains(matrix.target, 'android')"
        uses: actions/upload-artifact@v4
        with:
          name: npm-${{ matrix.npm_package }}
          path: packages/${{ matrix.npm_package }}/

+  build-mcp:
+    name: Build MCP ${{ matrix.target }}
+    runs-on: ${{ matrix.os }}
+    permissions:
+      contents: read
+    strategy:
+      matrix:
+        include:
+          ## Linux builds (using cargo-zigbuild)
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-gnu
+            zigbuild_target: x86_64-unknown-linux-gnu.2.17
+            artifact_name: target/x86_64-unknown-linux-gnu/release/fff-mcp
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-gnu
+            zigbuild_target: aarch64-unknown-linux-gnu.2.17
+            artifact_name: target/aarch64-unknown-linux-gnu/release/fff-mcp
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-musl
+            artifact_name: target/x86_64-unknown-linux-musl/release/fff-mcp
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-musl
+            artifact_name: target/aarch64-unknown-linux-musl/release/fff-mcp
+
+          ## macOS builds
+          - os: macos-latest
+            target: x86_64-apple-darwin
+            artifact_name: target/x86_64-apple-darwin/release/fff-mcp
+          - os: macos-latest
+            target: aarch64-apple-darwin
+            artifact_name: target/aarch64-apple-darwin/release/fff-mcp
+
+          ## Windows builds
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+            artifact_name: target/x86_64-pc-windows-msvc/release/fff-mcp.exe
+          - os: windows-latest
+            target: aarch64-pc-windows-msvc
+            artifact_name: target/aarch64-pc-windows-msvc/release/fff-mcp.exe
+
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          persist-credentials: false
+
+      - name: Install Rust
+        run: rustup target add ${{ matrix.target }}
+
+      - name: Install Zig
+        uses: goto-bus-stop/setup-zig@v2
+        with:
+          version: 0.15.2
+
+      - name: Install cargo-zigbuild
+        if: contains(matrix.os, 'ubuntu')
+        run: cargo install cargo-zigbuild
+
+      - name: Build for Linux
+        if: contains(matrix.os, 'ubuntu')
+        run: |
+          cargo zigbuild --release --target ${{ matrix.zigbuild_target || matrix.target }} -p fff-mcp --features zlob
+          cp "${{ matrix.artifact_name }}" "fff-mcp-${{ matrix.target }}"
+
+      - name: Build for macOS
+        if: contains(matrix.os, 'macos')
+        run: |
+          MACOSX_DEPLOYMENT_TARGET="13" cargo build --release --target ${{ matrix.target }} -p fff-mcp --features zlob
+          cp "${{ matrix.artifact_name }}" "fff-mcp-${{ matrix.target }}"
+
+      - name: Ad-hoc sign macOS binary
+        if: contains(matrix.os, 'macos')
+        run: codesign --force --sign - "fff-mcp-${{ matrix.target }}"
+
+      - name: Build for Windows
+        if: contains(matrix.os, 'windows')
+        shell: bash
+        run: |
+          cargo build --release --target ${{ matrix.target }} -p fff-mcp --features zlob
+          cp "${{ matrix.artifact_name }}" "fff-mcp-${{ matrix.target }}.exe"
+
+      - name: Upload artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: mcp-${{ matrix.target }}
+          path: fff-mcp-${{ matrix.target }}*
+
  release:
    name: Release
-    needs: [build-nvim, build-c]
+    needs: [build-nvim, build-c, build-mcp]
    runs-on: ubuntu-latest
+    # do not create releases on the forks (no permissions)
+    if: github.event_name != 'pull_request' || github.event.pull_request.user.login == 'dmtrKovalenko'
    permissions:
      contents: write
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Download artifacts
        uses: actions/download-artifact@v4
@@ -259,6 +402,19 @@ jobs:
            rmdir "$dir" 2>/dev/null || true
          done

+      - name: Flatten MCP artifacts
+        working-directory: ./binaries
+        run: |
+          for dir in mcp-*/; do
+            for file in "$dir"*; do
+              if [ -f "$file" ]; then
+                filename=$(basename "$file")
+                mv "$file" "./$filename"
+              fi
+            done
+            rmdir "$dir" 2>/dev/null || true
+          done
+
      - name: Remove npm package artifacts from release binaries
        working-directory: ./binaries
        run: |
@@ -274,25 +430,22 @@ jobs:
            fi
          done

-      - name: Prepare tag
-        id: vars
-        shell: bash
-        run: |
-          sha="$(git rev-parse --short HEAD)"
-          echo "tag=$sha" >> $GITHUB_OUTPUT
+      - name: Determine version
+        id: version
+        run: ./scripts/determine-version.sh

      - name: Upload Release Assets
        uses: softprops/action-gh-release@v2
        with:
-          name: "${{ steps.vars.outputs.tag }}"
-          tag_name: "${{ steps.vars.outputs.tag }}"
+          name: "${{ steps.version.outputs.version }}"
+          tag_name: "${{ steps.version.outputs.is_release == 'true' && format('v{0}', steps.version.outputs.version) || steps.version.outputs.version }}"
          token: ${{ github.token }}
          files: ./binaries/*
          draft: false
-          prerelease: true
-          generate_release_notes: false
+          prerelease: ${{ steps.version.outputs.is_release != 'true' }}
+          generate_release_notes: ${{ steps.version.outputs.is_release == 'true' }}
          body: |
-            Nightly release from commit: ${{ github.sha }}
+            ${{ steps.version.outputs.is_release == 'true' && format('Release {0}', steps.version.outputs.version) || format('Nightly release from commit: {0}', github.sha) }}

            ## Neovim Plugin
            - `{target}.so` / `.dylib` / `.dll` - Lua module for Neovim
@@ -300,39 +453,60 @@ jobs:
            ## C FFI Library (for Bun/Node/Python)
            - `c-lib-{target}.so` / `.dylib` / `.dll` - C FFI library

+            ## MCP Server
+            - `fff-mcp-{target}` - MCP server binary
+
+            Install with:
+            ```sh
+            curl -fsSL https://raw.githubusercontent.com/dmtrKovalenko/fff.nvim/main/install-mcp.sh | bash
+            ```
+
+  crates-publish:
+    name: Publish Rust crates
+    needs: [build-nvim, build-c, build-mcp]
+    runs-on: ubuntu-latest
+    if: >-
+      (github.event_name == 'push' && (github.ref == 'refs/heads/main' || github.ref == 'refs/heads/node' || startsWith(github.ref, 'refs/tags/v')))
+      || (github.event_name == 'pull_request' && (github.head_ref == 'main' || github.head_ref == 'node'))
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install Rust
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Install cargo-edit
+        run: cargo install cargo-edit
+
+      - name: Determine version
+        id: version
+        run: ./scripts/determine-version.sh
+
+      - name: Publish crates
+        env:
+          CARGO_REGISTRY_TOKEN: ${{ secrets.CARGO_REGISTRY_TOKEN }}
+        run: make publish-crates V="${{ steps.version.outputs.version }}"
+
  npm-publish:
    name: Publish npm packages
    needs: [build-c]
    runs-on: ubuntu-latest
    if: >-
-      (github.event_name == 'push' && (github.ref == 'refs/heads/main' || github.ref == 'refs/heads/feat/binaries'))
-      || (github.event_name == 'pull_request' && (github.head_ref == 'main' || github.head_ref == 'feat/binaries'))
+      (github.event_name == 'push' && (github.ref == 'refs/heads/main' || github.ref == 'refs/heads/node' || startsWith(github.ref, 'refs/tags/v')))
+      || (github.event_name == 'pull_request' && (github.head_ref == 'main' || github.head_ref == 'node'))
    permissions:
      contents: read
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Setup Node.js
        uses: actions/setup-node@v4
        with:
-          node-version: "20"
+          node-version: "25"
          registry-url: "https://registry.npmjs.org"

      - name: Determine version
        id: version
-        run: |
-          # Read the base version from fff-core Cargo.toml (single source of truth)
-          base_version=$(grep '^version' crates/fff-core/Cargo.toml | head -1 | sed 's/version = "\(.*\)"/\1/')
-          short_sha=$(git rev-parse --short HEAD)
-          
-          # Always publish as nightly prerelease: X.Y.Z-nightly.<short-sha>
-          echo "version=${base_version}-nightly.${short_sha}" >> $GITHUB_OUTPUT
-          
-          if [ "${{ github.ref }}" = "refs/heads/main" ]; then
-            echo "tag=nightly" >> $GITHUB_OUTPUT
-          else
-            echo "tag=dev" >> $GITHUB_OUTPUT
-          fi
+        run: ./scripts/determine-version.sh

      - name: Download npm package artifacts
        uses: actions/download-artifact@v4
@@ -345,93 +519,45 @@ jobs:
          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
        run: |
          VERSION="${{ steps.version.outputs.version }}"
-          TAG="${{ steps.version.outputs.tag }}"
-          
+          TAG="${{ steps.version.outputs.npm_tag }}"
+
          for pkg_dir in ./npm-packages/npm-*/; do
            if [ -d "$pkg_dir" ]; then
              pkg_name=$(node -p "require('${pkg_dir}package.json').name")
              echo "Publishing ${pkg_name}@${VERSION} with tag ${TAG}..."
-              
-              # Update version in package.json
-              node -e "
-                const fs = require('fs');
-                const pkg = JSON.parse(fs.readFileSync('${pkg_dir}package.json', 'utf8'));
-                pkg.version = '${VERSION}';
-                fs.writeFileSync('${pkg_dir}package.json', JSON.stringify(pkg, null, 2) + '\n');
-              "
-              
+
+              make set-npm-version PKG="$pkg_dir" VERSION="$VERSION"
+
              cd "$pkg_dir"
              npm publish --tag "$TAG" --access public || echo "Failed to publish ${pkg_name} (may already exist)"
              cd -
            fi
          done

-      - name: Publish main package
+      - name: Publish bun package
        env:
          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
        run: |
          VERSION="${{ steps.version.outputs.version }}"
-          TAG="${{ steps.version.outputs.tag }}"
-          
+          TAG="${{ steps.version.outputs.npm_tag }}"
+
          echo "Publishing @ff-labs/fff-bun@${VERSION} with tag ${TAG}..."
-          
-          # Update version and optionalDependencies versions in the main package
-          node -e "
-            const fs = require('fs');
-            const pkg = JSON.parse(fs.readFileSync('./packages/fff-bun/package.json', 'utf8'));
-            pkg.version = '${VERSION}';
-            if (pkg.optionalDependencies) {
-              for (const dep of Object.keys(pkg.optionalDependencies)) {
-                pkg.optionalDependencies[dep] = '${VERSION}';
-              }
-            }
-            fs.writeFileSync('./packages/fff-bun/package.json', JSON.stringify(pkg, null, 2) + '\n');
-          "
-          
+          make set-npm-version PKG=packages/fff-bun VERSION="$VERSION"
+
          cd packages/fff-bun
          npm publish --tag "$TAG" --access public || echo "Failed to publish @ff-labs/fff-bun (may already exist)"

-  comment-on-pr:
-    name: Comment on PR
-    needs: [build-nvim, build-c]
-    runs-on: ubuntu-latest
-    if: github.event_name == 'pull_request'
-    permissions:
-      pull-requests: write
-    steps:
-      - name: Get short SHA
-        id: vars
-        run: echo "short_sha=${GITHUB_SHA::7}" >> $GITHUB_OUTPUT
+      - name: Publish Node.js package
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+        run: |
+          VERSION="${{ steps.version.outputs.version }}"
+          TAG="${{ steps.version.outputs.npm_tag }}"

-      - name: Find existing comment
-        uses: peter-evans/find-comment@v3
-        id: find-comment
-        with:
-          issue-number: ${{ github.event.pull_request.number }}
-          comment-author: "github-actions[bot]"
-          body-includes: "<!-- fff-nvim-build-comment -->"
+          echo "Publishing @ff-labs/fff-node@${VERSION} with tag ${TAG}..."
+          make set-npm-version PKG=packages/fff-node VERSION="$VERSION"

-      - name: Create or update PR comment
-        uses: peter-evans/create-or-update-comment@v4
-        with:
-          comment-id: ${{ steps.find-comment.outputs.comment-id }}
-          issue-number: ${{ github.event.pull_request.number }}
-          edit-mode: replace
-          body: |
-            <!-- fff-nvim-build-comment -->
-            ## Build Artifacts for your PR
-
-            ### Neovim Plugin
-            Test with lazy.nvim:
-            ```lua
-            {
-              "dmtrKovalenko/fff.nvim",
-              tag = "${{ steps.vars.outputs.short_sha }}",
-            }
-            ```
-
-            ### Bun/TypeScript Package
-            The `fff` npm package will download binaries from this release automatically.
-
-            ---
-            *Built from ${{ github.sha }}*
+          cd packages/fff-node
+          npm install
+          npx tsc
+          npm publish --tag "$TAG" --access public || echo "Failed to publish @ff-labs/fff-node (may already exist)"
@@ -20,16 +20,16 @@ jobs:
      matrix:
        os: [ubuntu-latest, macos-latest]
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
      
      # Zig is required to compile zlob
      - name: Install Zig
-        uses: mlugg/setup-zig@v2
+        uses: goto-bus-stop/setup-zig@v2
        with:
          version: 0.15.2

      - name: Install Rust
-        uses: actions-rust-lang/setup-rust-toolchain@v1
+        uses: actions-rust-lang/setup-rust-toolchain@v1.15.4
        with:
          cache: true
          cache-on-failure: true
@@ -37,13 +37,13 @@ jobs:
          components: rustfmt, clippy

      - name: Run tests
-        run: cargo test --verbose --workspace --exclude fff-nvim
+        run: cargo test --features zlob --workspace --exclude fff-nvim

  fmt:
    name: cargo fmt
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
      - name: Install Rust
        uses: dtolnay/rust-toolchain@master
        with:
@@ -56,11 +56,11 @@ jobs:
    name: cargo clippy
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
      
      # Zig is required to compile zlob
      - name: Install Zig
-        uses: mlugg/setup-zig@v2
+        uses: goto-bus-stop/setup-zig@v2
        with:
          version: 0.15.2

@@ -17,7 +17,7 @@ jobs:
    name: Spell Check with Typos
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
        with:
          persist-credentials: false

@@ -25,7 +25,7 @@ jobs:
    name: Check lua files using Stylua
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
        with:
          persist-credentials: false

@@ -13,3 +13,12 @@ result
 big-repo/
 # all the perf like utility files
 *.data
+node_modules/
+
+dist/
+scripts/benchmark-results/
+
+# Native binaries (downloaded at install)
+*.dylib
+*.so
+*.dll
@@ -0,0 +1,25 @@
+-- luacheck configuration for fff.nvim
+-- https://luacheck.readthedocs.io/en/stable/config.html
+
+-- Neovim globals
+globals = { "vim" }
+
+-- Standard library
+std = "luajit"
+
+-- Ignore line length (handled by stylua)
+max_line_length = false
+
+-- Ignore unused self argument in methods
+self = false
+
+-- Files/directories to ignore
+exclude_files = {
+  ".luarocks/",
+}
+
+-- Warn about unused variables, but allow _ prefix convention
+unused_args = true
+ignore = {
+  "212", -- unused argument (too noisy for callback-heavy code)
+}
@@ -0,0 +1,44 @@
+{
+  "$schema": "https://raw.githubusercontent.com/LuaLS/vscode-lua/master/setting/schema.json",
+  "runtime": {
+    "version": "LuaJIT",
+    "pathStrict": true
+  },
+  "workspace": {
+    "library": [
+      "/opt/nvim/share/nvim/runtime/lua/vim/_meta",
+      "/opt/nvim/share/nvim/runtime/lua/vim/shared.lua",
+      "${3rd}/luv/library",
+      "${3rd}/busted/library",
+      "/opt/snacks.nvim/lua"
+    ],
+    "checkThirdParty": false
+  },
+  "diagnostics": {
+    "severity": {
+      "undefined-global": "Error",
+      "undefined-field": "Warning",
+      "missing-return": "Warning",
+      "redundant-parameter": "Warning",
+      "param-type-mismatch": "Warning",
+      "assign-type-mismatch": "Warning",
+      "cast-type-mismatch": "Warning",
+      "deprecated": "Warning",
+      "undefined-doc-param": "Warning"
+    },
+    "neededFileStatus": {
+      "undefined-global": "Any",
+      "undefined-field": "Any",
+      "missing-return": "Any",
+      "redundant-parameter": "Any",
+      "param-type-mismatch": "Any",
+      "assign-type-mismatch": "Any",
+      "cast-type-mismatch": "Any",
+      "deprecated": "Any",
+      "undefined-doc-param": "Any"
+    }
+  },
+  "type": {
+    "checkTableShape": true
+  }
+}
@@ -0,0 +1,42 @@
+{
+  "$schema": "https://raw.githubusercontent.com/LuaLS/vscode-lua/master/setting/schema.json",
+  "runtime": {
+    "version": "LuaJIT"
+  },
+  "workspace": {
+    "library": [
+      "/opt/homebrew/share/nvim/runtime/lua/vim/_meta",
+      "/opt/homebrew/share/nvim/runtime/lua/vim/shared.lua",
+      "${3rd}/luv/library",
+      "${3rd}/busted/library"
+    ],
+    "checkThirdParty": false
+  },
+  "diagnostics": {
+    "severity": {
+      "undefined-global": "Error",
+      "undefined-field": "Warning",
+      "missing-return": "Warning",
+      "redundant-parameter": "Warning",
+      "param-type-mismatch": "Warning",
+      "assign-type-mismatch": "Warning",
+      "cast-type-mismatch": "Warning",
+      "deprecated": "Warning",
+      "undefined-doc-param": "Warning"
+    },
+    "neededFileStatus": {
+      "undefined-global": "Any",
+      "undefined-field": "Any",
+      "missing-return": "Any",
+      "redundant-parameter": "Any",
+      "param-type-mismatch": "Any",
+      "assign-type-mismatch": "Any",
+      "cast-type-mismatch": "Any",
+      "deprecated": "Any",
+      "undefined-doc-param": "Any"
+    }
+  },
+  "type": {
+    "checkTableShape": true
+  }
+}
@@ -0,0 +1,9 @@
+{
+  "mcpServers": {
+    "fff": {
+      "type": "stdio",
+      "command": "/Users/neogoose/dev/fff.nvim/target/release/fff-mcp",
+      "args": []
+    }
+  }
+}
@@ -2,13 +2,17 @@
 members = [
  "crates/fff-c",
  "crates/fff-core",
+  "crates/fff-mcp",
  "crates/fff-nvim",
  "crates/fff-query-parser",
-  "crates/fff-searcher",
+  "crates/fff-grep",
 ]
 resolver = "2"

 [workspace.dependencies]
+fff-grep = { version = "0.2.3", path = "crates/fff-grep" }
+fff-query-parser = { version = "0.2.3", path = "crates/fff-query-parser", default-features = false }
+
 # Shared dependencies
 ahash = "0.8"
 bindet = "0.3"
@@ -22,16 +26,16 @@ git2 = { version = "0.20.2", default-features = false, features = [
  "vendored-libgit2",
 ] }
 glidesort = "0.1"
+globset = "0.4"
 grep-matcher = "0.1.8"
-grep-searcher = { path = "crates/fff-searcher" }
 heed = "0.22.0"
 ignore = "0.4.22"
 memmap2 = "0.9"
 mimalloc = "0.1.47"
-zlob = "1.2.9"
+zlob = "1.3.0"

 mlua = { version = "0.11.1", features = ["module", "luajit"] }
-neo_frizbee = "0.8.1"
+neo_frizbee = "0.8.2"
 notify = "8.1.0"
 notify-debouncer-full = "0.7"
 once_cell = "1.20.2"
@@ -1,9 +1,12 @@
 PLENARY_DIR ?= ../plenary.nvim

-.PHONY: build test test-rust test-lua test-setup
+.PHONY: build test test-rust test-lua test-bun test-node prepare-bun prepare-node set-npm-version header

 build:
-	cargo build --release
+	cargo build --release --features zlob
+
+header:
+	cbindgen --config crates/fff-c/cbindgen.toml --crate fff-c --output crates/fff-c/include/fff.h

 test-setup:
 	@if [ ! -d "$(PLENARY_DIR)" ]; then \
@@ -12,17 +15,78 @@ test-setup:
 	fi

 test-rust:
-	cargo test --verbose --workspace --exclude fff-nvim
+	cargo test --workspace --features zlob --exclude fff-nvim

 test-lua: test-setup build
 	nvim --headless -u tests/minimal_init.lua \
 		-c "PlenaryBustedFile tests/fff_core_spec.lua" 2>&1

-test: test-rust test-lua
+prepare-bun: build
+	mkdir -p packages/fff-bun/bin
+	cp target/release/libfff_c.dylib packages/fff-bun/bin/ 2>/dev/null; \
+	cp target/release/libfff_c.so packages/fff-bun/bin/ 2>/dev/null; \
+	cp target/release/fff_c.dll packages/fff-bun/bin/ 2>/dev/null; \
+	true
+
+prepare-node: build
+	mkdir -p packages/fff-node/bin
+	cp target/release/libfff_c.dylib packages/fff-node/bin/ 2>/dev/null; \
+	cp target/release/libfff_c.so packages/fff-node/bin/ 2>/dev/null; \
+	cp target/release/fff_c.dll packages/fff-node/bin/ 2>/dev/null; \
+	true
+
+test-bun: prepare-bun
+	cd packages/fff-bun && bun test src/
+
+test-node: prepare-node
+	cd packages/fff-node && npm run build && node test/e2e.mjs
+
+test: test-rust test-lua test-bun test-node
+
+# Update version in a package.json, including optionalDependencies.
+# Usage: make set-npm-version PKG=packages/fff-bun VERSION=1.0.0-nightly.abc1234
+set-npm-version:
+	@test -n "$(PKG)" || (echo "PKG is required" && exit 1)
+	@test -n "$(VERSION)" || (echo "VERSION is required" && exit 1)
+	node -e " \
+		const fs = require('fs'); \
+		const pkg = JSON.parse(fs.readFileSync('$(PKG)/package.json', 'utf8')); \
+		pkg.version = '$(VERSION)'; \
+		if (pkg.optionalDependencies) { \
+			for (const dep of Object.keys(pkg.optionalDependencies)) { \
+				pkg.optionalDependencies[dep] = '$(VERSION)'; \
+			} \
+		} \
+		fs.writeFileSync('$(PKG)/package.json', JSON.stringify(pkg, null, 2) + '\n'); \
+	"
+	@echo "Set $(PKG) to $(VERSION)"

 format-rust:
 	cargo fmt --all
 format-lua:
 	stylua .
+format-ts:
+	bun format

-format: format-rust format-lua
+format: format-rust format-lua format-ts
+
+lint-rust:
+	cargo clippy --workspace --features zlob -- -D warnings
+lint-lua:
+	 ~/.luarocks/bin/luacheck .
+lint-ts:
+	bun lint
+
+lint: lint-rust lint-lua lint-ts
+
+check: format lint
+
+CRATES_TO_PUBLISH= fff-grep fff-query-parser fff-search
+
+publish-crates:
+	@test -n "$(V)" || (echo "V is required. Usage: make publish-crates V=0.2.0" && exit 1)
+	cargo install cargo-edit
+	cargo set-version $(V) || exit 1;
+	@for crate in $(CRATES_TO_PUBLISH); do \
+		cargo publish -p $$crate --allow-dirty $$(if [ -n "$$CI" ]; then echo "--no-verify"; fi) || exit 1; \
+	done
@@ -1,9 +1,13 @@
 <p align="center">
-  <h2 align="center">FFF.nvim</h2>
+  <h1 align="center">FFF</h1>
 </p>

 <p align="center">
-	Finally a smart fuzzy file picker for neovim.
+  <a href="#mcp"><strong>AI agents (MCP)</strong></a> &nbsp;&nbsp;|&nbsp;&nbsp; <a href="#neovim-guide"><strong>Neovim users</strong></a>
+</p>
+
+<p align="center">
+  <i>A fast file search for your AI and neovim, with memory built-in</i>
 </p>

 <p align="center" style="text-decoration: none; border: none;">
@@ -14,35 +18,44 @@
 	<a href="https://github.com/dmtrKovalenko/fff.nvim/contributors" style="text-decoration: none"> <img alt="Contributors" src="https://img.shields.io/github/contributors/dmtrKovalenko/fff.nvim?color=%23DDB6F2&label=CONTRIBUTORS&logo=git&style=for-the-badge&logoColor=D9E0EE&labelColor=302D41"/></a>
 </p>

-**FFF** stands for ~freakin fast fuzzy file finder~ (pick 3) and it is an opinionated fuzzy file picker for neovim. Just for files, but we'll try to solve file picking completely.
+---

-It comes with a dedicated rust backend runtime that keep tracks of the file index, your file access and modifications, git status, and provides a comprehensive typo-resistant fuzzy search experience.
+**FFF** stands for ~~freakin fast fuzzy file finder~~ (pick 3) and it is an opinionated fuzzy file picker for your AI agent and Neovim. Just for file search, but we do the file search really fff well.

-## Features
+FFF is a tool for grepping, fuzzy file matching, globbing, and multigrepping with a strong focus on performance and useful search results. For humans - provides an unbelievable typo-resistant experience, for AI agents - implements the fastest file search with additional free memory suggesting the best search results based on various factors like frecency, git status, file size, definition matches, and more.

- Works out of the box with no additional configuration
- [Typo resistant fuzzy search](https://github.com/saghen/frizbee)
- Git status integration allowing to take advantage of last modified times within a worktree
- Separate file index maintained by a dedicated backend allows <10 milliseconds search time for 50k files codebase
- Display images in previews (for now requires snacks.nvim)
- Smart in a plenty of different ways hopefully helpful for your workflow
- This plugin initializes itself lazily by default
+## MCP

-## Installation
+FFF is an amazing way to reduce the time and tokens by giving your AI agent a bit of memory built-in to their file search tools. It makes your AI harness to find the code faster and spend less tokens by doing less roundtrips and reading less useless files.

-> [!NOTE]
-> Although we'll try to make sure to keep 100% backward compatibility, by using you should understand that silly bugs and breaking changes may happen.
-> And also we hope for your contributions and feedback to make this plugin ideal for everyone.
+![Chart showing the superiority of fff.nvim over builtin claude code tools](./chart.png)

-### Prerequisites
+You can install FFF as a dependency for your AI agent using a simple bash script:

-FFF.nvim requires:
+```bash
+curl -L https://dmtrkovalenko.dev/install-fff-mcp.sh | bash
+```

- Neovim 0.10.0+
- [Rustup](https://rustup.rs/) (we require nightly for building the native backend rustup will handle toolchain automatically)
+> The installation script is here [./install-mcp.sh](./install-mcp.sh) if you want to review it before running.
+
+It will print out the instructions on how to connect it to your `Claude Code`, `Codex`, `OpenCode`, etc. Once you have it connected just ask your agent to "use fff".
+Here is an example addition to `CLAUDE.md` that works perfectly:
+
+```sh
+# CLAUDE.md
+For any file search or grep in the current git indexed directory use fff tools
+```
+
+## Neovim guide
+
+Here is some demo on the linux repository (100k files, 8GB) but you better fill it yourself and see the magic
+
+https://github.com/user-attachments/assets/5d0e1ce9-642c-4c44-aa88-01b05bb86abb

 ### Installation

+FFF.nvim requires neovim 0.10.0 or higher
+
 #### lazy.nvim

 ```lua
@@ -78,12 +91,17 @@ FFF.nvim requires:
    {
      "fz",
      function() require('fff').live_grep({
-	grep = {
-	  modes = { 'fuzzy', 'plain' }
-	}
+        grep = {
+          modes = { 'fuzzy', 'plain' }
+        }
      }) end,
      desc = 'Live fffuzy grep',
-    }
+    },
+    {
+      "fc",
+      function() require('fff').live_grep({ query = vim.fn.expand("<cword>") }) end,
+      desc = 'Search current word',
+    },
  }
 }
 ```
@@ -136,6 +154,10 @@ require('fff').setup({
      prompt_position = 'bottom', -- or 'top'
      preview_position = 'right', -- or 'left', 'right', 'top', 'bottom'
      preview_size = 0.5,
+      flex = { -- set to false to disable flex layout
+        size = 130, -- column threshold: if screen width >= size, use preview_position; otherwise use wrap
+        wrap = 'top', -- position to use when screen is narrower than size
+      },
      show_scrollbar = true, -- Show scrollbar for pagination
      -- How to shorten long directory paths in the file list:
      -- 'middle_number' (default): uses dots for 1-3 hidden (a/./b, a/../b, a/.../b)
@@ -151,6 +173,7 @@ require('fff').setup({
      binary_file_threshold = 1024, -- amount of bytes to scan for binary content (set 0 to disable)
      imagemagick_info_format_str = '%m: %wx%h, %[colorspace], %q-bit',
      line_numbers = false,
+      cursorlineopt = 'both', -- the cursorlineopt used for lines in grep file previews, see :h cursorlineopt
      wrap_lines = false,
      filetypes = {
        svg = { wrap_lines = true },
@@ -170,27 +193,29 @@ require('fff').setup({
      preview_scroll_up = '<C-u>',
      preview_scroll_down = '<C-d>',
      toggle_debug = '<F2>',
+      -- grep mode: cycle between plain text, regex, and fuzzy search
+      cycle_grep_modes = '<S-Tab>',
      -- goes to the previous query in history
      cycle_previous_query = '<C-Up>',
      -- multi-select keymaps for quickfix
      toggle_select = '<Tab>',
      send_to_quickfix = '<C-q>',
-      -- grep mode: cycle between plain text, regex, and fuzzy search
-      toggle_grep_regex = '<S-Tab>',
+      -- this are specific for the normal mode (you can exit it using any other keybind like jj)
+      focus_list = '<leader>l',
+      focus_preview = '<leader>p',
    },
    hl = {
      border = 'FloatBorder',
      normal = 'Normal',
-      cursor = 'CursorLine',
+      cursor = 'CursorLine',  -- Falls back to 'Visual' if CursorLine is not defined
      matched = 'IncSearch',
      title = 'Title',
      prompt = 'Question',
-      active_file = 'Visual',
      frecency = 'Number',
      debug = 'Comment',
      combo_header = 'Number',
-      scrollbar = 'Comment', -- Highlight for scrollbar thumb (track uses border)
-      directory_path = 'Comment', -- Highlight for directory path in file list
+      scrollbar = 'Comment',
+      directory_path = 'Comment',
      -- Multi-select highlights
      selected = 'FFFSelected',
      selected_active = 'FFFSelectedActive',
@@ -219,9 +244,10 @@ require('fff').setup({
      grep_match = 'IncSearch',               -- Highlight for matched text in grep results
      grep_line_number = 'LineNr',            -- Highlight for :line:col location
      grep_regex_active = 'DiagnosticInfo',   -- Highlight for keybind + label when regex is on
-      grep_regex_inactive = 'Comment',        -- Highlight for keybind + label when regex is off
+      grep_plain_active = 'Comment',        -- Highlight for keybind + label when regex is off
+      grep_fuzzy_active = 'DiagnosticHint',   -- Highlight for keybind + label when fuzzy is on
      -- Cross-mode suggestion highlights
-      suggestion_header = 'WarningMsg',       -- Highlight for the "No results found. Suggested..." banner
+      suggestion_header = 'WarningMsg', -- Highlight for the "No results found. Suggested..." banner
    },
    -- Store file open frecency
    frecency = {
@@ -232,32 +258,35 @@ require('fff').setup({
    history = {
      enabled = true,
      db_path = vim.fn.stdpath('data') .. '/fff_queries',
-      min_combo_count = 3, -- file will get a boost if it was selected 3 in a row times per specific query
-      combo_boost_score_multiplier = 100, -- Score multiplier for combo matches
+      min_combo_count = 3, -- Minimum selections before combo boost applies (3 = boost starts on 3rd selection)
+      combo_boost_score_multiplier = 100, -- Score multiplier for combo matches (files repeatedly opened with same query)
    },
    -- Git integration
    git = {
      status_text_color = false, -- Apply git status colors to filename text (default: false, only sign column)
    },
    debug = {
-      enabled = false, -- Set to true to show scores in the UI
-      show_scores = false,
-      show_file_info = false, -- Show file info panel in preview
+      enabled = false, -- Show file info panel in preview
+      show_scores = false, -- Show scores inline in the UI
    },
    logging = {
      enabled = true,
      log_file = vim.fn.stdpath('log') .. '/fff.log',
      log_level = 'info',
    },
-    -- Live grep search configuration
+    -- find_files settings
+    file_picker = {
+      current_file_label = '(current)',
+    },
+    -- grep settings
    grep = {
      max_file_size = 10 * 1024 * 1024, -- Skip files larger than 10MB
-      max_matches_per_file = 200, -- Maximum matches per file
+      max_matches_per_file = 100, -- Maximum matches per file (set 0 to unlimited)
      smart_case = true, -- Case-insensitive unless query has uppercase
      time_budget_ms = 150, -- Max search time in ms per call (prevents UI freeze, 0 = no limit)
      modes = { 'plain', 'regex', 'fuzzy' }, -- Available grep modes and their cycling order
-    }
-})
+    },
+  })
 ```

 ### Key Features
@@ -265,19 +294,19 @@ require('fff').setup({
 #### Available Methods

 ```lua
-require('fff').find_files()                         -- Find files in current directory
-require('fff').find_in_git_root()                   -- Find files in the current git repository
+require('fff').find_files()                         -- Find files in current repository
 require('fff').scan_files()                         -- Trigger rescan of files in the current directory
-require('fff').refresh_git_status()                 -- Refresh git status for the active file lock
+require('fff').refresh_git_status()                 -- Refresh git status for the active file list
 require('fff').find_files_in_dir(path)              -- Find files in a specific directory
 require('fff').change_indexing_directory(new_path)  -- Change the base directory for the file picker
 ```

+just jump to the definition and see what other APIs are exposed we have a plenty
+
 #### Commands

 FFF.nvim provides several commands for interacting with the file picker:

- `:FFFFind [path|query]` - Open file picker. Optional: provide directory path or search query
 - `:FFFScan` - Manually trigger a rescan of files in the current directory
 - `:FFFRefreshGit` - Manually refresh git status for all files
 - `:FFFClearCache [all|frecency|files]` - Clear various caches
@@ -285,10 +314,6 @@ FFF.nvim provides several commands for interacting with the file picker:
 - `:FFFDebug [on|off|toggle]` - Toggle debug scores display
 - `:FFFOpenLog` - Open the FFF log file in a new tab

-#### Multiline Paste Support
-
-The input field automatically handles multiline clipboard content by joining all lines into a single search query. This is particularly useful when copying file paths from terminal output.
-
 #### Debug Mode

 Toggle scoring information display:
@@ -349,6 +374,32 @@ require('fff').live_grep({ query = 'search term' })

 When only one mode is configured, the mode indicator is hidden completely and the cycle keybind does nothing.

+#### Constraints
+
+There are a number of constraints you can use to refine your search in both grep and file search mode:
+
+- `git:modified` - show only modified files (one of `modified`, `staged`, `deleted`, `renamed`, `untracked`, `ignored`)
+- `test/` - any deeply nested children of any test/ dir
+- `!something` - exclude results matching something
+- `!test/`, `!git:modified` - combining with any other constraint works as negation
+- `./**/*.{rs,lua}` - any valid glob expression via [the fastest globbing library](https://github.com/dmtrKovalenko/zlob)
+
+For grep only:
+
+- `*.md`, `*.{c,h}` - extension filtering
+- `src/main.rs` - grep in a single file
+
+In addition to that, all constraints can be combined together like:
+
+```
+git:modified src/**/*.rs !src/**/mod.rs user controller
+```
+
+This will find all the files that qualify the constraints and:
+
+- match **both** user and controller (for file mode)
+- match "user controller" (for grep mode)
+
 #### Cross-Mode Suggestions

 When a search returns no results, FFF automatically queries the opposite search mode and displays the results as suggestions:
@@ -442,41 +493,8 @@ Run `:FFFHealth` to check the status of FFF.nvim and its dependencies. This will

 If you encounter issues, check the log file:

-```vim
+```
 :FFFOpenLog
 ```

 Or manually open the log file at `~/.local/state/nvim/log/fff.log` (default location).
-
-#### Common Issues
-
-**File picker not initializing:**
-
- Ensure the Rust backend is compiled: `cargo build --release` in the plugin directory
- Check that your Neovim version is 0.10.0 or higher
-
-**Image previews not working:**
-
- Verify your terminal supports images (kitty, iTerm2, WezTerm, etc.)
- For terminals without native image support, install one of: `chafa`, `viu`, or `img2txt`
- If using snacks.nvim, ensure it's properly configured
-
-**Performance issues:**
-
- Adjust `max_threads` in configuration based on your system
- Reduce `preview.max_lines` and `preview.max_size` for large files
- Clear cache if it becomes too large: `:FFFClearCache all`
-
-**Files not being indexed:**
-
- Run `:FFFScan` to manually trigger a file scan
- Check that the `base_path` is correctly set
- Verify you have read permissions for the directory
-
-#### Debug Mode
-
-Enable debug mode to see scoring information and troubleshoot search results:
-
- Press `F2` while in the picker
- Run `:FFFDebug on` to enable permanently
- Set `debug.show_scores = true` in configuration
@@ -0,0 +1,35 @@
+{
+  "$schema": "https://biomejs.dev/schemas/2.4.4/schema.json",
+  "files": {
+    "includes": ["packages/**/*.ts", "!packages/*/dist"],
+    "ignoreUnknown": true
+  },
+  "formatter": {
+    "enabled": true,
+    "indentStyle": "space",
+    "indentWidth": 2,
+    "lineWidth": 90
+  },
+  "javascript": {
+    "formatter": {
+      "quoteStyle": "double",
+      "trailingCommas": "all",
+      "semicolons": "always"
+    }
+  },
+  "linter": {
+    "enabled": true,
+    "rules": {
+      "recommended": true,
+      "style": {
+        "noNonNullAssertion": "off"
+      },
+      "suspicious": {
+        "noExplicitAny": "off"
+      },
+      "complexity": {
+        "noForEach": "off"
+      }
+    }
+  }
+}
@@ -0,0 +1,148 @@
+{
+  "lockfileVersion": 1,
+  "configVersion": 1,
+  "workspaces": {
+    "": {
+      "devDependencies": {
+        "@biomejs/biome": "^2.4.4",
+      },
+    },
+    "packages/fff-bun": {
+      "name": "@ff-labs/fff-bun",
+      "version": "0.1.37",
+      "bin": {
+        "fff": "./scripts/cli.ts",
+        "fff-demo": "./examples/search.ts",
+        "fff-grep": "./examples/grep.ts",
+      },
+      "devDependencies": {
+        "@types/bun": "^1.3.8",
+        "typescript": "^5.0.0",
+      },
+      "optionalDependencies": {
+        "@ff-labs/fff-bun-darwin-arm64": "0.0.0",
+        "@ff-labs/fff-bun-darwin-x64": "0.0.0",
+        "@ff-labs/fff-bun-linux-arm64-gnu": "0.0.0",
+        "@ff-labs/fff-bun-linux-arm64-musl": "0.0.0",
+        "@ff-labs/fff-bun-linux-x64-gnu": "0.0.0",
+        "@ff-labs/fff-bun-linux-x64-musl": "0.0.0",
+        "@ff-labs/fff-bun-win32-arm64": "0.0.0",
+        "@ff-labs/fff-bun-win32-x64": "0.0.0",
+      },
+      "peerDependencies": {
+        "bun": ">=1.0.0",
+      },
+    },
+    "packages/fff-node": {
+      "name": "@ff-labs/fff-node",
+      "version": "0.1.37",
+      "bin": {
+        "fff-node": "./dist/scripts/cli.js",
+      },
+      "dependencies": {
+        "ffi-rs": "^1.0.0",
+      },
+      "devDependencies": {
+        "@types/node": "^22.0.0",
+        "typescript": "^5.0.0",
+      },
+      "optionalDependencies": {
+        "@ff-labs/fff-bun-darwin-arm64": "0.0.0",
+        "@ff-labs/fff-bun-darwin-x64": "0.0.0",
+        "@ff-labs/fff-bun-linux-arm64-gnu": "0.0.0",
+        "@ff-labs/fff-bun-linux-arm64-musl": "0.0.0",
+        "@ff-labs/fff-bun-linux-x64-gnu": "0.0.0",
+        "@ff-labs/fff-bun-linux-x64-musl": "0.0.0",
+        "@ff-labs/fff-bun-win32-arm64": "0.0.0",
+        "@ff-labs/fff-bun-win32-x64": "0.0.0",
+      },
+    },
+  },
+  "packages": {
+    "@biomejs/biome": ["@biomejs/biome@2.4.4", "", { "optionalDependencies": { "@biomejs/cli-darwin-arm64": "2.4.4", "@biomejs/cli-darwin-x64": "2.4.4", "@biomejs/cli-linux-arm64": "2.4.4", "@biomejs/cli-linux-arm64-musl": "2.4.4", "@biomejs/cli-linux-x64": "2.4.4", "@biomejs/cli-linux-x64-musl": "2.4.4", "@biomejs/cli-win32-arm64": "2.4.4", "@biomejs/cli-win32-x64": "2.4.4" }, "bin": { "biome": "bin/biome" } }, "sha512-tigwWS5KfJf0cABVd52NVaXyAVv4qpUXOWJ1rxFL8xF1RVoeS2q/LK+FHgYoKMclJCuRoCWAPy1IXaN9/mS61Q=="],
+
+    "@biomejs/cli-darwin-arm64": ["@biomejs/cli-darwin-arm64@2.4.4", "", { "os": "darwin", "cpu": "arm64" }, "sha512-jZ+Xc6qvD6tTH5jM6eKX44dcbyNqJHssfl2nnwT6vma6B1sj7ZLTGIk6N5QwVBs5xGN52r3trk5fgd3sQ9We9A=="],
+
+    "@biomejs/cli-darwin-x64": ["@biomejs/cli-darwin-x64@2.4.4", "", { "os": "darwin", "cpu": "x64" }, "sha512-Dh1a/+W+SUCXhEdL7TiX3ArPTFCQKJTI1mGncZNWfO+6suk+gYA4lNyJcBB+pwvF49uw0pEbUS49BgYOY4hzUg=="],
+
+    "@biomejs/cli-linux-arm64": ["@biomejs/cli-linux-arm64@2.4.4", "", { "os": "linux", "cpu": "arm64" }, "sha512-V/NFfbWhsUU6w+m5WYbBenlEAz8eYnSqRMDMAW3K+3v0tYVkNyZn8VU0XPxk/lOqNXLSCCrV7FmV/u3SjCBShg=="],
+
+    "@biomejs/cli-linux-arm64-musl": ["@biomejs/cli-linux-arm64-musl@2.4.4", "", { "os": "linux", "cpu": "arm64" }, "sha512-+sPAXq3bxmFwhVFJnSwkSF5Rw2ZAJMH3MF6C9IveAEOdSpgajPhoQhbbAK12SehN9j2QrHpk4J/cHsa/HqWaYQ=="],
+
+    "@biomejs/cli-linux-x64": ["@biomejs/cli-linux-x64@2.4.4", "", { "os": "linux", "cpu": "x64" }, "sha512-R4+ZCDtG9kHArasyBO+UBD6jr/FcFCTH8QkNTOCu0pRJzCWyWC4EtZa2AmUZB5h3e0jD7bRV2KvrENcf8rndBg=="],
+
+    "@biomejs/cli-linux-x64-musl": ["@biomejs/cli-linux-x64-musl@2.4.4", "", { "os": "linux", "cpu": "x64" }, "sha512-gGvFTGpOIQDb5CQ2VC0n9Z2UEqlP46c4aNgHmAMytYieTGEcfqhfCFnhs6xjt0S3igE6q5GLuIXtdQt3Izok+g=="],
+
+    "@biomejs/cli-win32-arm64": ["@biomejs/cli-win32-arm64@2.4.4", "", { "os": "win32", "cpu": "arm64" }, "sha512-trzCqM7x+Gn832zZHgr28JoYagQNX4CZkUZhMUac2YxvvyDRLJDrb5m9IA7CaZLlX6lTQmADVfLEKP1et1Ma4Q=="],
+
+    "@biomejs/cli-win32-x64": ["@biomejs/cli-win32-x64@2.4.4", "", { "os": "win32", "cpu": "x64" }, "sha512-gnOHKVPFAAPrpoPt2t+Q6FZ7RPry/FDV3GcpU53P3PtLNnQjBmKyN2Vh/JtqXet+H4pme8CC76rScwdjDcT1/A=="],
+
+    "@ff-labs/fff-bun": ["@ff-labs/fff-bun@workspace:packages/fff-bun"],
+
+    "@ff-labs/fff-node": ["@ff-labs/fff-node@workspace:packages/fff-node"],
+
+    "@oven/bun-darwin-aarch64": ["@oven/bun-darwin-aarch64@1.3.10", "", { "os": "darwin", "cpu": "arm64" }, "sha512-PXgg5gqcS/rHwa1hF0JdM1y5TiyejVrMHoBmWY/DjtfYZoFTXie1RCFOkoG0b5diOOmUcuYarMpH7CSNTqwj+w=="],
+
+    "@oven/bun-darwin-x64": ["@oven/bun-darwin-x64@1.3.10", "", { "os": "darwin", "cpu": "x64" }, "sha512-Nhssuh7GBpP5PiDSOl3+qnoIG7PJo+ec2oomDevnl9pRY6x6aD2gRt0JE+uf+A8Om2D6gjeHCxjEdrw5ZHE8mA=="],
+
+    "@oven/bun-darwin-x64-baseline": ["@oven/bun-darwin-x64-baseline@1.3.10", "", { "os": "darwin", "cpu": "x64" }, "sha512-w1gaTlqU0IJCmJ1X+PGHkdNU1n8Gemx5YKkjhkJIguvFINXEBB5U1KG82QsT65Tk4KyNMfbLTlmy4giAvUoKfA=="],
+
+    "@oven/bun-linux-aarch64": ["@oven/bun-linux-aarch64@1.3.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-OUgPHfL6+PM2Q+tFZjcaycN3D7gdQdYlWnwMI31DXZKY1r4HINWk9aEz9t/rNaHg65edwNrt7dsv9TF7xK8xIA=="],
+
+    "@oven/bun-linux-aarch64-musl": ["@oven/bun-linux-aarch64-musl@1.3.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-Ui5pAgM7JE9MzHokF0VglRMkbak3lTisY4Mf1AZutPACXWgKJC5aGrgnHBfkl7QS6fEeYb0juy1q4eRznRHOsw=="],
+
+    "@oven/bun-linux-x64": ["@oven/bun-linux-x64@1.3.10", "", { "os": "linux", "cpu": "x64" }, "sha512-bzUgYj/PIZziB/ZesIP9HUyfvh6Vlf3od+TrbTTyVEuCSMKzDPQVW/yEbRp0tcHO3alwiEXwJDrWrHAguXlgiQ=="],
+
+    "@oven/bun-linux-x64-baseline": ["@oven/bun-linux-x64-baseline@1.3.10", "", { "os": "linux", "cpu": "x64" }, "sha512-oqvMDYpX6dGJO03HgO5bXuccEsH3qbdO3MaAiAlO4CfkBPLUXz3N0DDElg5hz0L6ktdDVKbQVE5lfe+LAUISQg=="],
+
+    "@oven/bun-linux-x64-musl": ["@oven/bun-linux-x64-musl@1.3.10", "", { "os": "linux", "cpu": "x64" }, "sha512-poVXvOShekbexHq45b4MH/mRjQKwACAC8lHp3Tz/hEDuz0/20oncqScnmKwzhBPEpqJvydXficXfBYuSim8opw=="],
+
+    "@oven/bun-linux-x64-musl-baseline": ["@oven/bun-linux-x64-musl-baseline@1.3.10", "", { "os": "linux", "cpu": "x64" }, "sha512-/hOZ6S1VsTX6vtbhWVL9aAnOrdpuO54mAGUWpTdMz7dFG5UBZ/VUEiK0pBkq9A1rlBk0GeD/6Y4NBFl8Ha7cRA=="],
+
+    "@oven/bun-windows-aarch64": ["@oven/bun-windows-aarch64@1.3.10", "", { "os": "win32", "cpu": "arm64" }, "sha512-GXbz2swvN2DLw2dXZFeedMxSJtI64xQ9xp9Eg7Hjejg6mS2E4dP1xoQ2yAo2aZPi/2OBPAVaGzppI2q20XumHA=="],
+
+    "@oven/bun-windows-x64": ["@oven/bun-windows-x64@1.3.10", "", { "os": "win32", "cpu": "x64" }, "sha512-qaS1In3yfC/Z/IGQriVmF8GWwKuNqiw7feTSJWaQhH5IbL6ENR+4wGNPniZSJFaM/SKUO0e/YCRdoVBvgU4C1g=="],
+
+    "@oven/bun-windows-x64-baseline": ["@oven/bun-windows-x64-baseline@1.3.10", "", { "os": "win32", "cpu": "x64" }, "sha512-gh3UAHbUdDUG6fhLc1Csa4IGdtghue6U8oAIXWnUqawp6lwb3gOCRvp25IUnLF5vUHtgfMxuEUYV7YA2WxVutw=="],
+
+    "@types/bun": ["@types/bun@1.3.9", "", { "dependencies": { "bun-types": "1.3.9" } }, "sha512-KQ571yULOdWJiMH+RIWIOZ7B2RXQGpL1YQrBtLIV3FqDcCu6FsbFUBwhdKUlCKUpS3PJDsHlJ1QKlpxoVR+xtw=="],
+
+    "@types/node": ["@types/node@22.19.15", "", { "dependencies": { "undici-types": "~6.21.0" } }, "sha512-F0R/h2+dsy5wJAUe3tAU6oqa2qbWY5TpNfL/RGmo1y38hiyO1w3x2jPtt76wmuaJI4DQnOBu21cNXQ2STIUUWg=="],
+
+    "@yuuang/ffi-rs-android-arm64": ["@yuuang/ffi-rs-android-arm64@1.3.1", "", { "os": "android", "cpu": "arm64" }, "sha512-V4nmlXdOYZEa7GOxSExVG95SLp8FE0iTq2yKeN54UlfNMr3Sik+1Ff57LcCv7qYcn4TBqnBAt5rT3FAM6T6caQ=="],
+
+    "@yuuang/ffi-rs-darwin-arm64": ["@yuuang/ffi-rs-darwin-arm64@1.3.1", "", { "os": "darwin", "cpu": "arm64" }, "sha512-YlnTMIyzfW3mAULC5ZA774nzQfFlYXM0rrfq/8ZzWt+IMbYk55a++jrI+6JeKV+1EqlDS3TFBEFtjdBNG94KzQ=="],
+
+    "@yuuang/ffi-rs-darwin-x64": ["@yuuang/ffi-rs-darwin-x64@1.3.1", "", { "os": "darwin", "cpu": "x64" }, "sha512-sI3LpQQ34SX4nyOHc5yxA7FSqs9qPEUMqW/y/wWo9cuyPpaHMFsi/BeOVYsnC0syp3FrY7gzn6RnD6PlXCktXg=="],
+
+    "@yuuang/ffi-rs-linux-arm-gnueabihf": ["@yuuang/ffi-rs-linux-arm-gnueabihf@1.3.1", "", { "os": "linux", "cpu": "arm" }, "sha512-1WkcGkJTlwh4ZA59htKI+RXhiL3oKiYwLv7PO8LUf6FuADK73s5GcXp67iakKu243uYu+qGYr4RHco4ySddYhQ=="],
+
+    "@yuuang/ffi-rs-linux-arm64-gnu": ["@yuuang/ffi-rs-linux-arm64-gnu@1.3.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-J2PwqviycZxaEVA0Bwv38LqGDGSB9A1DPN4iYginYJZSvTvKW8kh7Tis0HbZrX1YDKnY8hi3lt0N0tCTNPDH5Q=="],
+
+    "@yuuang/ffi-rs-linux-arm64-musl": ["@yuuang/ffi-rs-linux-arm64-musl@1.3.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-Hn1W1hBPssTaqikU1Bqp1XUdDdOgbnYVIOtR++LVx66hhrtjf/xrIUQOhTm+NmOFDG16JUKXe1skfM4gpaqYwg=="],
+
+    "@yuuang/ffi-rs-linux-x64-gnu": ["@yuuang/ffi-rs-linux-x64-gnu@1.3.1", "", { "os": "linux", "cpu": "x64" }, "sha512-kW6e+oCYZPvpH2ppPsffA18e1aLowtmWTRjVlyHtY04g/nQDepQvDUkkcvInh9fW5jLna7PjHvktW1tVgYIj2A=="],
+
+    "@yuuang/ffi-rs-linux-x64-musl": ["@yuuang/ffi-rs-linux-x64-musl@1.3.1", "", { "os": "linux", "cpu": "x64" }, "sha512-HTwblAzruUS16nQPrez3ozvEHm1Xxh8J8w7rZYrpmAcNl1hzyOT8z/hY70M9Rt9fOqQ4Ovgor9qVy/U3ZJo0ZA=="],
+
+    "@yuuang/ffi-rs-win32-arm64-msvc": ["@yuuang/ffi-rs-win32-arm64-msvc@1.3.1", "", { "os": "win32", "cpu": "arm64" }, "sha512-WeZkGl2BP1U4tRhEQH+FXLQS52N8obp74smK5AAGOfzPAT1pHkq6+dVkC1QCSIt7dHJs7SPtlnQw+5DkdZYlWA=="],
+
+    "@yuuang/ffi-rs-win32-ia32-msvc": ["@yuuang/ffi-rs-win32-ia32-msvc@1.3.1", "", { "os": "win32", "cpu": [ "x64", "ia32", ] }, "sha512-rNGgMeCH5mdeHiMiJgt7wWXovZ+FHEfXhU9p4zZBH4n8M1/QnEsRUwlapISPLpILSGpoYS6iBuq9/fUlZY8Mhg=="],
+
+    "@yuuang/ffi-rs-win32-x64-msvc": ["@yuuang/ffi-rs-win32-x64-msvc@1.3.1", "", { "os": "win32", "cpu": "x64" }, "sha512-dr2LcLD2CXo2a7BktlOpV68QhayqiI112KxIJC9tBgQO/Dkdg4CPsdqmvzzLhFo64iC5RLl2BT7M5lJImrfUWw=="],
+
+    "bun": ["bun@1.3.10", "", { "optionalDependencies": { "@oven/bun-darwin-aarch64": "1.3.10", "@oven/bun-darwin-x64": "1.3.10", "@oven/bun-darwin-x64-baseline": "1.3.10", "@oven/bun-linux-aarch64": "1.3.10", "@oven/bun-linux-aarch64-musl": "1.3.10", "@oven/bun-linux-x64": "1.3.10", "@oven/bun-linux-x64-baseline": "1.3.10", "@oven/bun-linux-x64-musl": "1.3.10", "@oven/bun-linux-x64-musl-baseline": "1.3.10", "@oven/bun-windows-aarch64": "1.3.10", "@oven/bun-windows-x64": "1.3.10", "@oven/bun-windows-x64-baseline": "1.3.10" }, "os": [ "linux", "win32", "darwin", ], "cpu": [ "x64", "arm64", ], "bin": { "bun": "bin/bun.exe", "bunx": "bin/bunx.exe" } }, "sha512-S/CXaXXIyA4CMjdMkYQ4T2YMqnAn4s0ysD3mlsY4bUiOCqGlv28zck4Wd4H4kpvbekx15S9mUeLQ7Uxd0tYTLA=="],
+
+    "bun-types": ["bun-types@1.3.9", "", { "dependencies": { "@types/node": "*" } }, "sha512-+UBWWOakIP4Tswh0Bt0QD0alpTY8cb5hvgiYeWCMet9YukHbzuruIEeXC2D7nMJPB12kbh8C7XJykSexEqGKJg=="],
+
+    "ffi-rs": ["ffi-rs@1.3.1", "", { "optionalDependencies": { "@yuuang/ffi-rs-android-arm64": "1.3.1", "@yuuang/ffi-rs-darwin-arm64": "1.3.1", "@yuuang/ffi-rs-darwin-x64": "1.3.1", "@yuuang/ffi-rs-linux-arm-gnueabihf": "1.3.1", "@yuuang/ffi-rs-linux-arm64-gnu": "1.3.1", "@yuuang/ffi-rs-linux-arm64-musl": "1.3.1", "@yuuang/ffi-rs-linux-x64-gnu": "1.3.1", "@yuuang/ffi-rs-linux-x64-musl": "1.3.1", "@yuuang/ffi-rs-win32-arm64-msvc": "1.3.1", "@yuuang/ffi-rs-win32-ia32-msvc": "1.3.1", "@yuuang/ffi-rs-win32-x64-msvc": "1.3.1" } }, "sha512-ZyNXL9fnclnZV+waQmWB9JrfbIEyxQa1OWtMrHOrAgcC04PgP5hBMG5TdhVN8N4uT/eul8zCFMVnJUukAFFlXA=="],
+
+    "typescript": ["typescript@5.9.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw=="],
+
+    "undici-types": ["undici-types@6.21.0", "", {}, "sha512-iwDZqg0QAGrg9Rav5H4n0M64c3mkR59cJ6wQp+7C4nI0gsmExaedaYLNO44eT4AtBBwjbTiGPMlt2Md0T9H9JQ=="],
+
+    "bun-types/@types/node": ["@types/node@25.3.3", "", { "dependencies": { "undici-types": "~7.18.0" } }, "sha512-DpzbrH7wIcBaJibpKo9nnSQL0MTRdnWttGyE5haGwK86xgMOkFLp7vEyfQPGLOJh5wNYiJ3V9PmUMDhV9u8kkQ=="],
+
+    "bun-types/@types/node/undici-types": ["undici-types@7.18.2", "", {}, "sha512-AsuCzffGHJybSaRrmr5eHr81mwJU3kjw6M+uprWvCXiNeN9SOGwQ3Jn8jb8m3Z6izVgknn1R0FTCEAP2QrLY/w=="],
+  }
+}
@@ -1,20 +1,23 @@
 [package]
 name = "fff-c"
-version = "0.1.0"
+version = "0.2.4"
 edition = "2024"
-description = "C FFI bindings for fff-core - use from any language with C FFI support"
+description = "Raw C api of FFF file finder"
 license = "MIT"

 [lib]
 crate-type = ["cdylib"]

+[features]
+default = []
+zlob = ["fff/zlob"]
+
 [dependencies]
 mimalloc.workspace = true
-once_cell.workspace = true
 tracing.workspace = true
 git2.workspace = true

-fff-core = { path = "../fff-core" }
-fff-query-parser = { path = "../fff-query-parser" }
+fff = { package = "fff-search", path = "../fff-core" , version = "0.2.2" }
+fff-query-parser = { path = "../fff-query-parser" , version = "0.2.4" }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
@@ -0,0 +1,28 @@
+language = "C"
+header = "/* Generated by cbindgen — do not edit manually. */"
+include_guard = "FFF_C_H"
+include_version = true
+no_includes = true
+sys_includes = ["stdint.h", "stdbool.h", "stddef.h"]
+
+[export]
+include = [
+  "FffResult",
+  "FffSearchResult", "FffFileItem", "FffScore", "FffLocation",
+  "FffGrepResult", "FffGrepMatch", "FffMatchRange",
+  "FffScanProgress",
+]
+
+[export.rename]
+"FffResult" = "FffResult"
+"FffSearchResult" = "FffSearchResult"
+"FffFileItem" = "FffFileItem"
+"FffScore" = "FffScore"
+"FffLocation" = "FffLocation"
+"FffGrepResult" = "FffGrepResult"
+"FffGrepMatch" = "FffGrepMatch"
+"FffMatchRange" = "FffMatchRange"
+"FffScanProgress" = "FffScanProgress"
+
+[fn]
+sort_by = "None"
@@ -0,0 +1,539 @@
+/* Generated by cbindgen — do not edit manually. */
+
+#ifndef FFF_C_H
+#define FFF_C_H
+
+/* Generated with cbindgen:0.29.2 */
+
+#include <stdint.h>
+#include <stdbool.h>
+#include <stddef.h>
+
+/**
+ * Result envelope returned by all `fff_*` functions.
+ *
+ * Heap-allocated — the caller must free it with `fff_free_result`.
+ *
+ * Depending on the function, the payload is delivered through different fields:
+ *
+ * | Function                   | Payload field | Type                          |
+ * |----------------------------|---------------|-------------------------------|
+ * | `fff_create_instance`      | `handle`      | opaque instance pointer       |
+ * | `fff_search`               | `handle`      | `*mut FffSearchResult`        |
+ * | `fff_live_grep`            | `handle`      | `*mut FffGrepResult`          |
+ * | `fff_multi_grep`           | `handle`      | `*mut FffGrepResult`          |
+ * | `fff_get_scan_progress`    | `handle`      | `*mut FffScanProgress`        |
+ * | `fff_health_check`         | `handle`      | `*mut c_char` (JSON string)   |
+ * | `fff_get_historical_query` | `handle`      | `*mut c_char` (string or null)|
+ * | `fff_wait_for_scan`        | `int_value`   | 1 = completed, 0 = timed out  |
+ * | `fff_track_query`          | `int_value`   | 1 = success, 0 = failure      |
+ * | `fff_refresh_git_status`   | `int_value`   | number of files updated       |
+ * | `fff_scan_files`           | (none)        | success flag only             |
+ * | `fff_restart_index`        | (none)        | success flag only             |
+ *
+ * On failure, `success` is false and `error` contains the message.
+ *
+ * **Important:** `fff_free_result` frees `error` but does **not** free `handle`.
+ * The caller must free the handle with the appropriate function
+ * (`fff_destroy`, `fff_free_search_result`, `fff_free_grep_result`,
+ *  `fff_free_string`, etc.).
+ */
+typedef struct FffResult {
+  /**
+   * Whether the operation succeeded.
+   */
+  bool success;
+  /**
+   * Error message on failure. Null on success.
+   */
+  char *error;
+  /**
+   * Opaque pointer payload (instance handle, typed result struct, or string). May be null.
+   */
+  void *handle;
+  /**
+   * Integer payload for simple return values (bool as 0/1, counts, etc.).
+   */
+  int64_t int_value;
+} FffResult;
+
+/**
+ * A file item returned by `fff_search`.
+ *
+ * All string fields are heap-allocated and owned by the parent `FffSearchResult`.
+ * Free the entire result with `fff_free_search_result`.
+ */
+typedef struct FffFileItem {
+  char *path;
+  char *relative_path;
+  char *file_name;
+  char *git_status;
+  uint64_t size;
+  uint64_t modified;
+  int64_t access_frecency_score;
+  int64_t modification_frecency_score;
+  int64_t total_frecency_score;
+  bool is_binary;
+} FffFileItem;
+
+/**
+ * Score breakdown for a search result.
+ */
+typedef struct FffScore {
+  int32_t total;
+  int32_t base_score;
+  int32_t filename_bonus;
+  int32_t special_filename_bonus;
+  int32_t frecency_boost;
+  int32_t distance_penalty;
+  int32_t current_file_penalty;
+  int32_t combo_match_boost;
+  bool exact_match;
+  char *match_type;
+} FffScore;
+
+/**
+ * Location parsed from a query string (e.g. `"file.ts:42:10"`).
+ *
+ * `tag` encodes the variant:
+ *   0 = no location,
+ *   1 = line only (`line` is set),
+ *   2 = position (`line` + `col`),
+ *   3 = range (`line`/`col` = start, `end_line`/`end_col` = end).
+ */
+typedef struct FffLocation {
+  uint8_t tag;
+  int32_t line;
+  int32_t col;
+  int32_t end_line;
+  int32_t end_col;
+} FffLocation;
+
+/**
+ * Search result returned by `fff_search`.
+ *
+ * The caller must free this with `fff_free_search_result`.
+ */
+typedef struct FffSearchResult {
+  /**
+   * Pointer to a heap-allocated array of `FffFileItem` (length = `count`).
+   */
+  struct FffFileItem *items;
+  /**
+   * Pointer to a heap-allocated array of `FffScore` (length = `count`).
+   */
+  struct FffScore *scores;
+  /**
+   * Number of items/scores in the arrays.
+   */
+  uint32_t count;
+  /**
+   * Total number of files that matched the query.
+   */
+  uint32_t total_matched;
+  /**
+   * Total number of indexed files.
+   */
+  uint32_t total_files;
+  /**
+   * Location parsed from the query string.
+   */
+  struct FffLocation location;
+} FffSearchResult;
+
+/**
+ * A byte range within a matched line, used for highlighting.
+ */
+typedef struct FffMatchRange {
+  uint32_t start;
+  uint32_t end;
+} FffMatchRange;
+
+/**
+ * A single grep match with file and line information.
+ *
+ * All string fields and arrays are heap-allocated. Free the parent
+ * `FffGrepResult` with `fff_free_grep_result` to release everything.
+ */
+typedef struct FffGrepMatch {
+  char *path;
+  char *relative_path;
+  char *file_name;
+  char *git_status;
+  char *line_content;
+  struct FffMatchRange *match_ranges;
+  char **context_before;
+  char **context_after;
+  uint64_t size;
+  uint64_t modified;
+  int64_t total_frecency_score;
+  int64_t access_frecency_score;
+  int64_t modification_frecency_score;
+  uint64_t line_number;
+  uint64_t byte_offset;
+  uint32_t col;
+  uint32_t match_ranges_count;
+  uint32_t context_before_count;
+  uint32_t context_after_count;
+  uint16_t fuzzy_score;
+  bool has_fuzzy_score;
+  bool is_binary;
+  bool is_definition;
+} FffGrepMatch;
+
+/**
+ * Grep result returned by `fff_live_grep` and `fff_multi_grep`.
+ *
+ * The caller must free this with `fff_free_grep_result`.
+ */
+typedef struct FffGrepResult {
+  /**
+   * Pointer to a heap-allocated array of `FffGrepMatch` (length = `count`).
+   */
+  struct FffGrepMatch *items;
+  /**
+   * Number of matches in the `items` array.
+   */
+  uint32_t count;
+  /**
+   * Total number of matches (always equal to `count`).
+   */
+  uint32_t total_matched;
+  /**
+   * Number of files actually opened and searched in this call.
+   */
+  uint32_t total_files_searched;
+  /**
+   * Total number of indexed files (before any filtering).
+   */
+  uint32_t total_files;
+  /**
+   * Number of files eligible for search after filtering.
+   */
+  uint32_t filtered_file_count;
+  /**
+   * File offset for the next page. 0 if all files have been searched.
+   */
+  uint32_t next_file_offset;
+  /**
+   * Regex compilation error when falling back to literal matching. Null if none.
+   */
+  char *regex_fallback_error;
+} FffGrepResult;
+
+/**
+ * Scan progress returned by `fff_get_scan_progress`.
+ *
+ * The caller must free this with `fff_free_scan_progress`.
+ */
+typedef struct FffScanProgress {
+  uint64_t scanned_files_count;
+  bool is_scanning;
+} FffScanProgress;
+
+/**
+ * Create a new file finder instance.
+ *
+ * Returns an opaque pointer that must be passed to all other `fff_*` calls
+ * and eventually freed with `fff_destroy`.
+ *
+ * # Parameters
+ *
+ * * `base_path`          – directory to index (required)
+ * * `frecency_db_path`   – path to frecency LMDB database (NULL/empty to skip)
+ * * `history_db_path`    – path to query history LMDB database (NULL/empty to skip)
+ * * `use_unsafe_no_lock` – use MDB_NOLOCK for LMDB (useful in single-process setups)
+ * * `warmup_mmap_cache`  – pre-populate mmap caches after the initial scan
+ * * `ai_mode`            – enable AI-agent optimizations (auto-track frecency on modifications)
+ *
+ * ## Safety
+ * String parameters must be valid null-terminated UTF-8 or NULL.
+ */
+struct FffResult *fff_create_instance(const char *base_path,
+                                      const char *frecency_db_path,
+                                      const char *history_db_path,
+                                      bool use_unsafe_no_lock,
+                                      bool warmup_mmap_cache,
+                                      bool ai_mode);
+
+/**
+ * Destroy a file finder instance and free all its resources.
+ *
+ * ## Safety
+ * `fff_handle` must be a valid pointer returned by `fff_create_instance`, or null (no-op).
+ */
+void fff_destroy(void *fff_handle);
+
+/**
+ * Perform fuzzy search on indexed files.
+ *
+ * # Parameters
+ *
+ * * `fff_handle`              – instance from `fff_create_instance`
+ * * `query`                   – search query string
+ * * `current_file`            – path of the currently open file for deprioritization (NULL/empty to skip)
+ * * `max_threads`             – maximum worker threads (0 = auto-detect)
+ * * `page_index`              – pagination offset (0 = first page)
+ * * `page_size`               – results per page (0 = default 100)
+ * * `combo_boost_multiplier`  – score multiplier for combo matches (0 = default 100)
+ * * `min_combo_count`         – minimum combo count before boost applies (0 = default 3)
+ *
+ * ## Safety
+ * * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ * * `query` and `current_file` must be valid null-terminated UTF-8 strings or NULL.
+ */
+struct FffResult *fff_search(void *fff_handle,
+                             const char *query,
+                             const char *current_file,
+                             uint32_t max_threads,
+                             uint32_t page_index,
+                             uint32_t page_size,
+                             int32_t combo_boost_multiplier,
+                             uint32_t min_combo_count);
+
+/**
+ * Perform content search (grep) across indexed files.
+ *
+ * # Parameters
+ *
+ * * `fff_handle`            – instance from `fff_create_instance`
+ * * `query`                 – search query (supports constraint syntax like `*.rs pattern`)
+ * * `mode`                  – 0 = plain text (SIMD), 1 = regex, 2 = fuzzy
+ * * `max_file_size`         – skip files larger than this in bytes (0 = default 10 MB)
+ * * `max_matches_per_file`  – max matches per file (0 = unlimited)
+ * * `smart_case`            – case-insensitive when query is all lowercase
+ * * `file_offset`           – file-based pagination offset (0 = start)
+ * * `page_limit`            – max matches to return (0 = default 50)
+ * * `time_budget_ms`        – wall-clock budget in ms (0 = unlimited)
+ * * `before_context`        – context lines before each match
+ * * `after_context`         – context lines after each match
+ * * `classify_definitions`  – tag matches that are code definitions
+ *
+ * ## Safety
+ * * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ * * `query` must be a valid null-terminated UTF-8 string.
+ */
+struct FffResult *fff_live_grep(void *fff_handle,
+                                const char *query,
+                                uint8_t mode,
+                                uint64_t max_file_size,
+                                uint32_t max_matches_per_file,
+                                bool smart_case,
+                                uint32_t file_offset,
+                                uint32_t page_limit,
+                                uint64_t time_budget_ms,
+                                uint32_t before_context,
+                                uint32_t after_context,
+                                bool classify_definitions);
+
+/**
+ * Perform multi-pattern OR search (Aho-Corasick) across indexed files.
+ *
+ * Searches for lines matching ANY of the provided patterns using
+ * SIMD-accelerated multi-needle matching.
+ *
+ * # Parameters
+ *
+ * * `fff_handle`              – instance from `fff_create_instance`
+ * * `patterns_joined`         – patterns separated by `\n` (e.g. `"foo\nbar\nbaz"`)
+ * * `constraints`             – file filter like `"*.rs"` or `"/src/"` (NULL/empty to skip)
+ * * `max_file_size`           – skip files larger than this in bytes (0 = default 10 MB)
+ * * `max_matches_per_file`    – max matches per file (0 = unlimited)
+ * * `smart_case`              – case-insensitive when all patterns are lowercase
+ * * `file_offset`             – file-based pagination offset (0 = start)
+ * * `page_limit`              – max matches to return (0 = default 50)
+ * * `time_budget_ms`          – wall-clock budget in ms (0 = unlimited)
+ * * `before_context`          – context lines before each match
+ * * `after_context`           – context lines after each match
+ * * `classify_definitions`    – tag matches that are code definitions
+ *
+ * ## Safety
+ * * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ * * `patterns_joined` and `constraints` must be valid null-terminated UTF-8 or NULL.
+ */
+struct FffResult *fff_multi_grep(void *fff_handle,
+                                 const char *patterns_joined,
+                                 const char *constraints,
+                                 uint64_t max_file_size,
+                                 uint32_t max_matches_per_file,
+                                 bool smart_case,
+                                 uint32_t file_offset,
+                                 uint32_t page_limit,
+                                 uint64_t time_budget_ms,
+                                 uint32_t before_context,
+                                 uint32_t after_context,
+                                 bool classify_definitions);
+
+/**
+ * Trigger a rescan of the file index.
+ *
+ * ## Safety
+ * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ */
+struct FffResult *fff_scan_files(void *fff_handle);
+
+/**
+ * Check if a scan is currently in progress.
+ *
+ * ## Safety
+ * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ */
+bool fff_is_scanning(void *fff_handle);
+
+/**
+ * Get scan progress information.
+ *
+ * ## Safety
+ * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ */
+struct FffResult *fff_get_scan_progress(void *fff_handle);
+
+/**
+ * Wait for initial scan to complete.
+ *
+ * ## Safety
+ * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ */
+struct FffResult *fff_wait_for_scan(void *fff_handle, uint64_t timeout_ms);
+
+/**
+ * Restart indexing in a new directory.
+ *
+ * ## Safety
+ * * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ * * `new_path` must be a valid null-terminated UTF-8 string.
+ */
+struct FffResult *fff_restart_index(void *fff_handle, const char *new_path);
+
+/**
+ * Refresh git status cache.
+ *
+ * ## Safety
+ * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ */
+struct FffResult *fff_refresh_git_status(void *fff_handle);
+
+/**
+ * Track query completion for smart suggestions.
+ *
+ * ## Safety
+ * * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ * * `query` and `file_path` must be valid null-terminated UTF-8 strings.
+ */
+struct FffResult *fff_track_query(void *fff_handle, const char *query, const char *file_path);
+
+/**
+ * Get historical query by offset (0 = most recent).
+ *
+ * ## Safety
+ * `fff_handle` must be a valid instance pointer from `fff_create_instance`.
+ */
+struct FffResult *fff_get_historical_query(void *fff_handle, uint64_t offset);
+
+/**
+ * Get health check information.
+ *
+ * ## Safety
+ * * `fff_handle` must be a valid instance pointer from `fff_create_instance`, or null for
+ *   a limited health check (version + git only).
+ * * `test_path` can be null or a valid null-terminated UTF-8 string.
+ */
+struct FffResult *fff_health_check(void *fff_handle, const char *test_path);
+
+/**
+ * Free a search result returned by `fff_search`.
+ *
+ * This frees the `FffSearchResult` struct, its `items` and `scores` arrays,
+ * and all heap-allocated strings within each item and score.
+ *
+ * ## Safety
+ * `result` must be a valid pointer previously returned via `FffResult.handle`
+ * from `fff_search`, or null (no-op).
+ */
+void fff_free_search_result(struct FffSearchResult *result);
+
+/**
+ * Get a pointer to the `index`-th `FffFileItem` in a search result.
+ *
+ * Returns null if `result` is null or `index >= result->count`.
+ * The returned pointer is valid until the search result is freed.
+ *
+ * ## Safety
+ * `result` must be a valid `FffSearchResult` pointer from `fff_search`.
+ */
+const struct FffFileItem *fff_search_result_get_item(const struct FffSearchResult *result,
+                                                     uint32_t index);
+
+/**
+ * Get a pointer to the `index`-th `FffScore` in a search result.
+ *
+ * Returns null if `result` is null or `index >= result->count`.
+ * The returned pointer is valid until the search result is freed.
+ *
+ * ## Safety
+ * `result` must be a valid `FffSearchResult` pointer from `fff_search`.
+ */
+const struct FffScore *fff_search_result_get_score(const struct FffSearchResult *result,
+                                                   uint32_t index);
+
+/**
+ * Free a grep result returned by `fff_live_grep` or `fff_multi_grep`.
+ *
+ * This frees the `FffGrepResult` struct, its `items` array, and all
+ * heap-allocated strings, match ranges, and context arrays within each match.
+ *
+ * ## Safety
+ * `result` must be a valid pointer previously returned via `FffResult.handle`
+ * from `fff_live_grep` or `fff_multi_grep`, or null (no-op).
+ */
+void fff_free_grep_result(struct FffGrepResult *result);
+
+/**
+ * Get a pointer to the `index`-th `FffGrepMatch` in a grep result.
+ *
+ * Returns null if `result` is null or `index >= result->count`.
+ * The returned pointer is valid until the grep result is freed.
+ *
+ * ## Safety
+ * `result` must be a valid `FffGrepResult` pointer from `fff_live_grep` or `fff_multi_grep`.
+ */
+const struct FffGrepMatch *fff_grep_result_get_match(const struct FffGrepResult *result,
+                                                     uint32_t index);
+
+/**
+ * Free a scan progress result returned by `fff_get_scan_progress`.
+ *
+ * ## Safety
+ * `result` must be a valid pointer previously returned via `FffResult.handle`
+ * from `fff_get_scan_progress`, or null (no-op).
+ */
+void fff_free_scan_progress(struct FffScanProgress *result);
+
+/**
+ * Offset a pointer by `byte_offset` bytes.
+ *
+ * General-purpose utility for FFI consumers that need pointer arithmetic
+ * (e.g. iterating over arrays). Returns null if `base` is null.
+ *
+ * ## Safety
+ * The resulting pointer must be within the bounds of the original allocation.
+ */
+const void *fff_ptr_offset(const void *base, uintptr_t byte_offset);
+
+/**
+ * Free a result returned by any `fff_*` function.
+ *
+ * ## Safety
+ * `result_ptr` must be a valid pointer returned by a `fff_*` function.
+ */
+void fff_free_result(struct FffResult *result_ptr);
+
+/**
+ * Free a string returned by `fff_*` functions.
+ *
+ * ## Safety
+ * `s` must be a valid C string allocated by this library.
+ */
+void fff_free_string(char *s);
+
+#endif  /* FFF_C_H */
@@ -1,133 +1,122 @@
 //! FFI-compatible type definitions
 //!
-//! These types use #[repr(C)] for C ABI compatibility and implement
-//! serde traits for JSON serialization.
+//! All result types use `#[repr(C)]` structs for direct memory access from any
+//! language with C FFI support. No JSON serialization is used for search or grep
+//! results — callers read struct fields directly.

-use std::ffi::{CString, c_char};
+use std::ffi::{CString, c_char, c_void};
 use std::ptr;

-use fff_core::git::format_git_status;
-use fff_core::{FileItem, GrepMatch, GrepResult, Location, Score, SearchResult};
-use serde::{Deserialize, Serialize};
+use fff::git::format_git_status;
+use fff::{FileItem, GrepMatch, GrepResult, Location, Score, SearchResult};

-/// Result type returned by all FFI functions
-/// Returned as a heap-allocated pointer that must be freed with fff_free_result
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+/// Allocate a heap CString from a `&str`, returning a raw pointer.
+fn cstring_new(s: &str) -> *mut c_char {
+    CString::new(s).unwrap_or_default().into_raw()
+}
+
+/// Convert a `Vec<T>` into a raw pointer + count, leaking the memory.
+fn vec_to_raw<T>(v: Vec<T>) -> (*mut T, u32) {
+    if v.is_empty() {
+        return (ptr::null_mut(), 0);
+    }
+    let count = v.len() as u32;
+    let mut boxed = v.into_boxed_slice();
+    let p = boxed.as_mut_ptr();
+    std::mem::forget(boxed);
+    (p, count)
+}
+
+/// Convert a `&[String]` into a heap-allocated array of C strings.
+fn strings_to_raw(v: &[String]) -> (*mut *mut c_char, u32) {
+    if v.is_empty() {
+        return (ptr::null_mut(), 0);
+    }
+    let ptrs: Vec<*mut c_char> = v.iter().map(|s| cstring_new(s)).collect();
+    vec_to_raw(ptrs)
+}
+
+/// Free a heap-allocated array of C strings.
+///
+/// ## Safety
+/// `arr` must have been produced by `strings_to_raw`.
+unsafe fn free_cstring_array(arr: *mut *mut c_char, count: u32) {
+    if arr.is_null() {
+        return;
+    }
+    unsafe {
+        let ptrs = Vec::from_raw_parts(arr, count as usize, count as usize);
+        for p in ptrs {
+            if !p.is_null() {
+                drop(CString::from_raw(p));
+            }
+        }
+    }
+}
+
+/// A file item returned by `fff_search`.
+///
+/// All string fields are heap-allocated and owned by the parent `FffSearchResult`.
+/// Free the entire result with `fff_free_search_result`.
 #[repr(C)]
-pub struct FffResult {
-    /// Whether the operation succeeded
-    pub success: bool,
-    /// JSON data on success (null-terminated string, caller must free)
-    pub data: *mut c_char,
-    /// Error message on failure (null-terminated string, caller must free)
-    pub error: *mut c_char,
-}
-
-impl FffResult {
-    /// Create a successful result with no data, returned as heap pointer
-    pub fn ok_empty() -> *mut Self {
-        Box::into_raw(Box::new(FffResult {
-            success: true,
-            data: ptr::null_mut(),
-            error: ptr::null_mut(),
-        }))
-    }
-
-    /// Create a successful result with data, returned as heap pointer
-    pub fn ok_data(data: &str) -> *mut Self {
-        Box::into_raw(Box::new(FffResult {
-            success: true,
-            data: CString::new(data).unwrap_or_default().into_raw(),
-            error: ptr::null_mut(),
-        }))
-    }
-
-    /// Create an error result, returned as heap pointer
-    pub fn err(error: &str) -> *mut Self {
-        Box::into_raw(Box::new(FffResult {
-            success: false,
-            data: ptr::null_mut(),
-            error: CString::new(error).unwrap_or_default().into_raw(),
-        }))
-    }
-}
-
-/// Initialization options (JSON-deserializable)
-#[derive(Debug, Deserialize)]
-pub struct InitOptions {
-    /// Base directory to index (required)
-    pub base_path: String,
-    /// Path to frecency database (optional, omit to skip frecency initialization)
-    pub frecency_db_path: Option<String>,
-    /// Path to query history database (optional, omit to skip query tracker initialization)
-    pub history_db_path: Option<String>,
-    /// Use unsafe no-lock mode for databases (optional, defaults to false)
-    #[serde(default)]
-    pub use_unsafe_no_lock: bool,
-    /// Pre-populate mmap caches for all files after initial scan so the first
-    /// grep search is as fast as subsequent ones (optional, defaults to false)
-    #[serde(default)]
-    pub warmup_mmap_cache: bool,
-}
-
-/// Search options (JSON-deserializable)
-#[derive(Debug, Default, Deserialize)]
-pub struct SearchOptions {
-    /// Maximum threads for parallel search (0 = auto)
-    pub max_threads: Option<usize>,
-    /// Current file path (for deprioritization)
-    pub current_file: Option<String>,
-    /// Combo boost score multiplier
-    pub combo_boost_multiplier: Option<i32>,
-    /// Minimum combo count for boost
-    pub min_combo_count: Option<u32>,
-    /// Page index for pagination
-    pub page_index: Option<usize>,
-    /// Page size for pagination
-    pub page_size: Option<usize>,
-}
-
-/// Scan progress (JSON-serializable)
-#[derive(Debug, Serialize)]
-pub struct ScanProgress {
-    pub scanned_files_count: usize,
-    pub is_scanning: bool,
-}
-
-/// File item for JSON serialization
-#[derive(Debug, Serialize)]
-pub struct FileItemJson {
-    pub path: String,
-    pub relative_path: String,
-    pub file_name: String,
+pub struct FffFileItem {
+    pub path: *mut c_char,
+    pub relative_path: *mut c_char,
+    pub file_name: *mut c_char,
+    pub git_status: *mut c_char,
    pub size: u64,
    pub modified: u64,
    pub access_frecency_score: i64,
    pub modification_frecency_score: i64,
    pub total_frecency_score: i64,
-    pub git_status: String,
    pub is_binary: bool,
 }

-impl FileItemJson {
-    pub fn from_file_item(item: &FileItem) -> Self {
-        FileItemJson {
-            path: item.path.to_string_lossy().to_string(),
-            relative_path: item.relative_path.clone(),
-            file_name: item.file_name.clone(),
+impl From<&FileItem> for FffFileItem {
+    fn from(item: &FileItem) -> Self {
+        FffFileItem {
+            path: cstring_new(&item.path.to_string_lossy()),
+            relative_path: cstring_new(&item.relative_path),
+            file_name: cstring_new(&item.file_name),
+            git_status: cstring_new(format_git_status(item.git_status)),
            size: item.size,
            modified: item.modified,
            access_frecency_score: item.access_frecency_score,
            modification_frecency_score: item.modification_frecency_score,
            total_frecency_score: item.total_frecency_score,
-            git_status: format_git_status(item.git_status).to_string(),
            is_binary: item.is_binary,
        }
    }
 }

-/// Score for JSON serialization
-#[derive(Debug, Serialize)]
-pub struct ScoreJson {
+impl FffFileItem {
+    /// ## Safety
+    /// All string pointers must have been allocated by `CString::into_raw`.
+    pub unsafe fn free_strings(&mut self) {
+        unsafe {
+            if !self.path.is_null() {
+                drop(CString::from_raw(self.path));
+            }
+            if !self.relative_path.is_null() {
+                drop(CString::from_raw(self.relative_path));
+            }
+            if !self.file_name.is_null() {
+                drop(CString::from_raw(self.file_name));
+            }
+            if !self.git_status.is_null() {
+                drop(CString::from_raw(self.git_status));
+            }
+        }
+    }
+}
+
+/// Score breakdown for a search result.
+#[repr(C)]
+pub struct FffScore {
    pub total: i32,
    pub base_score: i32,
    pub filename_bonus: i32,
@@ -137,12 +126,12 @@ pub struct ScoreJson {
    pub current_file_penalty: i32,
    pub combo_match_boost: i32,
    pub exact_match: bool,
-    pub match_type: String,
+    pub match_type: *mut c_char,
 }

-impl ScoreJson {
-    pub fn from_score(score: &Score) -> Self {
-        ScoreJson {
+impl From<&Score> for FffScore {
+    fn from(score: &Score) -> Self {
+        FffScore {
            total: score.total,
            base_score: score.base_score,
            filename_bonus: score.filename_bonus,
@@ -152,186 +141,389 @@ impl ScoreJson {
            current_file_penalty: score.current_file_penalty,
            combo_match_boost: score.combo_match_boost,
            exact_match: score.exact_match,
-            match_type: score.match_type.to_string(),
+            match_type: cstring_new(score.match_type),
        }
    }
 }

-/// Location for JSON serialization
-#[derive(Debug, Serialize)]
-#[serde(tag = "type")]
-pub enum LocationJson {
-    #[serde(rename = "line")]
-    Line { line: i32 },
-    #[serde(rename = "position")]
-    Position { line: i32, col: i32 },
-    #[serde(rename = "range")]
-    Range {
-        start: PositionJson,
-        end: PositionJson,
-    },
+impl FffScore {
+    /// ## Safety
+    /// `match_type` must have been allocated by `CString::into_raw`.
+    pub unsafe fn free_strings(&mut self) {
+        unsafe {
+            if !self.match_type.is_null() {
+                drop(CString::from_raw(self.match_type));
+            }
+        }
+    }
 }

-#[derive(Debug, Serialize)]
-pub struct PositionJson {
+/// Location parsed from a query string (e.g. `"file.ts:42:10"`).
+///
+/// `tag` encodes the variant:
+///   0 = no location,
+///   1 = line only (`line` is set),
+///   2 = position (`line` + `col`),
+///   3 = range (`line`/`col` = start, `end_line`/`end_col` = end).
+#[repr(C)]
+pub struct FffLocation {
+    pub tag: u8,
    pub line: i32,
    pub col: i32,
+    pub end_line: i32,
+    pub end_col: i32,
 }

-impl LocationJson {
-    pub fn from_location(loc: &Location) -> Self {
+impl From<Option<&Location>> for FffLocation {
+    fn from(loc: Option<&Location>) -> Self {
        match loc {
-            Location::Line(line) => LocationJson::Line { line: *line },
-            Location::Position { line, col } => LocationJson::Position {
+            None => FffLocation {
+                tag: 0,
+                line: 0,
+                col: 0,
+                end_line: 0,
+                end_col: 0,
+            },
+            Some(Location::Line(line)) => FffLocation {
+                tag: 1,
+                line: *line,
+                col: 0,
+                end_line: 0,
+                end_col: 0,
+            },
+            Some(Location::Position { line, col }) => FffLocation {
+                tag: 2,
                line: *line,
                col: *col,
+                end_line: 0,
+                end_col: 0,
            },
-            Location::Range { start, end } => LocationJson::Range {
-                start: PositionJson {
-                    line: start.0,
-                    col: start.1,
-                },
-                end: PositionJson {
-                    line: end.0,
-                    col: end.1,
-                },
+            Some(Location::Range { start, end }) => FffLocation {
+                tag: 3,
+                line: start.0,
+                col: start.1,
+                end_line: end.0,
+                end_col: end.1,
            },
        }
    }
 }

-/// Search result for JSON serialization
-#[derive(Debug, Serialize)]
-pub struct SearchResultJson {
-    pub items: Vec<FileItemJson>,
-    pub scores: Vec<ScoreJson>,
-    pub total_matched: usize,
-    pub total_files: usize,
-    pub location: Option<LocationJson>,
+/// Search result returned by `fff_search`.
+///
+/// The caller must free this with `fff_free_search_result`.
+#[repr(C)]
+pub struct FffSearchResult {
+    /// Pointer to a heap-allocated array of `FffFileItem` (length = `count`).
+    pub items: *mut FffFileItem,
+    /// Pointer to a heap-allocated array of `FffScore` (length = `count`).
+    pub scores: *mut FffScore,
+    /// Number of items/scores in the arrays.
+    pub count: u32,
+    /// Total number of files that matched the query.
+    pub total_matched: u32,
+    /// Total number of indexed files.
+    pub total_files: u32,
+    /// Location parsed from the query string.
+    pub location: FffLocation,
 }

-impl SearchResultJson {
-    pub fn from_search_result(result: &SearchResult) -> Self {
-        SearchResultJson {
-            items: result
-                .items
-                .iter()
-                .map(|item| FileItemJson::from_file_item(item))
-                .collect(),
-            scores: result.scores.iter().map(ScoreJson::from_score).collect(),
-            total_matched: result.total_matched,
-            total_files: result.total_files,
-            location: result.location.as_ref().map(LocationJson::from_location),
-        }
+impl FffSearchResult {
+    /// Convert a core `SearchResult` into a heap-allocated `FffSearchResult`.
+    pub fn from_core(result: &SearchResult) -> *mut Self {
+        let items: Vec<FffFileItem> = result.items.iter().map(|i| FffFileItem::from(*i)).collect();
+        let scores: Vec<FffScore> = result.scores.iter().map(FffScore::from).collect();
+        let count = items.len() as u32;
+
+        let (items_ptr, _) = vec_to_raw(items);
+        let (scores_ptr, _) = vec_to_raw(scores);
+
+        Box::into_raw(Box::new(FffSearchResult {
+            items: items_ptr,
+            scores: scores_ptr,
+            count,
+            total_matched: result.total_matched as u32,
+            total_files: result.total_files as u32,
+            location: FffLocation::from(result.location.as_ref()),
+        }))
    }
 }

-// ============================================================================
-// Grep (live search) types
-// ============================================================================
+// ---------------------------------------------------------------------------
+// Grep result types
+// ---------------------------------------------------------------------------

-/// Grep search options (JSON-deserializable)
-#[derive(Debug, Default, Deserialize)]
-pub struct GrepSearchOptionsJson {
-    /// Maximum file size to search (bytes, default: 10MB)
-    pub max_file_size: Option<u64>,
-    /// Maximum matches per file (default: 200)
-    pub max_matches_per_file: Option<usize>,
-    /// Smart case: case-insensitive if query is lowercase (default: true)
-    pub smart_case: Option<bool>,
-    /// File-based pagination offset (default: 0)
-    pub file_offset: Option<usize>,
-    /// Maximum matches to return (default: 50)
-    pub page_limit: Option<usize>,
-    /// Search mode: "plain", "regex", or "fuzzy" (default: "plain")
-    pub mode: Option<String>,
-    /// Time budget in milliseconds, 0 = unlimited (default: 0)
-    pub time_budget_ms: Option<u64>,
+/// A byte range within a matched line, used for highlighting.
+#[repr(C)]
+pub struct FffMatchRange {
+    pub start: u32,
+    pub end: u32,
 }

-/// A single grep match for JSON serialization
-#[derive(Debug, Serialize)]
-pub struct GrepMatchJson {
-    /// File metadata
-    pub path: String,
-    pub relative_path: String,
-    pub file_name: String,
-    pub git_status: String,
+/// A single grep match with file and line information.
+///
+/// All string fields and arrays are heap-allocated. Free the parent
+/// `FffGrepResult` with `fff_free_grep_result` to release everything.
+#[repr(C)]
+pub struct FffGrepMatch {
+    // -- pointers (8 bytes each) --
+    pub path: *mut c_char,
+    pub relative_path: *mut c_char,
+    pub file_name: *mut c_char,
+    pub git_status: *mut c_char,
+    pub line_content: *mut c_char,
+    pub match_ranges: *mut FffMatchRange,
+    pub context_before: *mut *mut c_char,
+    pub context_after: *mut *mut c_char,
+    // -- 8-byte numeric fields --
    pub size: u64,
    pub modified: u64,
-    pub is_binary: bool,
    pub total_frecency_score: i64,
    pub access_frecency_score: i64,
    pub modification_frecency_score: i64,
-    /// Match metadata
    pub line_number: u64,
-    pub col: usize,
    pub byte_offset: u64,
-    pub line_content: String,
-    /// Byte offset pairs (start, end) within line_content for highlighting
-    pub match_ranges: Vec<[u32; 2]>,
-    /// Fuzzy match score (only in fuzzy mode)
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub fuzzy_score: Option<u16>,
+    // -- 4-byte fields --
+    pub col: u32,
+    pub match_ranges_count: u32,
+    pub context_before_count: u32,
+    pub context_after_count: u32,
+    // -- 2-byte fields --
+    pub fuzzy_score: u16,
+    // -- 1-byte fields --
+    pub has_fuzzy_score: bool,
+    pub is_binary: bool,
+    pub is_definition: bool,
 }

-impl GrepMatchJson {
-    pub fn from_grep_match(m: &GrepMatch, file: &FileItem) -> Self {
-        GrepMatchJson {
-            path: file.path.to_string_lossy().to_string(),
-            relative_path: file.relative_path.clone(),
-            file_name: file.file_name.clone(),
-            git_status: format_git_status(file.git_status).to_string(),
+impl FffGrepMatch {
+    fn from_core_with_file(m: &GrepMatch, file: &FileItem) -> Self {
+        let ranges: Vec<FffMatchRange> = m
+            .match_byte_offsets
+            .iter()
+            .map(|&(start, end)| FffMatchRange { start, end })
+            .collect();
+        let (match_ranges, match_ranges_count) = vec_to_raw(ranges);
+        let (context_before, context_before_count) = strings_to_raw(&m.context_before);
+        let (context_after, context_after_count) = strings_to_raw(&m.context_after);
+        let (has_fuzzy_score, fuzzy_score) = match m.fuzzy_score {
+            Some(s) => (true, s),
+            None => (false, 0),
+        };
+
+        FffGrepMatch {
+            path: cstring_new(&file.path.to_string_lossy()),
+            relative_path: cstring_new(&file.relative_path),
+            file_name: cstring_new(&file.file_name),
+            git_status: cstring_new(format_git_status(file.git_status)),
+            line_content: cstring_new(&m.line_content),
+            match_ranges,
+            context_before,
+            context_after,
            size: file.size,
            modified: file.modified,
-            is_binary: file.is_binary,
            total_frecency_score: file.total_frecency_score,
            access_frecency_score: file.access_frecency_score,
            modification_frecency_score: file.modification_frecency_score,
            line_number: m.line_number,
-            col: m.col,
            byte_offset: m.byte_offset,
-            line_content: m.line_content.clone(),
-            match_ranges: m
-                .match_byte_offsets
-                .iter()
-                .map(|&(start, end)| [start, end])
-                .collect(),
-            fuzzy_score: m.fuzzy_score,
+            col: m.col as u32,
+            match_ranges_count,
+            context_before_count,
+            context_after_count,
+            fuzzy_score,
+            has_fuzzy_score,
+            is_binary: file.is_binary,
+            is_definition: m.is_definition,
+        }
+    }
+
+    /// ## Safety
+    /// All pointers must have been allocated by the corresponding `from_core`.
+    pub unsafe fn free_fields(&mut self) {
+        unsafe {
+            if !self.path.is_null() {
+                drop(CString::from_raw(self.path));
+            }
+            if !self.relative_path.is_null() {
+                drop(CString::from_raw(self.relative_path));
+            }
+            if !self.file_name.is_null() {
+                drop(CString::from_raw(self.file_name));
+            }
+            if !self.git_status.is_null() {
+                drop(CString::from_raw(self.git_status));
+            }
+            if !self.line_content.is_null() {
+                drop(CString::from_raw(self.line_content));
+            }
+            if !self.match_ranges.is_null() {
+                drop(Vec::from_raw_parts(
+                    self.match_ranges,
+                    self.match_ranges_count as usize,
+                    self.match_ranges_count as usize,
+                ));
+            }
+            free_cstring_array(self.context_before, self.context_before_count);
+            free_cstring_array(self.context_after, self.context_after_count);
        }
    }
 }

-/// Grep result for JSON serialization
-#[derive(Debug, Serialize)]
-pub struct GrepResultJson {
-    pub items: Vec<GrepMatchJson>,
-    pub total_matched: usize,
-    pub total_files_searched: usize,
-    pub total_files: usize,
-    pub filtered_file_count: usize,
-    pub next_file_offset: usize,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub regex_fallback_error: Option<String>,
+/// Grep result returned by `fff_live_grep` and `fff_multi_grep`.
+///
+/// The caller must free this with `fff_free_grep_result`.
+#[repr(C)]
+pub struct FffGrepResult {
+    /// Pointer to a heap-allocated array of `FffGrepMatch` (length = `count`).
+    pub items: *mut FffGrepMatch,
+    /// Number of matches in the `items` array.
+    pub count: u32,
+    /// Total number of matches (always equal to `count`).
+    pub total_matched: u32,
+    /// Number of files actually opened and searched in this call.
+    pub total_files_searched: u32,
+    /// Total number of indexed files (before any filtering).
+    pub total_files: u32,
+    /// Number of files eligible for search after filtering.
+    pub filtered_file_count: u32,
+    /// File offset for the next page. 0 if all files have been searched.
+    pub next_file_offset: u32,
+    /// Regex compilation error when falling back to literal matching. Null if none.
+    pub regex_fallback_error: *mut c_char,
 }

-impl GrepResultJson {
-    pub fn from_grep_result(result: &GrepResult) -> Self {
-        GrepResultJson {
-            items: result
-                .matches
-                .iter()
-                .map(|m| {
-                    let file = result.files[m.file_index];
-                    GrepMatchJson::from_grep_match(m, file)
-                })
-                .collect(),
-            total_matched: result.total_match_count,
-            total_files_searched: result.total_files_searched,
-            total_files: result.total_files,
-            filtered_file_count: result.filtered_file_count,
-            next_file_offset: result.next_file_offset,
-            regex_fallback_error: result.regex_fallback_error.clone(),
-        }
+impl FffGrepResult {
+    /// Convert a core `GrepResult` into a heap-allocated `FffGrepResult`.
+    pub fn from_core(result: &GrepResult) -> *mut Self {
+        let items: Vec<FffGrepMatch> = result
+            .matches
+            .iter()
+            .map(|m| {
+                let file = result.files[m.file_index];
+                FffGrepMatch::from_core_with_file(m, file)
+            })
+            .collect();
+        let (items_ptr, count) = vec_to_raw(items);
+
+        Box::into_raw(Box::new(FffGrepResult {
+            items: items_ptr,
+            count,
+            total_matched: result.matches.len() as u32,
+            total_files_searched: result.total_files_searched as u32,
+            total_files: result.total_files as u32,
+            filtered_file_count: result.filtered_file_count as u32,
+            next_file_offset: result.next_file_offset as u32,
+            regex_fallback_error: match &result.regex_fallback_error {
+                Some(e) => cstring_new(e),
+                None => ptr::null_mut(),
+            },
+        }))
    }
 }
+
+/// Result envelope returned by all `fff_*` functions.
+///
+/// Heap-allocated — the caller must free it with `fff_free_result`.
+///
+/// Depending on the function, the payload is delivered through different fields:
+///
+/// | Function                   | Payload field | Type                          |
+/// |----------------------------|---------------|-------------------------------|
+/// | `fff_create_instance`      | `handle`      | opaque instance pointer       |
+/// | `fff_search`               | `handle`      | `*mut FffSearchResult`        |
+/// | `fff_live_grep`            | `handle`      | `*mut FffGrepResult`          |
+/// | `fff_multi_grep`           | `handle`      | `*mut FffGrepResult`          |
+/// | `fff_get_scan_progress`    | `handle`      | `*mut FffScanProgress`        |
+/// | `fff_health_check`         | `handle`      | `*mut c_char` (JSON string)   |
+/// | `fff_get_historical_query` | `handle`      | `*mut c_char` (string or null)|
+/// | `fff_wait_for_scan`        | `int_value`   | 1 = completed, 0 = timed out  |
+/// | `fff_track_query`          | `int_value`   | 1 = success, 0 = failure      |
+/// | `fff_refresh_git_status`   | `int_value`   | number of files updated       |
+/// | `fff_scan_files`           | (none)        | success flag only             |
+/// | `fff_restart_index`        | (none)        | success flag only             |
+///
+/// On failure, `success` is false and `error` contains the message.
+///
+/// **Important:** `fff_free_result` frees `error` but does **not** free `handle`.
+/// The caller must free the handle with the appropriate function
+/// (`fff_destroy`, `fff_free_search_result`, `fff_free_grep_result`,
+///  `fff_free_string`, etc.).
+#[repr(C)]
+pub struct FffResult {
+    /// Whether the operation succeeded.
+    pub success: bool,
+    /// Error message on failure. Null on success.
+    pub error: *mut c_char,
+    /// Opaque pointer payload (instance handle, typed result struct, or string). May be null.
+    pub handle: *mut c_void,
+    /// Integer payload for simple return values (bool as 0/1, counts, etc.).
+    pub int_value: i64,
+}
+
+impl FffResult {
+    /// Create a successful result with no payload, returned as heap pointer.
+    pub fn ok_empty() -> *mut Self {
+        Box::into_raw(Box::new(FffResult {
+            success: true,
+            error: ptr::null_mut(),
+            handle: ptr::null_mut(),
+            int_value: 0,
+        }))
+    }
+
+    /// Create a successful result with an integer value.
+    pub fn ok_int(value: i64) -> *mut Self {
+        Box::into_raw(Box::new(FffResult {
+            success: true,
+            error: ptr::null_mut(),
+            handle: ptr::null_mut(),
+            int_value: value,
+        }))
+    }
+
+    /// Create a successful result carrying an opaque pointer (handle, typed struct, or string).
+    pub fn ok_handle(handle: *mut c_void) -> *mut Self {
+        Box::into_raw(Box::new(FffResult {
+            success: true,
+            error: ptr::null_mut(),
+            handle,
+            int_value: 0,
+        }))
+    }
+
+    /// Create a successful result carrying a C string in the `handle` field.
+    /// The caller must free it with `fff_free_string`.
+    pub fn ok_string(s: &str) -> *mut Self {
+        let cstr = CString::new(s).unwrap_or_default().into_raw();
+        Box::into_raw(Box::new(FffResult {
+            success: true,
+            error: ptr::null_mut(),
+            handle: cstr as *mut c_void,
+            int_value: 0,
+        }))
+    }
+
+    /// Create an error result, returned as heap pointer.
+    pub fn err(error: &str) -> *mut Self {
+        Box::into_raw(Box::new(FffResult {
+            success: false,
+            error: CString::new(error).unwrap_or_default().into_raw(),
+            handle: ptr::null_mut(),
+            int_value: 0,
+        }))
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Scan progress
+// ---------------------------------------------------------------------------
+
+/// Scan progress returned by `fff_get_scan_progress`.
+///
+/// The caller must free this with `fff_free_scan_progress`.
+#[repr(C)]
+pub struct FffScanProgress {
+    pub scanned_files_count: u64,
+    pub is_scanning: bool,
+}
@@ -1,9 +1,10 @@
 [package]
-name = "fff-core"
-version = "0.1.0"
+name = "fff-search"
+version = "0.2.4"
 edition = "2024"
-description = "High-performance file finder core library"
 license = "MIT"
+authors = ["Dmitriy Kovalenko <dmtr.kovalenko@outlook.com>"]
+description = "Faboulous & Fast File Finder - a fast and extremely correct file finder SDK with typo resistance, SIMD, prefiltering, and more"

 [lib]
 path = "src/lib.rs"
@@ -13,17 +14,17 @@ crate-type = ["rlib", "staticlib", "cdylib"]
 default = []
 # Enable C FFI exports
 ffi = []
+# Use zlob (Zig-compiled C globbing library) for glob matching.
+# Requires Zig to be installed. When disabled, falls back to globset (pure Rust).
+zlob = ["dep:zlob", "fff-query-parser/zlob"]

 [dependencies]
-# Workspace dependencies
 ahash = { workspace = true }
 rayon = { workspace = true }
 smallvec = { workspace = true }
 thiserror = { workspace = true }
 tracing = { workspace = true }
-
-# Local crates
-fff-query-parser = { path = "../fff-query-parser" }
+fff-query-parser = { workspace = true , version = "0.2.4" }

 # External dependencies
 bindet = { workspace = true }
@@ -32,8 +33,10 @@ chrono = { workspace = true }
 dirs = { workspace = true }
 git2 = { workspace = true }
 glidesort = { workspace = true }
+globset = { workspace = true }
 grep-matcher = { workspace = true }
-grep-searcher = { workspace = true }
+fff-grep = { workspace = true , version = "0.2.4" }
+aho-corasick = "1"
 memchr = "2"
 heed = { workspace = true }
 ignore = { workspace = true }
@@ -46,15 +49,12 @@ parking_lot = { workspace = true }
 pathdiff = { workspace = true }
 regex = { workspace = true }
 serde = { version = "1.0", features = ["derive"] }
+serde_json = "1.0"
 smartstring = { version = "1.0.1", features = ["serde"] }
+toml = "0.8"
 tracing-appender = "0.2"
 tracing-subscriber = { version = "0.3", features = ["env-filter"] }
-zlob = { workspace = true }
-
-# Platform-specific: Use vendored OpenSSL on non-Windows (Linux, macOS)
-[target.'cfg(not(windows))'.dependencies]
-openssl = { version = "0.10", features = ["vendored"] }
-
+zlob = { workspace = true, optional = true }
 # Platform-specific: dunce for Windows to avoid \\?\ extended path prefix
 [target.'cfg(windows)'.dependencies]
 dunce = { workspace = true }
@@ -0,0 +1,26 @@
+# fff
+
+fff is a file search toolkit. It is faster than ripgrep and fzf and designed for a long running applications like file editors, ai agents, or file exploerers.
+
+## Features
+
+- Fuzzy file name search 
+- Typo resistance
+- Frecency and query history ranking
+- Native git support via libgit
+- Advanced ranking 
+- Grep functionality with SIMD optimized plain matcher and regex
+- Multi grep using aho-corasick algorithm
+- Efficient memory mapping for file system
+- Cross platform support (Linux, Windows, MacOS)
+- Advnaced constraints syntax allowing to prefilter based on git status, glob, extension, size, timing and more
+
+## Performance
+
+FFF is designed for high performance and low latency. SIMD optimized where needed, parallelized for multi core systems, efficient sorting and ranking algorithms, memaps and much more.
+
+On MacOS FFF is about 20-50 times faster than ripgrep for content search and around 10 times faster than fzf for file name search.
+
+## Documentation
+
+Refer rust docs https://docs.rs/crate/fff-search/latest
@@ -1,16 +1,46 @@
 fn main() {
+    // When the `zlob` feature is enabled (Zig-compiled C library):
    // On Windows MSVC, explicitly link the C runtime libraries.
-    // This is needed because Zig-compiled static libraries (zlob) don't emit
-    // /DEFAULTLIB directives for the MSVC CRT. Without this, symbols like
-    // strcmp, memcpy, memchr etc. from vendored C libraries (libgit2, lmdb)
-    // are unresolved when linking the cdylib.
-    //
-    // We link both msvcrt (classic CRT) and ucrt (Universal CRT where memchr,
-    // strcmp etc. live on newer MSVC/ARM64 targets).
-    let target = std::env::var("TARGET").unwrap_or_default();
-    if target.contains("windows") && target.contains("msvc") {
-        println!("cargo:rustc-link-lib=msvcrt");
-        println!("cargo:rustc-link-lib=ucrt");
-        println!("cargo:rustc-link-lib=vcruntime");
+    // Zig-compiled static libraries don't emit /DEFAULTLIB directives for the
+    // MSVC CRT, so symbols like strcmp, memcpy etc. would be unresolved.
+    if std::env::var("CARGO_FEATURE_ZLOB").is_ok() {
+        let target = std::env::var("TARGET").unwrap_or_default();
+        if target.contains("windows") && target.contains("msvc") {
+            println!("cargo:rustc-link-lib=msvcrt");
+            println!("cargo:rustc-link-lib=ucrt");
+            println!("cargo:rustc-link-lib=vcruntime");
+        }
+    } else if std::env::var("CI").is_ok() {
+        // CI must always build with zlob for production-quality binaries.
+        if !zig_available() {
+            panic!(
+                "CI detected but Zig is not installed. \
+                 Please install Zig and build with `--features zlob`."
+            );
+        }
+        panic!(
+            "CI detected but `zlob` feature is not enabled. \
+             Build with `--features zlob`."
+        );
+    } else {
+        // Hint: if Zig is available but the zlob feature wasn't enabled,
+        // let the developer know they can get faster glob matching.
+        if zig_available() {
+            println!(
+                "cargo:warning=Zig detected but `zlob` feature is not enabled. \
+                 Build with `--features zlob` for faster glob matching."
+            );
+        }
    }
 }
+
+/// Probe the system for a working Zig installation.
+fn zig_available() -> bool {
+    std::process::Command::new("zig")
+        .arg("version")
+        .stdout(std::process::Stdio::null())
+        .stderr(std::process::Stdio::null())
+        .status()
+        .map(|s| s.success())
+        .unwrap_or(false)
+}
@@ -1,67 +1,131 @@
-use crate::FILE_PICKER;
 use crate::error::Error;
-use crate::file_picker::FilePicker;
+use crate::file_picker::{FFFMode, FilePicker};
 use crate::git::GitStatusCache;
 use crate::sort_buffer::sort_with_buffer;
+use crate::{SharedFrecency, SharedPicker};
 use git2::Repository;
 use notify::event::{AccessKind, AccessMode};
 use notify::{Config, EventKind, RecursiveMode};
-use notify_debouncer_full::{
-    DebounceEventResult, DebouncedEvent, RecommendedCache, new_debouncer_opt,
-};
+use notify_debouncer_full::{DebounceEventResult, DebouncedEvent, NoCache, new_debouncer_opt};
 use std::path::{Path, PathBuf};
-use std::sync::{Arc, Mutex};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
-use tracing::{Level, error, info, warn};
+use tracing::{Level, debug, error, info, warn};

-type Debouncer = notify_debouncer_full::Debouncer<notify::RecommendedWatcher, RecommendedCache>;
+type Debouncer = notify_debouncer_full::Debouncer<notify::RecommendedWatcher, NoCache>;

+/// Owns the file-system watcher and guarantees that all background threads
+/// are fully joined before `stop()` / `Drop` returns.
+///
+/// Architecture:
+///   - The debouncer (and its internal watcher) live inside an **owner thread**
+///     that we spawn and hold the `JoinHandle` for.
+///   - `stop()` sets a flag, unparks the owner thread, and **joins** it.
+///   - Inside the owner thread, `Debouncer::stop()` is called which joins the
+///     debouncer's event-processing thread.
+///   - On Windows an additional short sleep is added after `Debouncer::stop()`
+///     because `notify`'s `ReadDirectoryChangesWatcher` discards its thread
+///     `JoinHandle`, so we cannot join it directly. The watcher's `Drop` does
+///     signal the thread via semaphore so it exits almost immediately, but we
+///     need to give the OS a moment to reclaim it.
 pub struct BackgroundWatcher {
-    debouncer: Arc<Mutex<Option<Debouncer>>>,
+    stop_signal: Arc<AtomicBool>,
+    owner_thread: Option<std::thread::JoinHandle<()>>,
 }

 const DEBOUNCE_TIMEOUT: Duration = Duration::from_millis(250);
 const MAX_PATHS_THRESHOLD: usize = 1024;
 const MAX_SELECTIVE_WATCH_DIRS: usize = 100;
+/// Minimum seconds between frecency tracks of the same file in AI mode.
+/// Prevents score inflation from rapid burst edits by AI agents.
+const AI_MODE_COOLDOWN_SECS: u64 = 5 * 60;

 impl BackgroundWatcher {
-    pub fn new(base_path: PathBuf, git_workdir: Option<PathBuf>) -> Result<Self, Error> {
+    pub fn new(
+        base_path: PathBuf,
+        git_workdir: Option<PathBuf>,
+        shared_picker: SharedPicker,
+        shared_frecency: SharedFrecency,
+        mode: FFFMode,
+    ) -> Result<Self, Error> {
        info!(
-            "Initializing background watcher for path: {}",
-            base_path.display()
+            "Initializing background watcher for path: {}, mode: {:?}",
+            base_path.display(),
+            mode,
        );

-        let debouncer = Self::create_debouncer(base_path, git_workdir)?;
+        let debouncer =
+            Self::create_debouncer(base_path, git_workdir, shared_picker, shared_frecency, mode)?;
        info!("Background file watcher initialized successfully");

+        let stop_signal = Arc::new(AtomicBool::new(false));
+        let stop_clone = Arc::clone(&stop_signal);
+
+        // The owner thread keeps the debouncer alive and ensures proper
+        // cleanup: `Debouncer::stop()` joins its internal thread, then the
+        // watcher `Drop` signals its I/O thread to exit.
+        let owner_thread = std::thread::Builder::new()
+            .name("fff-watcher-owner".into())
+            .spawn(move || {
+                while !stop_clone.load(Ordering::Acquire) {
+                    std::thread::park_timeout(Duration::from_secs(1));
+                }
+                // Debouncer::stop() joins the debouncer's event thread, then
+                // drops the watcher (whose Drop signals the I/O thread).
+                debouncer.stop();
+                // On Windows the notify crate discards the ReadDirectoryChangesW
+                // thread's JoinHandle — we cannot join it. Its Drop signals the
+                // thread via semaphore so it exits almost immediately; give the
+                // OS a moment to fully reclaim it.
+                #[cfg(windows)]
+                std::thread::sleep(Duration::from_millis(250));
+            })
+            .expect("failed to spawn fff-watcher-owner thread");
+
        Ok(Self {
-            debouncer: Arc::new(Mutex::new(Some(debouncer))),
+            stop_signal,
+            owner_thread: Some(owner_thread),
        })
    }

    fn create_debouncer(
        base_path: PathBuf,
        git_workdir: Option<PathBuf>,
+        shared_picker: SharedPicker,
+        shared_frecency: SharedFrecency,
+        mode: FFFMode,
    ) -> Result<Debouncer, Error> {
        // do not follow symlinks as then notifiers spawns a bunch of events for symlinked
        // files that could be git ignored, we have to property differentiate those and if
        // the file was edited through a
        let config = Config::default().with_follow_symlinks(false);

+        let git_workdir_for_handler = git_workdir.clone();
        let mut debouncer = new_debouncer_opt(
            DEBOUNCE_TIMEOUT,
            Some(DEBOUNCE_TIMEOUT / 2), // tick rate for the event span
            {
                move |result: DebounceEventResult| match result {
                    Ok(events) => {
-                        handle_debounced_events(events, &git_workdir);
+                        handle_debounced_events(
+                            events,
+                            &git_workdir_for_handler,
+                            &shared_picker,
+                            &shared_frecency,
+                            mode,
+                        );
                    }
                    Err(errors) => {
                        error!("File watcher errors: {:?}", errors);
                    }
                }
            },
-            RecommendedCache::new(),
+            // There is an issue with recommended cache implementation on macos
+            // it keeps track of all the files added to the watcher which is not a problem
+            // for us because any rename to the file will anyway require the removing from the
+            // ordedred index and adding it back with the new name
+            NoCache::new(),
            config,
        )?;

@@ -91,6 +155,11 @@ impl BackgroundWatcher {
                    }
                }
            }
+
+            // In selective mode the .git directory is excluded from the non-ignored
+            // dirs, but we still need to observe changes that affect git status
+            // (staging, unstaging, committing, branch switches, merges, etc.).
+            watch_git_status_paths(&mut debouncer, git_workdir.as_ref());
        }

        info!(
@@ -102,30 +171,34 @@ impl BackgroundWatcher {
        Ok(debouncer)
    }

-    pub fn stop(&self) {
-        if let Ok(Some(debouncer)) = self.debouncer.lock().map(|mut debouncer| debouncer.take()) {
-            drop(debouncer);
-            info!("Background file watcher stopped successfully");
-        } else {
-            error!("Failed to stop background watcher");
+    pub fn stop(&mut self) {
+        self.stop_signal.store(true, Ordering::Release);
+        if let Some(handle) = self.owner_thread.take() {
+            handle.thread().unpark();
+
+            if let Err(e) = handle.join() {
+                error!("Watcher owner thread panicked: {:?}", e);
+            }
        }
+
+        info!("Background file watcher stopped successfully");
    }
 }

 impl Drop for BackgroundWatcher {
    fn drop(&mut self) {
-        if let Ok(mut debouncer_guard) = self.debouncer.lock() {
-            if let Some(debouncer) = debouncer_guard.take() {
-                drop(debouncer);
-            }
-        } else {
-            error!("Failed to acquire debouncer lock to drop");
-        }
+        self.stop();
    }
 }

-#[tracing::instrument(name = "fs_events", skip(events), level = Level::DEBUG)]
-fn handle_debounced_events(events: Vec<DebouncedEvent>, git_workdir: &Option<PathBuf>) {
+#[tracing::instrument(name = "fs_events", skip(events, shared_picker, shared_frecency), level = Level::DEBUG)]
+fn handle_debounced_events(
+    events: Vec<DebouncedEvent>,
+    git_workdir: &Option<PathBuf>,
+    shared_picker: &SharedPicker,
+    shared_frecency: &SharedFrecency,
+    mode: FFFMode,
+) {
    // this will be called very often, we have to minimiy the lock time for file picker
    let repo = git_workdir.as_ref().and_then(|p| Repository::open(p).ok());
    let mut need_full_rescan = false;
@@ -175,14 +248,30 @@ fn handle_debounced_events(events: Vec<DebouncedEvent>, git_workdir: &Option<Pat
                need_full_git_rescan = true;
            }

-            if !should_include_file(path, &repo) {
+            if is_git_file(path) {
                continue;
            }

-            if !path.exists() {
+            // Use a combination of event kind and filesystem state to decide
+            // whether a path is an addition/modification or a removal.
+            //
+            // We cannot rely on `path.exists()` alone because:
+            //   - A freshly created file might not be visible yet (race).
+            //   - macOS FSEvents uses Modify(Name(Any)) for both rename-in
+            //     and rename-out, so we must stat the path to disambiguate.
+            //
+            // We cannot rely on event kind alone because:
+            //   - Remove events are not always emitted (macOS often sends
+            //     Modify(Name(Any)) instead of Remove).
+            let is_removal = matches!(debounced_event.event.kind, EventKind::Remove(_));
+
+            if is_removal || !path.exists() {
                paths_to_remove.push(path.as_path());
            } else {
-                paths_to_add_or_modify.push(path.as_path());
+                // For additions/modifications, still filter gitignored files.
+                if should_include_file(path, &repo) {
+                    paths_to_add_or_modify.push(path.as_path());
+                }
            }
        }

@@ -204,7 +293,7 @@ fn handle_debounced_events(events: Vec<DebouncedEvent>, git_workdir: &Option<Pat

    if need_full_rescan {
        info!(?affected_paths_count, "Triggering full rescan");
-        trigger_full_rescan();
+        trigger_full_rescan(shared_picker, shared_frecency);
        return;
    }

@@ -220,99 +309,164 @@ fn handle_debounced_events(events: Vec<DebouncedEvent>, git_workdir: &Option<Pat
        paths_to_add_or_modify.len()
    );

+    // Apply file index updates (add/remove) unconditionally — these must
+    // happen even when there is no git repository.
+    let files_to_update_git_status =
+        if !paths_to_remove.is_empty() || !paths_to_add_or_modify.is_empty() {
+            debug!(
+                "Applying file index changes: {} to remove, {} to add/modify",
+                paths_to_remove.len(),
+                paths_to_add_or_modify.len(),
+            );
+
+            let apply_changes = |picker: &mut FilePicker| -> Vec<PathBuf> {
+                for path in &paths_to_remove {
+                    let removed = picker.remove_file_by_path(path);
+                    debug!("remove_file_by_path({:?}) -> {}", path, removed);
+                }
+
+                let mut files_to_update = Vec::with_capacity(paths_to_add_or_modify.len());
+                for path in &paths_to_add_or_modify {
+                    let result = picker.on_create_or_modify(path);
+                    match result {
+                        Some(file) => {
+                            debug!(
+                                "on_create_or_modify({:?}) -> Some({})",
+                                path,
+                                file.path.display()
+                            );
+                            files_to_update.push(file.path.clone());
+                        }
+                        None => {
+                            error!("on_create_or_modify({:?}) -> None (file not added!)", path);
+                        }
+                    }
+                }
+                info!(
+                    "apply_changes complete: {} files to update git status",
+                    files_to_update.len()
+                );
+                files_to_update
+            };
+
+            let Ok(mut guard) = shared_picker.write() else {
+                error!("Failed to acquire file picker write lock");
+                return;
+            };
+            let Some(ref mut picker) = *guard else {
+                error!("File picker not initialized");
+                return;
+            };
+            apply_changes(picker)
+        } else {
+            debug!("No file index changes to apply");
+            Vec::new()
+        };
+
+    // AI mode: auto-track frecency for all modified/created files.
+    // Uses a 5-minute cooldown per file to prevent score inflation from rapid
+    // burst edits (AI agents often edit the same file many times in minutes).
+    // This runs after apply_changes so the picker write lock is released.
+    if mode.is_ai() && !paths_to_add_or_modify.is_empty() {
+        let mut tracked_count = 0usize;
+        if let Ok(frecency_guard) = shared_frecency.read()
+            && let Some(ref frecency) = *frecency_guard
+        {
+            for path in &paths_to_add_or_modify {
+                // Skip if this file was tracked less than 5 minutes ago
+                let should_track = match frecency.seconds_since_last_access(path) {
+                    Ok(Some(secs)) => secs >= AI_MODE_COOLDOWN_SECS,
+                    Ok(None) => true, // Never tracked before
+                    Err(_) => true,   // DB error, track anyway
+                };
+                if !should_track {
+                    continue;
+                }
+
+                if let Err(e) = frecency.track_access(path) {
+                    error!("Failed to track frecency for {:?}: {:?}", path, e);
+                } else {
+                    tracked_count += 1;
+                }
+            }
+            if tracked_count > 0 {
+                info!("AI mode: tracked frecency for {} files", tracked_count);
+            }
+        }
+
+        // Update in-memory frecency scores for tracked files
+        if tracked_count > 0
+            && let Ok(mut picker_guard) = shared_picker.write()
+            && let Some(ref mut picker) = *picker_guard
+            && let Ok(frecency_guard) = shared_frecency.read()
+            && let Some(ref frecency) = *frecency_guard
+        {
+            for path in &paths_to_add_or_modify {
+                let _ = picker.update_single_file_frecency(path, frecency);
+            }
+        }
+    }
+
+    // Git status updates require a repository.
    let Some(repo) = repo.as_ref() else {
-        info!("No git repo, skipping git status updates");
+        debug!("No git repo available, skipping git status updates");
        return;
    };

    if need_full_git_rescan {
        info!("Triggering full git rescan");

-        if let Err(e) = FilePicker::refresh_git_status_global() {
+        let result = FilePicker::refresh_git_status(shared_picker, shared_frecency);
+        if let Err(e) = result {
            error!("Failed to refresh git status: {:?}", e);
        }
-
        return;
    }

-    if paths_to_remove.is_empty() && paths_to_add_or_modify.is_empty() {
-        return;
-    }
+    if !files_to_update_git_status.is_empty() {
+        info!(
+            "Fetching git status for {} files",
+            files_to_update_git_status.len()
+        );

-    let files_to_update_git_status = {
-        let Ok(mut file_picker_guard) = FILE_PICKER.write() else {
-            error!("Failed to acquire file picker write lock");
-            return;
-        };
-
-        let Some(ref mut picker) = *file_picker_guard else {
-            error!("File picker not initialized");
-            return;
-        };
-
-        // Apply file removals
-        for path in paths_to_remove {
-            picker.remove_file_by_path(path);
-            // No need to invalidate mmap — the FileItem (and its mmap) is dropped
-        }
-
-        // Apply file additions/modifications and collect paths for git status update
-        let mut files_to_update_git_status = Vec::with_capacity(paths_to_add_or_modify.len());
-        for path in paths_to_add_or_modify {
-            // on_create_or_modify clears the mmap internally when modified time changes
-            if let Some(file) = picker.on_create_or_modify(path) {
-                files_to_update_git_status.push(file.path.clone());
+        let status = match GitStatusCache::git_status_for_paths(repo, &files_to_update_git_status) {
+            Ok(status) => status,
+            Err(e) => {
+                tracing::error!(?e, "Failed to query git status");
+                return;
            }
-        }
+        };

-        files_to_update_git_status
-    };
-
-    info!(
-        "Fetching git status for {} files",
-        files_to_update_git_status.len()
-    );
-
-    let status = match GitStatusCache::git_status_for_paths(repo, &files_to_update_git_status) {
-        Ok(status) => status,
-        Err(e) => {
-            tracing::error!(?e, "Failed to query git statue");
-            return;
-        }
-    };
-
-    // only lock the picker for theshortest possitble time
-    if let Ok(mut file_picker_guard) = FILE_PICKER.write()
-        && let Some(ref mut picker) = *file_picker_guard
-    {
-        if let Err(e) = picker.update_git_statuses(status) {
-            error!("Failed to update git statuses: {:?}", e);
+        if let Ok(mut guard) = shared_picker.write()
+            && let Some(ref mut picker) = *guard
+        {
+            if let Err(e) = picker.update_git_statuses(status, shared_frecency) {
+                error!("Failed to update git statuses: {:?}", e);
+            } else {
+                info!("Successfully updated git statuses in picker");
+            }
        } else {
-            info!("Successfully updated git statuses in picker");
+            error!("Failed to acquire picker lock for git status update");
        }
-    } else {
-        error!("Failed to acquire picker lock for git status update");
    }
 }

-fn trigger_full_rescan() {
+fn trigger_full_rescan(shared_picker: &SharedPicker, shared_frecency: &SharedFrecency) {
    info!("Triggering full filesystem rescan");

    // Note: no need to clear mmaps — they are backed by the kernel page cache
    // and automatically reflect file changes. Old FileItems (and their mmaps)
    // are dropped when the picker rebuilds its file list.

-    let Ok(mut file_picker_guard) = FILE_PICKER.write() else {
+    let Ok(mut guard) = shared_picker.write() else {
        error!("Failed to acquire file picker write lock for full rescan");
        return;
    };
-
-    let Some(ref mut picker) = *file_picker_guard else {
+    let Some(ref mut picker) = *guard else {
        error!("File picker not initialized, cannot trigger rescan");
        return;
    };
-
-    if let Err(e) = picker.trigger_rescan() {
+    if let Err(e) = picker.trigger_rescan(shared_frecency) {
        error!("Failed to trigger full rescan: {:?}", e);
    } else {
        info!("Full filesystem rescan completed successfully");
@@ -320,12 +474,17 @@ fn trigger_full_rescan() {
 }

 fn should_include_file(path: &Path, repo: &Option<Repository>) -> bool {
-    if !path.is_file() || is_git_file(path) {
+    // Directories are not indexed — only regular files (and symlinks to files).
+    if path.is_dir() {
        return false;
    }

-    repo.as_ref()
-        .is_some_and(|repo| repo.is_path_ignored(path) == Ok(false))
+    // If there is a git repo, respect its ignore rules.
+    // If there is no repo (or the check fails), include the file.
+    match repo.as_ref() {
+        Some(repo) => repo.is_path_ignored(path) != Ok(true),
+        None => true,
+    }
 }

 #[inline]
@@ -375,6 +534,40 @@ fn is_ignore_definition_path(path: &Path) -> bool {
    )
 }

+fn watch_git_status_paths(debouncer: &mut Debouncer, git_workdir: Option<&PathBuf>) {
+    let Some(workdir) = git_workdir else {
+        return;
+    };
+
+    let git_dir = workdir.join(".git");
+    if !git_dir.is_dir() {
+        return;
+    }
+
+    // Watch .git/ non-recursively to catch top-level files:
+    // index, index.lock, HEAD, packed-refs, MERGE_HEAD, CHERRY_PICK_HEAD, REVERT_HEAD
+    if let Err(e) = debouncer.watch(&git_dir, RecursiveMode::NonRecursive) {
+        warn!("Failed to watch .git directory: {}", e);
+        return;
+    }
+
+    // Watch refs/ recursively to catch branch/tag changes
+    let refs_dir = git_dir.join("refs");
+    if refs_dir.is_dir()
+        && let Err(e) = debouncer.watch(&refs_dir, RecursiveMode::Recursive)
+    {
+        warn!("Failed to watch .git/refs: {}", e);
+    }
+
+    // Watch info/ non-recursively for exclude and sparse-checkout
+    let info_dir = git_dir.join("info");
+    if info_dir.is_dir()
+        && let Err(e) = debouncer.watch(&info_dir, RecursiveMode::NonRecursive)
+    {
+        warn!("Failed to watch .git/info: {}", e);
+    }
+}
+
 /// Collects immediate non-ignored subdirectories of `base_path` using the `ignore` crate
 /// to respect .gitignore, .ignore, and global gitignore rules. This is used to set up
 /// selective file watching — only non-ignored directories get a recursive watcher,
@@ -9,7 +9,6 @@
 use ahash::AHashSet;
 use fff_query_parser::{Constraint, GitStatusFilter};
 use smallvec::SmallVec;
-use zlob::{ZlobFlags, zlob_match_paths};

 use crate::git::is_modified_status;

@@ -33,6 +32,28 @@ pub trait Constrainable {
    fn git_status(&self) -> Option<git2::Status>;
 }

+/// Check if a relative path ends with the given suffix at a `/` boundary (case-insensitive).
+///
+/// Returns `true` when the path equals the suffix or the character before the suffix
+/// in the path is `/`. This ensures partial directory-name matches are rejected.
+///
+/// Examples:
+/// - `path_ends_with_suffix("libswscale/input.c", "libswscale/input.c")` → true (exact)
+/// - `path_ends_with_suffix("foo/libswscale/input.c", "libswscale/input.c")` → true (suffix)
+/// - `path_ends_with_suffix("xlibswscale/input.c", "libswscale/input.c")` → false (no boundary)
+#[inline]
+pub fn path_ends_with_suffix(path: &str, suffix: &str) -> bool {
+    if path.len() < suffix.len() {
+        return false;
+    }
+    let start = path.len() - suffix.len();
+    if !path[start..].eq_ignore_ascii_case(suffix) {
+        return false;
+    }
+    // Exact match, or the character before is /
+    start == 0 || path.as_bytes()[start - 1] == b'/'
+}
+
 /// Check if file extension matches (without allocation)
 #[inline]
 pub fn file_has_extension(file_name: &str, ext: &str) -> bool {
@@ -45,12 +66,14 @@ pub fn file_has_extension(file_name: &str, ext: &str) -> bool {
 }

 /// Check if path contains segment (without allocation)
+/// Supports both single segments ("src") and multi-segment paths ("libswscale/aarch64").
+/// For "libswscale/aarch64", checks that these appear as consecutive path components.
 #[inline]
 pub fn path_contains_segment(path: &str, segment: &str) -> bool {
    let path_bytes = path.as_bytes();
    let segment_len = segment.len();

-    // Check segment/ at start
+    // Check segment/ at start of path
    if path.len() > segment_len
        && path_bytes.get(segment_len) == Some(&b'/')
        && path[..segment_len].eq_ignore_ascii_case(segment)
@@ -102,6 +125,7 @@ fn item_matches_constraint_at_index<T: Constrainable>(
            return if negate { !result } else { result };
        }
        Constraint::PathSegment(segment) => path_contains_segment(item.relative_path(), segment),
+        Constraint::FilePath(suffix) => path_ends_with_suffix(item.relative_path(), suffix),
        Constraint::GitStatus(status_filter) => match (item.git_status(), status_filter) {
            (Some(status), GitStatusFilter::Modified) => is_modified_status(status),
            (Some(status), GitStatusFilter::Untracked) => status.contains(git2::Status::WT_NEW),
@@ -235,32 +259,8 @@ fn collect_glob_indices<'a>(
 ) {
    match constraint {
        Constraint::Glob(pattern) => {
-            if let Ok(Some(matches)) = zlob_match_paths(pattern, paths, ZlobFlags::RECOMMENDED) {
-                let matched_set: AHashSet<usize> =
-                    matches.iter().map(|s| s.as_ptr() as usize).collect();
-
-                let indices: AHashSet<usize> = if paths.len() >= PAR_THRESHOLD {
-                    use rayon::prelude::*;
-                    paths
-                        .par_iter()
-                        .enumerate()
-                        .filter(|(_, p)| matched_set.contains(&(p.as_ptr() as usize)))
-                        .map(|(i, _)| i)
-                        .collect::<Vec<_>>()
-                        .into_iter()
-                        .collect()
-                } else {
-                    paths
-                        .iter()
-                        .enumerate()
-                        .filter(|(_, p)| matched_set.contains(&(p.as_ptr() as usize)))
-                        .map(|(i, _)| i)
-                        .collect()
-                };
-                results.push((is_negated, indices));
-            } else {
-                results.push((is_negated, AHashSet::new()));
-            }
+            let indices = match_glob_pattern(pattern, paths);
+            results.push((is_negated, indices));
        }
        Constraint::Not(inner) => {
            collect_glob_indices(inner, paths, results, !is_negated);
@@ -269,6 +269,66 @@ fn collect_glob_indices<'a>(
    }
 }

+/// Match a glob pattern against a list of paths, returning the set of matching indices.
+///
+/// When the `zlob` feature is enabled, delegates to `zlob::zlob_match_paths` (Zig-compiled
+/// C library, fastest). Otherwise falls back to `globset::Glob` (pure Rust).
+#[cfg(feature = "zlob")]
+fn match_glob_pattern(pattern: &str, paths: &[&str]) -> AHashSet<usize> {
+    let Ok(Some(matches)) = zlob::zlob_match_paths(pattern, paths, zlob::ZlobFlags::RECOMMENDED)
+    else {
+        return AHashSet::new();
+    };
+
+    let matched_set: AHashSet<usize> = matches.iter().map(|s| s.as_ptr() as usize).collect();
+
+    if paths.len() >= PAR_THRESHOLD {
+        use rayon::prelude::*;
+        paths
+            .par_iter()
+            .enumerate()
+            .filter(|(_, p)| matched_set.contains(&(p.as_ptr() as usize)))
+            .map(|(i, _)| i)
+            .collect::<Vec<_>>()
+            .into_iter()
+            .collect()
+    } else {
+        paths
+            .iter()
+            .enumerate()
+            .filter(|(_, p)| matched_set.contains(&(p.as_ptr() as usize)))
+            .map(|(i, _)| i)
+            .collect()
+    }
+}
+
+#[cfg(not(feature = "zlob"))]
+fn match_glob_pattern(pattern: &str, paths: &[&str]) -> AHashSet<usize> {
+    let Ok(glob) = globset::Glob::new(pattern) else {
+        return AHashSet::new();
+    };
+    let matcher = glob.compile_matcher();
+
+    if paths.len() >= PAR_THRESHOLD {
+        use rayon::prelude::*;
+        paths
+            .par_iter()
+            .enumerate()
+            .filter(|(_, p)| matcher.is_match(p))
+            .map(|(i, _)| i)
+            .collect::<Vec<_>>()
+            .into_iter()
+            .collect()
+    } else {
+        paths
+            .iter()
+            .enumerate()
+            .filter(|(_, p)| matcher.is_match(p))
+            .map(|(i, _)| i)
+            .collect()
+    }
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;
@@ -312,8 +372,78 @@ mod tests {
        // Should not match filename
        assert!(!path_contains_segment("lib/src", "src"));

+        // Multi-segment constraints
+        assert!(path_contains_segment(
+            "libswscale/aarch64/input.S",
+            "libswscale/aarch64"
+        ));
+        assert!(path_contains_segment(
+            "foo/libswscale/aarch64/input.S",
+            "libswscale/aarch64"
+        ));
+        assert!(path_contains_segment(
+            "foo/LibSwscale/AArch64/input.S",
+            "libswscale/aarch64"
+        )); // case-insensitive
+        assert!(!path_contains_segment(
+            "xlibswscale/aarch64/input.S",
+            "libswscale/aarch64"
+        )); // partial match at start
+        assert!(!path_contains_segment(
+            "foo/libswscale/aarch64x/input.S",
+            "libswscale/aarch64"
+        )); // partial match at end
+        assert!(path_contains_segment(
+            "crates/fff-core/src/grep.rs",
+            "fff-core/src"
+        ));
+
        // Edge cases
        assert!(!path_contains_segment("", "src"));
        assert!(!path_contains_segment("src", "src")); // no trailing slash
    }
+
+    #[test]
+    fn test_path_ends_with_suffix() {
+        // Exact match
+        assert!(path_ends_with_suffix(
+            "libswscale/input.c",
+            "libswscale/input.c"
+        ));
+
+        // Suffix match at / boundary
+        assert!(path_ends_with_suffix(
+            "foo/libswscale/input.c",
+            "libswscale/input.c"
+        ));
+
+        // Deep nesting
+        assert!(path_ends_with_suffix(
+            "a/b/c/libswscale/input.c",
+            "libswscale/input.c"
+        ));
+
+        // No boundary — partial directory name
+        assert!(!path_ends_with_suffix(
+            "xlibswscale/input.c",
+            "libswscale/input.c"
+        ));
+
+        // Case insensitive
+        assert!(path_ends_with_suffix(
+            "foo/LibSwscale/Input.C",
+            "libswscale/input.c"
+        ));
+
+        // Single file name
+        assert!(path_ends_with_suffix("input.c", "input.c"));
+        assert!(!path_ends_with_suffix("xinput.c", "input.c"));
+
+        // Suffix longer than path
+        assert!(!path_ends_with_suffix("input.c", "foo/input.c"));
+
+        // Simple path
+        assert!(path_ends_with_suffix("src/main.rs", "src/main.rs"));
+        assert!(path_ends_with_suffix("crates/src/main.rs", "src/main.rs"));
+    }
 }
@@ -1,10 +1,44 @@
+//! Core file picker: filesystem indexing, background watching, and fuzzy search.
+//!
+//! [`FilePicker`] is the central component of fff-search. It:
+//!
+//! 1. **Indexes** a directory tree in a background thread, collecting every
+//!    non-ignored file into a path-sorted `Vec<FileItem>`.
+//! 2. **Watches** the filesystem via the `notify` crate, applying
+//!    create/modify/delete events to the index in real time.
+//! 3. **Owns files**: Provides a values for search and provides a good entry point for
+//!    fuzzy search and live grep
+//!
+//! # Lifecycle
+//!
+//! ```text
+//!   new_with_shared_state()
+//!     │
+//!     ├─> background scan thread ──> populates SharedPicker
+//!     └─> file-system watcher    ──> live updates SharedPicker
+//!
+//!   fuzzy_search()   <── static, borrows &[FileItem]
+//!   grep()           <── static, borrows &[FileItem] (live content search)
+//!   trigger_rescan() <── synchronous re-index
+//!   cancel()         <── shuts down background work
+//! ```
+//!
+//! # Thread Safety
+//!
+//! `FilePicker` itself is **not** `Sync`!
+//! all concurrent access goes through [`SharedPicker`](crate::SharedPicker) .
+//! The background scanner and watcher acquire write locks only when mutating
+//! the file index, so read-heavy search workloads rarely contend.
+
 use crate::background_watcher::BackgroundWatcher;
 use crate::error::Error;
 use crate::frecency::FrecencyTracker;
 use crate::git::GitStatusCache;
-use crate::query_tracker::QueryMatchEntry;
+use crate::grep::{GrepResult, GrepSearchOptions, grep_search};
+use crate::query_tracker::QueryTracker;
 use crate::score::match_and_score_files;
 use crate::types::{FileItem, PaginationArgs, ScoringContext, SearchResult};
+use crate::{SharedFrecency, SharedPicker};
 use fff_query_parser::FFFQuery;
 use git2::{Repository, Status, StatusOptions};
 use rayon::prelude::*;
@@ -18,33 +52,28 @@ use std::sync::{
 use std::time::SystemTime;
 use tracing::{Level, debug, error, info, warn};

-use crate::{FILE_PICKER, FRECENCY};
-
-/// Detect if a file is binary by checking for NUL bytes in the first 512 bytes.
-/// This is the same heuristic used by git and grep — simple, fast, and sufficient.
-#[inline]
-fn detect_binary(path: &Path, size: u64) -> bool {
-    // Empty files are not binary
-    if size == 0 {
-        return false;
-    }
-
-    let Ok(file) = std::fs::File::open(path) else {
-        return false;
-    };
-    let mut reader = std::io::BufReader::with_capacity(1024, file);
-
-    let mut buf = [0u8; 512];
-    let n = reader.read(&mut buf).unwrap_or(0);
-    buf[..n].contains(&0)
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum FFFMode {
+    #[default]
+    Neovim,
+    Ai,
 }

-#[derive(Debug, Clone, Copy)]
+impl FFFMode {
+    pub fn is_ai(self) -> bool {
+        self == FFFMode::Ai
+    }
+}
+
+/// Configuration for a single fuzzy search invocation.
+///
+/// Passed to [`FilePicker::fuzzy_search`] to control threading, pagination,
+/// and scoring behavior.
+#[derive(Debug, Clone, Copy, Default)]
 pub struct FuzzySearchOptions<'a> {
    pub max_threads: usize,
    pub current_file: Option<&'a str>,
    pub project_path: Option<&'a Path>,
-    pub last_same_query_match: Option<&'a QueryMatchEntry>,
    pub combo_boost_score_multiplier: i32,
    pub min_combo_count: u32,
    pub pagination: PaginationArgs,
@@ -52,7 +81,8 @@ pub struct FuzzySearchOptions<'a> {

 #[derive(Debug, Clone)]
 struct FileSync {
-    pub files: Vec<FileItem>,
+    /// Files sorted by path for binary search
+    files: Vec<FileItem>,
    pub git_workdir: Option<PathBuf>,
 }

@@ -64,9 +94,68 @@ impl FileSync {
        }
    }

+    /// Get all files (read-only). Files are sorted by path.
+    #[inline]
+    fn files(&self) -> &[FileItem] {
+        &self.files
+    }
+
+    fn get_file(&self, index: usize) -> Option<&FileItem> {
+        self.files.get(index)
+    }
+
+    /// Get mutable file at index
+    #[inline]
+    fn get_file_mut(&mut self, index: usize) -> Option<&mut FileItem> {
+        self.files.get_mut(index)
+    }
+
+    /// Find file index by path using binary search - O(log n)
+    #[inline]
    fn find_file_index(&self, path: &Path) -> Result<usize, usize> {
-        self.files
-            .binary_search_by(|file| file.path.as_os_str().cmp(path.as_os_str()))
+        self.files.binary_search_by(|f| f.path.as_path().cmp(path))
+    }
+
+    /// Get file count
+    #[inline]
+    #[allow(dead_code)]
+    fn len(&self) -> usize {
+        self.files.len()
+    }
+
+    /// Insert a file at position. Simple - no HashMap to maintain!
+    fn insert_file(&mut self, position: usize, file: FileItem) {
+        self.files.insert(position, file);
+    }
+
+    /// Remove file at index. Simple - no HashMap to maintain!
+    fn remove_file(&mut self, index: usize) {
+        if index < self.files.len() {
+            self.files.remove(index);
+        }
+    }
+
+    /// Remove files matching predicate.
+    /// Returns number of files removed.
+    fn retain_files<F>(&mut self, predicate: F) -> usize
+    where
+        F: FnMut(&FileItem) -> bool,
+    {
+        let initial_len = self.files.len();
+        self.files.retain(predicate);
+        initial_len - self.files.len()
+    }
+
+    /// Insert a file in sorted order (by path).
+    /// Returns true if inserted, false if file already exists.
+    fn insert_file_sorted(&mut self, file: FileItem) -> bool {
+        match self.find_file_index(&file.path) {
+            Ok(_) => false, // File already exists
+            Err(position) => {
+                self.insert_file(position, file);
+                true
+            }
+        }
    }
 }

@@ -110,26 +199,24 @@ impl FileItem {
        )
    }

-    pub fn update_frecency_scores(&mut self, tracker: &FrecencyTracker) -> Result<(), Error> {
-        self.access_frecency_score = tracker.get_access_score(&self.path);
+    pub fn update_frecency_scores(
+        &mut self,
+        tracker: &FrecencyTracker,
+        mode: FFFMode,
+    ) -> Result<(), Error> {
+        self.access_frecency_score = tracker.get_access_score(&self.path, mode);
        self.modification_frecency_score =
-            tracker.get_modification_score(self.modified, self.git_status);
+            tracker.get_modification_score(self.modified, self.git_status, mode);
        self.total_frecency_score = self.access_frecency_score + self.modification_frecency_score;

        Ok(())
    }
-
-    /// Locks the tracker and updates frecensy score for one file. If need multiple files updates
-    /// use `update_frecency_scores` instead.
-    pub fn update_frecency_scores_global(&mut self) -> Result<(), Error> {
-        let Some(ref frecency) = *FRECENCY.read().map_err(|_| Error::AcquireFrecencyLock)? else {
-            return Ok(());
-        };
-
-        self.update_frecency_scores(frecency)
-    }
 }

+/// The main file picker engine storage
+///
+/// It maintains an in memory index of all the files that are resent in the file system
+/// and borrows them to perform the search
 pub struct FilePicker {
    base_path: PathBuf,
    sync_data: FileSync,
@@ -137,6 +224,8 @@ pub struct FilePicker {
    scanned_files_count: Arc<AtomicUsize>,
    background_watcher: Option<BackgroundWatcher>,
    warmup_mmap_cache: bool,
+    cancelled: Arc<AtomicBool>,
+    mode: FFFMode,
 }

 impl std::fmt::Debug for FilePicker {
@@ -162,28 +251,38 @@ impl FilePicker {
        self.warmup_mmap_cache
    }

+    pub fn mode(&self) -> FFFMode {
+        self.mode
+    }
+
    pub fn git_root(&self) -> Option<&Path> {
        self.sync_data.git_workdir.as_deref()
    }

+    /// Get all indexed files sorted by path.
+    /// Note: Files are stored sorted by PATH for efficient insert/remove.
+    /// For frecency-sorted results, use search() which sorts matched results.
    pub fn get_files(&self) -> &[FileItem] {
-        &self.sync_data.files
+        self.sync_data.files()
    }

-    pub fn new(base_path: String) -> Result<Self, Error> {
-        Self::with_options(base_path, false)
-    }
-
-    /// Create a new FilePicker with explicit options.
+    /// Create a new FilePicker and place it into the provided shared handle.
    ///
-    /// When `warmup_mmap_cache` is `true`, all non-binary files will be mmap'd
-    /// and their pages paged in immediately after the initial scan completes.
-    /// This makes the first grep search as fast as subsequent ones at the cost
-    /// of a longer startup time and higher initial memory pressure.
-    pub fn with_options(base_path: String, warmup_mmap_cache: bool) -> Result<Self, Error> {
+    /// The background scan thread and file-system watcher write into the
+    /// provided `SharedPicker` and read frecency data from the provided
+    /// `SharedFrecency`.
+    ///
+    /// Multiple independent instances can coexist in the same process.
+    pub fn new_with_shared_state(
+        base_path: String,
+        warmup_mmap_cache: bool,
+        mode: FFFMode,
+        shared_picker: SharedPicker,
+        shared_frecency: SharedFrecency,
+    ) -> Result<(), Error> {
        info!(
-            "Initializing FilePicker with base_path: {}, warmup: {}",
-            base_path, warmup_mmap_cache
+            "Initializing FilePicker with base_path: {}, warmup: {}, mode: {:?}",
+            base_path, warmup_mmap_cache, mode
        );
        let path = PathBuf::from(&base_path);
        if !path.exists() {
@@ -191,51 +290,68 @@ impl FilePicker {
            return Err(Error::InvalidPath(path));
        }

-        let scan_signal = Arc::new(AtomicBool::new(false));
+        // Initialize scan_signal to `true` so that any `wait_for_scan` call
+        // that races with the background thread sees "scanning in progress"
+        // rather than a stale `false` (the thread hasn't started yet).
+        let scan_signal = Arc::new(AtomicBool::new(true));
        let synced_files_count = Arc::new(AtomicUsize::new(0));
+        let cancelled = Arc::new(AtomicBool::new(false));

-        let picker = Self {
+        let picker = FilePicker {
            base_path: path.clone(),
            sync_data: FileSync::new(),
            is_scanning: Arc::clone(&scan_signal),
            scanned_files_count: Arc::clone(&synced_files_count),
            background_watcher: None,
            warmup_mmap_cache,
+            cancelled: Arc::clone(&cancelled),
+            mode,
        };

+        // Place the picker into the shared handle before spawning the
+        // background thread so the thread can find it immediately.
+        {
+            let mut guard = shared_picker.write().map_err(|_| Error::AcquireItemLock)?;
+            *guard = Some(picker);
+        }
+
        spawn_scan_and_watcher(
            path.clone(),
            Arc::clone(&scan_signal),
            Arc::clone(&synced_files_count),
            warmup_mmap_cache,
+            mode,
+            shared_picker,
+            shared_frecency,
+            cancelled,
        );

-        Ok(picker)
+        Ok(())
    }

    /// Perform fuzzy search on files with a pre-parsed query.
    ///
-    /// The query should be parsed using `QueryParser::parse()` before calling this function.
-    /// This allows the caller to handle location parsing and other preprocessing.
+    /// The query should be parsed using [`FFFQuery`]::parse() before calling
+    /// this function. If a [`QueryTracker`] is provided, the search will
+    /// automatically look up the last selected file for this query and apply
+    /// combo-boost scoring.
    ///
-    /// # Arguments
-    /// * `files` - Slice of files to search
-    /// * `query` - The raw query string (used for max_typos calculation and debugging)
-    /// * `parsed` - Pre-parsed query result (can be None for simple single-token queries)
-    /// * `options` - Search options including pagination, threading, and scoring parameters
-    ///
-    /// # Returns
-    /// SearchResult containing matched files, scores, and location information
-    pub fn fuzzy_search<'a>(
+    pub fn fuzzy_search<'a, 'q>(
        files: &'a [FileItem],
-        query: &'a str,
-        parsed: Option<FFFQuery<'a>>,
-        options: FuzzySearchOptions<'a>,
+        query: &'q FFFQuery<'q>,
+        query_tracker: Option<&QueryTracker>,
+        options: FuzzySearchOptions<'q>,
    ) -> SearchResult<'a> {
-        let max_threads = options.max_threads.max(1);
+        let max_threads = if options.max_threads == 0 {
+            std::thread::available_parallelism()
+                .map(|n| n.get())
+                .unwrap_or(4)
+        } else {
+            options.max_threads
+        };
+
        debug!(
-            ?query,
-            parsed_is_some = parsed.is_some(),
+            raw_query = ?query.raw_query,
            pagination = ?options.pagination,
            ?max_threads,
            current_file = ?options.current_file,
@@ -243,41 +359,48 @@ impl FilePicker {
        );

        let total_files = files.len();
-
-        // Extract location from parsed query
-        let location = parsed.as_ref().and_then(|p| p.location);
+        let location = query.location;

        // Get effective query for max_typos calculation (without location suffix)
-        let effective_query = match &parsed {
-            Some(p) => match &p.fuzzy_query {
-                fff_query_parser::FuzzyQuery::Text(t) => *t,
-                fff_query_parser::FuzzyQuery::Parts(parts) if !parts.is_empty() => parts[0],
-                _ => query.trim(),
-            },
-            None => query.trim(),
+        let effective_query = match &query.fuzzy_query {
+            fff_query_parser::FuzzyQuery::Text(t) => *t,
+            fff_query_parser::FuzzyQuery::Parts(parts) if !parts.is_empty() => parts[0],
+            _ => query.raw_query.trim(),
        };

        // small queries with a large number of results can match absolutely everything
        let max_typos = (effective_query.len() as u16 / 4).clamp(2, 6);
+        // Look up the last file selected for this query (combo-boost scoring)
+        let last_same_query_entry =
+            query_tracker
+                .zip(options.project_path)
+                .and_then(|(tracker, project_path)| {
+                    tracker
+                        .get_last_query_entry(
+                            query.raw_query,
+                            project_path,
+                            options.min_combo_count,
+                        )
+                        .ok()
+                        .flatten()
+                });

        let context = ScoringContext {
-            raw_query: query,
-            parsed_query: parsed,
-            project_path: options.project_path,
+            query,
            max_typos,
            max_threads,
+            project_path: options.project_path,
            current_file: options.current_file,
-            last_same_query_match: options.last_same_query_match,
+            last_same_query_match: last_same_query_entry,
            combo_boost_score_multiplier: options.combo_boost_score_multiplier,
            min_combo_count: options.min_combo_count,
            pagination: options.pagination,
        };

        let time = std::time::Instant::now();
-
        let (items, scores, total_matched) = match_and_score_files(files, &context);

-        debug!(
+        info!(
            ?query,
            completed_in = ?time.elapsed(),
            total_matched,
@@ -295,6 +418,16 @@ impl FilePicker {
        }
    }

+    /// Perform a live grep search across indexed files with a pre-parsed query.
+    pub fn grep<'a>(
+        files: &'a [FileItem],
+        query: &FFFQuery<'_>,
+        options: &GrepSearchOptions,
+    ) -> GrepResult<'a> {
+        grep_search(files, query, options)
+    }
+
+    // Returns an ongoing or finisshed scan progress
    pub fn get_scan_progress(&self) -> ScanProgress {
        let scanned_count = self.scanned_files_count.load(Ordering::Relaxed);
        let is_scanning = self.is_scanning.load(Ordering::Relaxed);
@@ -304,38 +437,47 @@ impl FilePicker {
        }
    }

-    pub fn update_git_statuses(&mut self, status_cache: GitStatusCache) -> Result<(), Error> {
+    /// Update git statuses for files, using the provided shared frecency tracker.
+    pub fn update_git_statuses(
+        &mut self,
+        status_cache: GitStatusCache,
+        shared_frecency: &SharedFrecency,
+    ) -> Result<(), Error> {
        debug!(
            statuses_count = status_cache.statuses_len(),
            "Updating git status",
        );

-        let frecency = FRECENCY.read().map_err(|_| Error::AcquireFrecencyLock)?;
+        let mode = self.mode;
+        let frecency = shared_frecency
+            .read()
+            .map_err(|_| Error::AcquireFrecencyLock)?;
        status_cache
            .into_iter()
            .try_for_each(|(path, status)| -> Result<(), Error> {
                if let Some(file) = self.get_mut_file_by_path(&path) {
                    file.git_status = Some(status);
-
-                    if let Some(frecency) = frecency.as_ref() {
-                        file.update_frecency_scores(frecency)?;
+                    if let Some(ref f) = *frecency {
+                        file.update_frecency_scores(f, mode)?;
                    }
                } else {
                    error!(?path, "Couldn't update the git status for path");
                }
-
                Ok(())
            })?;

        Ok(())
    }

-    /// Fetches all the git statuses first and updates the global FILE_PICKER
-    /// with the new statuses with the smallest possible lock time.
-    pub fn refresh_git_status_global() -> Result<usize, Error> {
+    /// Refreshes git statuses using the provided shared picker and frecency handles.
+    pub fn refresh_git_status(
+        shared_picker: &SharedPicker,
+        shared_frecency: &SharedFrecency,
+    ) -> Result<usize, Error> {
        let git_status = {
-            let Some(ref picker) = *FILE_PICKER.read().map_err(|_| Error::AcquireItemLock)? else {
-                return Err(Error::FilePickerMissing)?;
+            let guard = shared_picker.read().map_err(|_| Error::AcquireItemLock)?;
+            let Some(ref picker) = *guard else {
+                return Err(Error::FilePickerMissing);
            };

            debug!(
@@ -343,29 +485,22 @@ impl FilePicker {
                picker.git_root()
            );

-            // we keep here readonly lock but allowing querying the index while it scan lasts
            GitStatusCache::read_git_status(
                picker.git_root(),
                StatusOptions::new()
                    .include_untracked(true)
                    .recurse_untracked_dirs(true)
-                    // when manually refreshing git status we want to include all unmodified file
-                    // to make sure that their status is correctly updated when user
-                    // commited/stashed/removed changes
                    .include_unmodified(true)
                    .exclude_submodules(true),
            )
        };

-        let mut file_picker = FILE_PICKER.write().map_err(|_| Error::AcquireItemLock)?;
-        let picker = file_picker
-            .as_mut()
-            .ok_or_else(|| Error::FilePickerMissing)?;
+        let mut guard = shared_picker.write().map_err(|_| Error::AcquireItemLock)?;
+        let picker = guard.as_mut().ok_or(Error::FilePickerMissing)?;

        let statuses_count = if let Some(git_status) = git_status {
            let count = git_status.statuses_len();
-            picker.update_git_statuses(git_status)?;
-
+            picker.update_git_statuses(git_status, shared_frecency)?;
            count
        } else {
            0
@@ -380,9 +515,9 @@ impl FilePicker {
        frecency_tracker: &FrecencyTracker,
    ) -> Result<(), Error> {
        if let Ok(index) = self.sync_data.find_file_index(file_path.as_ref())
-            && let Some(file) = self.sync_data.files.get_mut(index)
+            && let Some(file) = self.sync_data.get_file_mut(index)
        {
-            file.update_frecency_scores(frecency_tracker)?;
+            file.update_frecency_scores(frecency_tracker, self.mode)?;
        }

        Ok(())
@@ -392,35 +527,38 @@ impl FilePicker {
        self.sync_data
            .find_file_index(path.as_ref())
            .ok()
-            .and_then(|index| self.sync_data.files.get(index))
+            .and_then(|index| self.sync_data.files().get(index))
    }

    pub fn get_mut_file_by_path(&mut self, path: impl AsRef<Path>) -> Option<&mut FileItem> {
        self.sync_data
            .find_file_index(path.as_ref())
            .ok()
-            .and_then(|index| self.sync_data.files.get_mut(index))
+            .and_then(|index| self.sync_data.get_file_mut(index))
    }

    /// Add a file to the picker's files in sorted order (used by background watcher)
    pub fn add_file_sorted(&mut self, file: FileItem) -> Option<&FileItem> {
-        match self
-            .sync_data
-            .files
-            .binary_search_by(|f| f.relative_path.cmp(&file.relative_path))
-        {
-            Ok(position) => {
-                warn!(
-                    "Trying to insert a file that already exists: {}",
-                    file.relative_path
-                );
+        let path = file.path.clone();

-                self.sync_data.files.get(position)
-            }
-            Err(position) => {
-                self.sync_data.files.insert(position, file);
-                self.sync_data.files.get(position)
-            }
+        if self.sync_data.insert_file_sorted(file) {
+            // File was inserted, look it up
+            self.sync_data
+                .find_file_index(&path)
+                .ok()
+                .and_then(|idx| self.sync_data.get_file_mut(idx))
+                .map(|file_mut| &*file_mut) // Convert &mut to &
+        } else {
+            // File already exists
+            warn!(
+                "Trying to insert a file that already exists: {}",
+                path.display()
+            );
+            self.sync_data
+                .find_file_index(&path)
+                .ok()
+                .and_then(|idx| self.sync_data.get_file_mut(idx))
+                .map(|file_mut| &*file_mut) // Convert &mut to &
        }
    }

@@ -429,8 +567,12 @@ impl FilePicker {
        let path = path.as_ref();
        match self.sync_data.find_file_index(path) {
            Ok(pos) => {
-                // safe to read because we are in lock and binary search returned valid position
-                let file = &mut self.sync_data.files[pos];
+                debug!(
+                    "on_create_or_modify: file EXISTS at index {}, updating metadata",
+                    pos
+                );
+                // File exists - update its metadata (doesn't change indices, safe)
+                let file = self.sync_data.get_file_mut(pos)?;

                let modified = match std::fs::metadata(path) {
                    Ok(metadata) => metadata
@@ -456,21 +598,40 @@ impl FilePicker {
                    }
                }

-                Some(file)
+                Some(&*file) // Convert &mut to &
            }
            Err(pos) => {
-                let file_item = FileItem::new(path.to_path_buf(), &self.base_path, None);
-                self.sync_data.files.insert(pos, file_item);
+                debug!(
+                    "on_create_or_modify: file NEW, inserting at index {} (total files: {})",
+                    pos,
+                    self.sync_data.files().len()
+                );

-                self.sync_data.files.get(pos)
+                let file_item = FileItem::new(path.to_path_buf(), &self.base_path, None);
+                let path_buf = file_item.path.clone();
+
+                self.sync_data.insert_file(pos, file_item);
+                let result = self.sync_data.get_file(pos);
+
+                if result.is_none() {
+                    error!(
+                        "on_create_or_modify: FAILED to find file after insert! path={:?}",
+                        path_buf
+                    );
+                } else {
+                    debug!("on_create_or_modify: successfully inserted and found file");
+                }
+
+                result
            }
        }
    }

    pub fn remove_file_by_path(&mut self, path: impl AsRef<Path>) -> bool {
-        match self.sync_data.find_file_index(path.as_ref()) {
+        let path = path.as_ref();
+        match self.sync_data.find_file_index(path) {
            Ok(index) => {
-                self.sync_data.files.remove(index);
+                self.sync_data.remove_file(index);
                true
            }
            Err(_) => false,
@@ -480,22 +641,23 @@ impl FilePicker {
    // TODO make this O(n)
    pub fn remove_all_files_in_dir(&mut self, dir: impl AsRef<Path>) -> usize {
        let dir_path = dir.as_ref();
-        let initial_len = self.sync_data.files.len();
-
+        // Use the safe retain_files method which maintains both indices
        self.sync_data
-            .files
-            .retain(|file| !file.path.starts_with(dir_path));
+            .retain_files(|file| !file.path.starts_with(dir_path))
+    }

-        initial_len - self.sync_data.files.len()
+    /// Use this to prevent any substantial background threads from acquiring the locks
+    pub fn cancel(&self) {
+        self.cancelled.store(true, Ordering::Release);
    }

    pub fn stop_background_monitor(&mut self) {
-        if let Some(watcher) = self.background_watcher.take() {
+        if let Some(mut watcher) = self.background_watcher.take() {
            watcher.stop();
        }
    }

-    pub fn trigger_rescan(&mut self) -> Result<(), Error> {
+    pub fn trigger_rescan(&mut self, shared_frecency: &SharedFrecency) -> Result<(), Error> {
        if self.is_scanning.load(Ordering::Relaxed) {
            debug!("Scan already in progress, skipping trigger_rescan");
            return Ok(());
@@ -504,7 +666,12 @@ impl FilePicker {
        self.is_scanning.store(true, Ordering::Relaxed);
        self.scanned_files_count.store(0, Ordering::Relaxed);

-        let scan_result = scan_filesystem(&self.base_path, &self.scanned_files_count);
+        let scan_result = scan_filesystem(
+            &self.base_path,
+            &self.scanned_files_count,
+            shared_frecency,
+            self.mode,
+        );
        match scan_result {
            Ok(sync) => {
                info!(
@@ -515,7 +682,11 @@ impl FilePicker {
                self.sync_data = sync;

                if self.warmup_mmap_cache {
-                    warmup_mmaps(&self.sync_data.files);
+                    // Warmup in background to avoid blocking
+                    let files = self.sync_data.files().to_vec(); // Clone all files
+                    std::thread::spawn(move || {
+                        warmup_mmaps(&files);
+                    });
                }
            }
            Err(error) => error!(?error, "Failed to scan file system"),
@@ -525,45 +696,102 @@ impl FilePicker {
        Ok(())
    }

+    /// Quick way to check if scan is going without acquiring a lock for [Self::get_scan_progress]
    pub fn is_scan_active(&self) -> bool {
        self.is_scanning.load(Ordering::Relaxed)
    }
+
+    /// Return a clone of the scanning flag so callers can poll it without
+    /// holding a lock on the picker.
+    pub fn scan_signal(&self) -> Arc<AtomicBool> {
+        Arc::clone(&self.is_scanning)
+    }
+
+    /// Block the current thread until the background filesystem scan finishes.
+    ///
+    /// Briefly acquires a read lock on the shared picker to obtain the scan
+    /// signal, then drops the lock and polls without holding it — so the
+    /// background thread can still write to the index.
+    ///
+    /// Returns immediately if no picker has been initialised yet.
+    pub fn wait_for_scan(shared_picker: &SharedPicker) {
+        let signal = {
+            let guard = shared_picker.read().expect("shared picker lock poisoned");
+            match guard.as_ref() {
+                Some(picker) => picker.scan_signal(),
+                None => return,
+            }
+        };
+
+        while signal.load(Ordering::Acquire) {
+            std::thread::sleep(std::time::Duration::from_millis(10));
+        }
+    }
 }

+/// A point-in-time snapshot of the file-scanning progress.
+///
+/// Returned by [`FilePicker::get_scan_progress`]. Useful for displaying
+/// a progress indicator while the initial scan is running.
 #[allow(unused)]
 #[derive(Debug, Clone)]
 pub struct ScanProgress {
+    /// Number of files indexed so far.
    pub scanned_files_count: usize,
+    /// `true` while the background scan thread is still running.
    pub is_scanning: bool,
 }

+#[allow(clippy::too_many_arguments)]
 fn spawn_scan_and_watcher(
    base_path: PathBuf,
    scan_signal: Arc<AtomicBool>,
    synced_files_count: Arc<AtomicUsize>,
    warmup_mmap_cache: bool,
+    mode: FFFMode,
+    shared_picker: SharedPicker,
+    shared_frecency: SharedFrecency,
+    cancelled: Arc<AtomicBool>,
 ) {
    std::thread::spawn(move || {
-        scan_signal.store(true, Ordering::Relaxed);
+        // scan_signal is already `true` (set by the caller before spawning)
+        // so waiters see "scanning" even before this thread is scheduled.
        info!("Starting initial file scan");

        let mut git_workdir = None;
-        match scan_filesystem(&base_path, &synced_files_count) {
+        match scan_filesystem(&base_path, &synced_files_count, &shared_frecency, mode) {
            Ok(sync) => {
+                if cancelled.load(Ordering::Acquire) {
+                    info!("Scan completed but picker was replaced, discarding results");
+                    scan_signal.store(false, Ordering::Relaxed);
+                    return;
+                }
+
                info!(
                    "Initial filesystem scan completed: found {} files",
                    sync.files.len()
                );

                git_workdir = sync.git_workdir.clone();
-                if let Ok(mut file_picker_guard) = crate::FILE_PICKER.write()
-                    && let Some(ref mut picker) = *file_picker_guard
-                {
-                    picker.sync_data = sync;

-                    if warmup_mmap_cache {
-                        warmup_mmaps(&picker.sync_data.files);
+                // Write results into the provided shared handle.
+                let write_result = shared_picker.write().ok().map(|mut guard| {
+                    if let Some(ref mut picker) = *guard {
+                        picker.sync_data = sync;
                    }
+                });
+
+                if write_result.is_none() {
+                    error!("Failed to write scan results into picker");
+                }
+
+                // OPTIMIZATION: Warmup mmap cache in background to avoid blocking first grep.
+                if warmup_mmap_cache
+                    && !cancelled.load(Ordering::Acquire)
+                    && let Ok(guard) = shared_picker.read()
+                    && let Some(ref picker) = *guard
+                {
+                    warmup_mmaps(picker.sync_data.files());
                }
            }
            Err(e) => {
@@ -572,14 +800,39 @@ fn spawn_scan_and_watcher(
        }
        scan_signal.store(false, Ordering::Relaxed);

-        match BackgroundWatcher::new(base_path, git_workdir) {
+        // Don't create a watcher if this picker instance was already replaced
+        if cancelled.load(Ordering::Acquire) {
+            info!("Picker was replaced, skipping background watcher creation");
+            return;
+        }
+
+        match BackgroundWatcher::new(
+            base_path,
+            git_workdir,
+            shared_picker.clone(),
+            shared_frecency.clone(),
+            mode,
+        ) {
            Ok(watcher) => {
                info!("Background file watcher initialized successfully");

-                if let Ok(mut file_picker_guard) = crate::FILE_PICKER.write()
-                    && let Some(ref mut picker) = *file_picker_guard
-                {
-                    picker.background_watcher = Some(watcher);
+                // Final cancellation check: if the picker was replaced between
+                // watcher creation and this write, drop the watcher instead of
+                // storing it in the wrong picker.
+                if cancelled.load(Ordering::Acquire) {
+                    info!("Picker was replaced, dropping orphaned watcher");
+                    drop(watcher);
+                    return;
+                }
+
+                let write_result = shared_picker.write().ok().map(|mut guard| {
+                    if let Some(ref mut picker) = *guard {
+                        picker.background_watcher = Some(watcher);
+                    }
+                });
+
+                if write_result.is_none() {
+                    error!("Failed to store background watcher in picker");
                }
            }
            Err(e) => {
@@ -596,8 +849,8 @@ fn spawn_scan_and_watcher(
 ///
 /// Each file is mmap'd and a single byte is read to trigger the page fault.
 /// This runs in parallel using rayon.
+#[tracing::instrument(skip(files), name = "warmup_mmaps", level = Level::DEBUG)]
 fn warmup_mmaps(files: &[FileItem]) {
-    let warmup_start = std::time::Instant::now();
    let warmed = std::sync::atomic::AtomicUsize::new(0);

    files.par_iter().for_each(|file| {
@@ -605,27 +858,21 @@ fn warmup_mmaps(files: &[FileItem]) {
            return;
        }

-        if let Some(mmap) = file.get_mmap() {
-            // Read the first byte to trigger the initial page fault, which
-            // causes the kernel to start readahead for subsequent pages.
-            // This is cheaper than madvise and portable across all platforms.
-            let _ = std::hint::black_box(mmap.first());
+        if let Some(content) = file.get_mmap() {
+            // Read the first byte to trigger the initial page fault (mmap)
+            // or ensure the content is cached (Windows buffer).
+            let _ = std::hint::black_box(content.first());

            warmed.fetch_add(1, Ordering::Relaxed);
        }
    });
-
-    let warmed_count = warmed.load(Ordering::Relaxed);
-    info!(
-        "Mmap warmup completed: {warmed_count}/{} files in {:?}",
-        files.len(),
-        warmup_start.elapsed()
-    );
 }

 fn scan_filesystem(
    base_path: &Path,
    synced_files_count: &Arc<AtomicUsize>,
+    shared_frecency: &SharedFrecency,
+    mode: FFFMode,
 ) -> Result<FileSync, Error> {
    use ignore::{WalkBuilder, WalkState};
    use std::thread;
@@ -670,7 +917,7 @@ fn scan_filesystem(
            .build_parallel();

        let walker_start = std::time::Instant::now();
-        info!("SCAN: Starting file walker");
+        debug!("SCAN: Starting file walker");

        let files = Arc::new(std::sync::Mutex::new(Vec::new()));
        walker.run(|| {
@@ -712,7 +959,10 @@ fn scan_filesystem(
            Error::ThreadPanic
        })?;

-        let frecency = FRECENCY.read().map_err(|_| Error::AcquireFrecencyLock)?;
+        let frecency = shared_frecency
+            .read()
+            .map_err(|_| Error::AcquireFrecencyLock)?;
+
        files
            .par_iter_mut()
            .try_for_each(|file| -> Result<(), Error> {
@@ -721,7 +971,7 @@ fn scan_filesystem(
                }

                if let Some(frecency) = frecency.as_ref() {
-                    file.update_frecency_scores(frecency)?;
+                    file.update_frecency_scores(frecency, mode)?;
                }

                Ok(())
@@ -734,7 +984,6 @@ fn scan_filesystem(
            files.len()
        );

-        // Sort by OsStr instead of Path to avoid expensive component-by-component comparison
        files.par_sort_unstable_by(|a, b| a.path.as_os_str().cmp(b.path.as_os_str()));
        Ok(FileSync { files, git_workdir })
    })
@@ -750,3 +999,22 @@ fn is_git_file(path: &Path) -> bool {
        }
    })
 }
+
+/// Detect if a file is binary by checking for NUL bytes in the first 512 bytes.
+/// This is the same heuristic used by git and grep — simple, fast, and sufficient.
+#[inline]
+fn detect_binary(path: &Path, size: u64) -> bool {
+    // Empty files are not binary
+    if size == 0 {
+        return false;
+    }
+
+    let Ok(file) = std::fs::File::open(path) else {
+        return false;
+    };
+    let mut reader = std::io::BufReader::with_capacity(1024, file);
+
+    let mut buf = [0u8; 512];
+    let n = reader.read(&mut buf).unwrap_or(0);
+    buf[..n].contains(&0)
+}
@@ -1,11 +1,14 @@
 use crate::db_healthcheck::DbHealthChecker;
-use crate::{error::Error, git::is_modified_status};
+use crate::error::{Error, Result};
+use crate::file_picker::FFFMode;
+use crate::{SharedFrecency, git::is_modified_status};
 use heed::{Database, Env, EnvOpenOptions};
 use heed::{
    EnvFlags,
    types::{Bytes, SerdeBincode},
 };
 use std::fs;
+use std::path::PathBuf;
 use std::time::{SystemTime, UNIX_EPOCH};
 use std::{collections::VecDeque, path::Path};

@@ -13,6 +16,10 @@ const DECAY_CONSTANT: f64 = 0.0693; // ln(2)/10 for 10-day half-life
 const SECONDS_PER_DAY: f64 = 86400.0;
 const MAX_HISTORY_DAYS: f64 = 30.0; // Only consider accesses within 30 days

+// AI mode: faster decay since AI sessions are shorter and more intense
+const AI_DECAY_CONSTANT: f64 = 0.231; // ln(2)/3 for 3-day half-life
+const AI_MAX_HISTORY_DAYS: f64 = 7.0; // Only consider accesses within 7 days
+
 #[derive(Debug)]
 pub struct FrecencyTracker {
    env: Env,
@@ -27,12 +34,21 @@ const MODIFICATION_THRESHOLDS: [(i64, u64); 5] = [
    (1, 60 * 60 * 24 * 7), // 1 week
 ];

+// AI mode: compressed thresholds since AI edits happen in rapid bursts
+const AI_MODIFICATION_THRESHOLDS: [(i64, u64); 5] = [
+    (16, 30),         // 30 seconds
+    (8, 60 * 5),      // 5 minutes
+    (4, 60 * 15),     // 15 minutes
+    (2, 60 * 60),     // 1 hour
+    (1, 60 * 60 * 4), // 4 hours
+];
+
 impl DbHealthChecker for FrecencyTracker {
    fn get_env(&self) -> &heed::Env {
        &self.env
    }

-    fn count_entries(&self) -> Result<Vec<(&'static str, u64)>, Error> {
+    fn count_entries(&self) -> Result<Vec<(&'static str, u64)>> {
        let rtxn = self.env.read_txn().map_err(Error::DbStartReadTxn)?;
        let count = self.db.len(&rtxn).map_err(Error::DbRead)?;

@@ -41,10 +57,13 @@ impl DbHealthChecker for FrecencyTracker {
 }

 impl FrecencyTracker {
-    pub fn new(db_path: &str, use_unsafe_no_lock: bool) -> Result<Self, Error> {
+    pub fn new(db_path: impl AsRef<Path>, use_unsafe_no_lock: bool) -> Result<Self> {
+        let db_path = db_path.as_ref();
        fs::create_dir_all(db_path).map_err(Error::CreateDir)?;
+
        let env = unsafe {
            let mut opts = EnvOpenOptions::new();
+            opts.map_size(24 * 1024 * 1024); // 24 MiB
            if use_unsafe_no_lock {
                opts.flags(EnvFlags::NO_LOCK | EnvFlags::NO_SYNC | EnvFlags::NO_META_SYNC);
            }
@@ -65,7 +84,209 @@ impl FrecencyTracker {
        })
    }

-    fn get_accesses(&self, path: &Path) -> Result<Option<VecDeque<u64>>, Error> {
+    /// Spawns a background thread to purge stale frecency entries and compact the database.
+    /// Run it once in a while to purge old pages and keep DB file size reasonable.
+    ///
+    /// It's okay to not join this thread since it acquires locks for the db access
+    ///
+    /// ```
+    /// use fff_search::frecency::FrecencyTracker;
+    /// use fff_search::SharedFrecency;
+    /// let shared_frecency: SharedFrecency = Default::default();
+    /// let _ = FrecencyTracker::spawn_gc(shared_frecency, "/path/to/frecency_db".into(), true).ok();
+    /// ```
+    pub fn spawn_gc(
+        shared: SharedFrecency,
+        db_path: String,
+        use_unsafe_no_lock: bool,
+    ) -> Result<std::thread::JoinHandle<()>> {
+        Ok(std::thread::Builder::new()
+            .name("fff-frecency-gc".into())
+            .spawn(move || Self::run_frecency_gc(shared, db_path, use_unsafe_no_lock))?)
+    }
+
+    #[tracing::instrument(skip(shared), fields(db_path = %db_path))]
+    fn run_frecency_gc(shared: SharedFrecency, db_path: String, use_unsafe_no_lock: bool) {
+        let start = std::time::Instant::now();
+        let data_path = PathBuf::from(&db_path).join("data.mdb");
+
+        // Phase 1: Purge stale entries.
+        // The RwLock protects the Option<FrecencyTracker> (not the DB itself),
+        // so a read lock is sufficient — LMDB handles its own write serialization.
+        let (deleted, pruned) = {
+            let guard = match shared.read() {
+                Ok(g) => g,
+                Err(e) => {
+                    tracing::debug!("Failed to acquire read lock: {e}");
+                    return;
+                }
+            };
+            let Some(ref tracker) = *guard else {
+                return;
+            };
+            match tracker.purge_stale_entries() {
+                Ok(result) => result,
+                Err(e) => {
+                    tracing::debug!("Purge failed: {e}");
+                    return;
+                }
+            }
+        };
+
+        if deleted > 0 || pruned > 0 {
+            tracing::info!(deleted, pruned, elapsed = ?start.elapsed(), "Frecency GC purged entries");
+        }
+
+        // Compact if we purged entries OR the file has significant freelist bloat
+        let file_size = fs::metadata(&data_path).map(|m| m.len()).unwrap_or(0);
+        if deleted == 0 && pruned == 0 && file_size <= 512 * 1024 {
+            return;
+        }
+
+        // Phase 2: Manual compaction under a single write lock
+        let mut guard = match shared.write() {
+            Ok(g) => g,
+            Err(e) => {
+                tracing::debug!("Failed to acquire write lock: {e}");
+                return;
+            }
+        };
+
+        // Read all entries from current env
+        let entries: Vec<(Vec<u8>, VecDeque<u64>)> = match guard.as_ref() {
+            Some(tracker) => {
+                let rtxn = match tracker.env.read_txn() {
+                    Ok(t) => t,
+                    Err(e) => {
+                        tracing::debug!("Compaction read_txn failed: {e}");
+                        return;
+                    }
+                };
+                let iter = match tracker.db.iter(&rtxn) {
+                    Ok(i) => i,
+                    Err(e) => {
+                        tracing::debug!("Compaction iter failed: {e}");
+                        return;
+                    }
+                };
+                let mut entries = Vec::new();
+                let mut read_errors = 0u32;
+                for result in iter {
+                    match result {
+                        Ok((key, value)) => entries.push((key.to_vec(), value)),
+                        Err(_) => read_errors += 1,
+                    }
+                }
+                if read_errors > 0 {
+                    tracing::warn!(
+                        read_errors,
+                        "Skipped corrupted entries during compaction read"
+                    );
+                }
+                entries
+            }
+            None => return,
+        };
+
+        // Drop old tracker, delete files, create fresh env, write back
+        *guard = None;
+
+        let lock_path = PathBuf::from(&db_path).join("lock.mdb");
+        let _ = fs::remove_file(&data_path);
+        let _ = fs::remove_file(&lock_path);
+
+        let tracker = match FrecencyTracker::new(&db_path, use_unsafe_no_lock) {
+            Ok(t) => t,
+            Err(e) => {
+                tracing::error!("Compaction reopen failed, frecency disabled: {e}");
+                return;
+            }
+        };
+
+        let write_result = (|| -> std::result::Result<(), heed::Error> {
+            let mut wtxn = tracker.env.write_txn()?;
+            for (key, value) in &entries {
+                tracker.db.put(&mut wtxn, key.as_slice(), value)?;
+            }
+            wtxn.commit()?;
+            Ok(())
+        })();
+
+        match write_result {
+            Ok(()) => {
+                let new_size = fs::metadata(&data_path).map(|m| m.len()).unwrap_or(0);
+                *guard = Some(tracker);
+                tracing::debug!(
+                    entries = entries.len(),
+                    old_size = file_size,
+                    new_size,
+                    elapsed = ?start.elapsed(),
+                    "Frecency DB compacted"
+                );
+            }
+            Err(e) => {
+                tracing::error!("Compaction write failed, frecency data may be incomplete: {e}");
+                *guard = Some(tracker);
+            }
+        }
+    }
+
+    /// Removes entries where all timestamps are older than MAX_HISTORY_DAYS,
+    /// and prunes stale timestamps from entries that still have recent ones.
+    /// Returns (deleted_count, pruned_count).
+    fn purge_stale_entries(&self) -> Result<(usize, usize)> {
+        let now = self.get_now();
+        let cutoff_time = now.saturating_sub((MAX_HISTORY_DAYS * SECONDS_PER_DAY) as u64);
+
+        // Collect entries to delete or update
+        let rtxn = self.env.read_txn().map_err(Error::DbStartReadTxn)?;
+        let mut to_delete: Vec<Vec<u8>> = Vec::new();
+        let mut to_update: Vec<(Vec<u8>, VecDeque<u64>)> = Vec::new();
+
+        let iter = self.db.iter(&rtxn).map_err(Error::DbRead)?;
+        for result in iter {
+            let (key, accesses) = result.map_err(Error::DbRead)?;
+
+            // Timestamps are chronologically ordered (oldest at front).
+            // Find the first timestamp that is still within the retention window.
+            let fresh_start = accesses.iter().position(|&ts| ts >= cutoff_time);
+            match fresh_start {
+                None => {
+                    // All timestamps are stale — delete the entire entry
+                    to_delete.push(key.to_vec());
+                }
+                Some(0) => {
+                    // All timestamps are fresh — nothing to do
+                }
+                Some(start) => {
+                    // Some timestamps are stale — keep only the fresh ones
+                    let pruned: VecDeque<u64> = accesses.iter().skip(start).copied().collect();
+                    to_update.push((key.to_vec(), pruned));
+                }
+            }
+        }
+        drop(rtxn);
+
+        if to_delete.is_empty() && to_update.is_empty() {
+            return Ok((0, 0));
+        }
+
+        // Apply all changes in a single write transaction
+        let mut wtxn = self.env.write_txn().map_err(Error::DbStartWriteTxn)?;
+        for key in &to_delete {
+            self.db.delete(&mut wtxn, key).map_err(Error::DbWrite)?;
+        }
+        for (key, accesses) in &to_update {
+            self.db
+                .put(&mut wtxn, key, accesses)
+                .map_err(Error::DbWrite)?;
+        }
+        wtxn.commit().map_err(Error::DbCommit)?;
+
+        Ok((to_delete.len(), to_update.len()))
+    }
+
+    fn get_accesses(&self, path: &Path) -> Result<Option<VecDeque<u64>>> {
        let rtxn = self.env.read_txn().map_err(Error::DbStartReadTxn)?;

        let key_hash = Self::path_to_hash_bytes(path)?;
@@ -79,7 +300,7 @@ impl FrecencyTracker {
            .as_secs()
    }

-    fn path_to_hash_bytes(path: &Path) -> Result<[u8; 32], Error> {
+    fn path_to_hash_bytes(path: &Path) -> Result<[u8; 32]> {
        let Some(key) = path.to_str() else {
            return Err(Error::InvalidPath(path.to_path_buf()));
        };
@@ -87,7 +308,15 @@ impl FrecencyTracker {
        Ok(*blake3::hash(key.as_bytes()).as_bytes())
    }

-    pub fn track_access(&self, path: &Path) -> Result<(), Error> {
+    /// Returns seconds since the most recent tracked access, or `None` if the
+    /// file has never been tracked.
+    pub fn seconds_since_last_access(&self, path: &Path) -> Result<Option<u64>> {
+        let accesses = self.get_accesses(path)?;
+        let last = accesses.and_then(|a| a.back().copied());
+        Ok(last.map(|ts| self.get_now().saturating_sub(ts)))
+    }
+
+    pub fn track_access(&self, path: &Path) -> Result<()> {
        let mut wtxn = self.env.write_txn().map_err(Error::DbStartWriteTxn)?;

        let key_hash = Self::path_to_hash_bytes(path)?;
@@ -115,7 +344,7 @@ impl FrecencyTracker {
        Ok(())
    }

-    pub fn get_access_score(&self, file_path: &Path) -> i64 {
+    pub fn get_access_score(&self, file_path: &Path, mode: FFFMode) -> i64 {
        let accesses = self
            .get_accesses(file_path)
            .ok()
@@ -126,10 +355,21 @@ impl FrecencyTracker {
            return 0;
        }

+        let decay_constant = if mode.is_ai() {
+            AI_DECAY_CONSTANT
+        } else {
+            DECAY_CONSTANT
+        };
+        let max_history_days = if mode.is_ai() {
+            AI_MAX_HISTORY_DAYS
+        } else {
+            MAX_HISTORY_DAYS
+        };
+
        let now = self.get_now();
        let mut total_frecency = 0.0;

-        let cutoff_time = now.saturating_sub((MAX_HISTORY_DAYS * SECONDS_PER_DAY) as u64);
+        let cutoff_time = now.saturating_sub((max_history_days * SECONDS_PER_DAY) as u64);

        for &access_time in accesses.iter().rev() {
            if access_time < cutoff_time {
@@ -137,7 +377,7 @@ impl FrecencyTracker {
            }

            let days_ago = (now.saturating_sub(access_time) as f64) / SECONDS_PER_DAY;
-            let decay_factor = (-DECAY_CONSTANT * days_ago).exp();
+            let decay_factor = (-decay_constant * days_ago).exp();
            total_frecency += decay_factor;
        }

@@ -155,24 +395,31 @@ impl FrecencyTracker {
        &self,
        modified_time: u64,
        git_status: Option<git2::Status>,
+        mode: FFFMode,
    ) -> i64 {
        let is_modified_git_status = git_status.is_some_and(is_modified_status);
        if !is_modified_git_status {
            return 0;
        }

+        let thresholds = if mode.is_ai() {
+            &AI_MODIFICATION_THRESHOLDS
+        } else {
+            &MODIFICATION_THRESHOLDS
+        };
+
        let now = self.get_now();
        let duration_since = now.saturating_sub(modified_time);

-        for i in 0..MODIFICATION_THRESHOLDS.len() {
-            let (current_points, current_threshold) = MODIFICATION_THRESHOLDS[i];
+        for i in 0..thresholds.len() {
+            let (current_points, current_threshold) = thresholds[i];

            if duration_since <= current_threshold {
                if i == 0 || duration_since == current_threshold {
                    return current_points;
                }

-                let (prev_points, prev_threshold) = MODIFICATION_THRESHOLDS[i - 1];
+                let (prev_points, prev_threshold) = thresholds[i - 1];

                let time_range = current_threshold - prev_threshold;
                let time_offset = duration_since - prev_threshold;
@@ -192,6 +439,7 @@ impl FrecencyTracker {
 #[cfg(test)]
 mod tests {
    use super::*;
+    use crate::file_picker::FFFMode;

    fn calculate_test_frecency_score(access_timestamps: &[u64], current_time: u64) -> i64 {
        let mut total_frecency = 0.0;
@@ -269,34 +517,35 @@ mod tests {

        // At 5 minutes: should interpolate between 16 and 8 points
        let five_minutes_ago = current_time - (5 * 60);
-        let score = tracker.get_modification_score(five_minutes_ago, git_status);
+        let score = tracker.get_modification_score(five_minutes_ago, git_status, FFFMode::Neovim);

        // Expected: 16 - (8 * 3 / 13) = 16 - 1 = 15 points
        // (time_offset = 5-2 = 3, time_range = 15-2 = 13, points_diff = 16-8 = 8)
        assert_eq!(score, 15, "5 minutes should interpolate to 15 points");

        let two_minutes_ago = current_time - (2 * 60);
-        let score = tracker.get_modification_score(two_minutes_ago, git_status);
+        let score = tracker.get_modification_score(two_minutes_ago, git_status, FFFMode::Neovim);
        assert_eq!(score, 16, "2 minutes should be exactly 16 points");

        let fifteen_minutes_ago = current_time - (15 * 60);
-        let score = tracker.get_modification_score(fifteen_minutes_ago, git_status);
+        let score =
+            tracker.get_modification_score(fifteen_minutes_ago, git_status, FFFMode::Neovim);
        assert_eq!(score, 8, "15 minutes should be exactly 8 points");

        // At 12 hours: should interpolate between 4 and 2 points
        let twelve_hours_ago = current_time - (12 * 60 * 60);
-        let score = tracker.get_modification_score(twelve_hours_ago, git_status);
+        let score = tracker.get_modification_score(twelve_hours_ago, git_status, FFFMode::Neovim);
        // Expected: 4 - (2 * 11 / 23) = 4 - 0 = 4 points (integer division)
        // (time_offset = 12-1 = 11 hours, time_range = 24-1 = 23 hours, points_diff = 4-2 = 2)
        assert_eq!(score, 4, "12 hours should interpolate to 4 points");

        // at 18 hours for more significant interpolation
        let eighteen_hours_ago = current_time - (18 * 60 * 60);
-        let score = tracker.get_modification_score(eighteen_hours_ago, git_status);
+        let score = tracker.get_modification_score(eighteen_hours_ago, git_status, FFFMode::Neovim);
        // Expected: 4 - (2 * 17 / 23) = 4 - 1 = 3 points
        assert_eq!(score, 3, "18 hours should interpolate to 3 points");

-        let score = tracker.get_modification_score(five_minutes_ago, None);
+        let score = tracker.get_modification_score(five_minutes_ago, None, FFFMode::Neovim);
        assert_eq!(score, 0, "No git status should return 0");

        let _ = std::fs::remove_dir_all(&temp_dir);
@@ -125,31 +125,35 @@ pub fn is_modified_status(status: Status) -> bool {
    )
 }

-pub fn format_git_status(status: Option<Status>) -> &'static str {
+pub fn format_git_status_opt(status: Option<Status>) -> Option<&'static str> {
    match status {
-        None => "clear",
+        None => Some("clean"),
        Some(status) => {
            if status.contains(Status::WT_NEW) {
-                "untracked"
+                Some("untracked")
            } else if status.contains(Status::WT_MODIFIED) {
-                "modified"
+                Some("modified")
            } else if status.contains(Status::WT_DELETED) {
-                "deleted"
+                Some("deleted")
            } else if status.contains(Status::WT_RENAMED) {
-                "renamed"
+                Some("renamed")
            } else if status.contains(Status::INDEX_NEW) {
-                "staged_new"
+                Some("staged_new")
            } else if status.contains(Status::INDEX_MODIFIED) {
-                "staged_modified"
+                Some("staged_modified")
            } else if status.contains(Status::INDEX_DELETED) {
-                "staged_deleted"
+                Some("staged_deleted")
            } else if status.contains(Status::IGNORED) {
-                "ignored"
+                Some("ignored")
            } else if status.contains(Status::CURRENT) || status.is_empty() {
-                "clean"
+                Some("clean")
            } else {
-                "unknown"
+                None
            }
        }
    }
 }
+
+pub fn format_git_status(status: Option<Status>) -> &'static str {
+    format_git_status_opt(status).unwrap_or("unknown")
+}
@@ -1,43 +1,154 @@
-//! fff-core - High-performance file finder library
+//! # FFF Search — High-performance file finder core
 //!
-//! This crate provides the core file indexing and fuzzy search functionality.
-//! It maintains global state for the file picker, frecency tracker, and query tracker.
+//! This crate provides the core search engine for [FFF (Fast File Finder)](https://github.com/dmtrKovalenko/fff.nvim).
+//! It includes filesystem indexing with real-time watching, fuzzy matching powered
+//! by [frizbee](https://docs.rs/neo_frizbee), frecency scoring backed by LMDB,
+//! and multi-mode grep search.
+//!
+//! ## Architecture
+//!
+//! - [`file_picker::FilePicker`] — Main entry point. Indexes a directory tree in a
+//!   background thread, maintains a sorted file list, watches the filesystem for
+//!   changes, and performs fuzzy search with frecency-weighted scoring.
+//! - [`frecency::FrecencyTracker`] — LMDB-backed database that tracks file access
+//!   and modification patterns for intelligent result ranking.
+//! - [`query_tracker::QueryTracker`] — Tracks search query history and provides
+//!   "combo-boost" scoring for repeatedly matched files.
+//! - [`grep`] — Live grep search supporting regex, plain-text, and fuzzy modes
+//!   with optional constraint filtering.
+//! - [`git`] — Git status caching and repository detection.
+//!
+//! ## Shared State
+//!
+//! [`SharedPicker`], [`SharedFrecency`], and [`SharedQueryTracker`] are
+//! `Arc<RwLock<Option<T>>>` type aliases for thread-safe shared access. FFF
+//! is designed for long-running processes that keep the file index in global
+//! state, so these wrappers let background threads (scanner, watcher) share
+//! data with the calling code safely.
+//!
+//! ## Quick Start
+//!
+//! ```
+//! use fff_search::file_picker::FilePicker;
+//! use fff_search::frecency::FrecencyTracker;
+//! use fff_search::query_tracker::QueryTracker;
+//! use fff_search::{
+//!     FFFMode, FuzzySearchOptions, PaginationArgs, QueryParser,
+//!     SharedFrecency, SharedPicker, SharedQueryTracker,
+//! };
+//!
+//! let shared_picker: SharedPicker = Default::default();
+//! let shared_frecency: SharedFrecency = Default::default();
+//! let shared_query_tracker: SharedQueryTracker = Default::default();
+//!
+//! let tmp = std::env::temp_dir().join("fff-doctest");
+//! std::fs::create_dir_all(&tmp).unwrap();
+//!
+//! // 1. Optionally initialize frecency and query tracker databases
+//! let frecency = FrecencyTracker::new(tmp.join("frecency"), false)?;
+//! *shared_frecency.write().unwrap() = Some(frecency);
+//!
+//! let query_tracker = QueryTracker::new(tmp.join("queries"), false)?;
+//! *shared_query_tracker.write().unwrap() = Some(query_tracker);
+//!
+//! // 2. Init the file picker (spawns background scan + watcher)
+//! FilePicker::new_with_shared_state(
+//!     ".".into(),
+//!     /* warmup memap caches = */ false,
+//!     FFFMode::Ai, // use AI for ai agents, and Neovim for editors
+//!     shared_picker.clone(),
+//!     shared_frecency.clone(),
+//! )?;
+//!
+//! // 3. Wait for scan (in real app you would like to add some tokio flavor here)
+//! FilePicker::wait_for_scan(&shared_picker);
+//!
+//! // 4. Search: lock the picker and query tracker
+//! let picker_lock_guard = shared_picker.read().unwrap();
+//! let picker = picker_lock_guard.as_ref().unwrap();
+//! let query_tracker_lock_guard = shared_query_tracker.read().unwrap();
+//!
+//! // 5. Parse the query and perform fuzzy search with frecency and combo-boost scoring
+//! let parser = QueryParser::default();
+//! let query = parser.parse("lib.rs");
+//!
+//! let results = FilePicker::fuzzy_search(
+//!     picker.get_files(),
+//!     &query,
+//!     query_tracker_lock_guard.as_ref(),
+//!     FuzzySearchOptions {
+//!         max_threads: 0,
+//!         current_file: None,
+//!         pagination: PaginationArgs { offset: 0, limit: 50 },
+//!         ..Default::default()
+//!     },
+//! );
+//!
+//! assert!(results.total_matched > 0);
+//! assert!(results.items.first().unwrap().path.ends_with("lib.rs"));
+//!
+//! let _ = std::fs::remove_dir_all(&tmp);
+//! # Ok::<(), Box<dyn std::error::Error>>(())
+//! ```

 mod background_watcher;
-pub mod constraints;
+mod constraints;
 mod db_healthcheck;
 mod error;
-pub mod file_picker;
-pub mod frecency;
-pub mod git;
-pub mod grep;
-pub mod path_utils;
-pub mod query_tracker;
-pub mod score;
+mod score;
 mod sort_buffer;
+
+/// Core file picker: filesystem indexing, background watching, and fuzzy search.
+///
+/// See [`FilePicker`](file_picker::FilePicker) for the main entry point.
+pub mod file_picker;
+
+/// Frecency (frequency + recency) database for file access scoring.
+///
+/// Backed by LMDB for persistent, crash-safe storage.
+pub mod frecency;
+
+/// Git status caching and repository detection utilities.
+pub mod git;
+
+/// Live grep search with regex, plain-text, and fuzzy matching modes.
+///
+/// Supports constraint filtering (file extensions, path segments, globs)
+/// and parallel execution via rayon.
+pub mod grep;
+
+/// Tracing/logging initialization and panic hook setup.
+pub mod log;
+
+/// Path manipulation utilities: cross platform canonicalization, tilde expansion, and
+/// directory distance penalties for search scoring.
+pub mod path_utils;
+
+/// Search query history tracker for combo-boost scoring.
+///
+/// Records which files a user selects for each query, enabling the scorer
+/// to boost files that were previously chosen for similar searches.
+pub mod query_tracker;
+
+/// Core data types shared across the crate.
 pub mod types;

-use file_picker::FilePicker;
-use frecency::FrecencyTracker;
-use once_cell::sync::Lazy;
-use query_tracker::QueryTracker;
-use std::sync::RwLock;
+use std::sync::{Arc, RwLock};

-// Global state - same pattern as fff-nvim
-pub static FRECENCY: Lazy<RwLock<Option<FrecencyTracker>>> = Lazy::new(|| RwLock::new(None));
-pub static FILE_PICKER: Lazy<RwLock<Option<FilePicker>>> = Lazy::new(|| RwLock::new(None));
-pub static QUERY_TRACKER: Lazy<RwLock<Option<QueryTracker>>> = Lazy::new(|| RwLock::new(None));
+/// Thread-safe shared handle to the [`FilePicker`] instance.
+pub type SharedPicker = Arc<RwLock<Option<FilePicker>>>;
+
+/// Thread-safe shared handle to the [`FrecencyTracker`] instance.
+pub type SharedFrecency = Arc<RwLock<Option<FrecencyTracker>>>;
+
+/// Thread-safe shared handle to the [`QueryTracker`] instance.
+pub type SharedQueryTracker = Arc<RwLock<Option<QueryTracker>>>;

-// Re-export main types for convenience
 pub use db_healthcheck::{DbHealth, DbHealthChecker};
 pub use error::{Error, Result};
-pub use file_picker::{FuzzySearchOptions, ScanProgress};
-pub use types::{FileItem, PaginationArgs, Score, ScoringContext, SearchResult};
-
-// Re-export grep types
-pub use grep::{GrepMatch, GrepMode, GrepResult, GrepSearchOptions};
-
-// Re-export query parser types (including Location which moved there)
-pub use fff_query_parser::{
-    Constraint, FFFQuery, FuzzyQuery, Location, QueryParser, location::parse_location,
-};
+pub use fff_query_parser::*;
+pub use file_picker::*;
+pub use frecency::*;
+pub use grep::*;
+pub use query_tracker::*;
+pub use types::*;
@@ -0,0 +1,164 @@
+//! Shared logging utilities for FFF crates.
+//!
+//! Provides file-based tracing initialization and a panic hook that writes
+//! to both stderr and a fallback log file.
+
+use std::io;
+use std::path::Path;
+use tracing_appender::non_blocking;
+use tracing_subscriber::fmt::format::FmtSpan;
+use tracing_subscriber::{EnvFilter, fmt, prelude::*};
+
+static TRACING_INITIALIZED: std::sync::OnceLock<tracing_appender::non_blocking::WorkerGuard> =
+    std::sync::OnceLock::new();
+
+static PANIC_HOOK_INSTALLED: std::sync::OnceLock<()> = std::sync::OnceLock::new();
+
+/// Install panic hook that writes to both stderr and a fallback file.
+/// This is called separately from init_tracing to ensure panics are always logged.
+pub fn install_panic_hook() {
+    PANIC_HOOK_INSTALLED.get_or_init(|| {
+        let default_panic = std::panic::take_hook();
+
+        std::panic::set_hook(Box::new(move |panic_info| {
+            let payload = panic_info.payload();
+            let message = if let Some(s) = payload.downcast_ref::<&str>() {
+                s.to_string()
+            } else if let Some(s) = payload.downcast_ref::<String>() {
+                s.clone()
+            } else {
+                "Unknown panic payload".to_string()
+            };
+
+            let location = if let Some(location) = panic_info.location() {
+                format!(
+                    "{}:{}:{}",
+                    location.file(),
+                    location.line(),
+                    location.column()
+                )
+            } else {
+                "unknown location".to_string()
+            };
+
+            // Always log to tracing (if initialized)
+            tracing::error!(
+                panic.message = %message,
+                panic.location = %location,
+                "PANIC occurred in FFF"
+            );
+
+            // Always print to stderr
+            eprintln!("=== FFF PANIC ===");
+            eprintln!("Message: {}", message);
+            eprintln!("Location: {}", location);
+            eprintln!("=================");
+
+            // Try to write to fallback panic log file
+            if let Some(cache_dir) = dirs::cache_dir() {
+                let panic_log = cache_dir.join("fff_panic.log");
+                let timestamp = std::time::SystemTime::now()
+                    .duration_since(std::time::UNIX_EPOCH)
+                    .map(|d| d.as_secs())
+                    .unwrap_or(0);
+
+                let panic_entry = format!(
+                    "\n[{}] PANIC at {}\nMessage: {}\n",
+                    timestamp, location, message
+                );
+
+                let _ = std::fs::OpenOptions::new()
+                    .create(true)
+                    .append(true)
+                    .open(&panic_log)
+                    .and_then(|mut f| {
+                        use std::io::Write;
+                        f.write_all(panic_entry.as_bytes())
+                    });
+
+                eprintln!("Panic logged to: {}", panic_log.display());
+            }
+
+            default_panic(panic_info);
+        }));
+    });
+}
+
+/// Parse a log level string into a `tracing::Level`.
+///
+/// Accepts "trace", "debug", "info", "warn", "error" (case-insensitive).
+/// Returns `tracing::Level::INFO` for unrecognised values.
+pub fn parse_log_level(level: Option<&str>) -> tracing::Level {
+    match level.as_ref().map(|s| s.trim().to_lowercase()).as_deref() {
+        Some("trace") => tracing::Level::TRACE,
+        Some("debug") => tracing::Level::DEBUG,
+        Some("info") => tracing::Level::INFO,
+        Some("warn") => tracing::Level::WARN,
+        Some("error") => tracing::Level::ERROR,
+        _ => tracing::Level::INFO,
+    }
+}
+
+/// Initialize tracing with a single log file.
+///
+/// Creates the parent directory if it doesn't exist, truncates the log file,
+/// and sets up a non-blocking file appender with structured formatting.
+///
+/// # Arguments
+/// * `log_file_path` - Full path to the log file
+/// * `log_level` - Log level (trace, debug, info, warn, error)
+///
+/// # Returns
+/// * `Result<String, io::Error>` - Full path to the log file on success
+pub fn init_tracing(log_file_path: &str, log_level: Option<&str>) -> Result<String, io::Error> {
+    // Install panic hook first (does nothing if already installed)
+    install_panic_hook();
+
+    let log_path = Path::new(log_file_path);
+    if let Some(parent) = log_path.parent() {
+        std::fs::create_dir_all(parent)?;
+    }
+
+    let file_appender = std::fs::OpenOptions::new()
+        .create(true)
+        .write(true)
+        .truncate(true) // creates a new file on every setup
+        .open(log_path)?;
+
+    let level = parse_log_level(log_level);
+
+    TRACING_INITIALIZED.get_or_init(|| {
+        let (non_blocking_appender, guard) = non_blocking(file_appender);
+
+        let subscriber = tracing_subscriber::registry()
+            .with(
+                fmt::layer()
+                    .with_writer(non_blocking_appender)
+                    .with_target(true)
+                    .with_thread_ids(false)
+                    .with_thread_names(false)
+                    .with_file(true)
+                    .with_line_number(true)
+                    .with_ansi(false)
+                    .with_span_events(FmtSpan::NEW | FmtSpan::CLOSE),
+            )
+            .with(
+                EnvFilter::builder()
+                    .with_default_directive(level.into())
+                    .from_env_lossy(),
+            );
+
+        if let Err(e) = tracing::subscriber::set_global_default(subscriber) {
+            eprintln!("Failed to set tracing subscriber: {}", e);
+        } else {
+            tracing::info!(
+                "FFF tracing initialized with log file: {}",
+                log_path.display()
+            );
+        }
+
+        guard
+    });
+
+    Ok(log_file_path.to_string())
+}
@@ -1,12 +1,5 @@
-//! Path utility functions for file picker scoring
-
 use std::path::{Path, PathBuf};

-/// Canonicalize a path, resolving symlinks and producing an absolute path.
-///
-/// On Windows, uses `dunce::canonicalize` to avoid the `\\?\` extended-length path prefix
-/// that `std::fs::canonicalize` produces. Neovim cannot open paths with this prefix.
-/// On other platforms, delegates directly to `std::fs::canonicalize`.
 #[cfg(windows)]
 pub fn canonicalize(path: impl AsRef<Path>) -> std::io::Result<PathBuf> {
    dunce::canonicalize(path)
@@ -17,6 +10,22 @@ pub fn canonicalize(path: impl AsRef<Path>) -> std::io::Result<PathBuf> {
    std::fs::canonicalize(path)
 }

+#[cfg(windows)]
+pub fn expand_tilde(path: &str) -> PathBuf {
+    return PathBuf::from(path);
+}
+
+#[cfg(not(windows))]
+pub fn expand_tilde(path: &str) -> PathBuf {
+    if let Some(stripped) = path.strip_prefix("~/")
+        && let Some(home_dir) = dirs::home_dir()
+    {
+        return home_dir.join(stripped);
+    }
+
+    PathBuf::from(path)
+}
+
 /// Calculate distance penalty based on directory proximity
 /// Returns a negative penalty score based on how far the candidate is from the current file
 pub fn calculate_distance_penalty(current_file: Option<&str>, candidate_path: &str) -> i32 {
@@ -61,10 +61,13 @@ impl DbHealthChecker for QueryTracker {
 }

 impl QueryTracker {
-    pub fn new(db_path: &str, use_unsafe_no_lock: bool) -> Result<Self, Error> {
+    pub fn new(db_path: impl AsRef<Path>, use_unsafe_no_lock: bool) -> Result<Self, Error> {
+        let db_path = db_path.as_ref();
        fs::create_dir_all(db_path).map_err(Error::CreateDir)?;
+
        let env = unsafe {
            let mut opts = EnvOpenOptions::new();
+            opts.map_size(10 * 1024 * 1024); // 100 MiB
            opts.max_dbs(16); // Allow up to 16 databases per environment
            if use_unsafe_no_lock {
                opts.flags(EnvFlags::NO_LOCK | EnvFlags::NO_SYNC | EnvFlags::NO_META_SYNC);
@@ -131,39 +131,34 @@ pub fn match_and_score_files<'a>(
        return (vec![], vec![], 0);
    }

-    let parsed = &context.parsed_query;
-    let working_files: FileItems<'a> = match parsed.as_ref().and_then(|p| {
-        if p.constraints.is_empty() {
-            None
-        } else {
-            apply_constraints(files, &p.constraints)
+    let parsed = context.query;
+    let working_files: FileItems<'a> = if parsed.constraints.is_empty() {
+        FileItems::All(files)
+    } else {
+        match apply_constraints(files, &parsed.constraints) {
+            Some(filtered) if !filtered.is_empty() => FileItems::Filtered(filtered),
+            Some(_) => {
+                return (vec![], vec![], 0);
+            }
+            None => FileItems::All(files),
        }
-    }) {
-        Some(filtered) if !filtered.is_empty() => FileItems::Filtered(filtered),
-        Some(_) => {
-            return (vec![], vec![], 0);
-        }
-        None => FileItems::All(files),
    };

-    let query_trimmed: &str = context.raw_query.trim();
+    let query_trimmed: &str = parsed.raw_query.trim();
    let single_part_storage: [&str; 1] = [query_trimmed];

-    let fuzzy_parts: &[&str] = match parsed {
-        None => {
-            tracing::debug!("STEP 3: Query too short (<2 chars), returning frecency-sorted");
+    let fuzzy_parts: &[&str] = match &parsed.fuzzy_query {
+        FuzzyQuery::Text(t) if t.len() >= 2 => std::slice::from_ref(t),
+        FuzzyQuery::Parts(parts) if !parts.is_empty() => parts.as_slice(),
+        FuzzyQuery::Text(_) | FuzzyQuery::Parts(_) => {
+            return score_filtered_by_frecency(&working_files, context);
+        }
+        FuzzyQuery::Empty => {
            if query_trimmed.len() < 2 {
                return score_filtered_by_frecency(&working_files, context);
            }
            &single_part_storage
        }
-        Some(p) => match &p.fuzzy_query {
-            FuzzyQuery::Text(t) if t.len() >= 2 => std::slice::from_ref(t),
-            FuzzyQuery::Parts(parts) if !parts.is_empty() => parts.as_slice(),
-            _ => {
-                return score_filtered_by_frecency(&working_files, context);
-            }
-        },
    };

    let has_uppercase = fuzzy_parts
@@ -172,7 +167,6 @@ pub fn match_and_score_files<'a>(
    let query_contains_path_separator = fuzzy_parts.iter().any(|p| p.contains(MAIN_SEPARATOR));

    let options = neo_frizbee::Config {
-        prefilter: true,
        max_typos: Some(context.max_typos),
        sort: false,
        scoring: Scoring {
@@ -230,6 +224,14 @@ pub fn match_and_score_files<'a>(

            let mut base_score = path_match.score as i32;
            let frecency_boost = base_score.saturating_mul(file.total_frecency_score as i32) / 100;
+
+            // Give modified/dirty files a 15% boost to make them appear higher in results
+            let git_status_boost = if file.git_status.is_some_and(is_modified_status) {
+                base_score * 15 / 100
+            } else {
+                0
+            };
+
            let distance_penalty =
                calculate_distance_penalty(context.current_file, &file.relative_path);

@@ -279,6 +281,7 @@ pub fn match_and_score_files<'a>(
            let combo_match_boost = {
                let last_same_query_match = context
                    .last_same_query_match
+                    .as_ref()
                    .filter(|m| m.file_path.as_os_str() == file.path.as_os_str());

                match last_same_query_match {
@@ -296,6 +299,7 @@ pub fn match_and_score_files<'a>(

            let total = base_score
                .saturating_add(frecency_boost)
+                .saturating_add(git_status_boost)
                .saturating_add(distance_penalty)
                .saturating_add(filename_bonus)
                .saturating_add(current_file_penalty)
@@ -312,12 +316,14 @@ pub fn match_and_score_files<'a>(
                    0
                },
                frecency_boost,
+                git_status_boost,
                distance_penalty,
                combo_match_boost,
                exact_match: path_match.exact || filename_match.is_some_and(|m| m.exact),
                match_type: match filename_match {
                    Some(filename_match) if filename_match.exact => "exact_filename",
                    Some(_) => "fuzzy_filename",
+                    None if path_match.exact => "exact_path",
                    None => "fuzzy_path",
                },
            };
@@ -363,9 +369,18 @@ pub(crate) fn score_filtered_by_frecency<'a>(
        let total_frecency_score = file.access_frecency_score as i32
            + (file.modification_frecency_score as i32).saturating_mul(4);

+        // Give modified/dirty files a boost even in frecency-only mode
+        let git_status_boost = if file.git_status.is_some_and(is_modified_status) {
+            total_frecency_score * 15 / 100
+        } else {
+            0
+        };
+
        let current_file_penalty =
            calculate_current_file_penalty(file, total_frecency_score, context);
-        let total = total_frecency_score.saturating_add(current_file_penalty);
+        let total = total_frecency_score
+            .saturating_add(git_status_boost)
+            .saturating_add(current_file_penalty);

        let score = Score {
            total,
@@ -376,6 +391,7 @@ pub(crate) fn score_filtered_by_frecency<'a>(
            combo_match_boost: 0,
            current_file_penalty,
            frecency_boost: total_frecency_score,
+            git_status_boost,
            exact_match: false,
            match_type: "frecency",
        };
@@ -482,10 +498,11 @@ fn sort_and_paginate<'a>(
 mod tests {
    use super::*;
    use crate::types::PaginationArgs;
+    use fff_query_parser::QueryParser;
    use std::path::PathBuf;

    fn create_test_file(path: &str, score: i32, modified: u64) -> (FileItem, Score) {
-        let file_name = path.split('/').last().unwrap_or(path).to_string();
+        let file_name = path.split('/').next_back().unwrap_or(path).to_string();
        let file = FileItem::new_raw(
            PathBuf::from(path),
            path.to_string(),
@@ -503,6 +520,7 @@ mod tests {
            special_filename_bonus: 0,
            current_file_penalty: 0,
            frecency_boost: 0,
+            git_status_boost: 0,
            exact_match: false,
            match_type: "test",
            combo_match_boost: 0,
@@ -532,9 +550,11 @@ mod tests {
            .map(|(file, score)| (file, score.clone()))
            .collect();

+        let query_str = "test";
+        let parser = QueryParser::default();
+        let query = parser.parse(query_str);
        let context = ScoringContext {
-            raw_query: "test",
-            parsed_query: None,
+            query: &query,
            max_threads: 1,
            max_typos: 2,
            current_file: None,
@@ -568,7 +588,7 @@ mod tests {
    #[test]
    fn test_partial_sort_with_same_scores() {
        // Test tiebreaker with modified time
-        let test_data = vec![
+        let test_data = [
            create_test_file("file1.rs", 100, 5000), // Same score, older
            create_test_file("file2.rs", 100, 8000), // Same score, newer
            create_test_file("file3.rs", 100, 3000), // Same score, oldest
@@ -581,9 +601,11 @@ mod tests {
            .map(|(file, score)| (file, score.clone()))
            .collect();

+        let query_str = "test";
+        let parser = QueryParser::default();
+        let query = parser.parse(query_str);
        let context = ScoringContext {
-            raw_query: "test",
-            parsed_query: None,
+            query: &query,
            max_threads: 1,
            max_typos: 2,
            current_file: None,
@@ -617,7 +639,7 @@ mod tests {
    #[test]
    fn test_no_partial_sort_for_small_results() {
        // When results.len() <= threshold, should use regular sort
-        let test_data = vec![
+        let test_data = [
            create_test_file("file1.rs", 100, 1000),
            create_test_file("file2.rs", 200, 2000),
            create_test_file("file3.rs", 50, 3000),
@@ -628,9 +650,11 @@ mod tests {
            .map(|(file, score)| (file, score.clone()))
            .collect();

+        let query_str = "test";
+        let parser = QueryParser::default();
+        let query = parser.parse(query_str);
        let context = ScoringContext {
-            raw_query: "test",
-            parsed_query: None,
+            query: &query,
            max_threads: 1,
            max_typos: 2,
            current_file: None,
@@ -666,7 +690,6 @@ mod multi_part_tests {

        // Test with max_typos = 2 (safe for short needles)
        let options = neo_frizbee::Config {
-            prefilter: true,
            max_typos: Some(2),
            sort: false,
            ..Default::default()
@@ -698,7 +721,6 @@ mod multi_part_tests {
        let path = "core_workflow_service/kafka_event_consumer/src/ai_part_extraction_request/ai_part_extraction_request_handler.rs".to_lowercase();

        let options = neo_frizbee::Config {
-            prefilter: true,
            max_typos: Some(2),
            sort: false,
            ..Default::default()
@@ -1,18 +1,41 @@
 use std::path::{Path, PathBuf};
 use std::sync::OnceLock;

-use memmap2::Mmap;
-
 use crate::constraints::Constrainable;
 use crate::query_tracker::QueryMatchEntry;
 use fff_query_parser::{FFFQuery, FuzzyQuery, Location};

-/// A single indexed file with metadata, frecency scores, and lazy mmap.
+/// Cached file contents — mmap on Unix, heap buffer on Windows.
 ///
-/// The `mmap` field holds the memory-mapped file contents, initialized lazily
-/// on the first grep access and cached for subsequent searches. The mmap is
-/// backed by the kernel page cache and automatically reflects file modifications
-/// — no manual invalidation is needed.
+/// On Windows, memory-mapped files hold the file handle open and prevent
+/// editors from saving (writing/replacing) those files. Reading into a
+/// `Vec<u8>` releases the handle immediately after the read completes.
+#[derive(Debug)]
+#[allow(dead_code)] // variants are conditionally used per platform
+enum FileContent {
+    #[cfg(not(target_os = "windows"))]
+    Mmap(memmap2::Mmap),
+    #[cfg(target_os = "windows")]
+    Buffer(Vec<u8>),
+}
+
+impl std::ops::Deref for FileContent {
+    type Target = [u8];
+    fn deref(&self) -> &[u8] {
+        match self {
+            #[cfg(not(target_os = "windows"))]
+            FileContent::Mmap(m) => m,
+            #[cfg(target_os = "windows")]
+            FileContent::Buffer(b) => b,
+        }
+    }
+}
+
+/// A single indexed file with metadata, frecency scores, and lazy content cache.
+///
+/// File contents are initialized lazily on the first grep access and cached for
+/// subsequent searches. On Unix, uses mmap backed by the kernel page cache. On
+/// Windows, reads into a heap buffer to avoid holding file handles open.
 ///
 /// Thread-safety: `OnceLock` provides lock-free reads after initialization.
 /// Each file is only searched by one rayon worker at a time via `par_iter`.
@@ -30,10 +53,9 @@ pub struct FileItem {
    pub total_frecency_score: i64,
    pub git_status: Option<git2::Status>,
    pub is_binary: bool,
-    /// Lazily-initialized memory-mapped file contents for grep.
+    /// Lazily-initialized file contents for grep.
    /// Initialized on first grep access via `OnceLock`; lock-free on subsequent reads.
-    /// Automatically reflects file changes via the kernel page cache.
-    mmap: OnceLock<Mmap>,
+    content: OnceLock<FileContent>,
 }

 impl Clone for FileItem {
@@ -51,8 +73,8 @@ impl Clone for FileItem {
            total_frecency_score: self.total_frecency_score,
            git_status: self.git_status,
            is_binary: self.is_binary,
-            // Don't clone the mmap — the clone lazily re-creates it on demand
-            mmap: OnceLock::new(),
+            // Don't clone the content — the clone lazily re-creates it on demand
+            content: OnceLock::new(),
        }
    }
 }
@@ -83,46 +105,65 @@ impl FileItem {
            total_frecency_score: 0,
            git_status,
            is_binary,
-            mmap: OnceLock::new(),
+            content: OnceLock::new(),
        }
    }

-    /// Invalidate the cached mmap so the next `get_mmap()` call creates a fresh one.
+    /// Invalidate the cached content so the next `get_content()` call creates a fresh one.
    ///
    /// Call this when the background watcher detects that the file has been modified.
-    /// While the kernel page cache reflects content changes automatically, a file
-    /// that is truncated (made smaller) while mapped can cause SIGBUS if the search
-    /// accesses pages beyond the new file size. Invalidating the mmap ensures a
-    /// fresh mapping with the correct size is created on the next access.
+    /// On Unix, a file that is truncated while mapped can cause SIGBUS. On Windows,
+    /// the stale buffer simply won't reflect the new contents. In both cases,
+    /// invalidating ensures a fresh read on the next access.
    pub fn invalidate_mmap(&mut self) {
-        self.mmap = OnceLock::new();
+        self.content = OnceLock::new();
    }

-    /// Get the cached mmap or lazily create it. Returns `None` if the file
-    /// is too large, empty, or can't be opened/mapped.
+    /// Get the cached file contents or lazily load them. Returns `None` if the
+    /// file is too large, empty, or can't be opened.
    ///
    /// After the first call, this is lock-free (just an atomic load + pointer deref).
-    /// The mmap is backed by the kernel page cache and automatically reflects
-    /// file modifications — no manual invalidation is needed.
+    /// On Unix, uses mmap backed by the kernel page cache. On Windows, reads into
+    /// a heap buffer so the file handle is released immediately.
    #[inline]
-    pub fn get_mmap(&self) -> Option<&Mmap> {
-        if let Some(mmap) = self.mmap.get() {
-            return Some(mmap);
+    pub fn get_content(&self) -> Option<&[u8]> {
+        if let Some(content) = self.content.get() {
+            return Some(content);
        }

        if self.size == 0 || self.size > MAX_MMAP_FILE_SIZE {
            return None;
        }

-        let file = std::fs::File::open(&self.path).ok()?;
+        let content = load_file_content(&self.path)?;
+
+        // If another thread raced us, OnceLock discards ours and returns theirs.
+        Some(self.content.get_or_init(|| content))
+    }
+
+    /// Backward-compatible alias for `get_content`.
+    #[inline]
+    pub fn get_mmap(&self) -> Option<&[u8]> {
+        self.get_content()
+    }
+}
+
+/// Load file contents: mmap on Unix, heap buffer on Windows.
+fn load_file_content(path: &Path) -> Option<FileContent> {
+    #[cfg(not(target_os = "windows"))]
+    {
+        let file = std::fs::File::open(path).ok()?;
        // SAFETY: The mmap is backed by the kernel page cache and automatically
        // reflects file modifications. The only risk is SIGBUS if the file is
-        // truncated while mapped
-        let mmap = unsafe { Mmap::map(&file) }.ok()?;
+        // truncated while mapped.
+        let mmap = unsafe { memmap2::Mmap::map(&file) }.ok()?;
+        Some(FileContent::Mmap(mmap))
+    }

-        // If another thread raced us, OnceLock discards our mmap and returns theirs.
-        // This is fine — the duplicate mmap is just dropped.
-        Some(self.mmap.get_or_init(|| mmap))
+    #[cfg(target_os = "windows")]
+    {
+        let data = std::fs::read(path).ok()?;
+        Some(FileContent::Buffer(data))
    }
 }

@@ -148,13 +189,14 @@ impl Constrainable for FileItem {
    }
 }

-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, Default)]
 pub struct Score {
    pub total: i32,
    pub base_score: i32,
    pub filename_bonus: i32,
    pub special_filename_bonus: i32,
    pub frecency_boost: i32,
+    pub git_status_boost: i32,
    pub distance_penalty: i32,
    pub current_file_penalty: i32,
    pub combo_match_boost: i32,
@@ -168,38 +210,42 @@ pub struct PaginationArgs {
    pub limit: usize,
 }

+impl Default for PaginationArgs {
+    fn default() -> Self {
+        Self {
+            offset: 0,
+            limit: 100,
+        }
+    }
+}
+
 /// Context for scoring files during search.
 ///
-/// The `parsed_query` field contains the pre-parsed query with constraints,
+/// The `query` field contains the pre-parsed query with constraints,
 /// fuzzy parts, and location information. Parsing is done once at the API
 /// boundary and passed through.
 #[derive(Debug, Clone)]
 pub struct ScoringContext<'a> {
-    /// The original raw query string (for compatibility and debugging)
-    pub raw_query: &'a str,
-    /// Pre-parsed query containing constraints, fuzzy parts, and location
-    pub parsed_query: Option<FFFQuery<'a>>,
+    /// Parsed query containing raw text, constraints, fuzzy parts, and location
+    pub query: &'a FFFQuery<'a>,
    pub project_path: Option<&'a Path>,
    pub current_file: Option<&'a str>,
    pub max_typos: u16,
    pub max_threads: usize,
-    pub last_same_query_match: Option<&'a QueryMatchEntry>,
+    pub last_same_query_match: Option<QueryMatchEntry>,
    pub combo_boost_score_multiplier: i32,
    pub min_combo_count: u32,
    pub pagination: PaginationArgs,
 }

-impl<'a> ScoringContext<'a> {
+impl ScoringContext<'_> {
    /// Get the effective fuzzy query string for matching.
    /// Returns the first fuzzy part, or the raw query if no parsing was done.
-    pub fn effective_query(&self) -> &'a str {
-        match &self.parsed_query {
-            Some(p) => match &p.fuzzy_query {
-                FuzzyQuery::Text(t) => t,
-                FuzzyQuery::Parts(parts) if !parts.is_empty() => parts[0],
-                _ => self.raw_query.trim(),
-            },
-            None => self.raw_query.trim(),
+    pub fn effective_query(&self) -> &str {
+        match &self.query.fuzzy_query {
+            FuzzyQuery::Text(t) => t,
+            FuzzyQuery::Parts(parts) if !parts.is_empty() => parts[0],
+            _ => self.query.raw_query.trim(),
        }
    }
 }
@@ -2,8 +2,8 @@ use std::fs;
 use std::path::Path;
 use tempfile::TempDir;

-use fff_core::grep::{GrepMode, GrepSearchOptions, grep_search, parse_grep_query};
-use fff_core::types::FileItem;
+use fff_search::grep::{GrepMode, GrepSearchOptions, grep_search, parse_grep_query};
+use fff_search::types::FileItem;

 /// Create a file inside a temp dir and return its `FileItem`.
 fn create_file(base: &Path, relative: &str, contents: &str) -> FileItem {
@@ -25,6 +25,9 @@ fn plain_opts() -> GrepSearchOptions {
        page_limit: 200,
        mode: GrepMode::PlainText,
        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
    }
 }

@@ -38,6 +41,9 @@ fn regex_opts() -> GrepSearchOptions {
        page_limit: 200,
        mode: GrepMode::Regex,
        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
    }
 }

@@ -51,6 +57,9 @@ fn fuzzy_opts() -> GrepSearchOptions {
        page_limit: 200,
        mode: GrepMode::Fuzzy,
        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
    }
 }

@@ -63,7 +72,8 @@ fn plain_text_finds_exact_literal() {
        "Hello, World!\nGoodbye, World!\n",
    )];

-    let result = grep_search(&files, "Hello", None, &plain_opts());
+    let parsed = parse_grep_query("Hello");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
    assert_eq!(result.matches[0].line_number, 1);
@@ -80,7 +90,8 @@ fn plain_text_smart_case_insensitive() {
    )];

    // All lowercase query → smart case → case-insensitive
-    let result = grep_search(&files, "hello", None, &plain_opts());
+    let parsed = parse_grep_query("hello");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -99,7 +110,8 @@ fn plain_text_smart_case_sensitive_with_uppercase() {
    )];

    // Query has uppercase → smart case → case-sensitive
-    let result = grep_search(&files, "Hello", None, &plain_opts());
+    let parsed = parse_grep_query("Hello");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -119,13 +131,15 @@ fn plain_text_regex_metacharacters_are_literal() {
    )];

    // In plain text mode, these regex metacharacters should be literal
-    let result = grep_search(&files, "fn main()", None, &plain_opts());
+    let parsed = parse_grep_query("fn main()");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
    assert_eq!(result.matches[0].line_number, 1);

    // Parentheses should NOT be treated as regex groups
-    let result2 = grep_search(&files, "(\"test\")", None, &plain_opts());
+    let parsed2 = parse_grep_query("(\"test\")");
+    let result2 = grep_search(&files, &parsed2, &plain_opts());
    assert_eq!(result2.matches.len(), 1);
    assert_eq!(result2.matches[0].line_number, 2);
 }
@@ -140,7 +154,8 @@ fn plain_text_dot_is_literal() {
    )];

    // In plain text mode, dot should be literal, not "any char"
-    let result = grep_search(&files, "1.0", None, &plain_opts());
+    let parsed = parse_grep_query("1.0");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -159,7 +174,8 @@ fn plain_text_asterisk_is_literal() {
        "Use **bold** text\nUse *italic* text\nUse normal text\n",
    )];

-    let result = grep_search(&files, "**bold**", None, &plain_opts());
+    let parsed = parse_grep_query("**bold**");
+    let result = grep_search(&files, &parsed, &plain_opts());
    assert_eq!(result.matches.len(), 1);
    assert_eq!(result.matches[0].line_number, 1);
 }
@@ -173,7 +189,8 @@ fn plain_text_backslash_is_literal() {
        "C:\\Users\\foo\\bar\n/home/user/bin\n",
    )];

-    let result = grep_search(&files, "C:\\Users", None, &plain_opts());
+    let parsed = parse_grep_query("C:\\Users");
+    let result = grep_search(&files, &parsed, &plain_opts());
    assert_eq!(result.matches.len(), 1);
 }

@@ -186,9 +203,9 @@ fn plain_text_across_multiple_files() {
        create_file(tmp.path(), "c.txt", "no match here\n"),
    ];

-    let result = grep_search(&files, "use std", None, &plain_opts());
+    let parsed = parse_grep_query("use std");
+    let result = grep_search(&files, &parsed, &plain_opts());

-    assert_eq!(result.total_match_count, 3);
    assert_eq!(result.matches.len(), 3);
    // Should match in files a.txt and b.txt
    assert_eq!(result.files.len(), 2);
@@ -199,7 +216,8 @@ fn plain_text_highlight_offsets_are_correct() {
    let tmp = TempDir::new().unwrap();
    let files = vec![create_file(tmp.path(), "a.txt", "foo bar foo baz foo\n")];

-    let result = grep_search(&files, "foo", None, &plain_opts());
+    let parsed = parse_grep_query("foo");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
    let m = &result.matches[0];
@@ -216,7 +234,8 @@ fn plain_text_empty_query_returns_no_content_matches() {
    let tmp = TempDir::new().unwrap();
    let files = vec![create_file(tmp.path(), "a.txt", "some content\n")];

-    let result = grep_search(&files, "", None, &plain_opts());
+    let parsed = parse_grep_query("");
+    let result = grep_search(&files, &parsed, &plain_opts());

    // Empty query in grep returns git-modified welcome state (no content matches)
    // Since our test files have no git status, we expect 0 matches
@@ -237,7 +256,8 @@ fn plain_text_binary_files_are_skipped() {

    let files = vec![binary_file, text_file];

-    let result = grep_search(&files, "match this text", None, &plain_opts());
+    let parsed = parse_grep_query("match this text");
+    let result = grep_search(&files, &parsed, &plain_opts());

    // Only the text file should be searched, not the binary one
    assert_eq!(result.files.len(), 1);
@@ -256,7 +276,8 @@ fn plain_text_max_matches_per_file() {
    let mut opts = plain_opts();
    opts.max_matches_per_file = 5;

-    let result = grep_search(&files, "match_target", None, &opts);
+    let parsed = parse_grep_query("match_target");
+    let result = grep_search(&files, &parsed, &opts);

    assert_eq!(
        result.matches.len(),
@@ -277,19 +298,31 @@ fn plain_text_page_limit() {
    let mut opts = plain_opts();
    opts.page_limit = 10;

-    let result = grep_search(&files, "target", None, &opts);
+    let parsed = parse_grep_query("target");
+    let result = grep_search(&files, &parsed, &opts);

+    // page_limit is a soft minimum: we always finish the current file, so we
+    // get at least page_limit matches (no data loss) and at most
+    // max_matches_per_file (200) from a single file.
    assert!(
-        result.matches.len() <= 10,
-        "should respect page_limit: got {}",
+        result.matches.len() >= opts.page_limit,
+        "should return at least page_limit matches: got {}",
        result.matches.len()
    );
+    assert!(
+        result.matches.len() <= opts.max_matches_per_file,
+        "should never exceed max_matches_per_file: got {}",
+        result.matches.len()
+    );
+    // Single file with 100 lines all matching — all should be returned.
+    assert_eq!(result.matches.len(), 100, "all 100 lines must be returned");
 }

 #[test]
 fn plain_text_file_offset_pagination() {
    let tmp = TempDir::new().unwrap();
-    // Create many files so file-based pagination works
+    // Create many files (1 match per file) so file-based pagination exercises
+    // offset tracking across files with and without matches.
    let mut files = Vec::new();
    for i in 0..20 {
        files.push(create_file(
@@ -302,14 +335,52 @@ fn plain_text_file_offset_pagination() {
    let mut opts = plain_opts();
    opts.page_limit = 5;

-    let result1 = grep_search(&files, "unique_token", None, &opts);
-    assert!(result1.matches.len() > 0);
+    // Collect ALL matches across all pages and verify no duplicates and full coverage.
+    let mut all_line_texts: Vec<String> = Vec::new();
+    let mut pages = 0;
+    let max_pages = 20; // safety limit

-    if result1.next_file_offset > 0 {
-        opts.file_offset = result1.next_file_offset;
-        let result2 = grep_search(&files, "unique_token", None, &opts);
-        assert!(result2.matches.len() > 0, "second page should have results");
+    loop {
+        let parsed = parse_grep_query("unique_token");
+        let result = grep_search(&files, &parsed, &opts);
+
+        for m in &result.matches {
+            let text = m.line_content.trim().to_string();
+            assert!(
+                !all_line_texts.contains(&text),
+                "duplicate match across pages: '{}'",
+                text
+            );
+            all_line_texts.push(text);
+        }
+
+        pages += 1;
+        assert!(pages <= max_pages, "pagination did not terminate");
+
+        if result.next_file_offset == 0 {
+            break;
+        }
+
+        // Offset must strictly advance
+        assert!(
+            result.next_file_offset > opts.file_offset,
+            "next_file_offset ({}) did not advance past current ({})",
+            result.next_file_offset,
+            opts.file_offset
+        );
+        opts.file_offset = result.next_file_offset;
    }
+
+    assert_eq!(
+        all_line_texts.len(),
+        20,
+        "pagination should find all 20 matches across all pages, got {}",
+        all_line_texts.len()
+    );
+    assert!(
+        pages > 1,
+        "should require multiple pages with page_limit=5 and 20 files"
+    );
 }

 #[test]
@@ -321,7 +392,8 @@ fn plain_text_line_numbers_are_correct() {
        "line one\nline two\nline three\nline four\n",
    )];

-    let result = grep_search(&files, "line", None, &plain_opts());
+    let parsed = parse_grep_query("line");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 4);
    assert_eq!(result.matches[0].line_number, 1);
@@ -340,7 +412,8 @@ fn plain_text_max_file_size_filter() {
    let mut opts = plain_opts();
    opts.max_file_size = 100; // Only allow files up to 100 bytes

-    let result = grep_search(&files, "match_me", None, &opts);
+    let parsed = parse_grep_query("match_me");
+    let result = grep_search(&files, &parsed, &opts);

    assert_eq!(result.matches.len(), 0, "large file should be filtered out");
    assert_eq!(result.filtered_file_count, 0);
@@ -357,7 +430,8 @@ fn regex_basic_pattern() {
        "foo123\nbar456\nbaz789\nfoo_bar\n",
    )];

-    let result = grep_search(&files, "foo\\d+", None, &regex_opts());
+    let parsed = parse_grep_query("foo\\d+");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(result.matches.len(), 1);
    assert_eq!(result.matches[0].line_number, 1);
@@ -370,7 +444,8 @@ fn regex_capture_group_matching() {
    let files = vec![create_file(tmp.path(), "a.txt", "foobar\nfoobaz\nfoo123\n")];

    // Use a capturing group (not lookahead, which regex crate doesn't support)
-    let result = grep_search(&files, "foo(bar|baz)", None, &regex_opts());
+    let parsed = parse_grep_query("foo(bar|baz)");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(result.matches.len(), 2);
    let contents: Vec<&str> = result
@@ -392,7 +467,8 @@ fn regex_dot_matches_any_char() {
    )];

    // In regex mode, . matches any character, so v1.0 matches v1.0, v1x0, v1-0, and v100
-    let result = grep_search(&files, "v1.0", None, &regex_opts());
+    let parsed = parse_grep_query("v1.0");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(
        result.matches.len(),
@@ -410,7 +486,8 @@ fn regex_alternation() {
        "apple\nbanana\ncherry\napricot\n",
    )];

-    let result = grep_search(&files, "apple|cherry", None, &regex_opts());
+    let parsed = parse_grep_query("apple|cherry");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(result.matches.len(), 2);
    let lines: Vec<u64> = result.matches.iter().map(|m| m.line_number).collect();
@@ -427,7 +504,8 @@ fn regex_character_class() {
        "cat\ncut\ncot\ncit\ncxt\n",
    )];

-    let result = grep_search(&files, "c[aou]t", None, &regex_opts());
+    let parsed = parse_grep_query("c[aou]t");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(result.matches.len(), 3);
    let contents: Vec<&str> = result
@@ -449,7 +527,8 @@ fn regex_quantifiers() {
        "fo\nfoo\nfooo\nfoooo\nbar\n",
    )];

-    let result = grep_search(&files, "fo{2,3}", None, &regex_opts());
+    let parsed = parse_grep_query("fo{2,3}");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(result.matches.len(), 3, "should match foo, fooo, foooo");
 }
@@ -463,18 +542,42 @@ fn regex_anchors() {
        "start of line\nmiddle start end\nend of line\n",
    )];

-    let result = grep_search(&files, "^start", None, &regex_opts());
+    let parsed = parse_grep_query("^start");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(result.matches.len(), 1);
    assert_eq!(result.matches[0].line_number, 1);
 }

+#[test]
+fn regex_anchors_multiword() {
+    let tmp = TempDir::new().unwrap();
+    let files = vec![create_file(
+        tmp.path(),
+        "test.c",
+        "int ff_function(void);\nstatic int ff_other(void);\nint main(void);\nint ff_another(void);\n",
+    )];
+
+    // ^int ff_ should match lines starting with "int ff_"
+    let parsed = parse_grep_query("^int ff_");
+    let result = grep_search(&files, &parsed, &regex_opts());
+
+    assert_eq!(
+        result.matches.len(),
+        2,
+        "should match 2 lines starting with 'int ff_'"
+    );
+    assert!(result.matches[0].line_content.contains("ff_function"));
+    assert!(result.matches[1].line_content.contains("ff_another"));
+}
+
 #[test]
 fn regex_highlight_offsets_variable_length() {
    let tmp = TempDir::new().unwrap();
    let files = vec![create_file(tmp.path(), "a.txt", "aab aaab aaaab\n")];

-    let result = grep_search(&files, "a+b", None, &regex_opts());
+    let parsed = parse_grep_query("a+b");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(result.matches.len(), 1);
    let m = &result.matches[0];
@@ -496,7 +599,8 @@ fn regex_invalid_pattern_falls_back_to_literal() {
    )];

    // Invalid regex: unmatched group — should fall back to literal search
-    let result = grep_search(&files, "name(", None, &regex_opts());
+    let parsed = parse_grep_query("name(");
+    let result = grep_search(&files, &parsed, &regex_opts());

    // Fallback to literal: finds "name(" in "call name(arg)"
    assert_eq!(
@@ -511,7 +615,8 @@ fn regex_invalid_pattern_falls_back_to_literal() {
    assert!(result.matches[0].line_content.contains("name("));

    // A pattern that doesn't exist anywhere — still falls back but finds nothing
-    let result2 = grep_search(&files, "zzz(", None, &regex_opts());
+    let parsed2 = parse_grep_query("zzz(");
+    let result2 = grep_search(&files, &parsed2, &regex_opts());
    assert_eq!(result2.matches.len(), 0);
    assert!(result2.regex_fallback_error.is_some());
 }
@@ -526,11 +631,13 @@ fn regex_smart_case() {
    )];

    // Lowercase query → case-insensitive
-    let result_lower = grep_search(&files, "foo", None, &regex_opts());
+    let parsed_lower = parse_grep_query("foo");
+    let result_lower = grep_search(&files, &parsed_lower, &regex_opts());
    assert_eq!(result_lower.matches.len(), 3);

    // Query with uppercase → case-sensitive
-    let result_upper = grep_search(&files, "Foo", None, &regex_opts());
+    let parsed_upper = parse_grep_query("Foo");
+    let result_upper = grep_search(&files, &parsed_upper, &regex_opts());
    assert_eq!(result_upper.matches.len(), 1);
 }

@@ -551,7 +658,8 @@ fn regex_across_multiple_files() {
        create_file(tmp.path(), "readme.md", "# Title\nSome text\n"),
    ];

-    let result = grep_search(&files, "fn \\w+\\(\\)", None, &regex_opts());
+    let parsed = parse_grep_query("fn \\w+\\(\\)");
+    let result = grep_search(&files, &parsed, &regex_opts());

    // Should match: fn main(), fn helper(), fn test_one(), fn test_two()
    assert_eq!(result.matches.len(), 4);
@@ -569,8 +677,9 @@ fn plain_text_and_regex_agree_on_simple_literal() {
        "hello world\ngoodbye world\nhello again\n",
    )];

-    let plain_result = grep_search(&files, "hello", None, &plain_opts());
-    let regex_result = grep_search(&files, "hello", None, &regex_opts());
+    let parsed = parse_grep_query("hello");
+    let plain_result = grep_search(&files, &parsed, &plain_opts());
+    let regex_result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(plain_result.matches.len(), regex_result.matches.len());
    for (p, r) in plain_result.matches.iter().zip(regex_result.matches.iter()) {
@@ -589,8 +698,10 @@ fn plain_text_escapes_what_regex_does_not() {
    )];

    // "$100" — in plain text, $ is literal; in regex, $ is anchor
-    let plain_result = grep_search(&files, "$100", None, &plain_opts());
-    let regex_result = grep_search(&files, "\\$100", None, &regex_opts());
+    let parsed_plain = parse_grep_query("$100");
+    let plain_result = grep_search(&files, &parsed_plain, &plain_opts());
+    let parsed_regex = parse_grep_query("\\$100");
+    let regex_result = grep_search(&files, &parsed_regex, &regex_opts());

    // Plain text should find "$100" literally
    assert_eq!(plain_result.matches.len(), 1);
@@ -612,7 +723,7 @@ fn grep_with_extension_constraint() {
    ];

    let parsed = parse_grep_query("use std *.rs");
-    let result = grep_search(&files, "use std *.rs", parsed, &plain_opts());
+    let result = grep_search(&files, &parsed, &plain_opts());

    // Should only search .rs files
    for file in &result.files {
@@ -639,7 +750,8 @@ fn plain_text_bracket_is_literal() {
        "let x = arr[0];\nlet y = arr[1];\nlet z = something;\n",
    )];

-    let result = grep_search(&files, "arr[0]", None, &plain_opts());
+    let parsed = parse_grep_query("arr[0]");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -661,7 +773,7 @@ fn grep_backslash_escapes_extension_filter() {

    // Without escape: "*.rs" is an extension filter, so only .rs files are searched
    let parsed = parse_grep_query("pattern *.rs");
-    let result_filter = grep_search(&files, "pattern *.rs", parsed, &plain_opts());
+    let result_filter = grep_search(&files, &parsed, &plain_opts());
    assert_eq!(
        result_filter.files.len(),
        1,
@@ -670,7 +782,7 @@ fn grep_backslash_escapes_extension_filter() {

    // With escape: "\*.rs" is literal text, both files are searched
    let parsed_escaped = parse_grep_query("\\*.rs");
-    let result_literal = grep_search(&files, "\\*.rs", parsed_escaped, &plain_opts());
+    let result_literal = grep_search(&files, &parsed_escaped, &plain_opts());
    assert_eq!(
        result_literal.matches.len(),
        2,
@@ -687,7 +799,8 @@ fn grep_backslash_escapes_path_segment() {
    ];

    // With escape: "\\/src/" is literal text, not a path constraint
-    let result = grep_search(&files, "\\/src/", None, &plain_opts());
+    let parsed = parse_grep_query("\\/src/");
+    let result = grep_search(&files, &parsed, &plain_opts());
    assert_eq!(
        result.matches.len(),
        2,
@@ -705,7 +818,8 @@ fn grep_backslash_escapes_negation() {
    )];

    // With escape: "\\!test" is literal text "!test"
-    let result = grep_search(&files, "\\!test", None, &plain_opts());
+    let parsed = parse_grep_query("\\!test");
+    let result = grep_search(&files, &parsed, &plain_opts());
    assert_eq!(result.matches.len(), 1);
    assert!(result.matches[0].line_content.contains("!test"));
 }
@@ -720,7 +834,7 @@ fn grep_with_path_constraint() {
    ];

    let parsed = parse_grep_query("target_text /src/");
-    let result = grep_search(&files, "target_text /src/", parsed, &plain_opts());
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
    assert!(result.files[0].relative_path.starts_with("src/"));
@@ -739,7 +853,7 @@ fn grep_with_negated_extension_constraint() {

    let query = "target_text !*.rs";
    let parsed = parse_grep_query(query);
-    let result = grep_search(&files, query, parsed, &plain_opts());
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -765,7 +879,7 @@ fn grep_with_negated_path_constraint() {

    let query = "target_text !/src/";
    let parsed = parse_grep_query(query);
-    let result = grep_search(&files, query, parsed, &plain_opts());
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -791,7 +905,7 @@ fn grep_with_negated_text_constraint() {

    let query = "target_text !test";
    let parsed = parse_grep_query(query);
-    let result = grep_search(&files, query, parsed, &plain_opts());
+    let result = grep_search(&files, &parsed, &plain_opts());

    // "tests/helper.rs" contains "test" in path, should be excluded
    assert_eq!(
@@ -821,7 +935,8 @@ fn grep_empty_file_is_skipped() {
    let text_file = create_file(tmp.path(), "text.txt", "findme\n");

    let files = vec![empty_file, text_file];
-    let result = grep_search(&files, "findme", None, &plain_opts());
+    let parsed = parse_grep_query("findme");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
 }
@@ -831,7 +946,8 @@ fn grep_single_line_no_trailing_newline() {
    let tmp = TempDir::new().unwrap();
    let files = vec![create_file(tmp.path(), "a.txt", "no newline at end")];

-    let result = grep_search(&files, "no newline", None, &plain_opts());
+    let parsed = parse_grep_query("no newline");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
    assert_eq!(result.matches[0].line_number, 1);
@@ -846,11 +962,13 @@ fn grep_unicode_content() {
        "日本語テスト\nrégulière\nñoño\n",
    )];

-    let result = grep_search(&files, "régulière", None, &plain_opts());
+    let parsed = parse_grep_query("régulière");
+    let result = grep_search(&files, &parsed, &plain_opts());
    assert_eq!(result.matches.len(), 1);
    assert_eq!(result.matches[0].line_number, 2);

-    let result2 = grep_search(&files, "ñoño", None, &plain_opts());
+    let parsed2 = parse_grep_query("ñoño");
+    let result2 = grep_search(&files, &parsed2, &plain_opts());
    assert_eq!(result2.matches.len(), 1);
    assert_eq!(result2.matches[0].line_number, 3);
 }
@@ -861,7 +979,8 @@ fn grep_long_line_is_truncated() {
    let long_line = "x".repeat(1000) + "NEEDLE" + &"y".repeat(1000);
    let files = vec![create_file(tmp.path(), "long.txt", &long_line)];

-    let result = grep_search(&files, "NEEDLE", None, &plain_opts());
+    let parsed = parse_grep_query("NEEDLE");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
    // The line_content should be truncated to MAX_LINE_DISPLAY_LEN (512)
@@ -881,7 +1000,8 @@ fn regex_word_boundary() {
        "foo\nfoobar\nbarfoo\nfoo_baz\n",
    )];

-    let result = grep_search(&files, "\\bfoo\\b", None, &regex_opts());
+    let parsed = parse_grep_query("\\bfoo\\b");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(
        result.matches.len(),
@@ -900,7 +1020,8 @@ fn plain_text_question_mark_is_literal() {
        "what is this?\nhow does it work?\nno question here\nwhat?\n",
    )];

-    let result = grep_search(&files, "?", None, &plain_opts());
+    let parsed = parse_grep_query("?");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -918,7 +1039,8 @@ fn plain_text_query_with_question_mark_in_word() {
        "let x = foo?;\nlet y = bar.baz();\nfoo?.unwrap()\n",
    )];

-    let result = grep_search(&files, "foo?", None, &plain_opts());
+    let parsed = parse_grep_query("foo?");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(
        result.matches.len(),
@@ -933,7 +1055,8 @@ fn regex_question_mark_is_quantifier() {
    let files = vec![create_file(tmp.path(), "a.txt", "color\ncolour\ncolouur\n")];

    // In regex mode, ? means "zero or one of preceding"
-    let result = grep_search(&files, "colou?r", None, &regex_opts());
+    let parsed = parse_grep_query("colou?r");
+    let result = grep_search(&files, &parsed, &regex_opts());

    assert_eq!(
        result.matches.len(),
@@ -953,7 +1076,8 @@ fn fuzzy_finds_exact_substring() {
        "hello world\ngoodbye world\nhello again\n",
    )];

-    let result = grep_search(&files, "hello", None, &fuzzy_opts());
+    let parsed = parse_grep_query("hello");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    assert_eq!(
        result.matches.len(),
@@ -974,10 +1098,11 @@ fn fuzzy_finds_scattered_characters() {
    )];

    // "mutex" should fuzzy match "mutex_lock" (contiguous prefix)
-    let result = grep_search(&files, "mutex", None, &fuzzy_opts());
+    let parsed = parse_grep_query("mutex");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    assert!(
-        result.matches.len() >= 1,
+        !result.matches.is_empty(),
        "fuzzy should find 'mutex' in 'mutex_lock'"
    );
    assert!(result.matches[0].line_content.contains("mutex_lock"));
@@ -988,7 +1113,8 @@ fn fuzzy_highlight_offsets_correct() {
    let tmp = TempDir::new().unwrap();
    let files = vec![create_file(tmp.path(), "a.txt", "hello world\n")];

-    let result = grep_search(&files, "hell", None, &fuzzy_opts());
+    let parsed = parse_grep_query("hell");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    assert_eq!(result.matches.len(), 1);
    let m = &result.matches[0];
@@ -1012,11 +1138,12 @@ fn fuzzy_unicode_char_indices() {

    // Use "guli" which is a contiguous ASCII substring within "régulière"
    // (the chars g-u-l-i appear contiguously between the two accented chars)
-    let result = grep_search(&files, "guli", None, &fuzzy_opts());
+    let parsed = parse_grep_query("guli");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    // Should fuzzy match "régulière" (with multi-byte é and è)
    // This tests that character-to-byte offset conversion works with UTF-8
-    assert!(result.matches.len() >= 1);
+    assert!(!result.matches.is_empty());
    assert!(result.matches[0].line_content.contains("régulière"));
 }

@@ -1025,7 +1152,8 @@ fn fuzzy_empty_query_returns_empty() {
    let tmp = TempDir::new().unwrap();
    let files = vec![create_file(tmp.path(), "a.txt", "some content\n")];

-    let result = grep_search(&files, "", None, &fuzzy_opts());
+    let parsed = parse_grep_query("");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    // Empty query returns git-modified files, not fuzzy matches
    assert_eq!(result.matches.len(), 0);
@@ -1041,7 +1169,7 @@ fn fuzzy_with_extension_constraint() {
    ];

    let parsed = parse_grep_query("use std *.rs");
-    let result = grep_search(&files, "use std *.rs", parsed, &fuzzy_opts());
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    // Should only search .rs files
    for file in &result.files {
@@ -1064,14 +1192,30 @@ fn fuzzy_respects_page_limit() {

    let mut opts = fuzzy_opts();
    opts.page_limit = 10;
+    opts.max_matches_per_file = 50;

-    let result = grep_search(&files, "target", None, &opts);
+    let parsed = parse_grep_query("target");
+    let result = grep_search(&files, &parsed, &opts);

+    // page_limit is a soft minimum: we always finish the current file, so we
+    // get at least page_limit matches (no data loss) and at most
+    // max_matches_per_file (200) from a single file.
    assert!(
-        result.matches.len() <= 10,
-        "should respect page_limit: got {}",
+        result.matches.len() >= opts.page_limit,
+        "should return at least page_limit matches: got {}",
        result.matches.len()
    );
+    assert!(
+        result.matches.len() <= opts.max_matches_per_file,
+        "should never exceed max_matches_per_file: got {}",
+        result.matches.len()
+    );
+
+    assert_eq!(
+        result.matches.len(),
+        opts.max_matches_per_file,
+        "all limit of lines must be returned"
+    );
 }

 #[test]
@@ -1086,7 +1230,8 @@ fn fuzzy_respects_max_matches_per_file() {
    let mut opts = fuzzy_opts();
    opts.max_matches_per_file = 5;

-    let result = grep_search(&files, "match", None, &opts);
+    let parsed = parse_grep_query("match");
+    let result = grep_search(&files, &parsed, &opts);

    assert_eq!(
        result.matches.len(),
@@ -1106,7 +1251,8 @@ fn fuzzy_filters_low_quality_matches() {

    // Search for "abc" - should match "abc_def_ghi" and "abcdefghij" with high scores,
    // but NOT "xyz" (no relation) or "mutex_lock" (only weak letter overlap)
-    let result = grep_search(&files, "abc", None, &fuzzy_opts());
+    let parsed = parse_grep_query("abc");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    // Should only get high-quality matches
    assert!(
@@ -1135,7 +1281,8 @@ fn fuzzy_exact_match_always_passes() {
    )];

    // Exact matches should always pass regardless of score threshold
-    let result = grep_search(&files, "exact", None, &fuzzy_opts());
+    let parsed = parse_grep_query("exact");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    assert_eq!(
        result.matches.len(),
@@ -1154,7 +1301,8 @@ fn fuzzy_score_is_captured() {
        "hello world\ngoodbye world\n",
    )];

-    let result = grep_search(&files, "hello", None, &fuzzy_opts());
+    let parsed = parse_grep_query("hello");
+    let result = grep_search(&files, &parsed, &fuzzy_opts());

    assert_eq!(result.matches.len(), 1);
    let m = &result.matches[0];
@@ -1175,7 +1323,8 @@ fn fuzzy_score_is_none_in_plain_mode() {
    let tmp = TempDir::new().unwrap();
    let files = vec![create_file(tmp.path(), "test.txt", "hello world\n")];

-    let result = grep_search(&files, "hello", None, &plain_opts());
+    let parsed = parse_grep_query("hello");
+    let result = grep_search(&files, &parsed, &plain_opts());

    assert_eq!(result.matches.len(), 1);
    let m = &result.matches[0];
@@ -1,6 +1,9 @@
 [package]
-name = "grep-searcher"
-version = "0.1.16"
+name = "fff-grep"
+description = "File grepping logic for fff"
+license = "MIT"
+authors = ["Dmitriy Kovalenko <dmtr.kovalenko@outlok.com>"]
+version = "0.2.4"
 edition = "2024"

 [dependencies]
@@ -120,12 +120,7 @@ impl Searcher {

    /// Execute a search over the given slice and write the results to the
    /// given sink.
-    pub fn search_slice<M, S>(
-        &mut self,
-        matcher: M,
-        slice: &[u8],
-        write_to: S,
-    ) -> Result<(), S::Error>
+    pub fn search_slice<M, S>(&self, matcher: M, slice: &[u8], write_to: S) -> Result<(), S::Error>
    where
        M: Matcher,
        S: Sink,
@@ -0,0 +1,27 @@
+[package]
+name = "fff-mcp"
+version = "0.2.4"
+edition = "2024"
+description = "MCP server for FFF file finder - drop-in replacement for AI code assistant search tools"
+license = "MIT"
+
+[[bin]]
+name = "fff-mcp"
+path = "src/main.rs"
+
+[features]
+default = ["zlob"]
+zlob = ["fff/zlob"]
+
+[dependencies]
+fff = { package = "fff-search", path = "../fff-core", default-features = false , version = "0.2.2" }
+fff-query-parser = { path = "../fff-query-parser", default-features = false , version = "0.2.4" }
+mimalloc = { workspace = true }
+rmcp = { version = "1.1.0", features = ["server", "transport-io"] }
+schemars = "1.0"
+serde = { version = "1.0", features = ["derive"] }
+serde_json = "1.0"
+tokio = { version = "1", features = ["full"] }
+tracing = { workspace = true }
+git2 = { workspace = true }
+clap = { version = "4", features = ["derive"] }
@@ -0,0 +1,15 @@
+fn main() {
+    // Embed the git commit hash at build time for update checking.
+    let hash = std::process::Command::new("git")
+        .args(["rev-parse", "HEAD"])
+        .output()
+        .ok()
+        .filter(|o| o.status.success())
+        .and_then(|o| String::from_utf8(o.stdout).ok())
+        .map(|s| s.trim().to_string())
+        .unwrap_or_else(|| "unknown".to_string());
+
+    println!("cargo:rustc-env=FFF_GIT_HASH={}", hash);
+    println!("cargo:rerun-if-changed=../../.git/HEAD");
+    println!("cargo:rerun-if-changed=../../.git/refs/");
+}
@@ -0,0 +1,52 @@
+//! Cursor store for grep pagination.
+//!
+//! Maintains an in-memory map of opaque cursor IDs to file offsets.
+//! Cursors are evicted LRU-style when the store exceeds capacity.
+
+use std::collections::{HashMap, VecDeque};
+
+const MAX_CURSORS: usize = 20;
+
+/// Stores cursor state for paginated grep results.
+pub struct CursorStore {
+    counter: u64,
+    /// Map from cursor ID string → file offset for next page.
+    cursors: HashMap<String, usize>,
+    /// Insertion order for LRU eviction.
+    insertion_order: VecDeque<String>,
+}
+
+impl CursorStore {
+    pub fn new() -> Self {
+        Self {
+            counter: 0,
+            cursors: HashMap::new(),
+            insertion_order: VecDeque::new(),
+        }
+    }
+
+    /// Store a cursor and return its opaque ID string.
+    pub fn store(&mut self, file_offset: usize) -> String {
+        self.counter = self.counter.wrapping_add(1);
+        let id = self.counter.to_string();
+
+        self.cursors.insert(id.clone(), file_offset);
+        self.insertion_order.push_back(id.clone());
+
+        // Evict oldest cursors
+        while self.cursors.len() > MAX_CURSORS {
+            if let Some(oldest) = self.insertion_order.pop_front() {
+                self.cursors.remove(&oldest);
+            } else {
+                break;
+            }
+        }
+
+        id
+    }
+
+    /// Retrieve the file offset for a cursor ID.
+    pub fn get(&self, id: &str) -> Option<usize> {
+        self.cursors.get(id).copied()
+    }
+}
@@ -0,0 +1,296 @@
+//! FFF MCP Server — high-performance file finder for AI code assistants.
+//!
+//! Drop-in replacement for AI code assistant file search tools (Glob/Grep).
+//! Provides frecency-ranked, fuzzy-matched, git-aware file finding and
+//! code search via the Model Context Protocol (MCP).
+//!
+//! Uses `fff-core` directly (zero FFI overhead) for all search operations.
+
+mod cursor;
+mod output;
+mod server;
+mod update_check;
+
+use std::sync::{Arc, RwLock};
+
+use clap::Parser;
+use fff::file_picker::FilePicker;
+use fff::frecency::FrecencyTracker;
+use fff::{FFFMode, SharedFrecency, SharedPicker};
+use git2::Repository;
+use mimalloc::MiMalloc;
+use rmcp::{ServiceExt, transport::stdio};
+use server::FffServer;
+
+#[global_allocator]
+static GLOBAL: MiMalloc = MiMalloc;
+
+pub const MCP_INSTRUCTIONS: &str = concat!(
+    "FFF is a fast file finder with frecency-ranked results (frequent/recent files first, git-dirty files boosted).\n",
+    "\n",
+    "## Which Tool Should I Use?\n",
+    "\n",
+    "- **grep**: DEFAULT tool. Searches file CONTENTS -- definitions, usage, patterns. Use when you have a specific name or pattern.\n",
+    "- **find_files**: Explores which files/modules exist for a topic. Use when you DON'T have a specific identifier or LOOKING FOR A FILE.\n",
+    "- **multi_grep**: OR logic across multiple patterns. Use for case variants (e.g. ['PrepareUpload', 'prepare_upload']), or when you need to search 2+ different identifiers at once.\n",
+    "\n",
+    "## Core Rules\n",
+    "\n",
+    "### 1. Search BARE IDENTIFIERS only\n",
+    "Grep matches single lines. Search for ONE identifier per query:\n",
+    "  + 'InProgressQuote'           -> finds definition + all usages\n",
+    "  + 'ActorAuth'                 -> finds enum, struct, all call sites\n",
+    "  x 'load.*metadata.*InProgressQuote' -> regex spanning multiple tokens, 0 results\n",
+    "  x 'ctx.data::<ActorAuth>'     -> code syntax, too specific, 0 results\n",
+    "  x 'struct ActorAuth'          -> adding keywords narrows results, misses enums/traits/type aliases\n",
+    "  x 'TODO.*#\\d+'               -> complex regex, use simple 'TODO' then filter visually\n",
+    "\n",
+    "### 2. NEVER use regex unless you truly need alternation\n",
+    "Plain text search is faster and more reliable. Regex patterns like `.*`, `\\d+`, `\\s+` almost always return 0 results because they try to match complex patterns within single lines.\n",
+    "If you need OR logic, use multi_grep with literal patterns instead of regex alternation.\n",
+    "\n",
+    "### 3. Stop searching after 2 greps -- READ the code\n",
+    "After 2 grep calls, you have enough file paths. Read the top result to understand the code.\n",
+    "Do NOT keep grepping with variations. More greps != better understanding.\n",
+    "\n",
+    "### 4. Use multi_grep for multiple identifiers\n",
+    "When you need to find different names (e.g. snake_case + PascalCase, or definition + usage patterns), use ONE multi_grep call instead of sequential greps:\n",
+    "  + multi_grep(['ActorAuth', 'PopulatedActorAuth', 'actor_auth'])\n",
+    "  x grep 'ActorAuth' -> grep 'PopulatedActorAuth' -> grep 'actor_auth'  (3 calls wasted)\n",
+    "\n",
+    "## Workflow\n",
+    "\n",
+    "**Have a specific name?** -> grep the bare identifier.\n",
+    "**Need multiple name variants?** -> multi_grep with all variants in one call.\n",
+    "**Exploring a topic / finding files?** -> find_files.\n",
+    "**Got results?** -> Read the top file. Don't grep again.\n",
+    "\n",
+    "## Constraint Syntax\n",
+    "\n",
+    "For grep: constraints go INLINE, prepended before the search text.\n",
+    "For multi_grep: constraints go in the separate 'constraints' parameter.\n",
+    "\n",
+    "Constraints MUST match one of these formats:\n",
+    "  Extension: '*.rs', '*.{ts,tsx}'\n",
+    "  Directory: 'src/', 'quotes/'\n",
+    "  Filename: 'schema.rs', 'src/main.rs'\n",
+    "  Exclude: '!test/', '!*.spec.ts'\n",
+    "\n",
+    "! Bare words without extensions are NOT constraints. 'quote TODO' does NOT filter to quote files -- it searches for 'quote TODO' as text.\n",
+    "  + 'schema.rs TODO'   -> searches for 'TODO' in files schema.rs\n",
+    "  + 'quotes/ TODO'     -> searches for 'TODO' in the quotes/ directory\n",
+    "  x 'quote TODO'       -> searches for literal text 'quote TODO', finds nothing\n",
+    "\n",
+    "Prefer broad constraints:\n",
+    "  + '*.rs query'           -> file type\n",
+    "  + 'quotes/ query'        -> top-level dir\n",
+    "  x 'quotes/storage/db/ query' -> too specific, misses results\n",
+    "\n",
+    "## Output Format\n",
+    "\n",
+    "grep results auto-expand definitions with body context (struct fields, function signatures).\n",
+    "This often provides enough information WITHOUT a follow-up Read call.\n",
+    "Lines marked with | are definition body context. [def] marks definition files.\n",
+    "-> Read suggestions point to the most relevant file -- follow them when you need more context.\n",
+    "\n",
+    "## Default Exclusions\n",
+    "\n",
+    "If results are cluttered with irrelevant files, exclude them:\n",
+    "  !tests/ - exclude tests directory\n",
+    "  !*.spec.ts - exclude test files\n",
+    "  !generated/ - exclude generated code",
+);
+
+/// FFF MCP Server — high-performance file finder for AI code assistants.
+#[derive(Parser)]
+#[command(name = "fff-mcp", version = concat!(env!("CARGO_PKG_VERSION"), " (", env!("FFF_GIT_HASH"), ")"))]
+struct Args {
+    /// Base directory to index. Defaults to the current working directory.
+    #[arg(value_name = "PATH")]
+    base_path: Option<String>,
+
+    /// Path to the frecency database.
+    #[arg(long = "frecency-db")]
+    frecency_db_path: Option<String>,
+
+    /// Path to the query history database.
+    #[arg(long = "history-db")]
+    #[allow(dead_code)]
+    history_db_path: Option<String>,
+
+    /// Path to the log file.
+    #[arg(long = "log-file")]
+    log_file: Option<String>,
+
+    /// Log level (e.g. trace, debug, info, warn, error).
+    #[arg(long = "log-level")]
+    log_level: Option<String>,
+
+    /// Disable automatic update checks on startup.
+    #[arg(long = "no-update-check")]
+    no_update_check: bool,
+}
+
+/// Resolve default paths for frecency db, history db, and log file.
+/// Shares Neovim's standard data locations when they exist so the MCP
+/// server and fff.nvim plugin use the same databases.
+fn resolve_defaults(args: &mut Args) {
+    let home = dirs_home();
+    let is_windows = cfg!(target_os = "windows");
+
+    let nvim_cache_dir = if is_windows {
+        format!("{}\\AppData\\Local\\nvim-data", home)
+    } else {
+        format!("{}/.cache/nvim", home)
+    };
+    let nvim_data_dir = if is_windows {
+        format!("{}\\AppData\\Local\\nvim-data", home)
+    } else {
+        format!("{}/.local/share/nvim", home)
+    };
+
+    let use_nvim_paths = std::path::Path::new(&nvim_cache_dir).exists()
+        || std::path::Path::new(&nvim_data_dir).exists();
+
+    if args.frecency_db_path.is_none() {
+        args.frecency_db_path = Some(if use_nvim_paths {
+            format!("{}/fff_nvim", nvim_cache_dir)
+        } else {
+            format!("{}/.fff/frecency.mdb", home)
+        });
+    }
+    if args.history_db_path.is_none() {
+        args.history_db_path = Some(if use_nvim_paths {
+            format!("{}/fff_queries", nvim_data_dir)
+        } else {
+            format!("{}/.fff/history.mdb", home)
+        });
+    }
+
+    // Ensure parent directories exist for database paths
+    for path in [&args.frecency_db_path, &args.history_db_path]
+        .into_iter()
+        .flatten()
+    {
+        if let Some(parent) = std::path::Path::new(path).parent() {
+            let _ = std::fs::create_dir_all(parent);
+        }
+    }
+
+    if args.log_file.is_none() {
+        args.log_file = Some(if is_windows {
+            format!("{}\\AppData\\Local\\fff_mcp.log", home)
+        } else {
+            format!("{}/.cache/fff_mcp.log", home)
+        });
+    }
+}
+
+fn dirs_home() -> String {
+    std::env::var("HOME")
+        .or_else(|_| std::env::var("USERPROFILE"))
+        .unwrap_or_else(|_| "/tmp".to_string())
+}
+
+#[tokio::main]
+async fn main() -> Result<(), Box<dyn std::error::Error>> {
+    let mut args = Args::parse();
+    resolve_defaults(&mut args);
+
+    let log_file = args.log_file.as_deref().unwrap_or("");
+    if let Err(e) = fff::log::init_tracing(log_file, args.log_level.as_deref()) {
+        eprintln!("Warning: Failed to init tracing: {}", e);
+    }
+
+    let base_path = args.base_path.unwrap_or_else(|| {
+        std::env::current_dir()
+            .unwrap_or_default()
+            .to_string_lossy()
+            .to_string()
+    });
+
+    if Repository::discover(&base_path).is_err() {
+        tracing::error!("MCP server must be run within a Git repository");
+        return Err(format!("Not a Git repository: {}", base_path).into());
+    }
+
+    let frecency_db_path = args.frecency_db_path.unwrap_or_default();
+
+    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
+    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));
+    match FrecencyTracker::new(&frecency_db_path, false) {
+        Ok(tracker) => {
+            if let Ok(mut guard) = shared_frecency.write() {
+                *guard = Some(tracker);
+            }
+
+            let _ =
+                FrecencyTracker::spawn_gc(Arc::clone(&shared_frecency), frecency_db_path, false);
+        }
+        Err(e) => {
+            eprintln!("Warning: Failed to init frecency db: {}", e);
+        }
+    }
+
+    // Initialize file picker (spawns background scan + watcher)
+    FilePicker::new_with_shared_state(
+        base_path,
+        true, // warmup_mmap_cache
+        FFFMode::Ai,
+        Arc::clone(&shared_picker),
+        Arc::clone(&shared_frecency),
+    )
+    .map_err(|e| format!("Failed to init file picker: {}", e))?;
+
+    if !args.no_update_check {
+        update_check::spawn_update_check();
+    }
+
+    // Create and start the MCP server
+    let server = FffServer::new(shared_picker.clone(), shared_frecency.clone());
+
+    // Wait for initial scan in background — don't block server startup
+    let picker_clone_for_scan = Arc::clone(&shared_picker);
+    tokio::task::spawn_blocking(move || {
+        let start = std::time::Instant::now();
+        loop {
+            let is_scanning = picker_clone_for_scan
+                .read()
+                .ok()
+                .and_then(|g| g.as_ref().map(|p| p.is_scan_active()))
+                .unwrap_or(true);
+
+            if !is_scanning {
+                tracing::info!("Initial scan completed in {:?}", start.elapsed());
+                break;
+            }
+            std::thread::sleep(std::time::Duration::from_millis(50));
+        }
+    });
+
+    let service = server
+        .serve(stdio())
+        .await
+        .map_err(|e| format!("Failed to start MCP server: {}", e))?;
+
+    let picker_for_shutdown = shared_picker.clone();
+    tokio::spawn(async move {
+        tokio::signal::ctrl_c().await.ok();
+        if let Ok(mut guard) = picker_for_shutdown.write()
+            && let Some(ref mut picker) = *guard
+        {
+            picker.stop_background_monitor();
+        }
+        std::process::exit(0);
+    });
+
+    service.waiting().await?;
+
+    if let Ok(mut guard) = shared_picker.write()
+        && let Some(ref mut picker) = *guard
+    {
+        picker.stop_background_monitor();
+    }
+
+    Ok(())
+}
@@ -0,0 +1,596 @@
+//! Output formatting for MCP grep/search results.
+//!
+//! Port of `packages/fff-mcp/src/output.ts` — token-efficient formatting
+//! with definition auto-expansion, frecency/git annotations, and Read suggestions.
+
+use fff::GrepMatch;
+use fff::git::format_git_status_opt;
+use fff::grep::is_import_line;
+use fff::types::FileItem;
+
+use crate::cursor::CursorStore;
+
+/// Frecency score → single-token word. `None` for low-scoring files.
+fn frecency_word(score: i64) -> Option<&'static str> {
+    if score >= 100 {
+        Some("hot")
+    } else if score >= 50 {
+        Some("warm")
+    } else if score >= 10 {
+        Some("frequent")
+    } else {
+        None
+    }
+}
+
+/// Build " - hot git:modified" style suffix. Empty when nothing to report.
+pub fn file_suffix(git_status: Option<git2::Status>, frecency_score: i64) -> String {
+    match (
+        frecency_word(frecency_score),
+        format_git_status_opt(git_status),
+    ) {
+        (Some(f), Some(g)) => format!(" - {f} git:{g}"),
+        (Some(f), None) => format!(" - {f}"),
+        (None, Some(g)) => format!(" git:{g}"),
+        (None, None) => String::new(),
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum OutputMode {
+    Content,
+    FilesWithMatches,
+    Count,
+    Usage,
+}
+
+impl OutputMode {
+    pub fn new(s: Option<&str>) -> Self {
+        match s {
+            Some("files_with_matches") => Self::FilesWithMatches,
+            Some("count") => Self::Count,
+            Some("usage") => Self::Usage,
+            _ => Self::Content,
+        }
+    }
+}
+
+const LARGE_FILE_BYTES: u64 = 20_000;
+
+/// Tag for large files — nudges model to use offset/limit when reading.
+fn size_tag(bytes: u64) -> String {
+    if bytes < LARGE_FILE_BYTES {
+        String::new()
+    } else {
+        let kb = (bytes + 512) / 1024; // round
+        format!(" ({}KB - use offset to read relevant section)", kb)
+    }
+}
+
+const MAX_PREVIEW: usize = 120;
+const MAX_LINE_LEN: usize = 180;
+/// Max context lines to show when auto-expanding the first definition
+const MAX_DEF_EXPAND_FIRST: usize = 8;
+/// Max context lines for subsequent definitions
+const MAX_DEF_EXPAND: usize = 5;
+/// Max context lines for non-definition first match in small result sets
+const MAX_FIRST_MATCH_EXPAND: usize = 8;
+
+fn trauncate_line_for_ai(
+    line: &str,
+    match_ranges: Option<&[(u32, u32)]>,
+    max_len: usize,
+) -> String {
+    // Strip leading/trailing whitespace to save tokens — the LLM has file:line for location.
+    let trimmed = line.trim();
+    if trimmed.is_empty() {
+        return String::new();
+    }
+
+    let strip_offset = line.len() - line.trim_start().len();
+
+    if trimmed.len() <= max_len {
+        return trimmed.to_string();
+    }
+
+    // Adjust match ranges for the stripped leading whitespace
+    let adjusted: Vec<(u32, u32)>;
+    let ranges = match match_ranges {
+        Some(r) if strip_offset > 0 => {
+            let off = strip_offset as u32;
+            adjusted = r
+                .iter()
+                .map(|&(s, e)| (s.saturating_sub(off), e.saturating_sub(off)))
+                .collect();
+            Some(adjusted.as_slice())
+        }
+        other => other,
+    };
+
+    // Use first match range to center the window
+    if let Some(ranges) = ranges
+        && let Some(&(match_start, match_end)) = ranges.first()
+    {
+        let match_start = match_start as usize;
+        let match_end = match_end as usize;
+        let match_len = match_end.saturating_sub(match_start);
+
+        let budget = max_len.saturating_sub(match_len);
+        let before = budget / 3;
+        let after = budget - before;
+
+        let win_start = match_start.saturating_sub(before);
+        let win_end = (match_end + after).min(trimmed.len());
+
+        // Clamp to char boundaries
+        let win_start = floor_char_boundary(trimmed, win_start);
+        let win_end = ceil_char_boundary(trimmed, win_end);
+
+        let mut result = trimmed[win_start..win_end].to_string();
+        if win_start > 0 {
+            result.insert(0, '…');
+        }
+        if win_end < trimmed.len() {
+            result.push('…');
+        }
+        return result;
+    }
+
+    // No match ranges — truncate from start
+    let end = ceil_char_boundary(trimmed, max_len);
+    format!("{}…", &trimmed[..end])
+}
+
+/// Floor to a valid char boundary
+fn floor_char_boundary(s: &str, index: usize) -> usize {
+    if index >= s.len() {
+        return s.len();
+    }
+    let mut i = index;
+    while i > 0 && !s.is_char_boundary(i) {
+        i -= 1;
+    }
+    i
+}
+
+/// Ceil to a valid char boundary
+fn ceil_char_boundary(s: &str, index: usize) -> usize {
+    if index >= s.len() {
+        return s.len();
+    }
+    let mut i = index;
+    while i < s.len() && !s.is_char_boundary(i) {
+        i += 1;
+    }
+    i
+}
+
+/// Collected file metadata for the first match per file.
+struct FileMeta<'a> {
+    file: &'a FileItem,
+    line_number: u64,
+    line_content: String,
+    is_definition: bool,
+    match_ranges: Vec<(u32, u32)>,
+    context_after: Vec<String>,
+}
+
+/// Parameters for [`format_grep_results`].
+///
+/// Groups the read-only inputs so callers don't juggle 10 positional args.
+pub struct GrepFormatter<'a> {
+    pub matches: &'a [GrepMatch],
+    pub files: &'a [&'a FileItem],
+    pub total_matched: usize,
+    pub next_file_offset: usize,
+    pub regex_fallback_error: Option<&'a str>,
+    pub output_mode: OutputMode,
+    pub max_results: usize,
+    pub show_context: bool,
+    pub auto_expand_defs: bool,
+}
+
+impl GrepFormatter<'_> {
+    pub fn format(&self, cursor_store: &mut CursorStore) -> String {
+        let GrepFormatter {
+            matches,
+            files,
+            total_matched,
+            next_file_offset,
+            regex_fallback_error,
+            output_mode,
+            max_results,
+            show_context,
+            auto_expand_defs,
+        } = *self;
+
+        let items = if matches.len() > max_results {
+            &matches[..max_results]
+        } else {
+            matches
+        };
+
+        if output_mode == OutputMode::FilesWithMatches {
+            return format_files_with_matches(
+                items,
+                files,
+                next_file_offset,
+                auto_expand_defs,
+                cursor_store,
+            );
+        }
+
+        if output_mode == OutputMode::Count {
+            return format_count(items, files, next_file_offset, cursor_store);
+        }
+
+        // output_mode == usage
+        let mut lines: Vec<String> = Vec::new();
+        let unique_files = {
+            let mut seen = std::collections::HashSet::new();
+            for m in items {
+                seen.insert(m.file_index);
+            }
+            seen.len()
+        };
+
+        let max_output_chars: usize = if output_mode == OutputMode::Usage || unique_files <= 3 {
+            5000
+        } else if unique_files <= 8 {
+            3500
+        } else {
+            2500
+        };
+
+        if let Some(err) = regex_fallback_error {
+            lines.push(format!("! regex failed: {}, using literal match", err));
+        }
+
+        // File overview: collect first match per file
+        let file_preview = collect_file_preview(items, files);
+        let mut content_def_file = "";
+        let mut content_first_file = "";
+        for fm in &file_preview {
+            if content_first_file.is_empty() {
+                content_first_file = &fm.file.relative_path;
+            }
+            if content_def_file.is_empty() && fm.is_definition {
+                content_def_file = &fm.file.relative_path;
+            }
+        }
+
+        let content_suggest = if !content_def_file.is_empty() {
+            content_def_file
+        } else {
+            content_first_file
+        };
+        if !content_suggest.is_empty() {
+            let file_count = file_preview.len();
+            if file_count == 1 {
+                lines.push(format!("→ Read {} (only match)", content_suggest));
+            } else if !content_def_file.is_empty() {
+                lines.push(format!("→ Read {} [def]", content_suggest));
+            } else if file_count <= 3 {
+                lines.push(format!("→ Read {} (best match)", content_suggest));
+            }
+        }
+
+        if total_matched > items.len() {
+            lines.push(format!("{}/{} matches shown", items.len(), total_matched));
+        }
+
+        // Track which files already had a definition expanded
+        let mut def_expanded_files = std::collections::HashSet::new();
+
+        // Detailed content (subject to budget)
+        let mut char_count = 0usize;
+        let mut shown_count = 0usize;
+        let mut current_file = "";
+
+        // Reorder: definitions first, then usages, then imports (when auto-expanding)
+        let sorted_items: Vec<usize> = if auto_expand_defs {
+            let mut indices: Vec<usize> = (0..items.len()).collect();
+            indices.sort_unstable_by_key(|&i| {
+                if items[i].is_definition {
+                    0
+                } else if is_import_line(&items[i].line_content) {
+                    2
+                } else {
+                    1
+                }
+            });
+
+            indices
+        } else {
+            (0..items.len()).collect()
+        };
+
+        for &idx in &sorted_items {
+            let m = &items[idx];
+            let file = files[m.file_index];
+            let mut match_lines: Vec<String> = Vec::new();
+
+            if file.relative_path.as_str() != current_file {
+                current_file = &file.relative_path;
+                match_lines.push(current_file.to_string());
+            }
+
+            // Skip import-only lines when we already have definitions
+            if auto_expand_defs && is_import_line(&m.line_content) && !def_expanded_files.is_empty()
+            {
+                continue;
+            }
+
+            // Context before (only when explicitly requested)
+            if show_context && !m.context_before.is_empty() {
+                let start_line = m.line_number.saturating_sub(m.context_before.len() as u64);
+                for (i, ctx) in m.context_before.iter().enumerate() {
+                    match_lines.push(format!(
+                        " {}-{}",
+                        start_line + i as u64,
+                        trauncate_line_for_ai(ctx, None, MAX_LINE_LEN)
+                    ));
+                }
+            }
+
+            // Match line
+            match_lines.push(format!(
+                " {}: {}",
+                m.line_number,
+                trauncate_line_for_ai(
+                    &m.line_content,
+                    Some(m.match_byte_offsets.as_ref()),
+                    MAX_LINE_LEN
+                )
+            ));
+
+            // Context after (only when explicitly requested via context parameter)
+            if show_context && !m.context_after.is_empty() {
+                let start_line = m.line_number + 1;
+                for (i, ctx) in m.context_after.iter().enumerate() {
+                    match_lines.push(format!(
+                        " {}-{}",
+                        start_line + i as u64,
+                        trauncate_line_for_ai(ctx, None, MAX_LINE_LEN)
+                    ));
+                }
+                match_lines.push("--".to_string());
+            }
+
+            // Auto-expand definitions with body context
+            if auto_expand_defs
+                && !show_context
+                && m.is_definition
+                && !m.context_after.is_empty()
+                && !def_expanded_files.contains(file.relative_path.as_str())
+            {
+                let expand_limit = if def_expanded_files.is_empty() {
+                    MAX_DEF_EXPAND_FIRST
+                } else {
+                    MAX_DEF_EXPAND
+                };
+                def_expanded_files.insert(file.relative_path.as_str());
+                let start_line = m.line_number + 1;
+                for (i, ctx) in m.context_after.iter().take(expand_limit).enumerate() {
+                    if ctx.trim().is_empty() {
+                        break;
+                    }
+                    match_lines.push(format!(
+                        "  {}| {}",
+                        start_line + i as u64,
+                        trauncate_line_for_ai(ctx, None, MAX_LINE_LEN)
+                    ));
+                }
+            }
+
+            let chunk = match_lines.join("\n");
+            if char_count + chunk.len() > max_output_chars && shown_count > 0 {
+                break;
+            }
+
+            char_count += chunk.len();
+            lines.push(chunk);
+            shown_count += 1;
+        }
+
+        if next_file_offset > 0 {
+            let cursor_id = cursor_store.store(next_file_offset);
+            lines.push(format!("\ncursor: {}", cursor_id));
+        }
+
+        lines.join("\n")
+    }
+}
+
+fn format_files_with_matches(
+    items: &[GrepMatch],
+    files: &[&FileItem],
+    next_file_offset: usize,
+    auto_expand_defs: bool,
+    cursor_store: &mut CursorStore,
+) -> String {
+    let file_map = collect_file_preview(items, files);
+
+    let mut lines: Vec<String> = Vec::new();
+    let file_count = file_map.len();
+
+    // Find best Read target
+    let mut first_def_file = "";
+    let mut first_file = "";
+    for fm in &file_map {
+        if first_file.is_empty() {
+            first_file = &fm.file.relative_path;
+        }
+        if first_def_file.is_empty() && fm.is_definition {
+            first_def_file = &fm.file.relative_path;
+        }
+    }
+    let suggest_path = if !first_def_file.is_empty() {
+        first_def_file
+    } else {
+        first_file
+    };
+
+    if !suggest_path.is_empty() {
+        if file_count == 1 {
+            lines.push(format!(
+                "→ Read {} (only match — no need to search further)",
+                suggest_path
+            ));
+        } else if !first_def_file.is_empty() && file_count <= 5 {
+            lines.push(format!("→ Read {} (definition found)", suggest_path));
+        } else if !first_def_file.is_empty() {
+            lines.push(format!("→ Read {} (definition)", suggest_path));
+        } else if file_count <= 3 {
+            lines.push(format!("→ Read {} (best match)", suggest_path));
+        } else {
+            lines.push(format!("→ Read {}", suggest_path));
+        }
+    }
+
+    let is_small_set = file_count <= 5;
+    let mut def_expanded_count = 0usize;
+
+    for (file_idx, fm) in file_map.iter().enumerate() {
+        let is_def = fm.is_definition;
+        let def_tag = if is_def { " [def]" } else { "" };
+        lines.push(format!(
+            "{}{}{}",
+            fm.file.relative_path,
+            def_tag,
+            size_tag(fm.file.size)
+        ));
+
+        // Show preview
+        if !fm.line_content.is_empty() && (is_def || file_idx == 0 || is_small_set) {
+            let ranges_ref: Option<&[(u32, u32)]> = if fm.match_ranges.is_empty() {
+                None
+            } else {
+                Some(&fm.match_ranges)
+            };
+            lines.push(format!(
+                "  {}: {}",
+                fm.line_number,
+                trauncate_line_for_ai(&fm.line_content, ranges_ref, MAX_PREVIEW)
+            ));
+
+            // Auto-expand body context
+            if auto_expand_defs && !fm.context_after.is_empty() {
+                let expand_limit = if is_def {
+                    let limit = if def_expanded_count == 0 {
+                        MAX_DEF_EXPAND_FIRST
+                    } else {
+                        MAX_DEF_EXPAND
+                    };
+                    def_expanded_count += 1;
+                    limit
+                } else if is_small_set && file_idx == 0 {
+                    MAX_FIRST_MATCH_EXPAND
+                } else if is_small_set {
+                    MAX_DEF_EXPAND
+                } else {
+                    0
+                };
+
+                if expand_limit > 0 {
+                    let start_line = fm.line_number + 1;
+                    for (i, ctx) in fm.context_after.iter().take(expand_limit).enumerate() {
+                        if ctx.trim().is_empty() {
+                            break;
+                        }
+                        lines.push(format!(
+                            "  {}| {}",
+                            start_line + i as u64,
+                            trauncate_line_for_ai(ctx, None, MAX_PREVIEW)
+                        ));
+                    }
+                }
+            }
+        }
+    }
+
+    if next_file_offset > 0 {
+        let cursor_id = cursor_store.store(next_file_offset);
+        lines.push(format!("\ncursor: {}", cursor_id));
+    }
+
+    lines.join("\n")
+}
+
+fn format_count(
+    items: &[GrepMatch],
+    files: &[&FileItem],
+    next_file_offset: usize,
+    cursor_store: &mut CursorStore,
+) -> String {
+    let mut counts: std::collections::HashMap<&str, usize> = std::collections::HashMap::new();
+    let mut order: Vec<&str> = Vec::new();
+    for m in items {
+        let path = files[m.file_index].relative_path.as_str();
+        let count = counts.entry(path).or_insert_with(|| {
+            order.push(path);
+            0
+        });
+        *count += 1;
+    }
+
+    let mut lines: Vec<String> = Vec::new();
+    for path in &order {
+        lines.push(format!("{}: {}", path, counts[*path]));
+    }
+    if next_file_offset > 0 {
+        let cursor_id = cursor_store.store(next_file_offset);
+        lines.push(format!("\ncursor: {}", cursor_id));
+    }
+    lines.join("\n")
+}
+
+fn collect_file_preview<'a>(items: &[GrepMatch], files: &[&'a FileItem]) -> Vec<FileMeta<'a>> {
+    let mut file_preview: Vec<FileMeta<'a>> = Vec::new();
+    let mut seen = std::collections::HashSet::new();
+    for m in items {
+        let file = files[m.file_index];
+        if seen.insert(&file.relative_path) {
+            file_preview.push(FileMeta {
+                file,
+                line_number: m.line_number,
+                line_content: m.line_content.clone(),
+                is_definition: m.is_definition,
+                match_ranges: m.match_byte_offsets.iter().copied().collect(),
+                context_after: m.context_after.clone(),
+            });
+        }
+    }
+    file_preview
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn trunc_strips_whitespace() {
+        assert_eq!(trauncate_line_for_ai("    foo()", None, 180), "foo()");
+        assert_eq!(trauncate_line_for_ai("  bar  ", None, 180), "bar");
+        assert_eq!(trauncate_line_for_ai("   ", None, 180), "");
+    }
+
+    #[test]
+    fn trunc_adjusts_match_ranges_after_strip() {
+        // "    hello" — match on "hello" at bytes 4..9
+        let line = "    hello";
+        let ranges = [(4, 9)];
+        let result = trauncate_line_for_ai(line, Some(&ranges), 180);
+        // After stripping 4 leading spaces, the trimmed line is "hello"
+        assert_eq!(result, "hello");
+    }
+
+    #[test]
+    fn trunc_long_line_centered() {
+        let line = format!("{}match_here{}", " ".repeat(8), "x".repeat(200));
+        let ranges = [(8u32, 18u32)];
+        let result = trauncate_line_for_ai(&line, Some(&ranges), 50);
+        assert!(result.contains("match_here"));
+        assert!(result.len() <= 55); // budget + ellipsis chars
+    }
+}
@@ -0,0 +1,682 @@
+//! FFF MCP server — tool definitions and handlers.
+//!
+//! Uses the `rmcp` crate's `#[tool_router]` / `#[tool_handler]` macros
+//! for declarative tool registration. Each tool method directly calls
+//! `fff-core` APIs (no C FFI overhead).
+
+use std::borrow::Cow;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, Mutex};
+
+use fff::file_picker::FilePicker;
+use fff::grep::{self, GrepMode, GrepSearchOptions, has_regex_metacharacters};
+use fff::types::{FileItem, PaginationArgs};
+use fff::{FuzzySearchOptions, QueryParser, SharedFrecency, SharedPicker};
+use fff_query_parser::AiGrepConfig;
+use rmcp::handler::server::router::tool::ToolRouter;
+use rmcp::handler::server::wrapper::Parameters;
+use rmcp::model::*;
+use rmcp::{ServerHandler, schemars, tool, tool_handler, tool_router};
+
+use crate::cursor::CursorStore;
+use crate::output::{GrepFormatter, OutputMode, file_suffix};
+
+/// Strip common delimiters and lowercase for fuzzy fallback queries.
+fn cleanup_fuzzy_query(s: &str) -> String {
+    let mut out = String::with_capacity(s.len());
+    for c in s.chars() {
+        if !matches!(c, ':' | '-' | '_') {
+            out.extend(c.to_lowercase());
+        }
+    }
+    out
+}
+
+/// Compute grep search options from output mode and context settings.
+fn make_grep_options(
+    output_mode: OutputMode,
+    mode: GrepMode,
+    file_offset: usize,
+    context: Option<usize>,
+) -> (GrepSearchOptions, bool) {
+    let is_usage = output_mode == OutputMode::Usage;
+    let matches_per_file = match output_mode {
+        OutputMode::FilesWithMatches => 1,
+        _ if is_usage => 8,
+        _ => 10,
+    };
+    let ctx_lines = if is_usage {
+        context.unwrap_or(1)
+    } else {
+        context.unwrap_or(0)
+    };
+    let auto_expand = !is_usage && ctx_lines == 0;
+    let after_ctx = if auto_expand { 8 } else { ctx_lines };
+
+    (
+        GrepSearchOptions {
+            max_file_size: 10 * 1024 * 1024,
+            max_matches_per_file: matches_per_file,
+            smart_case: true,
+            file_offset,
+            page_limit: 50,
+            mode,
+            time_budget_ms: 0,
+            before_context: ctx_lines,
+            after_context: after_ctx,
+            classify_definitions: true,
+        },
+        auto_expand,
+    )
+}
+
+#[derive(Debug, serde::Deserialize, schemars::JsonSchema)]
+pub struct FindFilesParams {
+    /// Fuzzy search query. Supports path prefixes and glob constraints.
+    pub query: String,
+    /// Max results (default 20).
+    #[serde(rename = "maxResults")]
+    pub max_results: Option<usize>,
+    /// Cursor from previous result. Only use if previous results weren't sufficient.
+    pub cursor: Option<String>,
+}
+
+#[derive(Debug, serde::Deserialize, schemars::JsonSchema)]
+pub struct GrepParams {
+    /// Search text or regex query with optional constraint prefixes.
+    /// Matches within single lines only — use ONE specific term, not multiple words.
+    pub query: String,
+    /// Max matching lines (default 20).
+    #[serde(rename = "maxResults")]
+    pub max_results: Option<usize>,
+    /// Cursor from previous result. Only use if previous results weren't sufficient.
+    pub cursor: Option<String>,
+    /// Output format (default 'content').
+    pub output_mode: Option<String>,
+}
+
+fn deserialize_patterns<'de, D>(deserializer: D) -> Result<Vec<String>, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    use serde::de;
+
+    struct PatternsVisitor;
+
+    impl<'de> de::Visitor<'de> for PatternsVisitor {
+        type Value = Vec<String>;
+
+        fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
+            formatter.write_str("a string, an array of strings, or a stringified JSON array")
+        }
+
+        fn visit_str<E: de::Error>(self, v: &str) -> Result<Self::Value, E> {
+            // Try to parse as JSON array first
+            if v.starts_with('[')
+                && let Ok(parsed) = serde_json::from_str::<Vec<String>>(v)
+            {
+                return Ok(parsed);
+            }
+            Ok(vec![v.to_string()])
+        }
+
+        fn visit_string<E: de::Error>(self, v: String) -> Result<Self::Value, E> {
+            if v.starts_with('[')
+                && let Ok(parsed) = serde_json::from_str::<Vec<String>>(&v)
+            {
+                return Ok(parsed);
+            }
+            Ok(vec![v])
+        }
+
+        fn visit_seq<A: de::SeqAccess<'de>>(self, mut seq: A) -> Result<Self::Value, A::Error> {
+            let mut values = Vec::new();
+            while let Some(value) = seq.next_element::<String>()? {
+                values.push(value);
+            }
+            Ok(values)
+        }
+    }
+
+    deserializer.deserialize_any(PatternsVisitor)
+}
+
+#[derive(Debug, serde::Deserialize, schemars::JsonSchema)]
+pub struct MultiGrepParams {
+    /// Patterns to match (OR logic). Include all naming conventions: snake_case, PascalCase, camelCase.
+    #[serde(deserialize_with = "deserialize_patterns")]
+    pub patterns: Vec<String>,
+    /// File constraints (e.g. '*.{ts,tsx} !test/'). ALWAYS provide when possible.
+    pub constraints: Option<String>,
+    /// Max matching lines (default 20).
+    #[serde(rename = "maxResults")]
+    pub max_results: Option<usize>,
+    /// Cursor from previous result.
+    pub cursor: Option<String>,
+    /// Output format (default 'content').
+    pub output_mode: Option<String>,
+    /// Context lines before/after each match.
+    pub context: Option<usize>,
+}
+
+#[derive(Clone)]
+pub struct FffServer {
+    picker: SharedPicker,
+    #[allow(dead_code)]
+    frecency: SharedFrecency,
+    cursor_store: Arc<Mutex<CursorStore>>,
+    update_notice_sent: Arc<AtomicBool>,
+    tool_router: ToolRouter<Self>,
+}
+
+impl FffServer {
+    pub fn new(picker: SharedPicker, frecency: SharedFrecency) -> Self {
+        Self {
+            picker,
+            frecency,
+            cursor_store: Arc::new(Mutex::new(CursorStore::new())),
+            update_notice_sent: Arc::new(AtomicBool::new(false)),
+            tool_router: Self::tool_router(),
+        }
+    }
+
+    /// Wait for the initial file scan to complete.
+    #[allow(dead_code)]
+    pub fn wait_for_scan(&self) {
+        loop {
+            let guard = self.picker.read().ok();
+            let is_scanning = guard
+                .as_ref()
+                .and_then(|g| g.as_ref())
+                .map(|p| p.is_scan_active())
+                .unwrap_or(true);
+
+            if !is_scanning {
+                break;
+            }
+            std::thread::sleep(std::time::Duration::from_millis(50));
+        }
+    }
+
+    /// Lock the cursor store, returning an MCP error on poisoned mutex.
+    fn lock_cursors(&self) -> Result<std::sync::MutexGuard<'_, CursorStore>, ErrorData> {
+        self.cursor_store.lock().map_err(|e| {
+            ErrorData::internal_error(format!("Failed to acquire cursor store lock: {e}"), None)
+        })
+    }
+
+    /// If an update notice is available and hasn't been sent yet, append it
+    /// to the tool result. Called once per server lifetime (first tool call).
+    fn maybe_append_update_notice(&self, result: &mut CallToolResult) {
+        if self.update_notice_sent.swap(true, Ordering::Relaxed) {
+            return;
+        }
+        let notice = crate::update_check::get_update_notice();
+        if notice.is_empty() {
+            // Reset so the next call can try again (check may still be in flight)
+            self.update_notice_sent.store(false, Ordering::Relaxed);
+            return;
+        }
+        result.content.push(Content::text(notice));
+    }
+
+    /// Perform grep with auto-retry logic.
+    ///
+    /// Acquires the picker read-lock once and holds it for the entire
+    /// operation, so `GrepResult` references are used directly — no cloning.
+    /// Always uses AI query parsing since this is an MCP server for AI agents.
+    fn perform_grep(
+        &self,
+        query: &str,
+        mode: GrepMode,
+        max_results: usize,
+        cursor_id: Option<&str>,
+        output_mode: OutputMode,
+        context: Option<usize>,
+    ) -> Result<CallToolResult, ErrorData> {
+        let file_offset = cursor_id
+            .and_then(|id| self.cursor_store.lock().ok()?.get(id))
+            .unwrap_or(0);
+
+        let (options, auto_expand) = make_grep_options(output_mode, mode, file_offset, context);
+        let ctx_lines = options.before_context;
+
+        // Acquire picker lock once for the entire operation.
+        let guard = self.picker.read().map_err(|e| {
+            ErrorData::internal_error(format!("Failed to acquire picker lock: {e}"), None)
+        })?;
+        let picker = guard
+            .as_ref()
+            .ok_or_else(|| ErrorData::internal_error("File picker not initialized", None))?;
+
+        let files = picker.get_files();
+
+        let parser = QueryParser::new(AiGrepConfig);
+        let parsed = parser.parse(query);
+        let result = grep::grep_search(files, &parsed, &options);
+
+        if result.matches.is_empty() && file_offset == 0 {
+            // Auto-retry: try broadening multi-word queries by dropping first non-constraint word
+            let parts: Vec<&str> = query.split_whitespace().collect();
+            if parts.len() >= 2 {
+                let first_word = parts[0];
+                let is_valid_constraint = first_word.starts_with('!')
+                    || first_word.starts_with('*')
+                    || first_word.ends_with('/');
+
+                if !is_valid_constraint {
+                    let rest_query = parts[1..].join(" ");
+                    let rest_parsed = parser.parse(&rest_query);
+
+                    let rest_text = rest_parsed.grep_text();
+                    let retry_mode = if has_regex_metacharacters(&rest_text) {
+                        GrepMode::Regex
+                    } else {
+                        mode
+                    };
+
+                    let (retry_options, _) = make_grep_options(output_mode, retry_mode, 0, context);
+                    let retry_result = grep::grep_search(files, &rest_parsed, &retry_options);
+
+                    if !retry_result.matches.is_empty() && retry_result.matches.len() <= 10 {
+                        let mut cs = self.lock_cursors()?;
+                        let text = &GrepFormatter {
+                            matches: &retry_result.matches,
+                            files: &retry_result.files,
+                            total_matched: retry_result.matches.len(),
+                            next_file_offset: retry_result.next_file_offset,
+                            regex_fallback_error: retry_result.regex_fallback_error.as_deref(),
+                            output_mode,
+                            max_results,
+                            show_context: ctx_lines > 0,
+                            auto_expand_defs: auto_expand,
+                        }
+                        .format(&mut cs);
+                        return Ok(CallToolResult::success(vec![Content::text(format!(
+                            "0 matches for '{}'. Auto-broadened to '{}':\n{}",
+                            query, rest_query, text
+                        ))]));
+                    }
+                }
+            }
+
+            // Fuzzy fallback for typo tolerance
+            let fuzzy_query = cleanup_fuzzy_query(query);
+            let (fuzzy_options, _) = make_grep_options(output_mode, GrepMode::Fuzzy, 0, Some(0));
+            let fuzzy_parsed = parser.parse(&fuzzy_query);
+            let fuzzy_result = grep::grep_search(files, &fuzzy_parsed, &fuzzy_options);
+
+            if !fuzzy_result.matches.is_empty() {
+                let mut lines: Vec<String> = Vec::new();
+                lines.push(format!(
+                    "0 exact matches. {} approximate:",
+                    fuzzy_result.matches.len()
+                ));
+                let mut current_file = "";
+                for m in fuzzy_result.matches.iter().take(3) {
+                    let file = fuzzy_result.files[m.file_index];
+                    if file.relative_path.as_str() != current_file {
+                        current_file = &file.relative_path;
+                        lines.push(current_file.to_string());
+                    }
+                    lines.push(format!(" {}: {}", m.line_number, m.line_content));
+                }
+                return Ok(CallToolResult::success(vec![Content::text(
+                    lines.join("\n"),
+                )]));
+            }
+
+            // File path fallback: if query looks like a path, suggest the matching file
+            if query.contains('/') {
+                let file_parser = QueryParser::default();
+                let file_query = file_parser.parse(query);
+                let file_opts = FuzzySearchOptions {
+                    max_threads: 0,
+                    current_file: None,
+                    project_path: Some(picker.base_path()),
+                    combo_boost_score_multiplier: 100,
+                    min_combo_count: 3,
+                    pagination: PaginationArgs {
+                        offset: 0,
+                        limit: 1,
+                    },
+                };
+                let file_result = FilePicker::fuzzy_search(files, &file_query, None, file_opts);
+                if let (Some(top), Some(score)) =
+                    (file_result.items.first(), file_result.scores.first())
+                {
+                    // Only suggest when the match is strong enough.
+                    let query_len = query.len() as i32;
+                    if score.base_score > query_len * 10 {
+                        return Ok(CallToolResult::success(vec![Content::text(format!(
+                            "0 content matches. But there is a relevant file path: {}",
+                            top.relative_path
+                        ))]));
+                    }
+                }
+            }
+
+            return Ok(CallToolResult::success(vec![Content::text(
+                "0 matches.".to_string(),
+            )]));
+        }
+
+        if result.matches.is_empty() {
+            return Ok(CallToolResult::success(vec![Content::text(
+                "0 matches.".to_string(),
+            )]));
+        }
+
+        let mut cs = self.lock_cursors()?;
+        let text = &GrepFormatter {
+            matches: &result.matches,
+            files: &result.files,
+            total_matched: result.matches.len(),
+            next_file_offset: result.next_file_offset,
+            regex_fallback_error: result.regex_fallback_error.as_deref(),
+            output_mode,
+            max_results,
+            show_context: ctx_lines > 0,
+            auto_expand_defs: auto_expand,
+        }
+        .format(&mut cs);
+
+        Ok(CallToolResult::success(vec![Content::text(text)]))
+    }
+}
+
+#[tool_router]
+impl FffServer {
+    /// Fuzzy file search by name. Searches FILE NAMES, not file contents.
+    /// Use it when you need to find a file, not a definition.
+    /// Use grep instead for searching code content (definitions, usage patterns).
+    /// Supports fuzzy matching, path prefixes ('shc/'), and glob constraints.
+    /// IMPORTANT: Keep queries SHORT — prefer 1-2 terms max.
+    #[tool(
+        name = "find_files",
+        description = "Fuzzy file search by name. Searches FILE NAMES, not file contents. Use it when you need to find a file, not a definition. Use grep instead for searching code content (definitions, usage patterns). Supports fuzzy matching, path prefixes ('src/'), and glob constraints ('name **/src/*.{ts,tsx} !test/'). IMPORTANT: Keep queries SHORT — prefer 1-2 terms max. Multiple words are a waterfall (each narrows results), NOT OR. If unsure, start broad with 1 term and refine."
+    )]
+    fn find_files(
+        &self,
+        Parameters(params): Parameters<FindFilesParams>,
+    ) -> Result<CallToolResult, ErrorData> {
+        let max_results = params.max_results.unwrap_or(20);
+        let query = &params.query;
+
+        let page_offset = params
+            .cursor
+            .as_deref()
+            .and_then(|id| self.cursor_store.lock().ok()?.get(id))
+            .unwrap_or(0);
+
+        let guard = self.picker.read().map_err(|e| {
+            ErrorData::internal_error(format!("Failed to acquire picker lock: {e}"), None)
+        })?;
+        let picker = guard
+            .as_ref()
+            .ok_or_else(|| ErrorData::internal_error("File picker not initialized", None))?;
+
+        let files = picker.get_files();
+        let base_path = picker.base_path();
+        let make_opts = |offset: usize| FuzzySearchOptions {
+            max_threads: 0,
+            current_file: None,
+            project_path: Some(base_path),
+            combo_boost_score_multiplier: 100,
+            min_combo_count: 3,
+            pagination: PaginationArgs {
+                offset,
+                limit: max_results,
+            },
+        };
+
+        let parser = QueryParser::default();
+        let fff_query = parser.parse(query);
+        let result = FilePicker::fuzzy_search(files, &fff_query, None, make_opts(page_offset));
+        let total_files = result.total_files;
+
+        // Auto-retry with fewer terms if 3+ words return 0 results
+        let words: Vec<&str> = query.split_whitespace().collect();
+        let shorter = words.get(..2).map(|w| w.join(" "));
+
+        let (items, scores, total_matched) =
+            if result.items.is_empty() && words.len() >= 3 && page_offset == 0 {
+                if let Some(shorter) = &shorter {
+                    let shorter_query = parser.parse(shorter);
+                    let retry = FilePicker::fuzzy_search(
+                        files,
+                        &shorter_query,
+                        /*query_tracker=*/ None,
+                        make_opts(0),
+                    );
+
+                    (retry.items, retry.scores, retry.total_matched)
+                } else {
+                    (result.items, result.scores, result.total_matched)
+                }
+            } else {
+                (result.items, result.scores, result.total_matched)
+            };
+
+        if items.is_empty() {
+            return Ok(CallToolResult::success(vec![Content::text(format!(
+                "0 results ({} indexed)",
+                total_files
+            ))]));
+        }
+
+        let mut lines: Vec<String> = Vec::new();
+        let top_item = items[0];
+        let is_exact_match = scores[0].exact_match;
+
+        if page_offset == 0 {
+            if is_exact_match {
+                lines.push(format!("→ Read {} (exact match!)", top_item.relative_path));
+            } else if scores.len() < 2 || scores[0].total > scores[1].total.saturating_mul(2) {
+                lines.push(format!(
+                    "→ Read {} (best match — Read this file directly)",
+                    top_item.relative_path
+                ));
+            }
+        }
+
+        let next_offset = page_offset + items.len();
+        let has_more = next_offset < total_matched;
+
+        if has_more {
+            lines.push(format!("{}/{} matches", items.len(), total_matched));
+        }
+
+        for item in &items {
+            lines.push(format!(
+                "{}{}",
+                item.relative_path,
+                file_suffix(item.git_status, item.total_frecency_score)
+            ));
+        }
+
+        if has_more {
+            let mut cs = self.lock_cursors()?;
+            let cursor_id = cs.store(next_offset);
+            lines.push(format!("cursor: {}", cursor_id));
+        }
+
+        let mut result = CallToolResult::success(vec![Content::text(lines.join("\n"))]);
+        self.maybe_append_update_notice(&mut result);
+        Ok(result)
+    }
+
+    /// Search file contents for text patterns. This is the DEFAULT search tool.
+    /// Prefer plain text over regex. Filter files with constraints.
+    #[tool(
+        name = "grep",
+        description = "Search file contents. Search for bare identifiers (e.g. 'InProgressQuote', 'ActorAuth'), NOT code syntax or regex. Filter files with constraints (e.g. '*.rs query', 'src/ query'). Use filename, directory (ending with /) or glob expressions to prefilter. See server instructions for constraint syntax and core rules."
+    )]
+    fn grep(
+        &self,
+        Parameters(params): Parameters<GrepParams>,
+    ) -> Result<CallToolResult, ErrorData> {
+        let max_results = params.max_results.unwrap_or(20);
+        let output_mode = OutputMode::new(params.output_mode.as_deref());
+
+        let parsed = QueryParser::new(AiGrepConfig).parse(&params.query);
+        let grep_text = parsed.grep_text();
+
+        let mode = if has_regex_metacharacters(&grep_text) {
+            GrepMode::Regex
+        } else {
+            GrepMode::PlainText
+        };
+
+        let mut result = self.perform_grep(
+            &params.query,
+            mode,
+            max_results,
+            params.cursor.as_deref(),
+            output_mode,
+            None,
+        )?;
+        self.maybe_append_update_notice(&mut result);
+        Ok(result)
+    }
+
+    /// Search file contents for lines matching ANY of multiple patterns (OR logic).
+    /// Patterns are literal text — NEVER escape special characters.
+    #[tool(
+        name = "multi_grep",
+        description = "Search file contents for lines matching ANY of multiple patterns (OR logic). IMPORTANT: This returns files where ANY query matches, NOT all patterns. Patterns are literal text — NEVER escape special characters (no \\( \\) \\. etc). Faster than regex alternation for literal text. See server instructions for constraint syntax."
+    )]
+    fn multi_grep(
+        &self,
+        Parameters(params): Parameters<MultiGrepParams>,
+    ) -> Result<CallToolResult, ErrorData> {
+        let mut result = self.multi_grep_inner(params)?;
+        self.maybe_append_update_notice(&mut result);
+        Ok(result)
+    }
+}
+
+impl FffServer {
+    fn multi_grep_inner(&self, params: MultiGrepParams) -> Result<CallToolResult, ErrorData> {
+        let max_results = params.max_results.unwrap_or(20);
+        let output_mode = OutputMode::new(params.output_mode.as_deref());
+
+        let file_offset = params
+            .cursor
+            .as_deref()
+            .and_then(|id| self.cursor_store.lock().ok()?.get(id))
+            .unwrap_or(0);
+
+        let (options, auto_expand) = make_grep_options(
+            output_mode,
+            GrepMode::PlainText,
+            file_offset,
+            params.context,
+        );
+
+        let ctx_lines = options.before_context;
+        let constraint_query = params.constraints.as_deref().unwrap_or("");
+        let guard = self.picker.read().map_err(|e| {
+            ErrorData::internal_error(format!("Failed to acquire picker lock: {e}"), None)
+        })?;
+        let picker = guard
+            .as_ref()
+            .ok_or_else(|| ErrorData::internal_error("File picker not initialized", None))?;
+
+        let patterns_refs: Vec<&str> = params.patterns.iter().map(|s| s.as_str()).collect();
+
+        let parser = fff_query_parser::QueryParser::new(fff_query_parser::AiGrepConfig);
+        let parsed_constraints = parser.parse(constraint_query);
+        let constraints = parsed_constraints.constraints.as_slice();
+
+        let files = picker.get_files();
+        let result = grep::multi_grep_search(files, &patterns_refs, constraints, &options);
+        let file_refs: Vec<&FileItem> = result.files.to_vec();
+
+        if result.matches.is_empty() && file_offset == 0 {
+            // Fallback: try individual patterns with plain grep
+            let (fallback_options, _) =
+                make_grep_options(output_mode, GrepMode::PlainText, 0, params.context);
+
+            let fallback_options = GrepSearchOptions {
+                time_budget_ms: 3000,
+                before_context: 0,
+                ..fallback_options
+            };
+
+            for pat in &params.patterns {
+                let full_query: Cow<str> = if !constraint_query.is_empty() {
+                    Cow::Owned(format!("{} {}", constraint_query, pat))
+                } else {
+                    Cow::Borrowed(pat)
+                };
+
+                let parsed = parser.parse(&full_query);
+                let fb_result = grep::grep_search(files, &parsed, &fallback_options);
+
+                if !fb_result.matches.is_empty() {
+                    let fb_file_refs: Vec<&FileItem> = fb_result.files.to_vec();
+                    let mut cs = self.lock_cursors()?;
+                    let text = &GrepFormatter {
+                        matches: &fb_result.matches,
+                        files: &fb_file_refs,
+                        total_matched: fb_result.matches.len(),
+                        next_file_offset: fb_result.next_file_offset,
+                        regex_fallback_error: None,
+                        output_mode,
+                        max_results,
+                        show_context: false,
+                        auto_expand_defs: auto_expand,
+                    }
+                    .format(&mut cs);
+                    return Ok(CallToolResult::success(vec![Content::text(format!(
+                        "0 multi-pattern matches. Plain grep fallback for \"{}\":\n{}",
+                        pat, text
+                    ))]));
+                }
+            }
+
+            return Ok(CallToolResult::success(vec![Content::text(
+                "0 matches.".to_string(),
+            )]));
+        }
+
+        if result.matches.is_empty() {
+            return Ok(CallToolResult::success(vec![Content::text(
+                "0 matches.".to_string(),
+            )]));
+        }
+
+        let mut cs = self.lock_cursors()?;
+        let text = &GrepFormatter {
+            matches: &result.matches,
+            files: &file_refs,
+            total_matched: result.matches.len(),
+            next_file_offset: result.next_file_offset,
+            regex_fallback_error: None,
+            output_mode,
+            max_results,
+            show_context: ctx_lines > 0,
+            auto_expand_defs: auto_expand,
+        }
+        .format(&mut cs);
+
+        Ok(CallToolResult::success(vec![Content::text(text)]))
+    }
+}
+
+#[tool_handler]
+impl ServerHandler for FffServer {
+    fn get_info(&self) -> ServerInfo {
+        let notice = crate::update_check::get_update_notice();
+        let instructions = if notice.is_empty() {
+            crate::MCP_INSTRUCTIONS.to_string()
+        } else {
+            format!("{}{}", crate::MCP_INSTRUCTIONS, notice)
+        };
+
+        ServerInfo::new(ServerCapabilities::builder().enable_tools().build())
+            .with_server_info(Implementation::new("fff", env!("CARGO_PKG_VERSION")))
+            .with_instructions(instructions)
+    }
+}
@@ -0,0 +1,78 @@
+//! Background update checker — compares the embedded build hash against
+//! the latest GitHub release tag to surface upgrade notices in MCP instructions.
+
+use std::sync::OnceLock;
+
+const REPO: &str = "dmtrKovalenko/fff.nvim";
+const BUILD_HASH: &str = env!("FFF_GIT_HASH");
+
+/// Holds the result of the update check (empty string = up to date or check failed).
+static UPDATE_NOTICE: OnceLock<String> = OnceLock::new();
+
+/// Returns the update notice if the check has completed, empty string otherwise.
+pub fn get_update_notice() -> &'static str {
+    UPDATE_NOTICE.get().map(|s| s.as_str()).unwrap_or("")
+}
+
+/// Kick off the update check in a background thread so it never blocks the server.
+pub fn spawn_update_check() {
+    std::thread::spawn(|| {
+        let notice = check_latest_release();
+        let _ = UPDATE_NOTICE.set(notice);
+    });
+}
+
+/// Fetch the latest release tag from GitHub and compare against the build hash.
+fn check_latest_release() -> String {
+    match fetch_latest_tag() {
+        Ok(tag) => compare_versions(BUILD_HASH, &tag),
+        Err(_) => String::new(),
+    }
+}
+
+/// Compare a build hash against a release tag.
+/// Returns an update notice string, or empty if up-to-date.
+fn compare_versions(build_hash: &str, release_tag: &str) -> String {
+    let tag = release_tag.trim();
+    if tag.is_empty() || build_hash == "unknown" {
+        return String::new();
+    }
+
+    let our_short = &build_hash[..build_hash.len().min(tag.len())];
+    if our_short == tag {
+        return String::new();
+    }
+
+    format!(
+        "\n[fff update available: `curl -fsSL https://raw.githubusercontent.com/{REPO}/main/install-mcp.sh | bash`]\n"
+    )
+}
+
+/// Shell out to curl to fetch the latest release tag name from GitHub API.
+fn fetch_latest_tag() -> Result<String, Box<dyn std::error::Error>> {
+    let output = std::process::Command::new("curl")
+        .args([
+            "-fsSL",
+            "--max-time",
+            "5",
+            "-H",
+            "Accept: application/vnd.github.v3+json",
+            &format!("https://api.github.com/repos/{REPO}/releases?per_page=1"),
+        ])
+        .output()?;
+
+    if !output.status.success() {
+        return Err("curl failed".into());
+    }
+
+    let body = String::from_utf8(output.stdout)?;
+    let releases: Vec<serde_json::Value> = serde_json::from_str(&body)?;
+    let tag = releases
+        .first()
+        .and_then(|r| r.get("tag_name"))
+        .and_then(|v| v.as_str())
+        .unwrap_or("")
+        .to_string();
+
+    Ok(tag)
+}
@@ -1,12 +1,16 @@
 [package]
 name = "fff-nvim"
-version = "0.1.0"
+version = "0.2.4"
 edition = "2024"

 [lib]
 path = "src/lib.rs"
 crate-type = ["cdylib", "rlib"]

+[features]
+default = []
+zlob = ["fff/zlob"]
+
 [[bin]]
 name = "test_watcher"
 path = "src/bin/test_watcher.rs"
@@ -40,8 +44,8 @@ thiserror = { workspace = true }
 tracing = { workspace = true }

 # Local crates
-fff-core = { path = "../fff-core" }
-fff-query-parser = { path = "../fff-query-parser" }
+fff = { package = "fff-search", path = "../fff-core" , version = "0.2.2" }
+fff-query-parser = { path = "../fff-query-parser" , version = "0.2.4" }

 # External dependencies
 blake3 = "1.8.2"
@@ -61,9 +65,6 @@ once_cell = "1.20.2"
 pathdiff = "0.2.1"
 serde = { version = "1.0", features = ["derive"] }
 smartstring = { version = "1.0.1", features = ["serde"] }
-tracing-appender = "0.2"
-tracing-subscriber = { version = "0.3", features = ["env-filter"] }
-zlob = { workspace = true }

 [dev-dependencies]
 criterion = { version = "0.5", features = ["html_reports"] }
@@ -78,7 +79,3 @@ harness = false
 name = "query_tracker_bench"
 harness = false

-# Platform-specific: Use vendored OpenSSL on non-Windows (Linux, macOS)
-# On Windows, git2 uses the native SChannel TLS backend
-[target.'cfg(not(windows))'.dependencies]
-openssl = { version = "0.10", features = ["vendored"] }
@@ -1,8 +1,9 @@
 use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
-use fff_nvim::FILE_PICKER;
-use fff_nvim::file_picker::{FilePicker, FuzzySearchOptions};
-use fff_nvim::types::PaginationArgs;
+use fff::file_picker::{FFFMode, FilePicker};
+use fff::types::{FileItem, PaginationArgs};
+use fff::{FuzzySearchOptions, QueryParser, SharedFrecency, SharedPicker};
 use std::path::PathBuf;
+use std::sync::{Arc, RwLock};
 use std::time::Duration;

 /// Initialize tracing to output to console
@@ -19,20 +20,27 @@ fn init_tracing() {
    //     .try_init();
 }

-/// Initialize FilePicker and insert into global state
-fn init_file_picker_internal(path: &str) -> Result<(), String> {
-    let picker = FilePicker::new(path.to_string())
-        .map_err(|e| format!("Failed to create FilePicker: {:?}", e))?;
-
-    let mut picker_guard = FILE_PICKER
-        .write()
-        .map_err(|_| "Failed to acquire write lock")?;
-    *picker_guard = Some(picker);
-    Ok(())
+/// Initialize FilePicker using shared state
+fn init_file_picker_internal(
+    path: &str,
+    shared_picker: &SharedPicker,
+    shared_frecency: &SharedFrecency,
+) -> Result<(), String> {
+    FilePicker::new_with_shared_state(
+        path.to_string(),
+        false,
+        FFFMode::Neovim,
+        Arc::clone(shared_picker),
+        Arc::clone(shared_frecency),
+    )
+    .map_err(|e| format!("Failed to create FilePicker: {:?}", e))
 }

 /// Helper function to wait for scanning to complete and get file count
-fn wait_for_scan_completion(timeout_secs: u64) -> Result<usize, String> {
+fn wait_for_scan_completion(
+    shared_picker: &SharedPicker,
+    timeout_secs: u64,
+) -> Result<usize, String> {
    let start = std::time::Instant::now();
    let timeout = Duration::from_secs(timeout_secs);
    let mut last_log = std::time::Instant::now();
@@ -42,7 +50,7 @@ fn wait_for_scan_completion(timeout_secs: u64) -> Result<usize, String> {
        iteration += 1;

        {
-            let picker_guard = FILE_PICKER
+            let picker_guard = shared_picker
                .read()
                .map_err(|_| "Failed to acquire read lock")?;
            if let Some(ref picker) = *picker_guard {
@@ -91,9 +99,9 @@ fn wait_for_scan_completion(timeout_secs: u64) -> Result<usize, String> {
    }
 }

-/// Get files from the global FILE_PICKER
-fn get_files_snapshot() -> Result<Vec<fff_nvim::types::FileItem>, String> {
-    let picker_guard = FILE_PICKER
+/// Get files from the shared picker
+fn get_files_snapshot(shared_picker: &SharedPicker) -> Result<Vec<FileItem>, String> {
+    let picker_guard = shared_picker
        .read()
        .map_err(|_| "Failed to acquire read lock")?;
    if let Some(ref picker) = *picker_guard {
@@ -103,9 +111,9 @@ fn get_files_snapshot() -> Result<Vec<fff_nvim::types::FileItem>, String> {
    }
 }

-/// Clean up global state
-fn cleanup_global_state() {
-    if let Ok(mut picker_guard) = FILE_PICKER.write() {
+/// Clean up shared state
+fn cleanup_shared_state(shared_picker: &SharedPicker) {
+    if let Ok(mut picker_guard) = shared_picker.write() {
        if let Some(mut picker) = picker_guard.take() {
            picker.stop_background_monitor();
        }
@@ -113,7 +121,7 @@ fn cleanup_global_state() {
 }

 /// Initialize FilePicker once and return files snapshot
-fn setup_once() -> Result<Vec<fff_nvim::types::FileItem>, String> {
+fn setup_once() -> Result<(Vec<FileItem>, SharedPicker, SharedFrecency), String> {
    init_tracing();

    let big_repo_path = PathBuf::from("./big-repo");
@@ -121,36 +129,28 @@ fn setup_once() -> Result<Vec<fff_nvim::types::FileItem>, String> {
        return Err("./big-repo directory does not exist. Run git clone https://github.com/torvalds/linux.git big-repo".to_string());
    }

-    let canonical_path = fff_core::path_utils::canonicalize(&big_repo_path)
+    let canonical_path = fff::path_utils::canonicalize(&big_repo_path)
        .map_err(|e| format!("Failed to canonicalize path: {}", e))?;
    eprintln!("  Path: {:?}", canonical_path);

-    {
-        let picker_guard = FILE_PICKER
-            .read()
-            .map_err(|_| "Failed to acquire read lock")?;
-        if let Some(ref picker) = *picker_guard {
-            let files = picker.get_files();
-            if !files.is_empty() {
-                eprintln!("  ℹ Reusing existing index with {} files", files.len());
-                return Ok(files.to_vec());
-            }
-        }
-    }
+    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
+    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));

-    cleanup_global_state();
-    std::thread::sleep(Duration::from_millis(500));
-
-    init_file_picker_internal(&canonical_path.to_string_lossy())?;
+    init_file_picker_internal(
+        &canonical_path.to_string_lossy(),
+        &shared_picker,
+        &shared_frecency,
+    )?;

    eprintln!("  Waiting for background scan to complete...");
-    let file_count = wait_for_scan_completion(120)?;
+    let file_count = wait_for_scan_completion(&shared_picker, 120)?;
    eprintln!(
        "  ✓ Indexed {} files (will be reused for all benchmarks)\n",
        file_count
    );

-    get_files_snapshot()
+    let files = get_files_snapshot(&shared_picker)?;
+    Ok((files, shared_picker, shared_frecency))
 }

 /// Benchmark for indexing the big-repo directory
@@ -165,7 +165,7 @@ fn bench_indexing(c: &mut Criterion) {
        return;
    }

-    let canonical_path = match fff_core::path_utils::canonicalize(&big_repo_path) {
+    let canonical_path = match fff::path_utils::canonicalize(&big_repo_path) {
        Ok(p) => p,
        Err(e) => {
            eprintln!("⚠ Failed to canonicalize path: {}", e);
@@ -179,21 +179,23 @@ fn bench_indexing(c: &mut Criterion) {

    group.bench_function("index_big_repo", |b| {
        b.iter(|| {
-            cleanup_global_state();
-            std::thread::sleep(Duration::from_millis(500));
+            let sp: SharedPicker = Arc::new(RwLock::new(None));
+            let sf: SharedFrecency = Arc::new(RwLock::new(None));

            let start = std::time::Instant::now();
-            init_file_picker_internal(black_box(&canonical_path.to_string_lossy()))
+            init_file_picker_internal(black_box(&canonical_path.to_string_lossy()), &sp, &sf)
                .expect("Failed to init FilePicker");

-            match wait_for_scan_completion(120) {
+            match wait_for_scan_completion(&sp, 120) {
                Ok(file_count) => {
                    let elapsed = start.elapsed();
                    eprintln!("  ✓ Indexed {} files in {:?}", file_count, elapsed);
+                    cleanup_shared_state(&sp);
                    file_count
                }
                Err(e) => {
                    eprintln!("  ✗ Error: {}", e);
+                    cleanup_shared_state(&sp);
                    0
                }
            }
@@ -205,8 +207,8 @@ fn bench_indexing(c: &mut Criterion) {

 /// Benchmark for searching with various query patterns
 fn bench_search_queries(c: &mut Criterion) {
-    let files = match setup_once() {
-        Ok(files) => files,
+    let (files, _sp, _sf) = match setup_once() {
+        Ok(result) => result,
        Err(e) => {
            eprint!("Failed to setup picker {e:?}");
            return;
@@ -224,16 +226,18 @@ fn bench_search_queries(c: &mut Criterion) {
        ("partial", "src/lib"),
    ];

+    let parser = QueryParser::default();
+
    for (name, query) in test_queries {
-        group.bench_with_input(BenchmarkId::new("query", name), &query, |b, &query| {
+        let parsed = parser.parse(query);
+        group.bench_with_input(BenchmarkId::new("query", name), &query, |b, &_query| {
            b.iter(|| {
                let results = FilePicker::fuzzy_search(
                    black_box(&files),
-                    black_box(query),
+                    black_box(&parsed),
                    FuzzySearchOptions {
                        max_threads: 4,
                        current_file: None,
-
                        project_path: None,
                        last_same_query_match: None,
                        combo_boost_score_multiplier: 100,
@@ -254,8 +258,8 @@ fn bench_search_queries(c: &mut Criterion) {

 /// Benchmark search with different thread counts
 fn bench_search_thread_scaling(c: &mut Criterion) {
-    let files = match setup_once() {
-        Ok(files) => files,
+    let (files, _sp, _sf) = match setup_once() {
+        Ok(result) => result,
        Err(e) => {
            eprintln!("⚠ Skipping thread scaling benchmarks: {}", e);
            return;
@@ -266,6 +270,8 @@ fn bench_search_thread_scaling(c: &mut Criterion) {
    group.sample_size(100);

    let query = "controller";
+    let parser = QueryParser::default();
+    let parsed = parser.parse(query);
    let thread_counts = vec![1, 2, 4, 8];

    for threads in thread_counts {
@@ -276,11 +282,10 @@ fn bench_search_thread_scaling(c: &mut Criterion) {
                b.iter(|| {
                    let results = FilePicker::fuzzy_search(
                        black_box(&files),
-                        black_box(query),
+                        black_box(&parsed),
                        FuzzySearchOptions {
                            max_threads: threads,
                            current_file: None,
-
                            project_path: None,
                            last_same_query_match: None,
                            combo_boost_score_multiplier: 100,
@@ -302,8 +307,8 @@ fn bench_search_thread_scaling(c: &mut Criterion) {

 /// Benchmark search with different result limits
 fn bench_search_result_limits(c: &mut Criterion) {
-    let files = match setup_once() {
-        Ok(files) => files,
+    let (files, _sp, _sf) = match setup_once() {
+        Ok(result) => result,
        Err(e) => {
            eprintln!("⚠ Skipping result limit benchmarks: {}", e);
            return;
@@ -314,6 +319,8 @@ fn bench_search_result_limits(c: &mut Criterion) {
    group.sample_size(100);

    let query = "mod";
+    let parser = QueryParser::default();
+    let parsed = parser.parse(query);
    let result_limits = vec![10, 50, 100, 500];

    for limit in result_limits {
@@ -321,11 +328,10 @@ fn bench_search_result_limits(c: &mut Criterion) {
            b.iter(|| {
                let results = FilePicker::fuzzy_search(
                    black_box(&files),
-                    black_box(query),
+                    black_box(&parsed),
                    FuzzySearchOptions {
                        max_threads: 4,
                        current_file: None,
-
                        project_path: None,
                        last_same_query_match: None,
                        combo_boost_score_multiplier: 100,
@@ -346,8 +352,8 @@ fn bench_search_result_limits(c: &mut Criterion) {

 /// Benchmark search algorithm performance scaling with file count
 fn bench_search_scalability(c: &mut Criterion) {
-    let all_files = match setup_once() {
-        Ok(files) => files,
+    let (all_files, _sp, _sf) = match setup_once() {
+        Ok(result) => result,
        Err(e) => {
            eprintln!("⚠ Skipping scalability benchmarks: {}", e);
            return;
@@ -366,6 +372,8 @@ fn bench_search_scalability(c: &mut Criterion) {
    group.sample_size(50);

    let query = "controller";
+    let parser = QueryParser::default();
+    let parsed = parser.parse(query);
    let file_counts = vec![100, 1000, 5000, 10000, all_files.len().min(50000)];

    for count in file_counts {
@@ -378,11 +386,10 @@ fn bench_search_scalability(c: &mut Criterion) {
            b.iter(|| {
                let results = FilePicker::fuzzy_search(
                    black_box(subset),
-                    black_box(query),
+                    black_box(&parsed),
                    FuzzySearchOptions {
                        max_threads: 4,
                        current_file: None,
-
                        project_path: None,
                        last_same_query_match: None,
                        combo_boost_score_multiplier: 100,
@@ -403,8 +410,8 @@ fn bench_search_scalability(c: &mut Criterion) {

 /// Benchmark search performance with different ordering modes
 fn bench_search_ordering(c: &mut Criterion) {
-    let files = match setup_once() {
-        Ok(files) => files,
+    let (files, _sp, _sf) = match setup_once() {
+        Ok(result) => result,
        Err(e) => {
            eprintln!("⚠ Skipping ordering benchmarks: {}", e);
            return;
@@ -414,18 +421,19 @@ fn bench_search_ordering(c: &mut Criterion) {
    let mut group = c.benchmark_group("ordering");
    group.sample_size(100);

-    let query = "controller";
+    let parser = QueryParser::default();
+    let parsed_controller = parser.parse("controller");
+    let parsed_mod = parser.parse("mod");

    // Benchmark normal order (descending)
    group.bench_function("normal_order", |b| {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box(query),
+                black_box(&parsed_controller),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -445,11 +453,10 @@ fn bench_search_ordering(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box(query),
+                black_box(&parsed_controller),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -469,11 +476,10 @@ fn bench_search_ordering(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box("mod"),
+                black_box(&parsed_mod),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -492,11 +498,10 @@ fn bench_search_ordering(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box("mod"),
+                black_box(&parsed_mod),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -516,11 +521,10 @@ fn bench_search_ordering(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box("controller"),
+                black_box(&parsed_controller),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -539,11 +543,10 @@ fn bench_search_ordering(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box("controller"),
+                black_box(&parsed_controller),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -563,8 +566,8 @@ fn bench_search_ordering(c: &mut Criterion) {

 /// Benchmark pagination: first page vs deep page
 fn bench_pagination_performance(c: &mut Criterion) {
-    let files = match setup_once() {
-        Ok(files) => files,
+    let (files, _sp, _sf) = match setup_once() {
+        Ok(result) => result,
        Err(e) => {
            eprintln!("⚠ Skipping pagination benchmarks: {}", e);
            return;
@@ -575,6 +578,8 @@ fn bench_pagination_performance(c: &mut Criterion) {
    group.sample_size(100);

    let query = "mod";
+    let parser = QueryParser::default();
+    let parsed = parser.parse(query);
    let page_size = 40;

    // Benchmark first page (uses partial sort optimization)
@@ -582,11 +587,10 @@ fn bench_pagination_performance(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box(query),
+                black_box(&parsed),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -606,11 +610,10 @@ fn bench_pagination_performance(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box(query),
+                black_box(&parsed),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -630,11 +633,10 @@ fn bench_pagination_performance(c: &mut Criterion) {
        b.iter(|| {
            let results = FilePicker::fuzzy_search(
                black_box(&files),
-                black_box(query),
+                black_box(&parsed),
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
-
                    project_path: None,
                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
@@ -1,5 +1,5 @@
 use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
-use fff_nvim::query_tracker::QueryTracker;
+use fff::query_tracker::QueryTracker;
 use rand::distributions::Alphanumeric;
 use rand::prelude::*;
 use std::path::PathBuf;
@@ -1,6 +1,6 @@
 /// Simple search profiler that directly uses scan_filesystem without background thread overhead
-use fff_core::file_picker::FilePicker;
-use fff_core::{FileItem, FuzzySearchOptions, PaginationArgs, QueryParser};
+use fff::file_picker::FilePicker;
+use fff::{FileItem, FuzzySearchOptions, PaginationArgs, QueryParser};
 use std::time::Instant;

 fn main() {
@@ -14,7 +14,7 @@ fn main() {
    }

    let canonical_path =
-        fff_core::path_utils::canonicalize(&big_repo_path).expect("Failed to canonicalize path");
+        fff::path_utils::canonicalize(&big_repo_path).expect("Failed to canonicalize path");

    eprintln!("Loading files from: {:?}", canonical_path);

@@ -87,13 +87,12 @@ fn main() {
            let parsed = parser.parse(query);
            let results = FilePicker::fuzzy_search(
                &files,
-                query,
-                parsed,
+                &parsed,
+                None,
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
                    project_path: None,
-                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
                    min_combo_count: 3,
                    pagination: PaginationArgs {
@@ -1,4 +1,4 @@
-use fff_core::FileItem;
+use fff::FileItem;
 /// Fuzzy grep quality test against ~/dev/lightsource
 ///
 /// Runs queries through the fuzzy grep pipeline and prints results
@@ -7,7 +7,7 @@ use fff_core::FileItem;
 /// Usage:
 ///   cargo run --release --bin fuzzy_grep_test              # runs default test queries
 ///   cargo run --release --bin fuzzy_grep_test -- "query"   # runs a single user query
-use fff_core::grep::{GrepMode, GrepSearchOptions, grep_search, parse_grep_query};
+use fff::grep::{GrepMode, GrepSearchOptions, grep_search, parse_grep_query};
 use std::io::Read;
 use std::path::Path;
 use std::time::Instant;
@@ -71,11 +71,14 @@ fn run_fuzzy_query(files: &[FileItem], query: &str, label: &str) {
        page_limit: 100, // Get plenty of results
        mode: GrepMode::Fuzzy,
        time_budget_ms: 0, // No time limit — search all files
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
    };

    let parsed = parse_grep_query(query);
    let start = Instant::now();
-    let result = grep_search(files, query, parsed, &options);
+    let result = grep_search(files, &parsed, &options);
    let elapsed = start.elapsed();

    eprintln!("══════════════════════════════════════════════════════════════");
@@ -163,8 +166,7 @@ fn main() {
        std::process::exit(1);
    }

-    let canonical =
-        fff_core::path_utils::canonicalize(&repo_path).expect("Failed to canonicalize path");
+    let canonical = fff::path_utils::canonicalize(&repo_path).expect("Failed to canonicalize path");
    eprintln!("=== Fuzzy Grep Quality Test ===");
    eprintln!("Repository: {:?}\n", canonical);

@@ -1,4 +1,4 @@
-use fff_core::FileItem;
+use fff::FileItem;
 /// Live grep benchmark profiler for fff.nvim
 ///
 /// Benchmarks the full grep pipeline against a large repository (Linux kernel).
@@ -10,7 +10,7 @@ use fff_core::FileItem;
 /// Usage:
 ///   cargo build --release --bin grep_profiler
 ///   ./target/release/grep_profiler [--path /path/to/repo]
-use fff_core::grep::{GrepSearchOptions, grep_search, parse_grep_query};
+use fff::grep::{GrepMode, GrepSearchOptions, grep_search, parse_grep_query};
 use std::io::Read;
 use std::path::Path;
 use std::time::{Duration, Instant};
@@ -119,6 +119,10 @@ struct GrepBench<'a> {

 impl<'a> GrepBench<'a> {
    fn new(files: &'a [FileItem]) -> Self {
+        Self::with_mode(files, GrepMode::PlainText)
+    }
+
+    fn with_mode(files: &'a [FileItem], mode: GrepMode) -> Self {
        Self {
            files,
            options: GrepSearchOptions {
@@ -127,8 +131,11 @@ impl<'a> GrepBench<'a> {
                smart_case: true,
                file_offset: 0,
                page_limit: 50,
-                mode: Default::default(),
+                mode,
                time_budget_ms: 0,
+                before_context: 0,
+                after_context: 0,
+                classify_definitions: false,
            },
        }
    }
@@ -137,13 +144,9 @@ impl<'a> GrepBench<'a> {
    fn run_once(&self, query: &str) -> (Duration, usize, usize) {
        let parsed = parse_grep_query(query);
        let start = Instant::now();
-        let result = grep_search(self.files, query, parsed, &self.options);
+        let result = grep_search(self.files, &parsed, &self.options);
        let elapsed = start.elapsed();
-        (
-            elapsed,
-            result.total_match_count,
-            result.total_files_searched,
-        )
+        (elapsed, result.matches.len(), result.total_files_searched)
    }

    /// Benchmark a query with multiple iterations
@@ -219,12 +222,12 @@ fn main() {
        std::process::exit(1);
    }

-    let canonical = fff_core::path_utils::canonicalize(&repo).expect("Failed to canonicalize path");
+    let canonical = fff::path_utils::canonicalize(&repo).expect("Failed to canonicalize path");
    eprintln!("=== FFF Live Grep Profiler ===");
    eprintln!("Repository: {:?}", canonical);

    // Direct file loading (no background thread)
-    eprintln!("\n[1/5] Loading files...");
+    eprintln!("\n[1/7] Loading files...");
    let load_start = Instant::now();
    let files = load_files(&canonical);
    let load_time = load_start.elapsed();
@@ -240,7 +243,7 @@ fn main() {

    let bench = GrepBench::new(&files);

-    eprintln!("[2/5] Cold cache benchmarks (first search, mmap not yet loaded)");
+    eprintln!("[2/7] Cold cache benchmarks (first search, mmap not yet loaded)");
    eprintln!("  Each query runs once with fresh FileItem mmaps.\n");
    print_header();

@@ -263,7 +266,7 @@ fn main() {
        print_row(name, &stats, matches, files_searched, 1);
    }

-    eprintln!("\n[3/5] Warm cache benchmarks (mmap cache populated)");
+    eprintln!("\n[3/7] Warm cache benchmarks (plain text, mmap cache populated)");
    eprintln!("  Running 3 warmup iterations, then measuring.\n");
    print_header();

@@ -297,7 +300,84 @@ fn main() {
        print_row(name, &stats, matches, files_searched, *iters);
    }

-    eprintln!("\n[4/5] Incremental typing simulation");
+    // ── Fuzzy grep benchmarks ─────────────────────────────────────────────
+    eprintln!("\n[4/7] Fuzzy grep warm benchmarks");
+    eprintln!("  Running 3 warmup iterations, then measuring.\n");
+    print_header();
+
+    let fuzzy_bench = GrepBench::with_mode(&files, GrepMode::Fuzzy);
+
+    let fuzzy_queries: Vec<(&str, &str, usize)> = vec![
+        ("fuzzy_exact", "mutex_lock", 15),
+        ("fuzzy_typo", "mutx_lock", 15),
+        ("fuzzy_camel", "InodeOps", 15),
+        ("fuzzy_abbrev", "sched_rt", 15),
+        ("fuzzy_short", "kfr", 15),
+        ("fuzzy_common", "return", 10),
+        ("fuzzy_define", "MODULE_LICENSE", 15),
+        ("fuzzy_struct", "file_operations", 15),
+        ("fuzzy_long", "static_int_init", 15),
+        ("fuzzy_path", "printk *.c", 15),
+    ];
+
+    // Warmup
+    for (_, query, _) in &fuzzy_queries {
+        for _ in 0..3 {
+            fuzzy_bench.run_once(query);
+        }
+    }
+
+    for (name, query, iters) in &fuzzy_queries {
+        let (stats, matches, files_searched) = fuzzy_bench.bench_query(query, *iters);
+        print_row(name, &stats, matches, files_searched, *iters);
+    }
+
+    // ── Fuzzy incremental typing ────────────────────────────────────────
+    eprintln!("\n[5/7] Fuzzy incremental typing simulation");
+    eprintln!("  Simulates user typing character by character (fuzzy mode).\n");
+
+    let fuzzy_typing_sequences: Vec<(&str, Vec<&str>)> = vec![
+        (
+            "mutex_lock",
+            vec![
+                "m",
+                "mu",
+                "mut",
+                "mute",
+                "mutex",
+                "mutex_",
+                "mutex_l",
+                "mutex_lo",
+                "mutex_loc",
+                "mutex_lock",
+            ],
+        ),
+        ("printk", vec!["p", "pr", "pri", "prin", "print", "printk"]),
+        ("kfree", vec!["k", "kf", "kfr", "kfre", "kfree"]),
+    ];
+
+    for (name, sequence) in &fuzzy_typing_sequences {
+        eprintln!("  Typing '{}' ({} keystrokes):", name, sequence.len());
+        eprintln!(
+            "    {:>16} | {:>8} | {:>6} | {:>6}",
+            "Query", "Latency", "Match", "Files"
+        );
+        eprintln!("    {:-<16}-+-{:-<8}-+-{:-<6}-+-{:-<6}", "", "", "", "");
+
+        for prefix in sequence {
+            let (elapsed, matches, files_searched) = fuzzy_bench.run_once(prefix);
+            eprintln!(
+                "    {:>16} | {:>8} | {:>6} | {:>6}",
+                format!("\"{}\"", prefix),
+                fmt_dur(elapsed),
+                matches,
+                files_searched,
+            );
+        }
+        eprintln!();
+    }
+
+    eprintln!("[6/7] Incremental typing simulation (plain text)");
    eprintln!("  Simulates user typing character by character.\n");

    let typing_sequences: Vec<(&str, Vec<&str>)> = vec![
@@ -342,7 +422,7 @@ fn main() {
        eprintln!();
    }

-    eprintln!("[5/5] Pagination benchmark");
+    eprintln!("[7/7] Pagination benchmark");
    eprintln!("  Testing page_offset performance for common query.\n");

    let pagination_query = "return";
@@ -367,9 +447,12 @@ fn main() {
            page_limit: 50,
            mode: Default::default(),
            time_budget_ms: 0,
+            before_context: 0,
+            after_context: 0,
+            classify_definitions: false,
        };
        let start = Instant::now();
-        let result = grep_search(&files, pagination_query, parsed, &opts);
+        let result = grep_search(&files, &parsed, &opts);
        let elapsed = start.elapsed();
        eprintln!(
            "    {:>6} | {:>12} | {:>8} | {:>6} | {:>12}",
@@ -1,4 +1,5 @@
-use fff_core::FileItem;
+use fff::FFFQuery;
+use fff::FileItem;
 /// FFF vs ripgrep comparison benchmark
 ///
 /// Demonstrates why a persistent in-process search engine (fff) is fundamentally
@@ -20,7 +21,7 @@ use fff_core::FileItem;
 /// Usage:
 ///   cargo build --release --bin grep_vs_rg
 ///   ./target/release/grep_vs_rg [--path /path/to/repo] [--iters 5]
-use fff_core::grep::{GrepSearchOptions, grep_search, parse_grep_query};
+use fff::grep::{GrepSearchOptions, grep_search, parse_grep_query};
 use std::io::Read;
 use std::path::Path;
 use std::process::Command;
@@ -204,11 +205,34 @@ fn run_fff_full(files: &[FileItem], query: &str) -> (usize, Duration) {
        page_limit: usize::MAX,
        mode: Default::default(),
        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
    };
    let start = Instant::now();
-    let result = grep_search(files, query, parsed, &options);
+    let result = grep_search(files, &parsed, &options);
    let elapsed = start.elapsed();
-    (result.total_match_count, elapsed)
+    (result.matches.len(), elapsed)
+}
+
+#[allow(dead_code)]
+fn benchmark_fff_smart_case(files: &[FileItem], parsed: &FFFQuery<'_>) -> (usize, Duration) {
+    let options = GrepSearchOptions {
+        max_file_size: 10 * 1024 * 1024,
+        max_matches_per_file: usize::MAX,
+        smart_case: true,
+        file_offset: 0,
+        page_limit: 5000,
+        mode: Default::default(),
+        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
+    };
+    let start = Instant::now();
+    let result = grep_search(files, parsed, &options);
+    let elapsed = start.elapsed();
+    (result.matches.len(), elapsed)
 }

 /// fff paginated: first 50 results only (real UI scenario).
@@ -222,12 +246,13 @@ fn run_fff_page(files: &[FileItem], query: &str) -> (usize, Duration) {
        page_limit: 50,
        mode: Default::default(),
        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
    };
    let start = Instant::now();
-    let result = grep_search(files, query, parsed, &options);
+    let result = grep_search(files, &parsed, &options);
    let elapsed = start.elapsed();
-    // Use matches.len() — the actual truncated page the UI would display,
-    // not total_match_count which includes overshoot from parallel batches.
    (result.matches.len(), elapsed)
 }

@@ -294,7 +319,7 @@ fn main() {
        std::process::exit(1);
    }

-    let canonical = fff_core::path_utils::canonicalize(&repo).expect("Failed to canonicalize path");
+    let canonical = fff::path_utils::canonicalize(&repo).expect("Failed to canonicalize path");

    let rg_version = Command::new("rg")
        .arg("--version")
@@ -1,6 +1,7 @@
-use fff_core::file_picker::FilePicker;
-use fff_core::{FILE_PICKER, FuzzySearchOptions, PaginationArgs, QueryParser};
+use fff::file_picker::{FFFMode, FilePicker};
+use fff::{FuzzySearchOptions, PaginationArgs, QueryParser, SharedFrecency, SharedPicker};
 use std::env;
+use std::sync::{Arc, RwLock};
 use std::thread;
 use std::time::Duration;

@@ -60,6 +61,7 @@ fn format_bytes(bytes: usize) -> String {
 }

 fn test_search_memory_pattern(
+    shared_picker: &SharedPicker,
    name: &str,
    iterations: usize,
    query_pattern: impl Fn(usize) -> String,
@@ -82,19 +84,18 @@ fn test_search_memory_pattern(
        let query = query_pattern(i);

        let (result_count, _total_matched) = {
-            let file_picker_guard = FILE_PICKER.read().unwrap();
-            if let Some(ref picker) = *file_picker_guard {
+            let guard = shared_picker.read().unwrap();
+            if let Some(ref picker) = *guard {
                let parser = QueryParser::default();
                let parsed = parser.parse(&query);
                let search_result = FilePicker::fuzzy_search(
                    picker.get_files(),
-                    &query,
-                    parsed,
+                    &parsed,
+                    None,
                    FuzzySearchOptions {
                        max_threads: 1 + (i % 4),
                        current_file: None,
                        project_path: None,
-                        last_same_query_match: None,
                        combo_boost_score_multiplier: 100,
                        min_combo_count: 3,
                        pagination: PaginationArgs {
@@ -177,20 +178,25 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    println!("Test directory: {}", base_path);
    println!();

+    // Create shared state
+    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
+    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));
+
    // Initialize FilePicker
-    {
-        let mut file_picker_guard = FILE_PICKER.write().unwrap();
-        if file_picker_guard.is_none() {
-            println!("Initializing FilePicker...");
-            *file_picker_guard = Some(FilePicker::new(base_path.clone())?);
-        }
-    }
+    println!("Initializing FilePicker...");
+    FilePicker::new_with_shared_state(
+        base_path.clone(),
+        false,
+        FFFMode::Neovim,
+        Arc::clone(&shared_picker),
+        Arc::clone(&shared_frecency),
+    )?;

    // Wait for initial scan
    println!("Waiting for file scan...");
    loop {
-        if let Ok(file_picker_guard) = FILE_PICKER.read()
-            && let Some(ref picker) = *file_picker_guard
+        if let Ok(guard) = shared_picker.read()
+            && let Some(ref picker) = *guard
            && !picker.is_scan_active()
            && !picker.get_files().is_empty()
        {
@@ -200,8 +206,8 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    }

    let file_count = {
-        let file_picker_guard = FILE_PICKER.read()?;
-        file_picker_guard.as_ref().unwrap().get_files().len()
+        let guard = shared_picker.read().unwrap();
+        guard.as_ref().unwrap().get_files().len()
    };

    println!("📊 Found {} files", file_count);
@@ -216,10 +222,12 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Test different memory patterns

    // 1. Repeated same query - should have minimal growth if caching works
-    test_search_memory_pattern("Same Query Repeated (1000x)", 1000, |_| "test".to_string())?;
+    test_search_memory_pattern(&shared_picker, "Same Query Repeated (1000x)", 1000, |_| {
+        "test".to_string()
+    })?;

    // 2. Cycling through different queries
-    test_search_memory_pattern("Cycling Queries (1000x)", 1000, |i| {
+    test_search_memory_pattern(&shared_picker, "Cycling Queries (1000x)", 1000, |i| {
        let queries = [
            "test", "main", "lib", "src", "mod", "file", "picker", "fuzzy", "search",
        ];
@@ -227,24 +235,25 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    })?;

    // 3. Unique queries each time - worst case for any caching
-    test_search_memory_pattern("Unique Queries (500x)", 500, |i| {
+    test_search_memory_pattern(&shared_picker, "Unique Queries (500x)", 500, |i| {
        format!("unique_query_{}", i)
    })?;

    // 4. Queries that return many results
    test_search_memory_pattern(
+        &shared_picker,
        "High Result Count (500x)",
        500,
        |_| "a".to_string(), // Single character likely to match many files
    )?;

    // 5. Queries with no results
-    test_search_memory_pattern("No Results (500x)", 500, |_| {
+    test_search_memory_pattern(&shared_picker, "No Results (500x)", 500, |_| {
        "zzzz_no_match_expected".to_string()
    })?;

    // 6. Long intensive test
-    test_search_memory_pattern("Long Intensive Test (2000x)", 2000, |i| {
+    test_search_memory_pattern(&shared_picker, "Long Intensive Test (2000x)", 2000, |i| {
        let patterns = [
            "rs", "lua", "toml", "mod", "lib", "main", "test", "src", "file",
        ];
@@ -1,9 +1,12 @@
-use fff_core::file_picker::FilePicker;
-use fff_core::{FILE_PICKER, FileItem, FuzzySearchOptions, PaginationArgs, QueryParser};
+use fff::file_picker::{FFFMode, FilePicker};
+use fff::{
+    FileItem, FuzzySearchOptions, PaginationArgs, QueryParser, SharedFrecency, SharedPicker,
+};
+use std::sync::{Arc, RwLock};
 use std::time::{Duration, Instant};

 /// Wait for background scan to complete
-fn wait_for_scan(timeout_secs: u64) -> Result<usize, String> {
+fn wait_for_scan(shared_picker: &SharedPicker, timeout_secs: u64) -> Result<usize, String> {
    let start = Instant::now();
    let timeout = Duration::from_secs(timeout_secs);
    let mut iteration = 0;
@@ -11,7 +14,7 @@ fn wait_for_scan(timeout_secs: u64) -> Result<usize, String> {
    loop {
        iteration += 1;

-        let picker_guard = FILE_PICKER
+        let picker_guard = shared_picker
            .read()
            .map_err(|_| "Failed to acquire read lock")?;
        if let Some(ref picker) = *picker_guard {
@@ -45,21 +48,9 @@ fn wait_for_scan(timeout_secs: u64) -> Result<usize, String> {
    }
 }

-/// Initialize FilePicker and insert into global state
-fn init_file_picker(path: &str) -> Result<(), String> {
-    let picker = FilePicker::new(path.to_string())
-        .map_err(|e| format!("Failed to create FilePicker: {:?}", e))?;
-
-    let mut picker_guard = FILE_PICKER
-        .write()
-        .map_err(|_| "Failed to acquire write lock")?;
-    *picker_guard = Some(picker);
-    Ok(())
-}
-
-/// Get files snapshot from global state
-fn get_files() -> Result<Vec<FileItem>, String> {
-    let picker_guard = FILE_PICKER
+/// Get files snapshot from shared state
+fn get_files(shared_picker: &SharedPicker) -> Result<Vec<FileItem>, String> {
+    let picker_guard = shared_picker
        .read()
        .map_err(|_| "Failed to acquire read lock")?;
    if let Some(ref picker) = *picker_guard {
@@ -80,19 +71,30 @@ fn main() {
    }

    let canonical_path =
-        fff_core::path_utils::canonicalize(&big_repo_path).expect("Failed to canonicalize path");
+        fff::path_utils::canonicalize(&big_repo_path).expect("Failed to canonicalize path");
+
+    // Create shared state
+    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
+    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));

    eprintln!("Initializing FilePicker for: {:?}", canonical_path);
-    init_file_picker(&canonical_path.to_string_lossy()).expect("Failed to init FilePicker");
+    FilePicker::new_with_shared_state(
+        canonical_path.to_string_lossy().to_string(),
+        false,
+        FFFMode::Neovim,
+        Arc::clone(&shared_picker),
+        Arc::clone(&shared_frecency),
+    )
+    .expect("Failed to init FilePicker");

    // Give background thread time to start
    std::thread::sleep(Duration::from_millis(200));

    eprintln!("Waiting for scan to complete...");
-    let file_count = wait_for_scan(120).expect("Failed to wait for scan");
+    let file_count = wait_for_scan(&shared_picker, 120).expect("Failed to wait for scan");
    eprintln!("✓ Indexed {} files\n", file_count);

-    let files = get_files().expect("Failed to get files");
+    let files = get_files(&shared_picker).expect("Failed to get files");

    // Test queries representing different search patterns
    let test_queries = vec![
@@ -124,13 +126,12 @@ fn main() {
            let parsed = parser.parse(query);
            let results = FilePicker::fuzzy_search(
                &files,
-                query,
-                parsed,
+                &parsed,
+                None,
                FuzzySearchOptions {
                    max_threads: 4,
                    current_file: None,
                    project_path: None,
-                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
                    min_combo_count: 3,
                    pagination: PaginationArgs {
@@ -1,7 +1,8 @@
-use fff_core::file_picker::FilePicker;
-use fff_core::{FILE_PICKER, FuzzySearchOptions, PaginationArgs, QueryParser};
+use fff::file_picker::{FFFMode, FilePicker};
+use fff::{FuzzySearchOptions, PaginationArgs, QueryParser, SharedFrecency, SharedPicker};
 use std::env;
 use std::io::{self, Write};
+use std::sync::{Arc, RwLock};
 use std::thread;
 use std::time::{Duration, Instant};

@@ -77,24 +78,19 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    println!("Test directory: {}", base_path);
    println!();

-    // Initialize the file picker directly
+    // Create shared state
+    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
+    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));
+
+    // Initialize the file picker
    println!("📁 Initializing FilePicker...");
-    {
-        let mut file_picker_guard = FILE_PICKER.write().unwrap();
-        if file_picker_guard.is_none() {
-            println!("Creating new FilePicker for path: {}", base_path);
-            match FilePicker::new(base_path.clone()) {
-                Ok(picker) => {
-                    println!("FilePicker created successfully");
-                    *file_picker_guard = Some(picker);
-                }
-                Err(e) => {
-                    eprintln!("Failed to create FilePicker: {:?}", e);
-                    std::process::exit(1);
-                }
-            }
-        }
-    }
+    FilePicker::new_with_shared_state(
+        base_path.clone(),
+        false,
+        FFFMode::Neovim,
+        Arc::clone(&shared_picker),
+        Arc::clone(&shared_frecency),
+    )?;

    // Wait for initial scan to complete
    println!("⏳ Waiting for initial file scan to complete...");
@@ -102,8 +98,8 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut scan_completed = false;

    loop {
-        if let Ok(file_picker_guard) = FILE_PICKER.read()
-            && let Some(ref picker) = *file_picker_guard
+        if let Ok(guard) = shared_picker.read()
+            && let Some(ref picker) = *guard
        {
            if !picker.is_scan_active() {
                println!("Scan inactive, checking file count...");
@@ -128,10 +124,10 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    // If async scan didn't work, trigger a manual scan
    if !scan_completed {
        println!("Triggering manual rescan...");
-        if let Ok(mut file_picker_guard) = FILE_PICKER.write()
-            && let Some(ref mut picker) = *file_picker_guard
+        if let Ok(mut guard) = shared_picker.write()
+            && let Some(ref mut picker) = *guard
        {
-            match picker.trigger_rescan() {
+            match picker.trigger_rescan(&shared_frecency) {
                Ok(_) => println!("Manual rescan completed"),
                Err(e) => println!("Manual rescan failed: {:?}", e),
            }
@@ -139,8 +135,8 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    }

    let initial_file_count = {
-        let file_picker_guard = FILE_PICKER.read()?;
-        if let Some(ref picker) = *file_picker_guard {
+        let guard = shared_picker.read().unwrap();
+        if let Some(ref picker) = *guard {
            let files = picker.get_files();
            println!("Found {} files in picker", files.len());
            if !files.is_empty() {
@@ -151,7 +147,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
            }
            files.len()
        } else {
-            println!("No picker found in FILE_PICKER static!");
+            println!("No picker found!");
            0
        }
    };
@@ -199,18 +195,17 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
        let search_start = Instant::now();
        let parser = QueryParser::default();
        let (result_count, search_duration) = {
-            let file_picker_guard = FILE_PICKER.read().unwrap();
-            if let Some(ref picker) = *file_picker_guard {
+            let guard = shared_picker.read().unwrap();
+            if let Some(ref picker) = *guard {
                let parsed = parser.parse(query);
                let search_result = FilePicker::fuzzy_search(
                    picker.get_files(),
-                    query,
-                    parsed,
+                    &parsed,
+                    None,
                    FuzzySearchOptions {
                        max_threads,
                        current_file: None,
                        project_path: None,
-                        last_same_query_match: None,
                        combo_boost_score_multiplier: 100,
                        min_combo_count: 3,
                        pagination: PaginationArgs {
@@ -2,35 +2,16 @@
 #![allow(dead_code)]
 #![allow(clippy::enum_variant_names)]

-use fff_core::file_picker::FilePicker;
-use fff_core::git::format_git_status;
-use fff_core::{FILE_PICKER, FRECENCY, FuzzySearchOptions, PaginationArgs, QueryParser};
+use fff::file_picker::FilePicker;
+use fff::git::format_git_status;
+use fff::{FFFMode, FuzzySearchOptions, PaginationArgs, QueryParser, SharedFrecency, SharedPicker};
 use std::env;
 use std::io::{self, Write};
-use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, RwLock};
 use std::thread;
 use std::time::Duration;

-fn cleanup_global_state() {
-    // Clean up file picker
-    {
-        let mut file_picker = FILE_PICKER.write().unwrap();
-        if let Some(mut picker) = file_picker.take() {
-            picker.stop_background_monitor();
-            drop(picker);
-            println!("🧹 FilePicker cleaned up");
-        }
-    }
-
-    // Clean up frecency tracker
-    {
-        let mut frecency = FRECENCY.write().unwrap();
-        *frecency = None;
-        println!("🧹 Frecency tracker cleaned up");
-    }
-}
-
 fn main() -> Result<(), Box<dyn std::error::Error>> {
    let args: Vec<String> = env::args().collect();
    let base_path = if args.len() > 1 {
@@ -43,28 +24,39 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
    let running = Arc::new(AtomicBool::new(true));
    let r = running.clone();

+    // Create shared state
+    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
+    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));
+
+    // Clone for signal handler
+    let picker_for_cleanup = Arc::clone(&shared_picker);
    ctrlc::set_handler(move || {
        println!("\n🛑 Received interrupt signal, shutting down...");
-        cleanup_global_state();
+        if let Ok(mut guard) = picker_for_cleanup.write() {
+            if let Some(mut picker) = guard.take() {
+                picker.stop_background_monitor();
+                println!("🧹 FilePicker cleaned up");
+            }
+        }
        r.store(false, Ordering::SeqCst);
        std::process::exit(0);
    })?;

    let mut git_stats = std::collections::HashMap::new();
-    // Initialize the global file picker using lib.rs function
-    {
-        let mut file_picker = FILE_PICKER.write().unwrap();
-        if file_picker.is_some() {
-            eprintln!("❌ FilePicker already initialized");
-            std::process::exit(1);
-        }
-        *file_picker = Some(FilePicker::new(base_path.clone())?);
-    }

-    // Get initial file count from global state
+    // Initialize the file picker using shared state
+    FilePicker::new_with_shared_state(
+        base_path.clone(),
+        false,
+        FFFMode::default(),
+        Arc::clone(&shared_picker),
+        Arc::clone(&shared_frecency),
+    )?;
+
+    // Get initial file count from shared state
    let initial_count = {
-        let file_picker = FILE_PICKER.read().unwrap();
-        let files = file_picker.as_ref().unwrap().get_files();
+        let guard = shared_picker.read().unwrap();
+        let files = guard.as_ref().unwrap().get_files();
        println!("Initial file count: {}", files.len());

        if !files.is_empty() {
@@ -96,8 +88,8 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
        iteration += 1;

        let current_count = {
-            let file_picker = FILE_PICKER.read().unwrap();
-            file_picker.as_ref().unwrap().get_files().len()
+            let guard = shared_picker.read().unwrap();
+            guard.as_ref().unwrap().get_files().len()
        };

        if current_count != last_count {
@@ -111,13 +103,12 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                );

                // Show some recently added files
-                let file_picker = FILE_PICKER.read().unwrap();
-                let files = file_picker.as_ref().unwrap().get_files();
+                let guard = shared_picker.read().unwrap();
+                let files = guard.as_ref().unwrap().get_files();
                let newest_files = files.iter().rev().take(added.min(3));
                for file in newest_files {
                    println!("   ➕ {}", file.relative_path);
                }
-                drop(file_picker);
            } else {
                let removed = last_count - current_count;
                println!(
@@ -136,8 +127,8 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                timestamp, current_count
            );

-            let file_picker = FILE_PICKER.read().unwrap();
-            let current_files = file_picker.as_ref().unwrap().get_files();
+            let guard = shared_picker.read().unwrap();
+            let current_files = guard.as_ref().unwrap().get_files();

            git_stats.clear();
            for file in current_files {
@@ -156,19 +147,18 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {

        if iteration % 40 == 0 {
            let timestamp = chrono::Local::now().format("%H:%M:%S");
-            let file_picker = FILE_PICKER.read().unwrap();
-            let files = file_picker.as_ref().unwrap().get_files();
+            let guard = shared_picker.read().unwrap();
+            let files = guard.as_ref().unwrap().get_files();
            let parser = QueryParser::default();
            let parsed = parser.parse("rs");
            let search_results = FilePicker::fuzzy_search(
                files,
-                "rs",
-                parsed,
+                &parsed,
+                None,
                FuzzySearchOptions {
                    max_threads: 2,
                    current_file: None,
                    project_path: None,
-                    last_same_query_match: None,
                    combo_boost_score_multiplier: 100,
                    min_combo_count: 3,
                    pagination: PaginationArgs {
@@ -197,13 +187,16 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                    score.total
                );
            }
-            drop(file_picker);
        }

        io::stdout().flush().unwrap();
    }

    // Clean up before exit
-    cleanup_global_state();
+    if let Ok(mut guard) = shared_picker.write() {
+        if let Some(mut picker) = guard.take() {
+            picker.stop_background_monitor();
+        }
+    }
    Ok(())
 }
@@ -1,10 +1,10 @@
 //! Error handling for fff-nvim
 //!
-//! This module provides utilities for converting fff_core errors to mlua errors.
+//! This module provides utilities for converting fff errors to mlua errors.

-use fff_core::Error as CoreError;
+use fff::Error as CoreError;

-/// Convert a fff_core::Error to mlua::Error
+/// Convert a fff::Error to mlua::Error
 ///
 /// This function is used because we can't implement From<CoreError> for mlua::Error
 /// due to Rust's orphan rules (both types are foreign to this crate).
@@ -14,7 +14,7 @@ pub fn to_lua_error(err: CoreError) -> mlua::Error {
    mlua::Error::RuntimeError(string_value)
 }

-/// Extension trait for Result<T, fff_core::Error> to convert to LuaResult<T>
+/// Extension trait for Result<T, fff::Error> to convert to LuaResult<T>
 pub trait IntoLuaResult<T> {
    fn into_lua_result(self) -> mlua::Result<T>;
 }
@@ -0,0 +1,190 @@
+use mlua::prelude::*;
+use std::fmt::Write as _;
+use std::io::{Read, Seek, SeekFrom};
+
+// Byte category colors (matching hexyl's default theme)
+const COLOR_OFFSET: &str = "#888888";
+const COLOR_NULL: &str = "#555753";
+const COLOR_ASCII_PRINTABLE: &str = "#06989a";
+const COLOR_ASCII_WHITESPACE: &str = "#4e9a06";
+const COLOR_ASCII_OTHER: &str = "#4e9a06";
+const COLOR_NON_ASCII: &str = "#c4a000";
+
+fn byte_color(b: u8) -> &'static str {
+    match b {
+        0x00 => COLOR_NULL,
+        0x20 | 0x09 | 0x0a | 0x0d => COLOR_ASCII_WHITESPACE,
+        0x21..=0x7e => COLOR_ASCII_PRINTABLE,
+        0x01..=0x1f | 0x7f => COLOR_ASCII_OTHER,
+        _ => COLOR_NON_ASCII,
+    }
+}
+
+fn byte_char(b: u8) -> char {
+    match b {
+        0x20..=0x7e => b as char,
+        _ => '.',
+    }
+}
+
+const BYTES_PER_LINE: usize = 16;
+
+struct Span {
+    line: usize,
+    col_start: usize,
+    col_end: usize,
+    color: &'static str,
+}
+
+/// Push a span, merging with the previous one if same line and color.
+fn push_span(
+    spans: &mut Vec<Span>,
+    line: usize,
+    col_start: usize,
+    col_end: usize,
+    color: &'static str,
+) {
+    if let Some(last) = spans.last_mut() {
+        // Merge if same line, same color, and adjacent (allow small gaps for spaces between hex pairs)
+        if last.line == line && std::ptr::eq(last.color, color) && col_start <= last.col_end + 1 {
+            last.col_end = col_end;
+            return;
+        }
+    }
+    spans.push(Span {
+        line,
+        col_start,
+        col_end,
+        color,
+    });
+}
+
+/// Format raw bytes into hex dump lines with coalesced highlight spans.
+///
+/// Layout per line:
+/// ```text
+/// XXXXXXXX  HH HH HH HH HH HH HH HH  HH HH HH HH HH HH HH HH  CCCCCCCCCCCCCCCC
+/// ```
+fn format_hex_dump(raw_bytes: &[u8], base_offset: u64) -> (Vec<String>, Vec<Span>) {
+    let mut lines = Vec::new();
+    let mut spans = Vec::new();
+
+    for (chunk_idx, chunk) in raw_bytes.chunks(BYTES_PER_LINE).enumerate() {
+        let addr = base_offset + (chunk_idx * BYTES_PER_LINE) as u64;
+        let mut line = format!("{addr:08x}  ");
+
+        // Offset label highlight
+        push_span(&mut spans, chunk_idx, 0, 8, COLOR_OFFSET);
+
+        // Hex pairs with a gap after 8 bytes
+        for (i, &b) in chunk.iter().enumerate() {
+            if i == 8 {
+                line.push(' ');
+            }
+            let col = line.len();
+            push_span(&mut spans, chunk_idx, col, col + 2, byte_color(b));
+            write!(line, "{b:02x} ").unwrap();
+        }
+
+        // Pad if the last line is short
+        if chunk.len() < BYTES_PER_LINE {
+            let missing = BYTES_PER_LINE - chunk.len();
+            let mut pad = missing * 3;
+            if chunk.len() <= 8 {
+                pad += 1;
+            }
+            for _ in 0..pad {
+                line.push(' ');
+            }
+        }
+
+        // Separator before char panel
+        line.push(' ');
+
+        // Character panel — consecutive same-color chars merge automatically
+        let char_start = line.len();
+        for (i, &b) in chunk.iter().enumerate() {
+            let col = char_start + i;
+            push_span(&mut spans, chunk_idx, col, col + 1, byte_color(b));
+            line.push(byte_char(b));
+        }
+
+        lines.push(line);
+    }
+
+    (lines, spans)
+}
+
+/// Generate a hex dump for a binary file with paging support and highlight data.
+///
+/// Returns a Lua table:
+/// ```text
+/// {
+///   lines: string[],
+///   highlights: {line_0idx, col_start, col_end, color}[],
+///   has_more: bool,
+///   next_offset: number,
+/// }
+/// ```
+pub fn hex_dump(
+    lua: &Lua,
+    (file_path, offset, length): (String, Option<u64>, Option<u64>),
+) -> LuaResult<LuaValue> {
+    let offset = offset.unwrap_or(0);
+    let length = length.unwrap_or(4096);
+
+    let file = std::fs::File::open(&file_path)
+        .map_err(|e| LuaError::RuntimeError(format!("Failed to open file: {e}")))?;
+
+    let file_size = file
+        .metadata()
+        .map_err(|e| LuaError::RuntimeError(format!("Failed to get metadata: {e}")))?
+        .len();
+
+    let table = lua.create_table()?;
+
+    if offset >= file_size {
+        table.set("lines", lua.create_table()?)?;
+        table.set("highlights", lua.create_table()?)?;
+        table.set("has_more", false)?;
+        table.set("next_offset", file_size)?;
+        return Ok(LuaValue::Table(table));
+    }
+
+    let mut reader = std::io::BufReader::new(file);
+    reader
+        .seek(SeekFrom::Start(offset))
+        .map_err(|e| LuaError::RuntimeError(format!("Failed to seek: {e}")))?;
+    let mut raw_bytes = Vec::with_capacity(length as usize);
+    reader
+        .by_ref()
+        .take(length)
+        .read_to_end(&mut raw_bytes)
+        .map_err(|e| LuaError::RuntimeError(format!("Failed to read: {e}")))?;
+
+    let (plain_lines, hl_spans) = format_hex_dump(&raw_bytes, offset);
+
+    let lines_table = lua.create_table()?;
+    for (i, line) in plain_lines.iter().enumerate() {
+        lines_table.set(i + 1, line.as_str())?;
+    }
+    table.set("lines", lines_table)?;
+
+    let highlights_table = lua.create_table()?;
+    for (i, span) in hl_spans.iter().enumerate() {
+        let hl = lua.create_table()?;
+        hl.raw_set(1, span.line)?;
+        hl.raw_set(2, span.col_start)?;
+        hl.raw_set(3, span.col_end)?;
+        hl.raw_set(4, span.color)?;
+        highlights_table.raw_set(i + 1, hl)?;
+    }
+    table.set("highlights", highlights_table)?;
+
+    let bytes_read = raw_bytes.len() as u64;
+    let next_offset = offset + bytes_read;
+    table.set("has_more", next_offset < file_size)?;
+    table.set("next_offset", next_offset)?;
+
+    Ok(LuaValue::Table(table))
+}
@@ -1,17 +1,23 @@
 use crate::path_shortening::shorten_path_with_cache;
 use error::{IntoCoreError, IntoLuaResult};
-use fff_core::file_picker::FilePicker;
-use fff_core::frecency::FrecencyTracker;
-use fff_core::query_tracker::QueryTracker;
-use fff_core::{DbHealthChecker, Error, FuzzySearchOptions, PaginationArgs, QueryParser};
-use fff_core::{FILE_PICKER, FRECENCY, QUERY_TRACKER};
+use fff::file_picker::FilePicker;
+use fff::frecency::FrecencyTracker;
+use fff::path_utils::expand_tilde;
+use fff::query_tracker::QueryTracker;
+use fff::{
+    DbHealthChecker, Error, FFFMode, FileSearchConfig, FuzzySearchOptions, PaginationArgs,
+    QueryParser, Score, SearchResult, SharedFrecency, SharedPicker, SharedQueryTracker,
+};
 use mimalloc::MiMalloc;
 use mlua::prelude::*;
+use once_cell::sync::Lazy;
 use path_shortening::PathShortenStrategy;
 use std::path::{Path, PathBuf};
+use std::sync::{Arc, RwLock};
 use std::time::Duration;

 mod error;
+mod hex_dump;
 mod log;
 mod lua_types;
 mod path_shortening;
@@ -19,6 +25,12 @@ mod path_shortening;
 #[global_allocator]
 static GLOBAL: MiMalloc = MiMalloc;

+// the global state for neovim lives here for efficiency
+// lua ffi is pretty bad with the overhead of converting raw pointer into tables
+pub static FILE_PICKER: Lazy<SharedPicker> = Lazy::new(|| Arc::new(RwLock::new(None)));
+pub static FRECENCY: Lazy<SharedFrecency> = Lazy::new(|| Arc::new(RwLock::new(None)));
+pub static QUERY_TRACKER: Lazy<SharedQueryTracker> = Lazy::new(|| Arc::new(RwLock::new(None)));
+
 pub fn init_db(
    _: &Lua,
    (frecency_db_path, history_db_path, use_unsafe_no_lock): (String, String, bool),
@@ -33,6 +45,10 @@ pub fn init_db(
    *frecency =
        Some(FrecencyTracker::new(&frecency_db_path, use_unsafe_no_lock).into_lua_result()?);
    tracing::info!("Frecency database initialized at {}", frecency_db_path);
+    drop(frecency);
+
+    // Spawn background GC to purge stale entries without blocking startup
+    let _ = FrecencyTracker::spawn_gc(Arc::clone(&FRECENCY), frecency_db_path, use_unsafe_no_lock);

    let mut query_tracker = QUERY_TRACKER
        .write()
@@ -42,10 +58,10 @@ pub fn init_db(
        *query_tracker = None;
    }

-    let tracker = QueryTracker::new(&history_db_path, use_unsafe_no_lock).into_lua_result()?;
-    *query_tracker = Some(tracker);
-    tracing::info!("Query tracker database initialized at {}", history_db_path);
+    *query_tracker =
+        Some(QueryTracker::new(&history_db_path, use_unsafe_no_lock).into_lua_result()?);

+    tracing::info!("Query tracker database initialized at {}", history_db_path);
    Ok(true)
 }

@@ -68,31 +84,54 @@ pub fn destroy_query_db(_: &Lua, _: ()) -> LuaResult<bool> {
 }

 pub fn init_file_picker(_: &Lua, base_path: String) -> LuaResult<bool> {
-    let mut file_picker = FILE_PICKER
-        .write()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
-    if file_picker.is_some() {
-        return Ok(false);
+    {
+        let guard = FILE_PICKER
+            .read()
+            .with_lock_error(Error::AcquireItemLock)
+            .into_lua_result()?;
+        if guard.is_some() {
+            return Ok(false);
+        }
    }

-    let picker = FilePicker::new(base_path).into_lua_result()?;
-    *file_picker = Some(picker);
+    FilePicker::new_with_shared_state(
+        base_path,
+        false,
+        FFFMode::Neovim,
+        Arc::clone(&FILE_PICKER),
+        Arc::clone(&FRECENCY),
+    )
+    .into_lua_result()?;
+
    Ok(true)
 }

 fn reinit_file_picker_internal(path: &Path) -> Result<(), Error> {
-    let mut file_picker = FILE_PICKER
-        .write()
-        .with_lock_error(Error::AcquireItemLock)?;
-
-    // drop should clean it anyway but just to be extra sure
-    if let Some(mut picker) = file_picker.take() {
-        picker.stop_background_monitor();
+    // Cancel and stop the old picker under a single write lock to avoid
+    // a window where FILE_PICKER is None (which causes FilePickerMissing
+    // errors if the UI is searching concurrently).
+    {
+        let mut guard = FILE_PICKER
+            .write()
+            .with_lock_error(Error::AcquireItemLock)?;
+        if let Some(ref mut picker) = *guard {
+            // Signal cancellation BEFORE stopping — this tells any orphaned
+            // scan threads from this picker to discard their results.
+            picker.cancel();
+            picker.stop_background_monitor();
+        }
+        // Don't take() here — leave the old picker in place so searches
+        // still work until new_with_shared_state replaces it atomically.
    }

-    let new_picker = FilePicker::new(path.to_string_lossy().to_string())?;
-    *file_picker = Some(new_picker);
+    // Create new picker — this atomically replaces the old one via write lock
+    FilePicker::new_with_shared_state(
+        path.to_string_lossy().to_string(),
+        false,
+        FFFMode::Neovim,
+        Arc::clone(&FILE_PICKER),
+        Arc::clone(&FRECENCY),
+    )?;

    Ok(())
 }
@@ -106,10 +145,16 @@ pub fn restart_index_in_path(_: &Lua, new_path: String) -> LuaResult<()> {
        )));
    }

-    let canonical_path = fff_core::path_utils::canonicalize(&path).map_err(|e| {
+    let canonical_path = fff::path_utils::canonicalize(&path).map_err(|e| {
        LuaError::RuntimeError(format!("Failed to canonicalize path '{}': {}", new_path, e))
    })?;

+    if let Ok(Some(picker)) = FILE_PICKER.read().as_deref()
+        && picker.base_path() == canonical_path
+    {
+        return Ok(()); // same dir
+    }
+
    // Spawn a background thread to avoid blocking Lua/UI thread
    std::thread::spawn(move || {
        if let Err(e) = reinit_file_picker_internal(&canonical_path) {
@@ -136,7 +181,7 @@ pub fn scan_files(_: &Lua, _: ()) -> LuaResult<()> {
        .ok_or(Error::FilePickerMissing)
        .into_lua_result()?;

-    picker.trigger_rescan().into_lua_result()?;
+    picker.trigger_rescan(&FRECENCY).into_lua_result()?;
    ::tracing::info!("scan_files trigger_rescan completed");
    Ok(())
 }
@@ -173,26 +218,16 @@ pub fn fuzzy_search_files(
    let base_path = picker.base_path();
    let min_combo_count = min_combo_count.unwrap_or(3);

-    let last_same_query_entry = {
-        let query_tracker = QUERY_TRACKER
-            .read()
-            .with_lock_error(Error::AcquireFrecencyLock)
-            .into_lua_result()?;
+    let query_tracker_guard = QUERY_TRACKER
+        .read()
+        .with_lock_error(Error::AcquireFrecencyLock)
+        .into_lua_result()?;

-        if query_tracker.as_ref().is_none() {
-            tracing::warn!("Query tracker not initialized");
-        }
-
-        query_tracker
-            .as_ref()
-            .map(|tracker| tracker.get_last_query_entry(&query, base_path, min_combo_count))
-            .transpose()
-            .into_lua_result()?
-            .flatten()
-    };
+    if query_tracker_guard.as_ref().is_none() {
+        tracing::warn!("Query tracker not initialized");
+    }

    tracing::debug!(
-        ?last_same_query_entry,
        ?base_path,
        ?query,
        ?min_combo_count,
@@ -201,19 +236,18 @@ pub fn fuzzy_search_files(
        "Fuzzy search parameters"
    );

-    // Parse the query once at the API boundary
-    let parser = QueryParser::default();
+    let parser = QueryParser::new(FileSearchConfig);
    let parsed = parser.parse(&query);

+    let files = picker.get_files();
    let results = FilePicker::fuzzy_search(
-        picker.get_files(),
-        &query,
-        parsed,
+        files,
+        &parsed,
+        query_tracker_guard.as_ref(),
        FuzzySearchOptions {
            max_threads,
            current_file: current_file.as_deref(),
            project_path: Some(picker.base_path()),
-            last_same_query_match: last_same_query_entry.as_ref(),
            combo_boost_score_multiplier,
            min_combo_count,
            pagination: PaginationArgs {
@@ -223,6 +257,34 @@ pub fn fuzzy_search_files(
        },
    );

+    if results.items.is_empty() && query.contains(std::path::MAIN_SEPARATOR) {
+        let pure_query = match &parsed.fuzzy_query {
+            fff_query_parser::FuzzyQuery::Text(t) => t.trim(),
+            _ => query.trim(),
+        };
+
+        let path = expand_tilde(pure_query);
+        if path.is_absolute() && path.is_file() {
+            if let Ok(idx) = files.binary_search_by(|f| f.path.as_path().cmp(&path)) {
+                let found = SearchResult {
+                    items: vec![&files[idx]],
+                    scores: vec![Score {
+                        exact_match: true,
+                        match_type: "path",
+                        ..Default::default()
+                    }],
+                    total_matched: 1,
+                    total_files: results.total_files,
+                    location: parsed.location,
+                };
+
+                return lua_types::SearchResultLua::from(found).into_lua(lua);
+            }
+
+            return build_file_path_fallback(lua, &path, results.total_files);
+        }
+    }
+
    lua_types::SearchResultLua::from(results).into_lua(lua)
 }

@@ -257,15 +319,15 @@ pub fn live_grep(
        return Err(error::to_lua_error(Error::FilePickerMissing));
    };

-    let parsed = fff_core::grep::parse_grep_query(&query);
+    let parsed = fff::grep::parse_grep_query(&query);

    let mode = match grep_mode.as_deref() {
-        Some("regex") => fff_core::GrepMode::Regex,
-        Some("fuzzy") => fff_core::GrepMode::Fuzzy,
-        _ => fff_core::GrepMode::PlainText, // "plain" or nil or unknown
+        Some("regex") => fff::GrepMode::Regex,
+        Some("fuzzy") => fff::GrepMode::Fuzzy,
+        _ => fff::GrepMode::PlainText, // "plain" or nil or unknown
    };

-    let options = fff_core::GrepSearchOptions {
+    let options = fff::GrepSearchOptions {
        max_file_size: max_file_size.unwrap_or(10 * 1024 * 1024),
        max_matches_per_file: max_matches_per_file.unwrap_or(200),
        smart_case: smart_case.unwrap_or(true),
@@ -273,13 +335,66 @@ pub fn live_grep(
        page_limit: page_size.unwrap_or(50),
        mode,
        time_budget_ms: time_budget_ms.unwrap_or(0),
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
    };

-    let result = fff_core::grep::grep_search(picker.get_files(), &query, parsed, &options);
+    let result = fff::grep::grep_search(picker.get_files(), &parsed, &options);

    lua_types::GrepResultLua::from(result).into_lua(lua)
 }

+/// Build a file-picker result for an absolute path that exists on disk but
+/// isn't in the picker index (e.g. file from a different project).
+fn build_file_path_fallback(lua: &Lua, path: &Path, total_files: usize) -> LuaResult<LuaValue> {
+    let table = lua.create_table()?;
+
+    let name = path
+        .file_name()
+        .map(|n| n.to_string_lossy().to_string())
+        .unwrap_or_default();
+    let path_str = path.to_string_lossy().to_string();
+
+    let item = lua.create_table()?;
+    item.set("path", path_str.as_str())?;
+    item.set("relative_path", path_str.as_str())?;
+    item.set("name", name.as_str())?;
+    item.set("size", path.metadata().map(|m| m.len()).unwrap_or(0))?;
+    item.set("modified", 0u64)?;
+    item.set("access_frecency_score", 0i64)?;
+    item.set("modification_frecency_score", 0i64)?;
+    item.set("total_frecency_score", 0i64)?;
+    item.set("git_status", "")?;
+    item.set("is_binary", false)?;
+
+    let items_table = lua.create_table()?;
+    items_table.set(1, item)?;
+    table.set("items", items_table)?;
+
+    let score = lua.create_table()?;
+    score.set("total", 0)?;
+    score.set("base_score", 0)?;
+    score.set("filename_bonus", 0)?;
+    score.set("special_filename_bonus", 0)?;
+    score.set("frecency_boost", 0)?;
+    score.set("git_status_boost", 0)?;
+    score.set("distance_penalty", 0)?;
+    score.set("current_file_penalty", 0)?;
+    score.set("combo_match_boost", 0)?;
+    score.set("exact_match", true)?;
+    score.set("match_type", "path")?;
+
+    let scores_table = lua.create_table()?;
+    scores_table.set(1, score)?;
+    table.set("scores", scores_table)?;
+
+    table.set("total_matched", 1)?;
+    table.set("total_files", total_files)?;
+
+    Ok(LuaValue::Table(table))
+}
+
 pub fn track_access(_: &Lua, file_path: String) -> LuaResult<bool> {
    let file_path = PathBuf::from(&file_path);

@@ -362,7 +477,7 @@ pub fn get_git_root(_: &Lua, _: ()) -> LuaResult<Option<String>> {
 }

 pub fn refresh_git_status(_: &Lua, _: ()) -> LuaResult<usize> {
-    FilePicker::refresh_git_status_global().into_lua_result()
+    FilePicker::refresh_git_status(&FILE_PICKER, &FRECENCY).into_lua_result()
 }

 pub fn update_single_file_frecency(_: &Lua, file_path: String) -> LuaResult<bool> {
@@ -435,7 +550,7 @@ pub fn track_query_completion(_: &Lua, (query, file_path): (String, String)) ->
    };

    // Canonicalize the file path before spawning thread
-    let file_path = match fff_core::path_utils::canonicalize(&file_path) {
+    let file_path = match fff::path_utils::canonicalize(&file_path) {
        Ok(path) => path,
        Err(e) => {
            tracing::warn!(?file_path, error = ?e, "Failed to canonicalize file path for tracking");
@@ -444,8 +559,9 @@ pub fn track_query_completion(_: &Lua, (query, file_path): (String, String)) ->
    };

    // Spawn background thread to do the actual tracking (expensive DB write)
+    let query_tracker = Arc::clone(&QUERY_TRACKER);
    std::thread::spawn(move || {
-        if let Ok(Some(tracker)) = QUERY_TRACKER.write().as_deref_mut()
+        if let Ok(Some(tracker)) = query_tracker.write().as_deref_mut()
            && let Err(e) = tracker.track_query_completion(&query, &project_path, &file_path)
        {
            tracing::error!(
@@ -497,8 +613,9 @@ pub fn track_grep_query(_: &Lua, query: String) -> LuaResult<bool> {
        picker.base_path().to_path_buf()
    };

+    let query_tracker = Arc::clone(&QUERY_TRACKER);
    std::thread::spawn(move || {
-        if let Ok(Some(tracker)) = QUERY_TRACKER.write().as_deref_mut()
+        if let Ok(Some(tracker)) = query_tracker.write().as_deref_mut()
            && let Err(e) = tracker.track_grep_query(&query, &project_path)
        {
            tracing::error!(
@@ -538,21 +655,28 @@ pub fn get_historical_grep_query(_: &Lua, offset: usize) -> LuaResult<Option<Str
 }

 pub fn wait_for_initial_scan(_: &Lua, timeout_ms: Option<u64>) -> LuaResult<bool> {
-    let file_picker = FILE_PICKER
-        .read()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
-    let picker = file_picker
-        .as_ref()
-        .ok_or(Error::FilePickerMissing)
-        .into_lua_result()?;
+    // Extract the scan signal Arc WITHOUT holding the read lock, so the
+    // scan thread can acquire the write lock to store its results.
+    // Holding a read lock while polling would deadlock: the scan thread
+    // needs a write lock to finish, but can't acquire it while we hold the read lock.
+    let scan_signal = {
+        let file_picker = FILE_PICKER
+            .read()
+            .with_lock_error(Error::AcquireItemLock)
+            .into_lua_result()?;
+        let picker = file_picker
+            .as_ref()
+            .ok_or(Error::FilePickerMissing)
+            .into_lua_result()?;
+        picker.scan_signal()
+    }; // read lock released here

    let timeout_ms = timeout_ms.unwrap_or(500);
    let timeout_duration = Duration::from_millis(timeout_ms);
    let start_time = std::time::Instant::now();
    let mut sleep_duration = Duration::from_millis(1);

-    while picker.is_scan_active() {
+    while scan_signal.load(std::sync::atomic::Ordering::Relaxed) {
        if start_time.elapsed() >= timeout_duration {
            ::tracing::warn!("wait_for_initial_scan timed out after {}ms", timeout_ms);
            return Ok(false);
@@ -772,6 +896,7 @@ fn create_exports(lua: &Lua) -> LuaResult<LuaTable> {
    )?;
    exports.set("health_check", lua.create_function(health_check)?)?;
    exports.set("shorten_path", lua.create_function(shorten_path)?)?;
+    exports.set("hex_dump", lua.create_function(hex_dump::hex_dump)?)?;

    Ok(exports)
 }
@@ -1,152 +1,3 @@
-use std::io;
-use std::path::Path;
-use tracing_appender::non_blocking;
-use tracing_subscriber::fmt::format::FmtSpan;
-use tracing_subscriber::{EnvFilter, fmt, prelude::*};
+//! Logging setup for fff-nvim — delegates to the shared fff-core::log utilities.

-static TRACING_INITIALIZED: std::sync::OnceLock<tracing_appender::non_blocking::WorkerGuard> =
-    std::sync::OnceLock::new();
-
-static PANIC_HOOK_INSTALLED: std::sync::OnceLock<()> = std::sync::OnceLock::new();
-
-/// Install panic hook that writes to both stderr and a fallback file
-/// This is called separately from init_tracing to ensure panics are always logged
-pub fn install_panic_hook() {
-    PANIC_HOOK_INSTALLED.get_or_init(|| {
-        let default_panic = std::panic::take_hook();
-
-        std::panic::set_hook(Box::new(move |panic_info| {
-            let payload = panic_info.payload();
-            let message = if let Some(s) = payload.downcast_ref::<&str>() {
-                s.to_string()
-            } else if let Some(s) = payload.downcast_ref::<String>() {
-                s.clone()
-            } else {
-                "Unknown panic payload".to_string()
-            };
-
-            let location = if let Some(location) = panic_info.location() {
-                format!(
-                    "{}:{}:{}",
-                    location.file(),
-                    location.line(),
-                    location.column()
-                )
-            } else {
-                "unknown location".to_string()
-            };
-
-            // Always log to tracing (if initialized)
-            tracing::error!(
-                panic.message = %message,
-                panic.location = %location,
-                "PANIC occurred in FFF.nvim"
-            );
-
-            // Always print to stderr
-            eprintln!("=== FFF.nvim PANIC ===");
-            eprintln!("Message: {}", message);
-            eprintln!("Location: {}", location);
-            eprintln!("======================");
-
-            // Try to write to fallback panic log file
-            if let Some(cache_dir) = dirs::cache_dir() {
-                let panic_log = cache_dir.join("fff_nvim_panic.log");
-                let timestamp = std::time::SystemTime::now()
-                    .duration_since(std::time::UNIX_EPOCH)
-                    .map(|d| d.as_secs())
-                    .unwrap_or(0);
-
-                let panic_entry = format!(
-                    "\n[{}] PANIC at {}\nMessage: {}\n",
-                    timestamp, location, message
-                );
-
-                let _ = std::fs::OpenOptions::new()
-                    .create(true)
-                    .append(true)
-                    .open(&panic_log)
-                    .and_then(|mut f| {
-                        use std::io::Write;
-                        f.write_all(panic_entry.as_bytes())
-                    });
-
-                eprintln!("Panic logged to: {}", panic_log.display());
-            }
-
-            default_panic(panic_info);
-        }));
-    });
-}
-
-/// Initialize tracing with single log file
-///
-/// # Arguments
-/// * `log_file_path` - Full path to the log file
-/// * `log_level` - Log level (trace, debug, info, warn, error)
-///
-/// # Returns
-/// * `Result<String, io::Error>` - Full path to the log file on success
-pub fn init_tracing(log_file_path: &str, log_level: Option<&str>) -> Result<String, io::Error> {
-    // Install panic hook first (does nothing if already installed)
-    install_panic_hook();
-
-    let log_path = Path::new(log_file_path);
-    if let Some(parent) = log_path.parent() {
-        std::fs::create_dir_all(parent)?;
-    }
-
-    let file_appender = std::fs::OpenOptions::new()
-        .create(true)
-        .write(true)
-        .truncate(true) // creates a new file on every setup
-        .open(log_path)?;
-
-    let level = match log_level
-        .as_ref()
-        .map(|s| s.trim().to_lowercase())
-        .as_deref()
-    {
-        Some("trace") => tracing::Level::TRACE,
-        Some("debug") => tracing::Level::DEBUG,
-        Some("info") => tracing::Level::INFO,
-        Some("warn") => tracing::Level::WARN,
-        Some("error") => tracing::Level::ERROR,
-        _ => tracing::Level::INFO,
-    };
-
-    TRACING_INITIALIZED.get_or_init(|| {
-        let (non_blocking_appender, guard) = non_blocking(file_appender);
-
-        let subscriber = tracing_subscriber::registry()
-            .with(
-                fmt::layer()
-                    .with_writer(non_blocking_appender)
-                    .with_target(true)
-                    .with_thread_ids(false)
-                    .with_thread_names(false)
-                    .with_file(true)
-                    .with_line_number(true)
-                    .with_ansi(false)
-                    .with_span_events(FmtSpan::NEW | FmtSpan::CLOSE),
-            )
-            .with(
-                EnvFilter::builder()
-                    .with_default_directive(level.into())
-                    .from_env_lossy(),
-            );
-
-        if let Err(e) = tracing::subscriber::set_global_default(subscriber) {
-            eprintln!("Failed to set tracing subscriber: {}", e);
-        } else {
-            tracing::info!(
-                "FFF.nvim tracing initialized with log file: {}",
-                log_path.display()
-            );
-        }
-
-        guard
-    });
-
-    Ok(log_file_path.to_string())
-}
+pub use fff::log::{init_tracing, install_panic_hook};
@@ -1,12 +1,7 @@
-//! Lua type conversions for fff-core types
-//!
-//! This module provides IntoLua implementations for core types.
-
-use fff_core::git::format_git_status;
-use fff_core::{FileItem, GrepResult, Location, Score, SearchResult};
+use fff::git::format_git_status;
+use fff::{FileItem, GrepResult, Location, Score, SearchResult};
 use mlua::prelude::*;

-/// Wrapper for SearchResult that implements IntoLua
 pub struct SearchResultLua<'a> {
    inner: SearchResult<'a>,
 }
@@ -17,7 +12,6 @@ impl<'a> From<SearchResult<'a>> for SearchResultLua<'a> {
    }
 }

-/// Wrapper for GrepResult that implements IntoLua
 pub struct GrepResultLua<'a> {
    inner: GrepResult<'a>,
 }
@@ -167,7 +161,7 @@ impl IntoLua for GrepResultLua<'_> {
        }
        table.set("items", items_table)?;

-        table.set("total_matched", self.inner.total_match_count)?;
+        table.set("total_matched", self.inner.matches.len())?;
        table.set("total_files_searched", self.inner.total_files_searched)?;
        table.set("total_files", self.inner.total_files)?;
        table.set("filtered_file_count", self.inner.filtered_file_count)?;
@@ -34,7 +34,7 @@ impl PathCache {
        }
    }

-    #[tracing::instrument(skip(self), fields(path = %path.display(), max_size))]
+    #[tracing::instrument(skip(self), fields(path = %path.display(), max_size), level = tracing::Level::TRACE)]
    fn get(&self, path: &Path, max_size: usize) -> Option<&str> {
        self.map.get(path).and_then(|entry| {
            // Only return cached value if max_size matches
@@ -85,7 +85,7 @@ pub fn shorten_path_with_cache(
            .read()
            .map_err(|_| "Failed to acquire path cache lock".to_string())?;
        if let Some(cached) = cache.get(path, max_size) {
-            tracing::debug!("Cache hit for path '{}'", path.display());
+            tracing::trace!("Cache hit for path '{}'", path.display());
            return Ok(cached.to_string());
        }
    }
@@ -1,14 +1,20 @@
 [package]
 name = "fff-query-parser"
-version = "0.1.0"
+version = "0.2.4"
 edition = "2024"
+description = "Query parser for fff file finder - includes specific syntax for various constraints like globs, extensions, regex etc"
+license = "MIT"
+authors = ["Dmitriy Kovalenko <dmtr.kovalenko@outlok.com>"]

 [lib]
 path = "src/lib.rs"

+[features]
+default = []
+zlob = ["dep:zlob"]
+
 [dependencies]
-smallvec = { workspace = true }
-zlob = { version = "1.2.8" }
+zlob = { workspace = true, optional = true }

 [dev-dependencies]
 criterion = { version = "0.5", features = ["html_reports"] }
@@ -98,7 +98,7 @@ fn bench_parse_various_lengths(c: &mut Criterion) {
 }

 fn bench_config_comparison(c: &mut Criterion) {
-    let file_picker = QueryParser::new(FilePickerConfig);
+    let file_picker = QueryParser::new(FileSearchConfig);
    let grep = QueryParser::new(GrepConfig);

    let query = "src name *.rs !test";
@@ -1,5 +1,46 @@
 use crate::constraints::Constraint;
-use zlob::{ZlobFlags, has_wildcards};
+use crate::glob_detect::has_wildcards;
+
+/// Check if a token looks like a filename or file path for use as a `FilePath` constraint.
+///
+/// A token is a filename/path if ALL of:
+/// - Does NOT end with `/` (that's a directory/PathSegment)
+/// - Does NOT contain wildcards (`*`, `?`, `{`, `[`) — those are globs
+/// - Last component (after final `/`) contains `.` with a valid-looking extension
+///   (1–10 alphanumeric chars starting with a letter, e.g. `rs`, `json`, `tsx`)
+///
+/// This covers both bare filenames (`score.rs`) and path-prefixed ones (`src/main.rs`).
+#[inline]
+fn is_filename_constraint_token(token: &str) -> bool {
+    let bytes = token.as_bytes();
+
+    // Must NOT end with / (that's a PathSegment)
+    if bytes.last() == Some(&b'/') {
+        return false;
+    }
+
+    // Must NOT contain wildcards (those are globs)
+    if has_wildcards(token) {
+        return false;
+    }
+
+    // Get the filename component (after last /)
+    let filename = token.rsplit('/').next().unwrap_or(token);
+
+    // Extension must exist and look like a real file extension:
+    // starts with an ASCII letter (rejects version numbers like "v2.0"),
+    // followed by alphanumeric chars, max 10 chars total.
+    match filename.rfind('.') {
+        Some(dot_pos) => {
+            let ext = &filename[dot_pos + 1..];
+            !ext.is_empty()
+                && ext.len() <= 10
+                && ext.as_bytes()[0].is_ascii_alphabetic()
+                && ext.bytes().all(|b| b.is_ascii_alphanumeric())
+        }
+        None => false,
+    }
+}

 /// Parser configuration trait - allows different picker types to customize parsing
 pub trait ParserConfig {
@@ -32,6 +73,13 @@ pub trait ParserConfig {
        true
    }

+    /// Should parse location suffixes (e.g., file:12, file:12:4)
+    /// Disabled for grep modes where colon-number patterns like localhost:8080
+    /// are search text, not file locations.
+    fn enable_location(&self) -> bool {
+        true
+    }
+
    /// Determine whether a token should be treated as a glob constraint.
    ///
    /// The default implementation delegates to `zlob::has_wildcards` with
@@ -40,7 +88,7 @@ pub trait ParserConfig {
    /// Override this in configs where some wildcard characters are common
    /// in search text (e.g. grep mode where `?` and `[` appear in code).
    fn is_glob_pattern(&self, token: &str) -> bool {
-        has_wildcards(token, ZlobFlags::RECOMMENDED)
+        has_wildcards(token)
    }

    /// Custom constraint parsers for picker-specific needs
@@ -51,10 +99,19 @@ pub trait ParserConfig {

 /// Default configuration for file picker - all features enabled
 #[derive(Debug, Clone, Copy, Default)]
-pub struct FilePickerConfig;
+pub struct FileSearchConfig;

-impl ParserConfig for FilePickerConfig {
-    // All defaults enabled
+impl ParserConfig for FileSearchConfig {
+    /// Detect bare filenames (`score.rs`) and path-prefixed filenames (`src/main.rs`)
+    /// as `FilePath` constraints so that multi-token queries like `score.rs file_picker`
+    /// filter by filename first, then fuzzy-match the remaining text against the path.
+    fn parse_custom<'a>(&self, token: &'a str) -> Option<Constraint<'a>> {
+        if is_filename_constraint_token(token) {
+            Some(Constraint::FilePath(token))
+        } else {
+            None
+        }
+    }
 }

 /// Configuration for full-text search (grep) - file constraints enabled for
@@ -76,6 +133,10 @@ impl ParserConfig for GrepConfig {
        false
    }

+    fn enable_location(&self) -> bool {
+        false
+    }
+
    /// Only recognise globs that are clearly directory/path oriented.
    ///
    /// Characters like `?`, `[`, and bare `*` (without `/`) are extremely
@@ -87,7 +148,7 @@ impl ParserConfig for GrepConfig {
    /// - Contains `{…}` → brace expansion (e.g. `{src,lib}`)
    fn is_glob_pattern(&self, token: &str) -> bool {
        // Must contain at least one glob wildcard character
-        if !has_wildcards(token, ZlobFlags::RECOMMENDED) {
+        if !has_wildcards(token) {
            return false;
        }

@@ -98,12 +159,76 @@ impl ParserConfig for GrepConfig {
            return true;
        }

-        // Brace expansion → useful for directory alternatives
-        if bytes.contains(&b'{') && bytes.contains(&b'}') {
-            return true;
+        // Brace expansion → useful for directory alternatives.
+        // Require a comma between `{` and `}` AND at least one letter to
+        // distinguish real glob expansions like `{src,lib}` or `*.{ts,tsx}`
+        // from code patterns like `format!("{}")` and regex quantifiers `{2,3}`.
+        if let Some(open) = bytes.iter().position(|&b| b == b'{')
+            && let Some(close) = bytes.iter().rposition(|&b| b == b'}')
+        {
+            let inner = &bytes[open + 1..close];
+            if inner.contains(&b',') && inner.iter().any(|b| b.is_ascii_alphabetic()) {
+                return true;
+            }
        }

        // Everything else (?, [, bare * without /) → treat as literal text
        false
    }
 }
+
+/// Configuration for AI-mode grep — extends `GrepConfig` behavior with
+/// automatic file-path constraint detection.
+///
+/// Bare filenames with valid extensions (`schema.rs`) and path-prefixed
+/// filenames (`libswscale/input.c`) are detected as `FilePath` constraints
+/// so the search is scoped to matching files. The caller validates the
+/// constraint against the index and drops it if no files match (fallback).
+#[derive(Debug, Clone, Copy, Default)]
+pub struct AiGrepConfig;
+
+impl ParserConfig for AiGrepConfig {
+    fn enable_path_segments(&self) -> bool {
+        true
+    }
+
+    fn enable_git_status(&self) -> bool {
+        false
+    }
+
+    fn enable_location(&self) -> bool {
+        false
+    }
+
+    fn is_glob_pattern(&self, token: &str) -> bool {
+        // First check GrepConfig's strict rules (path globs, brace expansion)
+        if GrepConfig.is_glob_pattern(token) {
+            return true;
+        }
+
+        // AI agents use `*text*` to scope file searches (e.g. `*quote* TODO`).
+        // Recognise tokens that start AND end with `*` with non-empty text
+        // between them as glob constraints. Bare `*` or `**` are excluded.
+        if !has_wildcards(token) {
+            return false;
+        }
+        let bytes = token.as_bytes();
+        if bytes.len() >= 3
+            && bytes[0] == b'*'
+            && bytes[bytes.len() - 1] == b'*'
+            && bytes[1..bytes.len() - 1].iter().all(|&b| b != b'*')
+        {
+            return true;
+        }
+
+        false
+    }
+
+    fn parse_custom<'a>(&self, token: &'a str) -> Option<Constraint<'a>> {
+        if is_filename_constraint_token(token) {
+            Some(Constraint::FilePath(token))
+        } else {
+            None
+        }
+    }
+}
@@ -1,5 +1,3 @@
-use smallvec::SmallVec;
-
 /// Constraint types that can be extracted from a query
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum Constraint<'a> {
@@ -22,6 +20,10 @@ pub enum Constraint<'a> {
    /// Path constraint: /src/ -> PathSegment("src")
    PathSegment(&'a str),

+    /// File path constraint (AI mode): "libswscale/input.c" → FilePath("libswscale/input.c")
+    /// Matches files whose relative path ends with this suffix at a `/` boundary.
+    FilePath(&'a str),
+
    /// File type constraint: type:rust -> FileType("rust")
    FileType(&'a str),

@@ -41,5 +43,5 @@ pub enum GitStatusFilter {
    Unmodified,
 }

-/// Stack-allocated buffer for text parts (up to 16 parts without heap allocation)
-pub(crate) type TextPartsBuffer<'a> = SmallVec<[&'a str; 16]>;
+/// Buffer for text parts during query parsing.
+pub(crate) type TextPartsBuffer<'a> = Vec<&'a str>;
@@ -0,0 +1,18 @@
+//! Glob wildcard detection — delegates to zlob when available, pure-Rust fallback otherwise.
+//!
+//! All call sites use a single function: `has_wildcards(text) -> bool`.
+//! When the `zlob` feature is enabled this calls `zlob::has_wildcards` with
+//! `ZlobFlags::RECOMMENDED`; without it we check for the same set of wildcard
+//! characters (`*`, `?`, `[`, `{`) in pure Rust.
+
+#[cfg(feature = "zlob")]
+#[inline]
+pub fn has_wildcards(s: &str) -> bool {
+    zlob::has_wildcards(s, zlob::ZlobFlags::RECOMMENDED)
+}
+
+#[cfg(not(feature = "zlob"))]
+#[inline]
+pub fn has_wildcards(s: &str) -> bool {
+    s.bytes().any(|b| matches!(b, b'*' | b'?' | b'[' | b'{'))
+}
@@ -1,8 +1,7 @@
-//! Fast, zero-allocation query parser for file search
+//! Fast query parser for file search
 //!
 //! This parser takes a search query and extracts structured constraints
 //! while preserving text for fuzzy matching. Designed for maximum performance:
-//! - Zero allocations for queries with ≤8 constraints (SmallVec)
 //! - Single-pass parsing with minimal branching
 //! - Stack-allocated string buffers
 //!
@@ -13,12 +12,13 @@
 //!
 //! let parser = QueryParser::default();
 //!
-//! // Single-token queries return None (no parsing needed)
+//! // Single-token queries return FFFQuery with Text fuzzy query and no constraints
 //! let result = parser.parse("hello");
-//! assert!(result.is_none());
+//! assert!(result.constraints.is_empty());
+//! assert_eq!(result.fuzzy_query, FuzzyQuery::Text("hello"));
 //!
 //! // Multi-token queries are parsed
-//! let result = parser.parse("name *.rs").expect("Should parse");
+//! let result = parser.parse("name *.rs");
 //! match &result.fuzzy_query {
 //!     FuzzyQuery::Text(text) => assert_eq!(*text, "name"),
 //!     _ => panic!("Expected text"),
@@ -26,11 +26,11 @@
 //! assert!(matches!(result.constraints[0], Constraint::Extension("rs")));
 //!
 //! // Parse glob pattern with text
-//! let result = parser.parse("**/*.rs foo").expect("Should parse");
+//! let result = parser.parse("**/*.rs foo");
 //! assert!(matches!(result.constraints[0], Constraint::Glob("**/*.rs")));
 //!
 //! // Parse negation
-//! let result = parser.parse("!*.rs foo").expect("Should parse");
+//! let result = parser.parse("!*.rs foo");
 //! match &result.constraints[0] {
 //!     Constraint::Not(inner) => {
 //!         assert!(matches!(inner.as_ref(), Constraint::Extension("rs")));
@@ -41,19 +41,16 @@

 mod config;
 mod constraints;
+pub mod glob_detect;
 pub mod location;
 mod parser;

-pub use config::{FilePickerConfig, GrepConfig, ParserConfig};
+pub use config::{AiGrepConfig, FileSearchConfig, GrepConfig, ParserConfig};
 pub use constraints::{Constraint, GitStatusFilter};
 pub use location::Location;
 pub use parser::{FFFQuery, FuzzyQuery, QueryParser};

-// Re-export SmallVec for convenience
-pub use smallvec::SmallVec;
-
-/// Type alias for constraint vector - stack-allocated for ≤8 constraints
-pub type ConstraintVec<'a> = SmallVec<[Constraint<'a>; 8]>;
+pub type ConstraintVec<'a> = Vec<Constraint<'a>>;

 #[cfg(test)]
 mod tests {
@@ -63,32 +60,30 @@ mod tests {
    fn test_empty_query() {
        let parser = QueryParser::default();
        let result = parser.parse("");
-        // Empty query returns None (single-token behavior)
-        assert!(result.is_none());
+        assert!(result.constraints.is_empty());
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Empty);
    }

    #[test]
    fn test_whitespace_only() {
        let parser = QueryParser::default();
        let result = parser.parse("   ");
-        // Whitespace-only returns None
-        assert!(result.is_none());
+        assert!(result.constraints.is_empty());
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Empty);
    }

    #[test]
    fn test_single_token() {
        let parser = QueryParser::default();
        let result = parser.parse("hello");
-        // Single token returns None (no parsing needed)
-        assert!(result.is_none());
+        assert!(result.constraints.is_empty());
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("hello"));
    }

    #[test]
    fn test_simple_text() {
        let parser = QueryParser::default();
-        let result = parser
-            .parse("hello world")
-            .expect("Should parse multi-token");
+        let result = parser.parse("hello world");

        match &result.fuzzy_query {
            FuzzyQuery::Parts(parts) => {
@@ -106,9 +101,7 @@ mod tests {
    fn test_extension_only() {
        let parser = QueryParser::default();
        // Single constraint token - returns Some so constraint can be applied
-        let result = parser
-            .parse("*.rs")
-            .expect("Should parse single constraint");
+        let result = parser.parse("*.rs");
        assert!(matches!(result.fuzzy_query, FuzzyQuery::Empty));
        assert_eq!(result.constraints.len(), 1);
        assert!(matches!(result.constraints[0], Constraint::Extension("rs")));
@@ -117,9 +110,7 @@ mod tests {
    #[test]
    fn test_glob_pattern() {
        let parser = QueryParser::default();
-        let result = parser
-            .parse("**/*.rs foo")
-            .expect("Should parse multi-token");
+        let result = parser.parse("**/*.rs foo");
        assert_eq!(result.constraints.len(), 1);
        // Glob patterns with ** are treated as globs, not extensions
        match &result.constraints[0] {
@@ -131,7 +122,7 @@ mod tests {
    #[test]
    fn test_negation_pattern() {
        let parser = QueryParser::default();
-        let result = parser.parse("!test foo").expect("Should parse multi-token");
+        let result = parser.parse("!test foo");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -144,7 +135,7 @@ mod tests {
    #[test]
    fn test_path_segment() {
        let parser = QueryParser::default();
-        let result = parser.parse("/src/ foo").expect("Should parse multi-token");
+        let result = parser.parse("/src/ foo");
        assert_eq!(result.constraints.len(), 1);
        assert!(matches!(
            result.constraints[0],
@@ -155,9 +146,7 @@ mod tests {
    #[test]
    fn test_git_status() {
        let parser = QueryParser::default();
-        let result = parser
-            .parse("status:modified foo")
-            .expect("Should parse multi-token");
+        let result = parser.parse("status:modified foo");
        assert_eq!(result.constraints.len(), 1);
        assert!(matches!(
            result.constraints[0],
@@ -168,9 +157,7 @@ mod tests {
    #[test]
    fn test_file_type() {
        let parser = QueryParser::default();
-        let result = parser
-            .parse("type:rust foo")
-            .expect("Should parse multi-token");
+        let result = parser.parse("type:rust foo");
        assert_eq!(result.constraints.len(), 1);
        assert!(matches!(
            result.constraints[0],
@@ -181,9 +168,7 @@ mod tests {
    #[test]
    fn test_complex_query() {
        let parser = QueryParser::default();
-        let result = parser
-            .parse("src name *.rs !test /lib/ status:modified")
-            .expect("Should parse");
+        let result = parser.parse("src name *.rs !test /lib/ status:modified");

        // Verify we have fuzzy text
        match &result.fuzzy_query {
@@ -223,21 +208,16 @@ mod tests {
    }

    #[test]
-    fn test_no_heap_allocation_for_small_queries() {
+    fn test_small_constraint_count() {
        let parser = QueryParser::default();
-        let result = parser
-            .parse("*.rs *.toml !test")
-            .expect("Should parse multi-token");
-        // SmallVec should not have spilled to heap
-        assert!(!result.constraints.spilled());
+        let result = parser.parse("*.rs *.toml !test");
+        assert_eq!(result.constraints.len(), 3);
    }

    #[test]
    fn test_many_fuzzy_parts() {
        let parser = QueryParser::default();
-        let result = parser
-            .parse("one two three four five six")
-            .expect("Should parse");
+        let result = parser.parse("one two three four five six");

        match &result.fuzzy_query {
            FuzzyQuery::Parts(parts) => {
@@ -1,8 +1,8 @@
 use crate::ConstraintVec;
 use crate::config::ParserConfig;
 use crate::constraints::{Constraint, GitStatusFilter, TextPartsBuffer};
+use crate::glob_detect::has_wildcards;
 use crate::location::{Location, parse_location};
-use zlob::{ZlobFlags, has_wildcards};

 #[derive(Debug, Clone, PartialEq)]
 #[allow(clippy::large_enum_variant)]
@@ -14,6 +14,8 @@ pub enum FuzzyQuery<'a> {

 #[derive(Debug, Clone, PartialEq)]
 pub struct FFFQuery<'a> {
+    /// The original raw query string before parsing
+    pub raw_query: &'a str,
    /// Parsed constraints (stack-allocated for ≤8 constraints)
    pub constraints: ConstraintVec<'a>,
    pub fuzzy_query: FuzzyQuery<'a>,
@@ -32,8 +34,8 @@ impl<C: ParserConfig> QueryParser<C> {
        Self { config }
    }

-    pub fn parse<'a>(&self, query: &'a str) -> Option<FFFQuery<'a>> {
-        let query: &'a str = query;
+    pub fn parse<'a>(&self, query: &'a str) -> FFFQuery<'a> {
+        let raw_query = query;
        let config: &C = &self.config;
        let mut constraints = ConstraintVec::new();
        let query = query.trim();
@@ -44,34 +46,75 @@ impl<C: ParserConfig> QueryParser<C> {
        if whitespace_count == 0 {
            // Try to parse as constraint first
            if let Some(constraint) = parse_token(query, config) {
-                constraints.push(constraint);
-                return Some(FFFQuery {
-                    constraints,
-                    fuzzy_query: FuzzyQuery::Empty,
-                    location: None,
-                });
+                // Don't treat filename tokens (FilePath) as constraints in single-token
+                // queries — the user is fuzzy-searching, not filtering. FilePath constraints
+                // are only useful as filters in multi-token queries like "score.rs search".
+                //
+                // Also skip PathSegment constraints when the token looks like an absolute
+                // file path with a location suffix (e.g. /Users/.../file.rs:12). Without
+                // this, the leading `/` causes the entire path to be consumed as a
+                // PathSegment, preventing location parsing from running.
+                let has_location_suffix = matches!(constraint, Constraint::PathSegment(_))
+                    && query.bytes().any(|b| b == b':')
+                    && query
+                        .bytes()
+                        .rev()
+                        .take_while(|&b| b != b':')
+                        .all(|b| b.is_ascii_digit());
+                if !matches!(constraint, Constraint::FilePath(_)) && !has_location_suffix {
+                    constraints.push(constraint);
+                    return FFFQuery {
+                        raw_query,
+                        constraints,
+                        fuzzy_query: FuzzyQuery::Empty,
+                        location: None,
+                    };
+                }
            }

            // Try to extract location from single token (e.g., "file:12")
-            let (query_without_loc, location) = parse_location(query);
-            if location.is_some() {
-                return Some(FFFQuery {
-                    constraints,
-                    fuzzy_query: FuzzyQuery::Text(query_without_loc),
-                    location,
-                });
+            if config.enable_location() {
+                let (query_without_loc, location) = parse_location(query);
+                if location.is_some() {
+                    return FFFQuery {
+                        raw_query,
+                        constraints,
+                        fuzzy_query: FuzzyQuery::Text(query_without_loc),
+                        location,
+                    };
+                }
            }

-            // Plain text single token - return None (caller handles as simple fuzzy match)
-            return None;
+            // Plain text single token
+            return FFFQuery {
+                raw_query,
+                constraints,
+                fuzzy_query: if query.is_empty() {
+                    FuzzyQuery::Empty
+                } else {
+                    FuzzyQuery::Text(query)
+                },
+                location: None,
+            };
        }

-        // Stack-allocated buffer for text parts (up to 16 parts)
        let mut text_parts = TextPartsBuffer::new();
        let tokens = query.split_whitespace();

+        let mut has_file_path = false;
        for token in tokens {
            match parse_token(token, config) {
+                Some(Constraint::FilePath(_)) => {
+                    if has_file_path {
+                        // Only one FilePath constraint allowed; treat extra path
+                        // tokens as literal text (e.g. an import path the user is
+                        // searching for).
+                        text_parts.push(token);
+                    } else {
+                        constraints.push(Constraint::FilePath(token));
+                        has_file_path = true;
+                    }
+                }
                Some(constraint) => {
                    constraints.push(constraint);
                }
@@ -83,7 +126,7 @@ impl<C: ParserConfig> QueryParser<C> {

        // Try to extract location from the last fuzzy token
        // e.g., "search file:12" -> fuzzy="search file", location=Line(12)
-        let location = if !text_parts.is_empty() {
+        let location = if config.enable_location() && !text_parts.is_empty() {
            let last_idx = text_parts.len() - 1;
            let (without_loc, loc) = parse_location(text_parts[last_idx]);
            if loc.is_some() {
@@ -115,17 +158,59 @@ impl<C: ParserConfig> QueryParser<C> {
            }
        };

-        Some(FFFQuery {
+        FFFQuery {
+            raw_query,
            constraints,
            fuzzy_query,
            location,
-        })
+        }
    }
 }

-impl Default for QueryParser<crate::FilePickerConfig> {
+impl<'a> FFFQuery<'a> {
+    /// Returns the grep search text by joining all non-constraint text tokens.
+    ///
+    /// Backslash-escaped tokens (e.g. `\*.rs`) are included as literal text
+    /// with the leading `\` stripped, since the backslash is only an escape
+    /// signal to the parser and should not appear in the final pattern.
+    ///
+    /// `FuzzyQuery::Empty` → empty string  
+    /// `FuzzyQuery::Text("foo")` → `"foo"`  
+    /// `FuzzyQuery::Parts(["a", "\\*.rs", "b"])` → `"a *.rs b"`
+    pub fn grep_text(&self) -> String {
+        match &self.fuzzy_query {
+            FuzzyQuery::Empty => String::new(),
+            FuzzyQuery::Text(t) => strip_leading_backslash(t).to_string(),
+            FuzzyQuery::Parts(parts) => parts
+                .iter()
+                .map(|t| strip_leading_backslash(t))
+                .collect::<Vec<_>>()
+                .join(" "),
+        }
+    }
+}
+
+/// Strip the leading `\` from a backslash-escaped constraint token only.
+///
+/// We strip the backslash when the next character is a constraint trigger
+/// (`*`, `/`, `!`) — the user typed `\*.rs` to mean literal `*.rs`, not an
+/// extension constraint. For regex escape sequences like `\w`, `\b`, `\d`,
+/// `\s`, `\n` etc., the backslash is preserved so regex mode works correctly.
+#[inline]
+fn strip_leading_backslash(token: &str) -> &str {
+    if token.len() > 1 && token.starts_with('\\') {
+        let next = token.as_bytes()[1];
+        // Only strip if the backslash is escaping a constraint trigger character
+        if next == b'*' || next == b'/' || next == b'!' {
+            return &token[1..];
+        }
+    }
+    token
+}
+
+impl Default for QueryParser<crate::FileSearchConfig> {
    fn default() -> Self {
-        Self::new(crate::FilePickerConfig)
+        Self::new(crate::FileSearchConfig)
    }
 }

@@ -151,7 +236,7 @@ fn parse_token<'a, C: ParserConfig>(token: &'a str, config: &C) -> Option<Constr
                // Only return Extension if the rest doesn't have wildcards
                // e.g., *.rs is Extension, but *.test.* should be Glob
                let ext_part = &token[2..];
-                if !has_wildcards(ext_part, ZlobFlags::RECOMMENDED) {
+                if !has_wildcards(ext_part) {
                    return Some(constraint);
                }
            }
@@ -250,7 +335,7 @@ fn parse_token_without_negation<'a, C: ParserConfig>(
            // Try extension first (*.rs) - simple patterns without additional wildcards
            if let Some(constraint) = parse_extension(token) {
                let ext_part = &token[2..];
-                if !has_wildcards(ext_part, ZlobFlags::RECOMMENDED) {
+                if !has_wildcards(ext_part) {
                    return Some(constraint);
                }
            }
@@ -308,11 +393,12 @@ fn parse_path_segment(token: &str) -> Option<Constraint<'_>> {
 }

 /// Parse path segment with trailing slash: www/ -> PathSegment("www")
+/// Also supports multi-segment paths: libswscale/aarch64/ -> PathSegment("libswscale/aarch64")
 #[inline]
 fn parse_path_segment_trailing(token: &str) -> Option<Constraint<'_>> {
    if token.len() > 1 && token.ends_with('/') {
        let segment = token.trim_end_matches('/');
-        if !segment.is_empty() && !segment.contains('/') {
+        if !segment.is_empty() {
            Some(Constraint::PathSegment(segment))
        } else {
            None
@@ -351,7 +437,7 @@ fn parse_git_status(value: &str) -> Option<Constraint<'_>> {
 #[cfg(test)]
 mod tests {
    use super::*;
-    use crate::FilePickerConfig;
+    use crate::{FileSearchConfig, GrepConfig};

    #[test]
    fn test_parse_extension() {
@@ -366,7 +452,7 @@ mod tests {

    #[test]
    fn test_incomplete_patterns_ignored() {
-        let config = FilePickerConfig;
+        let config = FileSearchConfig;
        // Incomplete patterns should return None and be treated as noise
        assert_eq!(parse_token("*", &config), None);
        assert_eq!(parse_token("*.", &config), None);
@@ -395,18 +481,23 @@ mod tests {
            parse_path_segment_trailing("src/"),
            Some(Constraint::PathSegment("src"))
        );
-        // Should not match paths with multiple segments
-        assert_eq!(parse_path_segment_trailing("src/lib/"), None);
+        // Multi-segment paths should work
+        assert_eq!(
+            parse_path_segment_trailing("src/lib/"),
+            Some(Constraint::PathSegment("src/lib"))
+        );
+        assert_eq!(
+            parse_path_segment_trailing("libswscale/aarch64/"),
+            Some(Constraint::PathSegment("libswscale/aarch64"))
+        );
        // Should not match without trailing slash
        assert_eq!(parse_path_segment_trailing("www"), None);
    }

    #[test]
    fn test_trailing_slash_in_query() {
-        let parser = QueryParser::new(FilePickerConfig);
-        let result = parser
-            .parse("www/ test")
-            .expect("Should parse multi-token query");
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("www/ test");
        assert_eq!(result.constraints.len(), 1);
        assert!(matches!(
            result.constraints[0],
@@ -441,11 +532,9 @@ mod tests {

    #[test]
    fn test_negation_text() {
-        let parser = QueryParser::new(FilePickerConfig);
+        let parser = QueryParser::new(FileSearchConfig);
        // Need two tokens for parsing to return Some
-        let result = parser
-            .parse("!test foo")
-            .expect("Should parse multi-token query");
+        let result = parser.parse("!test foo");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -457,10 +546,8 @@ mod tests {

    #[test]
    fn test_negation_extension() {
-        let parser = QueryParser::new(FilePickerConfig);
-        let result = parser
-            .parse("!*.rs foo")
-            .expect("Should parse multi-token query");
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("!*.rs foo");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -472,10 +559,8 @@ mod tests {

    #[test]
    fn test_negation_path_segment() {
-        let parser = QueryParser::new(FilePickerConfig);
-        let result = parser
-            .parse("!/src/ foo")
-            .expect("Should parse multi-token query");
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("!/src/ foo");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -487,10 +572,8 @@ mod tests {

    #[test]
    fn test_negation_git_status() {
-        let parser = QueryParser::new(FilePickerConfig);
-        let result = parser
-            .parse("!status:modified foo")
-            .expect("Should parse multi-token query");
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("!status:modified foo");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -505,10 +588,8 @@ mod tests {

    #[test]
    fn test_backslash_escape_extension() {
-        let parser = QueryParser::new(FilePickerConfig);
-        let result = parser
-            .parse("\\*.rs foo")
-            .expect("Should parse multi-token query");
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("\\*.rs foo");
        // \*.rs should NOT be parsed as an Extension constraint
        assert_eq!(result.constraints.len(), 0);
        // Both tokens should be text
@@ -524,10 +605,8 @@ mod tests {

    #[test]
    fn test_backslash_escape_path_segment() {
-        let parser = QueryParser::new(FilePickerConfig);
-        let result = parser
-            .parse("\\/src/ foo")
-            .expect("Should parse multi-token query");
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("\\/src/ foo");
        assert_eq!(result.constraints.len(), 0);
        match result.fuzzy_query {
            FuzzyQuery::Parts(parts) => {
@@ -540,39 +619,99 @@ mod tests {

    #[test]
    fn test_backslash_escape_negation() {
-        let parser = QueryParser::new(FilePickerConfig);
-        let result = parser
-            .parse("\\!test foo")
-            .expect("Should parse multi-token query");
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("\\!test foo");
        assert_eq!(result.constraints.len(), 0);
    }

+    #[test]
+    fn test_grep_text_plain_text() {
+        // Multi-token plain text — no constraints
+        let q = QueryParser::new(GrepConfig).parse("name =");
+        assert_eq!(q.grep_text(), "name =");
+    }
+
+    #[test]
+    fn test_grep_text_strips_constraint() {
+        let q = QueryParser::new(GrepConfig).parse("name = *.rs someth");
+        assert_eq!(q.grep_text(), "name = someth");
+    }
+
+    #[test]
+    fn test_grep_text_leading_constraint() {
+        let q = QueryParser::new(GrepConfig).parse("*.rs name =");
+        assert_eq!(q.grep_text(), "name =");
+    }
+
+    #[test]
+    fn test_grep_text_only_constraints() {
+        let q = QueryParser::new(GrepConfig).parse("*.rs /src/");
+        assert_eq!(q.grep_text(), "");
+    }
+
+    #[test]
+    fn test_grep_text_path_constraint() {
+        let q = QueryParser::new(GrepConfig).parse("name /src/ value");
+        assert_eq!(q.grep_text(), "name value");
+    }
+
+    #[test]
+    fn test_grep_text_negation_constraint() {
+        let q = QueryParser::new(GrepConfig).parse("name !*.rs value");
+        assert_eq!(q.grep_text(), "name value");
+    }
+
+    #[test]
+    fn test_grep_text_backslash_escape_stripped() {
+        // \*.rs should be text with the leading \ removed
+        let q = QueryParser::new(GrepConfig).parse("\\*.rs foo");
+        assert_eq!(q.grep_text(), "*.rs foo");
+
+        let q = QueryParser::new(GrepConfig).parse("\\/src/ foo");
+        assert_eq!(q.grep_text(), "/src/ foo");
+
+        let q = QueryParser::new(GrepConfig).parse("\\!test foo");
+        assert_eq!(q.grep_text(), "!test foo");
+    }
+
+    #[test]
+    fn test_grep_text_question_mark_is_text() {
+        let q = QueryParser::new(GrepConfig).parse("foo? bar");
+        assert_eq!(q.grep_text(), "foo? bar");
+    }
+
+    #[test]
+    fn test_grep_text_bracket_is_text() {
+        let q = QueryParser::new(GrepConfig).parse("arr[0] more");
+        assert_eq!(q.grep_text(), "arr[0] more");
+    }
+
+    #[test]
+    fn test_grep_text_path_glob_is_constraint() {
+        let q = QueryParser::new(GrepConfig).parse("pattern src/**/*.rs");
+        assert_eq!(q.grep_text(), "pattern");
+    }
+
    #[test]
    fn test_grep_question_mark_is_text() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
-        // Single token "foo?" should return None (treated as plain text by caller)
        let result = parser.parse("foo?");
-        assert!(result.is_none(), "foo? should be plain text in grep mode");
+        assert!(result.constraints.is_empty());
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("foo?"));
    }

    #[test]
    fn test_grep_bracket_is_text() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
        let result = parser.parse("arr[0] something");
-        let result = result.expect("Should parse multi-token query");
        // arr[0] should NOT be a glob in grep mode
        assert_eq!(result.constraints.len(), 0);
    }

    #[test]
    fn test_grep_path_glob_is_constraint() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
-        let result = parser
-            .parse("pattern src/**/*.rs")
-            .expect("Should parse with path glob");
+        let result = parser.parse("pattern src/**/*.rs");
        // src/**/*.rs contains / so it should be treated as a glob
        assert_eq!(result.constraints.len(), 1);
        assert!(matches!(
@@ -583,11 +722,8 @@ mod tests {

    #[test]
    fn test_grep_brace_is_constraint() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
-        let result = parser
-            .parse("pattern {src,lib}")
-            .expect("Should parse with brace expansion");
+        let result = parser.parse("pattern {src,lib}");
        assert_eq!(result.constraints.len(), 1);
        assert!(matches!(
            result.constraints[0],
@@ -595,13 +731,51 @@ mod tests {
        ));
    }

+    #[test]
+    fn test_grep_text_preserves_backslash_escapes() {
+        // Regex patterns like \w+ and \bfoo\b must survive grep_text()
+        // The parser sees \w+ as a text token (not a constraint escape),
+        // but strip_leading_backslash was stripping the \ anyway.
+        let q = QueryParser::new(GrepConfig).parse("pub struct \\w+");
+        assert_eq!(
+            q.grep_text(),
+            "pub struct \\w+",
+            "Backslash-w in regex must be preserved"
+        );
+
+        let q = QueryParser::new(GrepConfig).parse("\\bword\\b more");
+        assert_eq!(
+            q.grep_text(),
+            "\\bword\\b more",
+            "Backslash-b word boundaries must be preserved"
+        );
+
+        // Single-token regex like "fn\\s+\\w+" returns FFFQuery with Text fuzzy query
+        let result = QueryParser::new(GrepConfig).parse("fn\\s+\\w+");
+        assert!(result.constraints.is_empty());
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("fn\\s+\\w+"));
+
+        // But the escaped constraint forms SHOULD still be stripped:
+        let q = QueryParser::new(GrepConfig).parse("\\*.rs foo");
+        assert_eq!(
+            q.grep_text(),
+            "*.rs foo",
+            "Escaped constraint \\*.rs should still have backslash stripped"
+        );
+
+        let q = QueryParser::new(GrepConfig).parse("\\/src/ foo");
+        assert_eq!(
+            q.grep_text(),
+            "/src/ foo",
+            "Escaped constraint \\/src/ should still have backslash stripped"
+        );
+    }
+
    #[test]
    fn test_grep_bare_star_is_text() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
        // "a*b" contains * but no / or {} — should be text in grep mode
        let result = parser.parse("a*b something");
-        let result = result.expect("Should parse");
        assert_eq!(
            result.constraints.len(),
            0,
@@ -611,11 +785,8 @@ mod tests {

    #[test]
    fn test_grep_negated_text() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
-        let result = parser
-            .parse("pattern !test")
-            .expect("Should parse negated text in grep mode");
+        let result = parser.parse("pattern !test");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -631,11 +802,8 @@ mod tests {

    #[test]
    fn test_grep_negated_path_segment() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
-        let result = parser
-            .parse("pattern !/src/")
-            .expect("Should parse negated path segment in grep mode");
+        let result = parser.parse("pattern !/src/");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -651,11 +819,8 @@ mod tests {

    #[test]
    fn test_grep_negated_extension() {
-        use crate::GrepConfig;
        let parser = QueryParser::new(GrepConfig);
-        let result = parser
-            .parse("pattern !*.rs")
-            .expect("Should parse negated extension in grep mode");
+        let result = parser.parse("pattern !*.rs");
        assert_eq!(result.constraints.len(), 1);
        match &result.constraints[0] {
            Constraint::Not(inner) => {
@@ -668,4 +833,345 @@ mod tests {
            other => panic!("Expected Not constraint, got {:?}", other),
        }
    }
+
+    // ── AI grep config tests ──────────────────────────────────────────
+
+    #[test]
+    fn test_ai_grep_detects_file_path() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("libswscale/input.c rgba32ToY");
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(
+                result.constraints[0],
+                Constraint::FilePath("libswscale/input.c")
+            ),
+            "Expected FilePath, got {:?}",
+            result.constraints[0]
+        );
+        assert_eq!(result.grep_text(), "rgba32ToY");
+    }
+
+    #[test]
+    fn test_ai_grep_detects_nested_file_path() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("src/main.rs fn main");
+        assert_eq!(result.constraints.len(), 1);
+        assert!(matches!(
+            result.constraints[0],
+            Constraint::FilePath("src/main.rs")
+        ));
+        assert_eq!(result.grep_text(), "fn main");
+    }
+
+    #[test]
+    fn test_ai_grep_no_false_positive_trailing_slash() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("src/ pattern");
+        // Should be PathSegment, NOT FilePath
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(result.constraints[0], Constraint::PathSegment("src")),
+            "Expected PathSegment, got {:?}",
+            result.constraints[0]
+        );
+    }
+
+    #[test]
+    fn test_ai_grep_bare_filename_is_file_path() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("main.rs pattern");
+        // Bare filename with valid extension → FilePath constraint
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(result.constraints[0], Constraint::FilePath("main.rs")),
+            "Expected FilePath, got {:?}",
+            result.constraints[0]
+        );
+        assert_eq!(result.grep_text(), "pattern");
+    }
+
+    #[test]
+    fn test_ai_grep_bare_filename_schema_rs() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("schema.rs part_revisions");
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(result.constraints[0], Constraint::FilePath("schema.rs")),
+            "Expected FilePath(schema.rs), got {:?}",
+            result.constraints[0]
+        );
+        assert_eq!(result.grep_text(), "part_revisions");
+    }
+
+    #[test]
+    fn test_ai_grep_bare_word_no_extension_not_constraint() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("schema pattern");
+        // No extension → not a file path, just text
+        assert_eq!(result.constraints.len(), 0);
+        assert_eq!(result.grep_text(), "schema pattern");
+    }
+
+    #[test]
+    fn test_ai_grep_no_false_positive_no_extension() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("src/utils pattern");
+        // No extension in last component → not a file path, just text
+        assert_eq!(result.constraints.len(), 0);
+        assert_eq!(result.grep_text(), "src/utils pattern");
+    }
+
+    #[test]
+    fn test_ai_grep_wildcard_not_filepath() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("src/**/*.rs pattern");
+        // Contains wildcards → should be a Glob, not FilePath
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(result.constraints[0], Constraint::Glob("src/**/*.rs")),
+            "Expected Glob, got {:?}",
+            result.constraints[0]
+        );
+    }
+
+    #[test]
+    fn test_ai_grep_star_text_star_is_glob() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("*quote* TODO");
+        // `*quote*` should be recognised as a glob constraint in AI mode
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(result.constraints[0], Constraint::Glob("*quote*")),
+            "Expected Glob(*quote*), got {:?}",
+            result.constraints[0]
+        );
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("TODO"));
+    }
+
+    #[test]
+    fn test_ai_grep_bare_star_not_glob() {
+        use crate::AiGrepConfig;
+        let parser = QueryParser::new(AiGrepConfig);
+        let result = parser.parse("* pattern");
+        // Bare `*` should NOT be treated as a glob (too broad)
+        assert!(
+            result.constraints.is_empty(),
+            "Expected no constraints, got {:?}",
+            result.constraints
+        );
+    }
+
+    #[test]
+    fn test_grep_no_location_parsing_single_token() {
+        let parser = QueryParser::new(GrepConfig);
+        // localhost:8080 should NOT be parsed as location -- it's a search pattern
+        let result = parser.parse("localhost:8080");
+        assert!(result.constraints.is_empty());
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("localhost:8080"));
+    }
+
+    #[test]
+    fn test_grep_no_location_parsing_multi_token() {
+        let q = QueryParser::new(GrepConfig).parse("*.rs localhost:8080");
+        assert_eq!(
+            q.grep_text(),
+            "localhost:8080",
+            "Colon-number suffix should be preserved in grep text"
+        );
+        assert!(
+            q.location.is_none(),
+            "Grep should not parse location from colon-number"
+        );
+    }
+
+    #[test]
+    fn test_grep_braces_without_comma_is_text() {
+        let parser = QueryParser::new(GrepConfig);
+        // Code patterns like format!("{}") should NOT be treated as brace expansion
+        let result = parser.parse(r#"format!("{}\\AppData", home)"#);
+        assert!(
+            result.constraints.is_empty(),
+            "Braces without comma should be text, got {:?}",
+            result.constraints
+        );
+        assert_eq!(result.grep_text(), r#"format!("{}\\AppData", home)"#);
+    }
+
+    #[test]
+    fn test_grep_format_braces_not_glob() {
+        let parser = QueryParser::new(GrepConfig);
+        // Code like format!("{}\\path", var) must not have tokens eaten as glob constraints.
+        // The trailing comma on the first token means both { } and , are present,
+        // but the comma is outside the braces so it should NOT trigger brace expansion.
+        let input = "format!(\"{}\\\\AppData\", home)";
+        let result = parser.parse(input);
+        assert!(
+            result.constraints.is_empty(),
+            "format! pattern should have no constraints, got {:?}",
+            result.constraints
+        );
+    }
+
+    #[test]
+    fn test_grep_config_star_text_star_not_glob() {
+        use crate::GrepConfig;
+        let parser = QueryParser::new(GrepConfig);
+        let result = parser.parse("*quote* TODO");
+        // Regular grep mode should NOT treat `*quote*` as a glob
+        assert!(
+            result.constraints.is_empty(),
+            "Expected no constraints in GrepConfig, got {:?}",
+            result.constraints
+        );
+    }
+
+    // ── File picker filename constraint tests ─────────────────────────
+
+    #[test]
+    fn test_file_picker_bare_filename_constraint() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("score.rs file_picker");
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(result.constraints[0], Constraint::FilePath("score.rs")),
+            "Expected FilePath(\"score.rs\"), got {:?}",
+            result.constraints[0]
+        );
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("file_picker"));
+    }
+
+    #[test]
+    fn test_file_picker_path_prefixed_filename_constraint() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("libswscale/slice.c lum_convert");
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(
+                result.constraints[0],
+                Constraint::FilePath("libswscale/slice.c")
+            ),
+            "Expected FilePath(\"libswscale/slice.c\"), got {:?}",
+            result.constraints[0]
+        );
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("lum_convert"));
+    }
+
+    #[test]
+    fn test_file_picker_single_token_filename_stays_fuzzy() {
+        let parser = QueryParser::new(FileSearchConfig);
+        // Single-token filename should NOT become a constraint -- it should
+        // return FFFQuery with Text fuzzy query so the caller uses it for fuzzy matching.
+        let result = parser.parse("score.rs");
+        assert!(result.constraints.is_empty());
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("score.rs"));
+    }
+
+    #[test]
+    fn test_absolute_path_with_location_not_path_segment() {
+        let parser = QueryParser::new(FileSearchConfig);
+        // Absolute file path with :line should parse as text + location,
+        // NOT as a PathSegment constraint (which would eat the whole token).
+        let result = parser.parse("/Users/neogoose/dev/fframes/src/renderer/concatenator.rs:12");
+        assert!(
+            result.constraints.is_empty(),
+            "Absolute path with location should not become a constraint, got {:?}",
+            result.constraints
+        );
+        assert_eq!(
+            result.fuzzy_query,
+            FuzzyQuery::Text("/Users/neogoose/dev/fframes/src/renderer/concatenator.rs")
+        );
+        assert_eq!(result.location, Some(Location::Line(12)));
+    }
+
+    #[test]
+    fn test_file_picker_filename_with_multiple_fuzzy_parts() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("main.rs src components");
+        assert_eq!(result.constraints.len(), 1);
+        assert!(matches!(
+            result.constraints[0],
+            Constraint::FilePath("main.rs")
+        ));
+        assert_eq!(
+            result.fuzzy_query,
+            FuzzyQuery::Parts(vec!["src", "components"])
+        );
+    }
+
+    #[test]
+    fn test_file_picker_version_number_not_filename() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("v2.0 release");
+        // v2.0 extension starts with digit → not a filename constraint
+        assert!(
+            result.constraints.is_empty(),
+            "v2.0 should not be a FilePath constraint, got {:?}",
+            result.constraints
+        );
+    }
+
+    #[test]
+    fn test_file_picker_only_one_filepath_constraint() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("main.rs score.rs");
+        // Only first filename becomes a constraint; second is text
+        assert_eq!(result.constraints.len(), 1);
+        assert!(matches!(
+            result.constraints[0],
+            Constraint::FilePath("main.rs")
+        ));
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("score.rs"));
+    }
+
+    #[test]
+    fn test_file_picker_filename_with_extension_constraint() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("main.rs *.lua");
+        // main.rs → FilePath, *.lua → Extension
+        assert_eq!(result.constraints.len(), 2);
+        assert!(matches!(
+            result.constraints[0],
+            Constraint::FilePath("main.rs")
+        ));
+        assert!(matches!(
+            result.constraints[1],
+            Constraint::Extension("lua")
+        ));
+    }
+
+    #[test]
+    fn test_file_picker_dotfile_is_filename() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse(".gitignore src");
+        assert_eq!(result.constraints.len(), 1);
+        assert!(
+            matches!(result.constraints[0], Constraint::FilePath(".gitignore")),
+            "Expected FilePath(\".gitignore\"), got {:?}",
+            result.constraints[0]
+        );
+        assert_eq!(result.fuzzy_query, FuzzyQuery::Text("src"));
+    }
+
+    #[test]
+    fn test_file_picker_no_extension_not_filename() {
+        let parser = QueryParser::new(FileSearchConfig);
+        let result = parser.parse("Makefile src");
+        // No dot → not a filename constraint
+        assert!(
+            result.constraints.is_empty(),
+            "Makefile should not be a FilePath constraint, got {:?}",
+            result.constraints
+        );
+    }
 }
@@ -1,52 +1,68 @@
-*fff.nvim.txt*        For Neovim >= 0.10.0       Last change: 2026 February 17
+*fff.nvim.txt*         For Neovim >= 0.10.0         Last change: 2026 March 18

 ==============================================================================
 Table of Contents                                 *fff.nvim-table-of-contents*

-  - Features                                               |fff.nvim-features|
-  - Installation                                       |fff.nvim-installation|
-FFF.nvimFinally a smart fuzzy file picker for neovim.
+  - MCP                                                         |fff.nvim-mcp|
+  - Neovim guide                                       |fff.nvim-neovim-guide|
+1. Links                                                      |fff.nvim-links|
+FFFAI agents (MCP)   |   Neovim usersA fast file search for your AI and neovim, with memory built-in



- **FFF** stands for ~freakin fast fuzzy file finder~ (pick 3) and it is an
-opinionated fuzzy file picker for neovim. Just for files, but we’ll try to
-solve file picking completely.
+ ------------------------------------------------------------------------------
+**FFF** stands for ~freakin fast fuzzy file finder~ (pick 3) and it is an
+opinionated fuzzy file picker for your AI agent and Neovim. Just for file
+search, but we do the file search really fff well.

-It comes with a dedicated rust backend runtime that keep tracks of the file
-index, your file access and modifications, git status, and provides a
-comprehensive typo-resistant fuzzy search experience.
+FFF is a tool for grepping, fuzzy file matching, globbing, and multigrepping
+with a strong focus on performance and useful search results. For humans -
+provides an unbelievable typo-resistant experience, for AI agents - implements
+the fastest file search with additional free memory suggesting the best search
+results based on various factors like frecency, git status, file size,
+definition matches, and more.


-FEATURES                                                   *fff.nvim-features*
+MCP                                                             *fff.nvim-mcp*

- Works out of the box with no additional configuration
- Typo resistant fuzzy search <https://github.com/saghen/frizbee>
- Git status integration allowing to take advantage of last modified times within a worktree
- Separate file index maintained by a dedicated backend allows <10 milliseconds search time for 50k files codebase
- Display images in previews (for now requires snacks.nvim)
- Smart in a plenty of different ways hopefully helpful for your workflow
- This plugin initializes itself lazily by default
+FFF is an amazing way to reduce the time and tokens by giving your AI agent a
+bit of memory built-in to their file search tools. It makes your AI harness to
+find the code faster and spend less tokens by doing less roundtrips and reading
+less useless files.
+
+You can install FFF as a dependency for your AI agent using a simple bash
+script:
+
+>bash
+    curl -L https://dmtrkovalenko.dev/install-fff-mcp.sh | bash
+<


-INSTALLATION                                           *fff.nvim-installation*
+  The installation script is here ./install-mcp.sh <./install-mcp.sh> if you want
+  to review it before running.
+It will print out the instructions on how to connect it to your `Claude Code`,
+`Codex`, `OpenCode`, etc. Once you have it connected just ask your agent to
+"use fff". Here is an example addition to `CLAUDE.md` that works perfectly:
+
+>sh
+    # CLAUDE.md
+    For any file search or grep in the current git indexed directory use fff tools
+<


-  [!NOTE] Although we’ll try to make sure to keep 100% backward compatibility,
-  by using you should understand that silly bugs and breaking changes may happen.
-  And also we hope for your contributions and feedback to make this plugin ideal
-  for everyone.
+NEOVIM GUIDE                                           *fff.nvim-neovim-guide*

-PREREQUISITES ~
+Here is some demo on the linux repository (100k files, 8GB) but you better fill
+it yourself and see the magic

-FFF.nvim requires:

- Neovim 0.10.0+
- Rustup <https://rustup.rs/> (we require nightly for building the native backend rustup will handle toolchain automatically)
+https://github.com/user-attachments/assets/5d0e1ce9-642c-4c44-aa88-01b05bb86abb


 INSTALLATION ~

+FFF.nvim requires neovim 0.10.0 or higher
+

 LAZY.NVIM

@@ -83,12 +99,17 @@ LAZY.NVIM
        {
          "fz",
          function() require('fff').live_grep({
-        grep = {
-          modes = { 'fuzzy', 'plain' }
-        }
+            grep = {
+              modes = { 'fuzzy', 'plain' }
+            }
          }) end,
          desc = 'Live fffuzy grep',
-        }
+        },
+        {
+          "fc",
+          function() require('fff').live_grep({ query = vim.fn.expand("<cword>") }) end,
+          desc = 'Search current word',
+        },
      }
    }
 <
@@ -144,6 +165,10 @@ all available options:
          prompt_position = 'bottom', -- or 'top'
          preview_position = 'right', -- or 'left', 'right', 'top', 'bottom'
          preview_size = 0.5,
+          flex = { -- set to false to disable flex layout
+            size = 130, -- column threshold: if screen width >= size, use preview_position; otherwise use wrap
+            wrap = 'top', -- position to use when screen is narrower than size
+          },
          show_scrollbar = true, -- Show scrollbar for pagination
          -- How to shorten long directory paths in the file list:
          -- 'middle_number' (default): uses dots for 1-3 hidden (a/./b, a/../b, a/.../b)
@@ -159,6 +184,7 @@ all available options:
          binary_file_threshold = 1024, -- amount of bytes to scan for binary content (set 0 to disable)
          imagemagick_info_format_str = '%m: %wx%h, %[colorspace], %q-bit',
          line_numbers = false,
+          cursorlineopt = 'both', -- the cursorlineopt used for lines in grep file previews, see :h cursorlineopt
          wrap_lines = false,
          filetypes = {
            svg = { wrap_lines = true },
@@ -178,27 +204,29 @@ all available options:
          preview_scroll_up = '<C-u>',
          preview_scroll_down = '<C-d>',
          toggle_debug = '<F2>',
+          -- grep mode: cycle between plain text, regex, and fuzzy search
+          cycle_grep_modes = '<S-Tab>',
          -- goes to the previous query in history
          cycle_previous_query = '<C-Up>',
          -- multi-select keymaps for quickfix
          toggle_select = '<Tab>',
          send_to_quickfix = '<C-q>',
-          -- grep mode: cycle between plain text, regex, and fuzzy search
-          toggle_grep_regex = '<S-Tab>',
+          -- this are specific for the normal mode (you can exit it using any other keybind like jj)
+          focus_list = '<leader>l',
+          focus_preview = '<leader>p',
        },
        hl = {
          border = 'FloatBorder',
          normal = 'Normal',
-          cursor = 'CursorLine',
+          cursor = 'CursorLine',  -- Falls back to 'Visual' if CursorLine is not defined
          matched = 'IncSearch',
          title = 'Title',
          prompt = 'Question',
-          active_file = 'Visual',
          frecency = 'Number',
          debug = 'Comment',
          combo_header = 'Number',
-          scrollbar = 'Comment', -- Highlight for scrollbar thumb (track uses border)
-          directory_path = 'Comment', -- Highlight for directory path in file list
+          scrollbar = 'Comment',
+          directory_path = 'Comment',
          -- Multi-select highlights
          selected = 'FFFSelected',
          selected_active = 'FFFSelectedActive',
@@ -227,9 +255,10 @@ all available options:
          grep_match = 'IncSearch',               -- Highlight for matched text in grep results
          grep_line_number = 'LineNr',            -- Highlight for :line:col location
          grep_regex_active = 'DiagnosticInfo',   -- Highlight for keybind + label when regex is on
-          grep_regex_inactive = 'Comment',        -- Highlight for keybind + label when regex is off
+          grep_plain_active = 'Comment',        -- Highlight for keybind + label when regex is off
+          grep_fuzzy_active = 'DiagnosticHint',   -- Highlight for keybind + label when fuzzy is on
          -- Cross-mode suggestion highlights
-          suggestion_header = 'WarningMsg',       -- Highlight for the "No results found. Suggested..." banner
+          suggestion_header = 'WarningMsg', -- Highlight for the "No results found. Suggested..." banner
        },
        -- Store file open frecency
        frecency = {
@@ -240,32 +269,35 @@ all available options:
        history = {
          enabled = true,
          db_path = vim.fn.stdpath('data') .. '/fff_queries',
-          min_combo_count = 3, -- file will get a boost if it was selected 3 in a row times per specific query
-          combo_boost_score_multiplier = 100, -- Score multiplier for combo matches
+          min_combo_count = 3, -- Minimum selections before combo boost applies (3 = boost starts on 3rd selection)
+          combo_boost_score_multiplier = 100, -- Score multiplier for combo matches (files repeatedly opened with same query)
        },
        -- Git integration
        git = {
          status_text_color = false, -- Apply git status colors to filename text (default: false, only sign column)
        },
        debug = {
-          enabled = false, -- Set to true to show scores in the UI
-          show_scores = false,
-          show_file_info = false, -- Show file info panel in preview
+          enabled = false, -- Show file info panel in preview
+          show_scores = false, -- Show scores inline in the UI
        },
        logging = {
          enabled = true,
          log_file = vim.fn.stdpath('log') .. '/fff.log',
          log_level = 'info',
        },
-        -- Live grep search configuration
+        -- find_files settings
+        file_picker = {
+          current_file_label = '(current)',
+        },
+        -- grep settings
        grep = {
          max_file_size = 10 * 1024 * 1024, -- Skip files larger than 10MB
-          max_matches_per_file = 200, -- Maximum matches per file
+          max_matches_per_file = 100, -- Maximum matches per file (set 0 to unlimited)
          smart_case = true, -- Case-insensitive unless query has uppercase
          time_budget_ms = 150, -- Max search time in ms per call (prevents UI freeze, 0 = no limit)
          modes = { 'plain', 'regex', 'fuzzy' }, -- Available grep modes and their cycling order
-        }
-    })
+        },
+      })
 <


@@ -275,20 +307,21 @@ KEY FEATURES ~
 AVAILABLE METHODS

 >lua
-    require('fff').find_files()                         -- Find files in current directory
-    require('fff').find_in_git_root()                   -- Find files in the current git repository
+    require('fff').find_files()                         -- Find files in current repository
    require('fff').scan_files()                         -- Trigger rescan of files in the current directory
-    require('fff').refresh_git_status()                 -- Refresh git status for the active file lock
+    require('fff').refresh_git_status()                 -- Refresh git status for the active file list
    require('fff').find_files_in_dir(path)              -- Find files in a specific directory
    require('fff').change_indexing_directory(new_path)  -- Change the base directory for the file picker
 <

+just jump to the definition and see what other APIs are exposed we have a
+plenty
+

 COMMANDS

 FFF.nvim provides several commands for interacting with the file picker:

- `:FFFFind [path|query]` - Open file picker. Optional: provide directory path or search query
 - `:FFFScan` - Manually trigger a rescan of files in the current directory
 - `:FFFRefreshGit` - Manually refresh git status for all files
 - `:FFFClearCache [all|frecency|files]` - Clear various caches
@@ -297,13 +330,6 @@ FFF.nvim provides several commands for interacting with the file picker:
 - `:FFFOpenLog` - Open the FFF log file in a new tab


-MULTILINE PASTE SUPPORT
-
-The input field automatically handles multiline clipboard content by joining
-all lines into a single search query. This is particularly useful when copying
-file paths from terminal output.
-
-
 DEBUG MODE

 Toggle scoring information display:
@@ -362,12 +388,43 @@ your configuration, or per-call when invoking `live_grep()`.
        modes = { 'fuzzy' },
      }
    })
+    
+    -- Pre-fill the search with an initial query
+    require('fff').live_grep({ query = 'search term' })
 <

 When only one mode is configured, the mode indicator is hidden completely and
 the cycle keybind does nothing.


+CONSTRAINTS
+
+There are a number of constraints you can use to refine your search in both
+grep and file search mode:
+
+- `git:modified` - show only modified files (one of `modified`, `staged`, `deleted`, `renamed`, `untracked`, `ignored`)
+- `test/` - any deeply nested children of any test/ dir
+- `!something` - exclude results matching something
+- `!test/`, `!git:modified` - combining with any other constraint works as negation
+- `./**/*.{rs,lua}` - any valid glob expression via the fastest globbing library <https://github.com/dmtrKovalenko/zlob>
+
+For grep only:
+
+- `*.md`, `*.{c,h}` - extension filtering
+- `src/main.rs` - grep in a single file
+
+In addition to that, all constraints can be combined together like:
+
+>
+    git:modified src/**/*.rs !src/**/mod.rs user controller
+<
+
+This will find all the files that qualify the constraints and:
+
+- match **both** user and controller (for file mode)
+- match "user controller" (for grep mode)
+
+
 CROSS-MODE SUGGESTIONS

 When a search returns no results, FFF automatically queries the opposite search
@@ -477,47 +534,17 @@ VIEWING LOGS

 If you encounter issues, check the log file:

->vim
+>
    :FFFOpenLog
 <

 Or manually open the log file at `~/.local/state/nvim/log/fff.log` (default
 location).

+==============================================================================
+1. Links                                                      *fff.nvim-links*

-COMMON ISSUES
-
-**File picker not initializing:**
-
- Ensure the Rust backend is compiled: `cargo build --release` in the plugin directory
- Check that your Neovim version is 0.10.0 or higher
-
-**Image previews not working:**
-
- Verify your terminal supports images (kitty, iTerm2, WezTerm, etc.)
- For terminals without native image support, install one of: `chafa`, `viu`, or `img2txt`
- If using snacks.nvim, ensure it’s properly configured
-
-**Performance issues:**
-
- Adjust `max_threads` in configuration based on your system
- Reduce `preview.max_lines` and `preview.max_size` for large files
- Clear cache if it becomes too large: `:FFFClearCache all`
-
-**Files not being indexed:**
-
- Run `:FFFScan` to manually trigger a file scan
- Check that the `base_path` is correctly set
- Verify you have read permissions for the directory
-
-
-DEBUG MODE
-
-Enable debug mode to see scoring information and troubleshoot search results:
-
- Press `F2` while in the picker
- Run `:FFFDebug on` to enable permanently
- Set `debug.show_scores = true` in configuration
+1. *Chart showing the superiority of fff.nvim over builtin claude code tools*: ./chart.png

 Generated by panvimdoc <https://github.com/kdheepak/panvimdoc>

@@ -2,11 +2,11 @@
  "nodes": {
    "crane": {
      "locked": {
-        "lastModified": 1767744144,
-        "narHash": "sha256-9/9ntI0D+HbN4G0TrK3KmHbTvwgswz7p8IEJsWyef8Q=",
+        "lastModified": 1773857772,
+        "narHash": "sha256-5xsK26KRHf0WytBtsBnQYC/lTWDhQuT57HJ7SzuqZcM=",
        "owner": "ipetkov",
        "repo": "crane",
-        "rev": "2fb033290bf6b23f226d4c8b32f7f7a16b043d7e",
+        "rev": "b556d7bbae5ff86e378451511873dfd07e4504cd",
        "type": "github"
      },
      "original": {
@@ -35,11 +35,11 @@
    },
    "nixpkgs": {
      "locked": {
-        "lastModified": 1767364772,
-        "narHash": "sha256-fFUnEYMla8b7UKjijLnMe+oVFOz6HjijGGNS1l7dYaQ=",
+        "lastModified": 1773628058,
+        "narHash": "sha256-hpXH0z3K9xv0fHaje136KY872VT2T5uwxtezlAskQgY=",
        "owner": "NixOS",
        "repo": "nixpkgs",
-        "rev": "16c7794d0a28b5a37904d55bcca36003b9109aaa",
+        "rev": "f8573b9c935cfaa162dd62cc9e75ae2db86f85df",
        "type": "github"
      },
      "original": {
@@ -64,11 +64,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1770865833,
-        "narHash": "sha256-oiARqnlvaW6pVGheVi4ye6voqCwhg5hCcGish2ZvQzI=",
+        "lastModified": 1773803479,
+        "narHash": "sha256-GD6i1F2vrSxbsmbS92+8+x3DbHOJ+yrS78Pm4xigW4M=",
        "owner": "oxalica",
        "repo": "rust-overlay",
-        "rev": "c8cfbe26238638e2f3a2c0ae7e8d240f5e4ded85",
+        "rev": "f17186f52e82ec5cf40920b58eac63b78692ac7c",
        "type": "github"
      },
      "original": {
@@ -0,0 +1,272 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# FFF MCP Server installer
+# Usage: curl -fsSL https://raw.githubusercontent.com/dmtrKovalenko/fff.nvim/main/install-mcp.sh | bash
+
+REPO="dmtrKovalenko/fff.nvim"
+BINARY_NAME="fff-mcp"
+INSTALL_DIR="${FFF_MCP_INSTALL_DIR:-$HOME/.local/bin}"
+
+info() { printf '\033[1;34m%s\033[0m\n' "$*"; }
+success() { printf '\033[1;38;5;208m%s\033[0m\n' "$*"; }
+warn() { printf '\033[1;33m%s\033[0m\n' "$*"; }
+error() { printf '\033[1;31mError: %s\033[0m\n' "$*" >&2; exit 1; }
+
+# Print JSON with syntax highlighting via jq if available, plain otherwise
+print_json() {
+    if command -v jq &>/dev/null; then
+        echo "$1" | jq .
+    else
+        echo "$1"
+    fi
+}
+
+detect_platform() {
+    local os arch target
+
+    os="$(uname -s)"
+    arch="$(uname -m)"
+
+    case "$os" in
+        Linux)
+            # Prefer musl (static) for maximum compatibility
+            case "$arch" in
+                x86_64)  target="x86_64-unknown-linux-musl" ;;
+                aarch64|arm64) target="aarch64-unknown-linux-musl" ;;
+                *) error "Unsupported architecture: $arch" ;;
+            esac
+            ;;
+        Darwin)
+            case "$arch" in
+                x86_64)  target="x86_64-apple-darwin" ;;
+                aarch64|arm64) target="aarch64-apple-darwin" ;;
+                *) error "Unsupported architecture: $arch" ;;
+            esac
+            ;;
+        MINGW*|MSYS*|CYGWIN*)
+            case "$arch" in
+                x86_64)  target="x86_64-pc-windows-msvc" ;;
+                aarch64|arm64) target="aarch64-pc-windows-msvc" ;;
+                *) error "Unsupported architecture: $arch" ;;
+            esac
+            ;;
+        *) error "Unsupported OS: $os" ;;
+    esac
+
+    echo "$target"
+}
+
+get_latest_release_tag() {
+    local target="$1"
+    local releases_json
+    releases_json=$(curl -fsSL "https://api.github.com/repos/${REPO}/releases") \
+        || error "Failed to fetch releases from https://github.com/${REPO}/releases"
+
+    # Find the first release that contains an fff-mcp binary for our platform
+    local tag
+    tag=$(echo "$releases_json" \
+        | grep -oE '"(tag_name|name)": *"[^"]*"' \
+        | awk -v target="fff-mcp-${target}" '
+            /"tag_name":/ { gsub(/.*": *"|"/, ""); current_tag = $0; next }
+            /"name":/ && index($0, target) { print current_tag; exit }
+        ')
+
+    if [ -z "$tag" ]; then
+        error "No release found containing fff-mcp binaries for ${target}. The MCP build may not have been released yet."
+    fi
+    echo "$tag"
+}
+
+download_binary() {
+    local target="$1"
+    local tag="$2"
+    local ext=""
+
+    case "$target" in
+        *windows*) ext=".exe" ;;
+    esac
+
+    local filename="${BINARY_NAME}-${target}${ext}"
+    local url="https://github.com/${REPO}/releases/download/${tag}/${filename}"
+    local checksum_url="${url}.sha256"
+
+    info "Downloading ${filename} from release ${tag}..."
+
+    local tmp_dir
+    tmp_dir="$(mktemp -d)"
+    trap 'rm -rf "$tmp_dir"' EXIT
+
+    if ! curl -fsSL -o "${tmp_dir}/${filename}" "$url" 2>/dev/null; then
+        echo "" >&2
+        printf '\033[1;31mError: Failed to download binary for your platform.\033[0m\n' >&2
+        echo "" >&2
+        echo "  URL: ${url}" >&2
+        echo "  Release: ${tag}" >&2
+        echo "  Platform: ${target}" >&2
+        echo "" >&2
+        echo "This likely means the MCP binary hasn't been built for this release yet." >&2
+        echo "Check available releases at: https://github.com/${REPO}/releases" >&2
+        exit 1
+    fi
+
+    # Verify checksum if sha256sum is available
+    if command -v sha256sum &>/dev/null; then
+        if curl -fsSL -o "${tmp_dir}/${filename}.sha256" "$checksum_url" 2>/dev/null; then
+            info "Verifying checksum..."
+            (cd "$tmp_dir" && sha256sum -c "${filename}.sha256") \
+                || error "Checksum verification failed!"
+        else
+            warn "Checksum file not available, skipping verification."
+        fi
+    fi
+
+    # Install
+    mkdir -p "$INSTALL_DIR"
+    mv "${tmp_dir}/${filename}" "${INSTALL_DIR}/${BINARY_NAME}${ext}"
+    chmod +x "${INSTALL_DIR}/${BINARY_NAME}${ext}"
+
+    if [ "$IS_UPDATE" != true ]; then
+        success "Installed ${BINARY_NAME} to ${INSTALL_DIR}/${BINARY_NAME}${ext}"
+    fi
+}
+
+check_path() {
+    case ":$PATH:" in
+        *":${INSTALL_DIR}:"*) return 0 ;;
+    esac
+
+    warn "${INSTALL_DIR} is not in your PATH."
+    echo ""
+    echo "Add it to your shell profile:"
+    echo ""
+
+    local shell_name
+    shell_name="$(basename "${SHELL:-bash}")"
+    case "$shell_name" in
+        zsh)
+            echo "  echo 'export PATH=\"${INSTALL_DIR}:\$PATH\"' >> ~/.zshrc"
+            echo "  source ~/.zshrc"
+            ;;
+        fish)
+            echo "  fish_add_path ${INSTALL_DIR}"
+            ;;
+        *)
+            echo "  echo 'export PATH=\"${INSTALL_DIR}:\$PATH\"' >> ~/.bashrc"
+            echo "  source ~/.bashrc"
+            ;;
+    esac
+    echo ""
+}
+
+print_setup_instructions() {
+    local binary_path="${INSTALL_DIR}/${BINARY_NAME}"
+    local found_any=false
+
+    echo ""
+    success "FFF MCP Server installed successfully!"
+    echo ""
+    info "Setup with your AI coding assistant:"
+    echo ""
+
+    # Claude Code
+    if command -v claude &>/dev/null; then
+        found_any=true
+        success "[Claude Code] detected"
+        echo ""
+        echo "Global (recommended):"
+        echo "claude mcp add -s user fff -- ${binary_path}"
+        echo ""
+        echo "Or project-level .mcp.json (uses PATH):"
+        echo ""
+        print_json '{
+  "mcpServers": {
+    "fff": {
+      "type": "stdio",
+      "command": "fff-mcp",
+      "args": []
+    }
+  }
+}'
+        echo ""
+    fi
+
+    # OpenCode
+    if command -v opencode &>/dev/null; then
+        found_any=true
+        success "[OpenCode] detected"
+        echo ""
+        echo "Add to ~/.config/opencode/opencode.json:"
+        echo ""
+        print_json '{
+  "mcp": {
+    "fff": {
+      "type": "local",
+      "command": ["fff-mcp"],
+      "enabled": true
+    }
+  }
+}'
+        echo ""
+    fi
+
+    # Codex
+    if command -v codex &>/dev/null; then
+        found_any=true
+        success "[Codex] detected"
+        echo ""
+        echo "codex mcp add fff -- fff-mcp"
+        echo ""
+    fi
+
+    if [ "$found_any" = false ]; then
+        echo "No AI coding assistants detected."
+        echo ""
+        echo "Binary path: ${binary_path}"
+        echo ""
+    fi
+
+    echo "Binary: ${binary_path}"
+    echo "Docs:   https://github.com/${REPO}"
+    echo ""
+    info "Tip: Add this to your CLAUDE.md or AGENTS.md to make AI use fff for all searches:"
+    echo "\""
+    echo "Use the fff MCP tools for all file search operations instead of default tools."
+    echo "\""
+
+
+}
+
+main() {
+    local target
+    target="$(detect_platform)"
+
+    local existing_binary="${INSTALL_DIR}/${BINARY_NAME}"
+    IS_UPDATE=false
+
+    if [ -x "$existing_binary" ]; then
+        IS_UPDATE=true
+        info "Updating FFF MCP Server..."
+    else
+        info "Installing FFF MCP Server..."
+    fi
+    echo ""
+
+    info "Detected platform: ${target}"
+
+    local tag
+    tag="$(get_latest_release_tag "$target")"
+
+    download_binary "$target" "$tag"
+
+    if [ "$IS_UPDATE" = true ]; then
+        echo ""
+        success "FFF MCP Server updated to ${tag}!"
+        echo ""
+    else
+        check_path
+        print_setup_instructions
+    fi
+}
+
+main
@@ -41,7 +41,7 @@ local function detect_combo_item(items, file_picker, combo_boost_score_multiplie
 end

 local function create_header_text(combo_count, win_width, disable_combo_display)
-  local combo_text = nil
+  local combo_text
  if disable_combo_display then
    combo_text = LAST_MATCH_TEXT_FORMAT
  else
@@ -57,32 +57,32 @@ end

 local function apply_header_highlights(buf, ns_id, line_idx, text_len, border_hl)
  local config = require('fff.conf').get()
-  vim.api.nvim_buf_add_highlight(buf, ns_id, border_hl, line_idx - 1, 0, -1)
-  vim.api.nvim_buf_add_highlight(
+  vim.api.nvim_buf_set_extmark(buf, ns_id, line_idx - 1, 0, { end_row = line_idx, end_col = 0, hl_group = border_hl })
+  vim.api.nvim_buf_set_extmark(
    buf,
    ns_id,
-    config.hl.combo_header,
    line_idx - 1,
    LEFT_HEADER_PADDING,
-    LEFT_HEADER_PADDING + text_len
+    { end_col = LEFT_HEADER_PADDING + text_len, hl_group = config.hl.combo_header }
  )
 end

 local function get_or_create_overlay_buf(state_key)
  if not overlay_state[state_key] or not vim.api.nvim_buf_is_valid(overlay_state[state_key]) then
+    ---@diagnostic disable-next-line: assign-type-mismatch
    overlay_state[state_key] = vim.api.nvim_create_buf(false, true)
-    vim.api.nvim_buf_set_option(overlay_state[state_key], 'bufhidden', 'wipe')
+    vim.api.nvim_set_option_value('bufhidden', 'wipe', { buf = overlay_state[state_key] })
  end
  return overlay_state[state_key]
 end

 local function update_overlay_content(buf, content, border_hl)
  -- Batch all buffer operations together for performance
-  vim.api.nvim_buf_set_option(buf, 'modifiable', true)
+  vim.api.nvim_set_option_value('modifiable', true, { buf = buf })
  vim.api.nvim_buf_set_lines(buf, 0, -1, false, { content })
  vim.api.nvim_buf_clear_namespace(buf, overlay_state.ns_id, 0, -1)
-  vim.api.nvim_buf_add_highlight(buf, overlay_state.ns_id, border_hl, 0, 0, -1)
-  vim.api.nvim_buf_set_option(buf, 'modifiable', false)
+  vim.api.nvim_buf_set_extmark(buf, overlay_state.ns_id, 0, 0, { end_row = 1, end_col = 0, hl_group = border_hl })
+  vim.api.nvim_set_option_value('modifiable', false, { buf = buf })
 end

 local function position_overlay_window(state_key, buf, width, row, col)
@@ -101,10 +101,11 @@ local function position_overlay_window(state_key, buf, width, row, col)
  if overlay_state[state_key] and vim.api.nvim_win_is_valid(overlay_state[state_key]) then
    vim.api.nvim_win_set_config(overlay_state[state_key], win_config)
  else
+    ---@diagnostic disable-next-line: assign-type-mismatch
    overlay_state[state_key] = vim.api.nvim_open_win(buf, false, win_config)
  end

-  vim.api.nvim_win_set_option(overlay_state[state_key], 'winhl', 'Normal:Normal')
+  vim.api.nvim_set_option_value('winhighlight', 'Normal:Normal', { win = overlay_state[state_key] })
 end

 local function update_overlays(list_win, combo_header_line, border_hl)
@@ -1,8 +1,82 @@
 local M = {}

--@class fff.conf.State
+--- @class FffLayoutConfig
+--- @field height number
+--- @field width number
+--- @field prompt_position string
+--- @field preview_position string
+--- @field preview_size number
+--- @field show_scrollbar boolean
+--- @field path_shorten_strategy string
+
+--- @class FffPreviewConfig
+--- @field enabled boolean
+--- @field max_size number
+--- @field chunk_size number
+--- @field binary_file_threshold number
+--- @field imagemagick_info_format_str string
+--- @field line_numbers boolean
+--- @field cursorlineopt string
+--- @field wrap_lines boolean
+--- @field filetypes table<string, table>
+
+--- @class FffKeymapsConfig
+--- @field close string
+--- @field select string
+--- @field select_split string
+--- @field select_vsplit string
+--- @field select_tab string
+--- @field move_up string|string[]
+--- @field move_down string|string[]
+--- @field preview_scroll_up string
+--- @field preview_scroll_down string
+--- @field toggle_debug string
+--- @field cycle_grep_modes string
+--- @field cycle_previous_query string
+--- @field toggle_select string
+--- @field send_to_quickfix string
+--- @field focus_list string
+--- @field focus_preview string
+
+--- @class FffFrecencyConfig
+--- @field enabled boolean
+--- @field db_path string
+
+--- @class FffHistoryConfig
+--- @field enabled boolean
+--- @field db_path string
+--- @field min_combo_count number
+--- @field combo_boost_score_multiplier number
+
+--- @class FffGrepConfig
+--- @field max_file_size number
+--- @field max_matches_per_file number
+--- @field smart_case boolean
+--- @field time_budget_ms number
+--- @field modes string[]
+
+--- @class FffConfig
+--- @field base_path string
+--- @field prompt string
+--- @field title string
+--- @field max_results number
+--- @field max_threads number
+--- @field lazy_sync boolean
+--- @field layout FffLayoutConfig
+--- @field preview FffPreviewConfig
+--- @field keymaps FffKeymapsConfig
+--- @field hl table<string, string>
+--- @field frecency FffFrecencyConfig
+--- @field history FffHistoryConfig
+--- @field git table
+--- @field debug table
+--- @field logging table
+--- @field file_picker table
+--- @field grep FffGrepConfig
+
+---@class fff.conf.State
 local state = {
-  ---@type table | nil
+  ---@type FffConfig|nil
  config = nil,
 }

@@ -99,6 +173,19 @@ local function handle_deprecated_config(user_config)
  return migrated_config
 end

+---@param name table list of highlight groups to choose from
+---@return string one of the provided groups
+local function fallback_hl(name)
+  local resolved_hl
+  for _, hl in ipairs(name) do
+    local resolved_group = vim.api.nvim_get_hl(0, { name = hl })
+
+    if not vim.tbl_isempty(resolved_group) then resolved_hl = hl end
+  end
+
+  return resolved_hl or name[#name]
+end
+
 local function init()
  local config = vim.g.fff or {}
  local default_config = {
@@ -107,14 +194,24 @@ local function init()
    title = 'FFFiles',
    max_results = 100,
    max_threads = 4,
+    lazy_sync = true, -- set to false if you want file indexing to start on open
    layout = {
      height = 0.8,
      width = 0.8,
      prompt_position = 'bottom', -- or 'top'
      preview_position = 'right', -- or 'left', 'right', 'top', 'bottom'
      preview_size = 0.5,
+      flex = { -- set to nil to disable flex layout
+        size = 130, -- column threshold: if screen width >= size, use preview_position; otherwise use wrap
+        wrap = 'top', -- position to use when screen is narrower than size
+      },
      show_scrollbar = true, -- Show scrollbar for pagination
-      path_shorten_strategy = 'middle_number', -- or 'middle', 'end'
+      -- How to shorten long directory paths in the file list:
+      -- 'middle_number' (default): uses dots for 1-3 hidden (a/./b, a/../b, a/.../b)
+      --                            and numbers for 4+ (a/.4./b, a/.5./b)
+      -- 'middle': always uses dots (a/./b, a/../b, a/.../b)
+      -- 'end': truncates from the end (home/user/projects)
+      path_shorten_strategy = 'middle_number',
    },
    preview = {
      enabled = true,
@@ -123,6 +220,7 @@ local function init()
      binary_file_threshold = 1024, -- amount of bytes to scan for binary content (set 0 to disable)
      imagemagick_info_format_str = '%m: %wx%h, %[colorspace], %q-bit',
      line_numbers = false,
+      cursorlineopt = 'both',
      wrap_lines = false,
      filetypes = {
        svg = { wrap_lines = true },
@@ -136,31 +234,35 @@ local function init()
      select_split = '<C-s>',
      select_vsplit = '<C-v>',
      select_tab = '<C-t>',
+      -- you can assign multiple keys to any action
      move_up = { '<Up>', '<C-p>' },
      move_down = { '<Down>', '<C-n>' },
      preview_scroll_up = '<C-u>',
      preview_scroll_down = '<C-d>',
      toggle_debug = '<F2>',
+      -- grep mode: cycle between plain text, regex, and fuzzy search
+      cycle_grep_modes = '<S-Tab>',
+      -- goes to the previous query in history
      cycle_previous_query = '<C-Up>',
+      -- multi-select keymaps for quickfix
      toggle_select = '<Tab>',
      send_to_quickfix = '<C-q>',
+      -- this are specific for the normal mode (you can exit it using any other keybind like jj)
      focus_list = '<leader>l',
      focus_preview = '<leader>p',
-      toggle_grep_regex = '<S-Tab>',
    },
    hl = {
      border = 'FloatBorder',
      normal = 'Normal',
-      cursor = 'CursorLine',
      matched = 'IncSearch',
      title = 'Title',
      prompt = 'Question',
-      active_file = 'Visual',
+      cursor = fallback_hl({ 'CursorLine', 'Visual' }),
      frecency = 'Number',
      debug = 'Comment',
      combo_header = 'Number',
      scrollbar = 'Comment',
-      directory_path = 'Comment', -- Highlight for directory path in file list
+      directory_path = 'Comment',
      -- Multi-select highlights
      selected = 'FFFSelected',
      selected_active = 'FFFSelectedActive',
@@ -187,39 +289,46 @@ local function init()
      git_sign_ignored_selected = 'FFFGitSignIgnoredSelected',
      -- Grep highlights
      grep_match = 'IncSearch', -- Highlight for matched text in grep results
-      grep_line_number = 'LineNr', -- Highlight for :line:col location in grep results
+      grep_line_number = 'LineNr', -- Highlight for :line:col location
      grep_regex_active = 'DiagnosticInfo', -- Highlight for keybind + label when regex is on
-      grep_regex_inactive = 'Comment', -- Highlight for keybind + label when regex is off (plain mode)
+      grep_plain_active = 'Comment', -- Highlight for keybind + label when regex is off
      grep_fuzzy_active = 'DiagnosticHint', -- Highlight for keybind + label when fuzzy is on
      -- Cross-mode suggestion highlights
      suggestion_header = 'WarningMsg', -- Highlight for the "No results found. Suggested..." banner
    },
+    -- Store file open frecency
    frecency = {
      enabled = true,
      db_path = vim.fn.stdpath('cache') .. '/fff_nvim',
    },
+    -- Store successfully opened queries with respective matches
    history = {
      enabled = true,
      db_path = vim.fn.stdpath('data') .. '/fff_queries',
      min_combo_count = 3, -- Minimum selections before combo boost applies (3 = boost starts on 3rd selection)
      combo_boost_score_multiplier = 100, -- Score multiplier for combo matches (files repeatedly opened with same query)
    },
+    -- Git integration
    git = {
      status_text_color = false, -- Apply git status colors to filename text (default: false, only sign column)
    },
    debug = {
-      enabled = false, -- Set to true to show scores in the UI
-      show_scores = false,
-      show_file_info = false, -- Show file info panel in preview
+      enabled = false, -- Show file info panel in preview
+      show_scores = false, -- Show scores inline in the UI
    },
    logging = {
      enabled = true,
      log_file = vim.fn.stdpath('log') .. '/fff.log',
      log_level = 'info',
    },
+    -- find_files settings
+    file_picker = {
+      current_file_label = '(current)',
+    },
+    -- grep settings
    grep = {
      max_file_size = 10 * 1024 * 1024, -- Skip files larger than 10MB
-      max_matches_per_file = 200, -- Maximum matches per file
+      max_matches_per_file = 100, -- Maximum matches per file (set 0 to unlimited)
      smart_case = true, -- Case-insensitive unless query has uppercase
      time_budget_ms = 150, -- Max search time in ms per call (prevents UI freeze, 0 = no limit)
      modes = { 'plain', 'regex', 'fuzzy' }, -- Available grep modes and their cycling order
@@ -233,10 +342,10 @@ local function init()
 end

 --- Setup the file picker with the given configuration
--- @param config table Configuration options
+--- @param config FffConfig Configuration options
 function M.setup(config) vim.g.fff = config end

--- @return table the fff configuration
+--- @return FffConfig the fff configuration
 function M.get()
  if not state.config then init() end
  return state.config
@@ -246,7 +355,7 @@ end
 function M.toggle_debug()
  local old_debug_state = state.config.debug.show_scores
  state.config.debug.show_scores = not state.config.debug.show_scores
-  state.config.debug.show_file_info = state.config.debug.show_scores
+  state.config.debug.enabled = state.config.debug.show_scores
  local status = state.config.debug.show_scores and 'enabled' or 'disabled'
  vim.notify('FFF debug scores ' .. status, vim.log.levels.INFO)
  return old_debug_state ~= state.config.debug.show_scores
@@ -65,12 +65,6 @@ local function setup_global_autocmds(config)
    end,
    desc = 'Automatically sync FFF directory changes',
  })
-
-  vim.api.nvim_create_autocmd('VimLeavePre', {
-    group = group,
-    callback = function() pcall(fuzzy.cleanup_file_picker) end,
-    desc = 'Cleanup FFF background threads on Neovim exit',
-  })
 end

 --- @return boolean
@@ -25,7 +25,26 @@ end
 local function binary_exists(plugin_dir)
  local binary_path = get_binary_path(plugin_dir)
  local stat = vim.uv.fs_stat(binary_path)
-  return stat and stat.type == 'file'
+  if stat and stat.type == 'file' then return true end
+
+  -- On Windows the rename over a loaded DLL fails, so a verified binary may be
+  -- left at binary_path .. '.tmp'. Promote it now that the old session is gone.
+  local tmp_path = binary_path .. '.tmp'
+  local tmp_stat = vim.uv.fs_stat(tmp_path)
+  if tmp_stat and tmp_stat.type == 'file' then
+    -- Verify the .tmp is a valid library before promoting it, in case the
+    -- process was killed between the loadlib check and the rename attempt
+    -- during a previous download, leaving a corrupt or partial .tmp on disk.
+    local loader = package.loadlib(tmp_path, 'luaopen_fff_nvim')
+    if not loader then
+      vim.uv.fs_unlink(tmp_path)
+      return false
+    end
+    local ok = vim.uv.fs_rename(tmp_path, binary_path)
+    return ok ~= nil
+  end
+
+  return false
 end

 local function download_file(url, output_path, opts, callback)
@@ -53,6 +72,12 @@ local function download_file(url, output_path, opts, callback)
      table.insert(curl_args, opts.proxy)
    end

+    if opts.extra_curl_args then
+      for _, arg in ipairs(opts.extra_curl_args) do
+        table.insert(curl_args, arg)
+      end
+    end
+
    table.insert(curl_args, url)
    vim.system(curl_args, {}, function(result)
      if result.code ~= 0 then
@@ -71,31 +96,66 @@ local function download_from_github(version, binary_path, opts, callback)
  local extension = system.get_lib_extension()
  local binary_name = triple .. '.' .. extension
  local url = string.format('https://github.com/%s/releases/download/%s/%s', GITHUB_REPO, version, binary_name)
+
  vim.schedule(function()
    vim.notify(string.format('Downloading fff.nvim binary for ' .. version), vim.log.levels.INFO)
    vim.notify(string.format('Do not open fff until you see a success notification.'), vim.log.levels.WARN)
  end)

-  download_file(url, binary_path, {
+  -- Download to a temp path first so we can validate before replacing the live binary.
+  -- If we wrote directly to binary_path and the current process already has the old
+  -- library loaded, package.loadlib() on the same path returns the *cached* handle —
+  -- meaning a truncated download would pass validation silently.
+  -- Using a distinct temp path forces dlopen to load the new file for real.
+  local tmp_path = binary_path .. '.tmp'
+
+  download_file(url, tmp_path, {
    proxy = opts.proxy,
    extra_curl_args = opts.extra_curl_args,
  }, function(success, err)
    if not success then
+      vim.uv.fs_unlink(tmp_path)
      callback(false, err)
      return
    end

-    -- Verify the binary can be loaded
-    local ok, err_msg = pcall(function() package.loadlib(binary_path, 'luaopen_fff_nvim') end)
+    vim.schedule(function()
+      -- Validate the downloaded binary by actually loading it (temp path is not yet
+      -- loaded by this process, so dlopen loads the new file for real and catches
+      -- truncated or corrupt downloads).
+      -- Note: package.loadlib returns (nil, error_string) on failure rather than throwing.
+      local loader, load_err = package.loadlib(tmp_path, 'luaopen_fff_nvim')

-    if not ok then
-      vim.uv.fs_unlink(binary_path)
-      callback(false, 'Downloaded binary is not valid: ' .. (err_msg or 'unknown error'))
-      return
-    end
+      if not loader then
+        vim.uv.fs_unlink(tmp_path)
+        callback(false, 'Downloaded binary is not valid: ' .. (load_err or 'unknown error'))
+        return
+      end

-    vim.schedule(function() vim.notify('fff.nvim binary downloaded successfully!', vim.log.levels.INFO) end)
-    callback(true, nil)
+      -- Atomically replace the live binary only after successful validation.
+      -- On Windows the old .dll may be locked by the current process, so rename can
+      -- fail if fff is already loaded. In that case, leave the verified .tmp on disk
+      -- so the next Neovim start can pick it up automatically.
+      local rename_ok, rename_err = vim.uv.fs_rename(tmp_path, binary_path)
+      if not rename_ok then
+        if vim.uv.os_uname().sysname:lower():match('windows') then
+          vim.notify(
+            'fff.nvim binary downloaded to '
+              .. tmp_path
+              .. '.\nThe live binary is locked by the current session — please restart Neovim to apply the update.',
+            vim.log.levels.WARN
+          )
+          callback(true, nil)
+        else
+          vim.uv.fs_unlink(tmp_path)
+          callback(false, 'Failed to install binary: ' .. (rename_err or 'unknown error'))
+        end
+        return
+      end
+
+      vim.notify('fff.nvim binary downloaded successfully!', vim.log.levels.INFO)
+      callback(true, nil)
+    end)
  end)
 end

@@ -131,7 +191,11 @@ function M.download_binary(callback)
      if callback then
        callback(false, err)
      else
-        error('Failed to download fff.nvim binary: ' .. (err or 'unknown error'))
+        vim.schedule(
+          function()
+            vim.notify('Failed to download fff.nvim binary: ' .. (err or 'unknown error'), vim.log.levels.ERROR)
+          end
+        )
      end
      return
    end
@@ -160,8 +224,14 @@ function M.build_binary(callback)
 end

 function M.download_or_build_binary()
+  local done = false
+  local fatal_error = nil
+
  M.ensure_downloaded({ force = true }, function(download_success, download_error)
-    if download_success then return end
+    if download_success then
+      done = true
+      return
+    end

    vim.schedule(
      function()
@@ -174,12 +244,25 @@ function M.download_or_build_binary()

    M.build_binary(function(build_success, build_error)
      if not build_success then
-        error('Failed to build fff.nvim binary. Build error: ' .. (build_error or 'unknown error'))
+        fatal_error = 'Failed to build fff.nvim binary. Build error: ' .. (build_error or 'unknown error')
      else
        vim.schedule(function() vim.notify('fff.nvim binary built successfully!', vim.log.levels.INFO) end)
      end
+      done = true
    end)
  end)
+
+  -- Block the caller (and keep the Neovim event loop alive) until the entire
+  -- download-or-build chain finishes.  This is critical for lazy.nvim build
+  -- hooks: lazy returns from the hook immediately after this function returns,
+  -- and if Neovim exits before the final rename(tmp → libfff_nvim.{dylib,so,dll})
+  -- executes, the binary is never written to disk.  vim.wait pumps the event
+  -- loop so all vim.system / vim.schedule callbacks can fire.
+  local timeout_ms = 1000 * 60 * 2 -- 2 minutes
+  local ok, wait_err = vim.wait(timeout_ms, function() return done end, 100)
+  if not ok and wait_err == -2 then error('fff.nvim: download_or_build_binary timed out') end
+
+  if fatal_error then error(fatal_error) end
 end

 function M.get_binary_path()
@@ -44,7 +44,7 @@ function M.get_directory_icon(dirname)
    end
  elseif M.provider_name == 'mini.icons' then
    if M.provider.get then
-      local icon, hl, is_default = M.provider.get('directory', basename)
+      local icon, hl, _ = M.provider.get('directory', basename)
      if icon and icon ~= '' and hl then return icon, hl end
    end
  end
@@ -25,11 +25,11 @@ local function reserve_image_buffer_space(bufnr, metadata_lines_count)
    table.insert(buffer_lines, '')
  end

-  local was_modifiable = vim.api.nvim_buf_get_option(bufnr, 'modifiable')
-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
+  local was_modifiable = vim.api.nvim_get_option_value('modifiable', { buf = bufnr })
+  vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })

  vim.api.nvim_buf_set_lines(bufnr, 0, -1, false, buffer_lines)
-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', was_modifiable)
+  vim.api.nvim_set_option_value('modifiable', was_modifiable, { buf = bufnr })

  return metadata_lines_count or 2
 end
@@ -44,9 +44,9 @@ local function update_metadata_lines(bufnr, info_lines, reserved_lines_count)
    metadata_lines[i] = info_lines[i] or ''
  end

-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
+  vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })
  vim.api.nvim_buf_set_lines(bufnr, 0, reserved_lines_count, false, metadata_lines)
-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
+  vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
 end

 local function identify_image_lines_async(file_path, bufnr, callback)
@@ -144,13 +144,12 @@ end
 --- Fully asynchronous
 --- @param file_path string Path to the image file
 --- @param bufnr number Buffer number to display in
--- @param max_width number Maximum width in characters
--- @param max_height number Maximum height in characters
 --- @return boolean
-function M.display_image(file_path, bufnr, max_width, max_height)
-  max_width = max_width or 80
-  max_height = max_height or 24
-  vim.api.nvim_buf_set_option(bufnr, 'number', false)
+function M.display_image(file_path, bufnr)
+  local wins = vim.fn.win_findbuf(bufnr)
+  for _, win in ipairs(wins) do
+    vim.api.nvim_set_option_value('number', false, { win = win })
+  end

  local reserved_metadata_lines = reserve_image_buffer_space(bufnr, 2)
  local image_content_starts_at_line = reserved_metadata_lines + 1
@@ -167,9 +166,9 @@ function M.display_image(file_path, bufnr, max_width, max_height)
      '',
      'snacks.nvim plugin is not installed or not available.',
    }
-    vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
+    vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })
    vim.api.nvim_buf_set_lines(bufnr, image_content_starts_at_line, -1, false, error_lines)
-    vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
+    vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
    return false
  end

@@ -180,9 +179,9 @@ function M.display_image(file_path, bufnr, max_width, max_height)
      'Terminal does not support image preview.',
      'Please use a terminal that supports images, such as Kitty, Wezterm or Alacritty.',
    }
-    vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
+    vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })
    vim.api.nvim_buf_set_lines(bufnr, image_content_starts_at_line, -1, false, error_lines)
-    vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
+    vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
    return false
  end

@@ -193,9 +192,9 @@ function M.display_image(file_path, bufnr, max_width, max_height)
      'File format is not supported for image preview.',
      'File: ' .. vim.fn.fnamemodify(file_path, ':t'),
    }
-    vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
+    vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })
    vim.api.nvim_buf_set_lines(bufnr, image_content_starts_at_line, -1, false, error_lines)
-    vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
+    vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
    return false
  end

@@ -51,7 +51,7 @@ end
 --- @param max_threads number|nil Maximum number of threads to use
 --- @param min_combo_count_override number|nil Optional override for min_combo_count (nil uses config)
 --- @param page_index number Page index (0-based: 0, 1, 2, ...)
--- @param page_size number Items per page
+--- @param page_size number|nil Items per page (nil uses config default)
 --- @return table List of matching files
 function M.search_files_paginated(query, current_file, max_threads, min_combo_count_override, page_index, page_size)
  local config = require('fff.conf').get()
@@ -2,51 +2,26 @@ local utils = require('fff.utils')
 local file_picker = require('fff.file_picker')
 local image = require('fff.file_picker.image')
 local location_utils = require('fff.location_utils')
+local rust = require('fff.rust')

 local M = {}

-- Additional fallback for certain ambiguous filetypes which vim.filetype.match is not handling correctly
-local function get_fixed_filetype_detection(extension)
-  local extension_map = {
-    ts = 'typescript',
-    tex = 'latex',
-    md = 'markdown',
-    txt = 'text',
-  }
-
-  return extension_map[extension]
-end
-
-local function detect_filetype(file_path)
-  local has_plenary, plenary_filetype = pcall(require, 'plenary.filetype')
-  if has_plenary then
-    local detected = plenary_filetype.detect(file_path)
-    if detected and detected ~= '' then return detected end
-  end
-
-  local builtin_filetype = vim.filetype.match({ filename = file_path })
-  if builtin_filetype and builtin_filetype ~= '' then return builtin_filetype end
-
-  local extension = vim.fn.fnamemodify(file_path, ':e'):lower()
-  return get_fixed_filetype_detection(extension)
-end
-
 local function set_buffer_lines(bufnr, lines)
  if not bufnr or not vim.api.nvim_buf_is_valid(bufnr) then return end

-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
+  vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })
  vim.api.nvim_buf_set_lines(bufnr, 0, -1, false, lines)
-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
+  vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
 end

 local function append_buffer_lines(bufnr, lines)
  if not bufnr or not vim.api.nvim_buf_is_valid(bufnr) then return end
  if not lines or #lines == 0 then return end

-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
+  vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })
  local current_lines = vim.api.nvim_buf_line_count(bufnr)
  vim.api.nvim_buf_set_lines(bufnr, current_lines, current_lines, false, lines)
-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
+  vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
 end

 local function find_existing_buffer(file_path)
@@ -183,7 +158,7 @@ end
 -- Forward declaration for ensure_content_loaded_async (used in read_file_streaming_async callback)
 local ensure_content_loaded_async

-local function read_file_streaming_async(file_path, bufnr, callback)
+local function read_file_streaming_async(file_path, callback)
  local generation = M.state.preview_generation

  init_dynamic_loading_async(file_path, function(success, error_msg)
@@ -241,6 +216,13 @@ end
 ensure_content_loaded_async = function(target_line)
  if not M.state.bufnr or not vim.api.nvim_buf_is_valid(M.state.bufnr) then return end
  if not M.state.has_more_content or M.state.is_loading then return end
+  -- Guard against missing file handle: without it load_next_chunk_async returns
+  -- synchronously with empty data, which triggers apply_location_highlighting
+  -- -> ensure_content_loaded_async again, causing infinite recursion (stack overflow).
+  if not M.state.file_operation then
+    M.state.has_more_content = false
+    return
+  end

  local current_buffer_lines = vim.api.nvim_buf_line_count(M.state.bufnr)
  local buffer_needed = target_line + 50
@@ -278,7 +260,10 @@ ensure_content_loaded_async = function(target_line)
        M.apply_location_highlighting(M.state.bufnr)
      end
    else
-      -- EOF with no additional data — apply highlighting with whatever we have
+      -- EOF with no additional data — mark loading as finished to prevent
+      -- apply_location_highlighting -> ensure_content_loaded_async recursion,
+      -- then apply highlighting with whatever content we have.
+      M.state.has_more_content = false
      M.apply_location_highlighting(M.state.bufnr)
    end
  end)
@@ -288,8 +273,8 @@ local function link_buffer_content(source_bufnr, target_bufnr)
  local lines = vim.api.nvim_buf_get_lines(source_bufnr, 0, -1, false)
  set_buffer_lines(target_bufnr, lines)

-  local source_ft = vim.api.nvim_buf_get_option(source_bufnr, 'filetype')
-  if source_ft ~= '' then vim.api.nvim_buf_set_option(target_bufnr, 'filetype', source_ft) end
+  local source_ft = vim.api.nvim_get_option_value('filetype', { buf = source_bufnr })
+  if source_ft ~= '' then vim.api.nvim_set_option_value('filetype', source_ft, { buf = target_bufnr }) end

  M.state.has_more_content = false
  M.state.total_file_lines = #lines
@@ -312,11 +297,13 @@ M.state = {
  loading_chunk_size = 1000,
  is_loading = false,
  has_more_content = true,
-  file_handle = nil,
+  file_handle = nil, ---@type uv.uv_fs_t|nil
  file_operation = nil, -- Ongoing file operation: {fd?: any, file_path?: string, position?: number}
  location = nil, -- Current location data for highlighting
  location_namespace = nil, -- Namespace for location highlighting
  preview_generation = 0, -- Monotonically increasing token to detect stale async callbacks
+  is_binary_preview = false, -- Whether the current preview is a hex dump
+  hex_byte_offset = 0, -- Next byte offset for hex dump paging
 }

 --- Setup preview configuration
@@ -331,9 +318,8 @@ end

 --- Check if file is too big for initial preview (inspired by snacks.nvim)
 --- @param file_path string Path to the file
--- @param bufnr number|nil Buffer number to check (unused with dynamic loading)
 --- @return boolean True if file is too big for initial preview
-function M.is_big_file(file_path, bufnr)
+function M.is_big_file(file_path)
  -- Only check file size for early detection - no line limits with dynamic loading
  local stat = vim.uv.fs_stat(file_path)
  if stat and stat.size > M.config.max_size then return true end
@@ -358,7 +344,7 @@ function M.get_file_info(file_path)
  }

  info.extension = vim.fn.fnamemodify(file_path, ':e'):lower()
-  info.filetype = detect_filetype(file_path) or 'text'
+  info.filetype = utils.detect_filetype(file_path) or 'text'
  info.size_formatted = utils.format_file_size(info.size)
  info.modified_formatted = os.date('%Y-%m-%d %H:%M:%S', info.modified)
  info.accessed_formatted = os.date('%Y-%m-%d %H:%M:%S', info.accessed)
@@ -369,7 +355,7 @@ end
 --- Create file info content without custom borders
 --- @param file table File information from search results
 --- @param info table File system information
--- @param file_index number Index of the file in search results (for score lookup)
+--- @param file_index number|nil Index of the file in search results (for score lookup)
 --- @return table Lines for the file info content
 function M.create_file_info_content(file, info, file_index)
  local lines = {}
@@ -474,7 +460,7 @@ end
 --- @return boolean Success status
 function M.preview_file(file_path, bufnr)
  -- Early size detection to prevent memory issues
-  if M.is_big_file(file_path, bufnr) then
+  if M.is_big_file(file_path) then
    local info = M.get_file_info(file_path)
    local lines = {
      'File too large for preview',
@@ -501,10 +487,12 @@ function M.preview_file(file_path, bufnr)
    if success then
      local file_config = M.get_file_config(file_path)

-      vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
-      vim.api.nvim_buf_set_option(bufnr, 'readonly', true)
-      vim.api.nvim_buf_set_option(bufnr, 'buftype', 'nofile')
-      vim.api.nvim_buf_set_option(bufnr, 'wrap', file_config.wrap_lines or M.config.wrap_lines)
+      vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
+      vim.api.nvim_set_option_value('readonly', true, { buf = bufnr })
+      vim.api.nvim_set_option_value('buftype', 'nofile', { buf = bufnr })
+      if M.state.winid and vim.api.nvim_win_is_valid(M.state.winid) then
+        vim.api.nvim_set_option_value('wrap', file_config.wrap_lines or M.config.wrap_lines, { win = M.state.winid })
+      end

      M.state.scroll_offset = 0

@@ -522,7 +510,7 @@ function M.preview_file(file_path, bufnr)
  M.state.bufnr = bufnr
  local generation = M.state.preview_generation

-  read_file_streaming_async(file_path, bufnr, function(content, err, loading_more)
+  read_file_streaming_async(file_path, function(content, err, loading_more)
    if M.state.preview_generation ~= generation then
      -- Preview moved on to a different file, discard
      cleanup_file_operation()
@@ -544,11 +532,13 @@ function M.preview_file(file_path, bufnr)
      set_buffer_lines(bufnr, content)

      local file_config = M.get_file_config(file_path)
-      vim.api.nvim_buf_set_option(bufnr, 'filetype', info.filetype)
-      vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
-      vim.api.nvim_buf_set_option(bufnr, 'readonly', true)
-      vim.api.nvim_buf_set_option(bufnr, 'buftype', 'nofile')
-      vim.api.nvim_buf_set_option(bufnr, 'wrap', file_config.wrap_lines or M.config.wrap_lines)
+      vim.api.nvim_set_option_value('filetype', info.filetype, { buf = bufnr })
+      vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
+      vim.api.nvim_set_option_value('readonly', true, { buf = bufnr })
+      vim.api.nvim_set_option_value('buftype', 'nofile', { buf = bufnr })
+      if M.state.winid and vim.api.nvim_win_is_valid(M.state.winid) then
+        vim.api.nvim_set_option_value('wrap', file_config.wrap_lines or M.config.wrap_lines, { win = M.state.winid })
+      end

      M.state.content_height = #content
      M.state.scroll_offset = 0
@@ -567,7 +557,69 @@ function M.preview_file(file_path, bufnr)
  return true
 end

--- Preview a binary file with async file type detection
+-- Hex preview highlight support: dynamically create hl groups from "#rrggbb"
+local hex_ns = nil
+local hex_hl_cache = {}
+
+local function ensure_hex_ns()
+  if not hex_ns then hex_ns = vim.api.nvim_create_namespace('fff_hex_preview') end
+  return hex_ns
+end
+
+local function get_hex_hl_group(hex_color)
+  local cached = hex_hl_cache[hex_color]
+  if cached then return cached end
+  local group = 'FffHex_' .. hex_color:sub(2)
+  vim.api.nvim_set_hl(0, group, { fg = hex_color })
+  hex_hl_cache[hex_color] = group
+  return group
+end
+
+--- Apply hex highlight spans to a buffer
+--- @param bufnr number Buffer number
+--- @param highlights table Array of {line_0idx, col_start, col_end, "#rrggbb"}
+--- @param line_offset number Lines to add to each highlight line index (for header)
+local function apply_hex_highlights(bufnr, highlights, line_offset)
+  if not highlights or not vim.api.nvim_buf_is_valid(bufnr) then return end
+  local ns = ensure_hex_ns()
+  for _, hl in ipairs(highlights) do
+    pcall(vim.api.nvim_buf_set_extmark, bufnr, ns, hl[1] + line_offset, hl[2], {
+      end_col = hl[3],
+      hl_group = get_hex_hl_group(hl[4]),
+    })
+  end
+end
+
+--- Load a page of hex dump content from the Rust backend
+--- @param file_path string Path to the binary file
+--- @param byte_offset number Byte offset to start reading from
+--- @return table|nil Result with lines, highlights, has_more, next_offset
+local function load_hex_page(file_path, byte_offset)
+  local ok, result = pcall(rust.hex_dump, file_path, byte_offset, 4096)
+  if ok and result then return result end
+  return nil
+end
+
+--- Load more hex content when scrolling near the end of the buffer
+local function load_more_hex_content()
+  if not M.state.bufnr or not vim.api.nvim_buf_is_valid(M.state.bufnr) then return end
+  if not M.state.has_more_content or not M.state.current_file then return end
+
+  local current_lines = vim.api.nvim_buf_line_count(M.state.bufnr)
+  local result = load_hex_page(M.state.current_file, M.state.hex_byte_offset)
+  if result and result.lines and #result.lines > 0 then
+    append_buffer_lines(M.state.bufnr, result.lines)
+    M.state.hex_byte_offset = result.next_offset
+    M.state.has_more_content = result.has_more
+    M.state.content_height = vim.api.nvim_buf_line_count(M.state.bufnr)
+    M.state.loaded_lines = M.state.content_height
+    apply_hex_highlights(M.state.bufnr, result.highlights, current_lines)
+  else
+    M.state.has_more_content = false
+  end
+end
+
+--- Preview a binary file using hexyl-powered hex dump with paging
 --- @param file_path string Path to the file
 --- @param bufnr number Buffer number for preview
 --- @return boolean Success status
@@ -575,49 +627,40 @@ function M.preview_binary_file(file_path, bufnr)
  local info = M.get_file_info(file_path)
  local lines = {}

-  set_buffer_lines(bufnr, lines)
-  vim.api.nvim_buf_set_option(bufnr, 'filetype', 'text')
-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
-  vim.api.nvim_buf_set_option(bufnr, 'readonly', true)
+  M.state.is_binary_preview = true
+  M.state.hex_byte_offset = 0

+  -- Build header synchronously (file -b is fast, typically <10ms)
  if vim.fn.executable('file') == 1 then
-    local cmd = { 'file', '-b', file_path }
-    vim.system(cmd, { text = true }, function(result)
-      vim.schedule(function()
-        if not vim.api.nvim_buf_is_valid(bufnr) then return end
+    local output = vim.fn.system({ 'file', '-b', file_path })
+    if vim.v.shell_error == 0 and output then
+      local file_type = output:gsub('\n', '')
+      table.insert(lines, 'Binary file: ' .. file_type)
+      if info and info.size_formatted then table.insert(lines, 'Size: ' .. info.size_formatted) end
+      table.insert(lines, '')
+    end
+  end

-        if result.code == 0 and result.stdout then
-          local file_type = result.stdout:gsub('\n', '')
-          table.insert(lines, 'Binary file: ' .. file_type)
-          if info and info.size_formatted then table.insert(lines, 'Size: ' .. info.size_formatted) end
+  local hex_result = load_hex_page(file_path, 0)
+  if hex_result and hex_result.lines then
+    for _, hex_line in ipairs(hex_result.lines) do
+      table.insert(lines, hex_line)
+    end
+    M.state.hex_byte_offset = hex_result.next_offset
+    M.state.has_more_content = hex_result.has_more
+  end

-          if vim.fn.executable('xxd') == 1 then
-            table.insert(lines, '')
-            set_buffer_lines(bufnr, lines)
+  set_buffer_lines(bufnr, lines)
+  vim.api.nvim_set_option_value('filetype', 'text', { buf = bufnr })
+  vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
+  vim.api.nvim_set_option_value('readonly', true, { buf = bufnr })

-            local hex_cmd = { 'xxd', '-l', '8192', file_path }
-            vim.system(hex_cmd, { text = true }, function(hex_result)
-              vim.schedule(function()
-                if not vim.api.nvim_buf_is_valid(bufnr) then return end
+  M.state.content_height = #lines
+  M.state.loaded_lines = #lines

-                if hex_result.code == 0 and hex_result.stdout then
-                  local hex_lines = vim.split(hex_result.stdout, '\n')
-                  for _, line in ipairs(hex_lines) do
-                    if line:match('%S') then table.insert(lines, line) end
-                  end
-                else
-                  table.insert(lines, 'Use a hex editor or appropriate application to view this file.')
-                end
-                set_buffer_lines(bufnr, lines)
-              end)
-            end)
-          else
-            table.insert(lines, 'Use a hex editor or appropriate application to view this file.')
-            set_buffer_lines(bufnr, lines)
-          end
-        end
-      end)
-    end)
+  if hex_result and hex_result.highlights then
+    local header_lines = #lines - (hex_result.lines and #hex_result.lines or 0)
+    apply_hex_highlights(bufnr, hex_result.highlights, header_lines)
  end

  return true
@@ -629,7 +672,7 @@ end
 function M.get_file_config(file_path)
  if not M.config or not M.config.filetypes then return {} end

-  local filetype = detect_filetype(file_path) or 'text'
+  local filetype = utils.detect_filetype(file_path) or 'text'
  return M.config.filetypes[filetype] or {}
 end

@@ -645,6 +688,7 @@ function M.preview(file_path, bufnr, location, is_binary)
  M.state.preview_generation = M.state.preview_generation + 1

  if M.state.file_handle then
+    ---@diagnostic disable-next-line: undefined-field
    M.state.file_handle:close()
    M.state.file_handle = nil
  end
@@ -653,6 +697,9 @@ function M.preview(file_path, bufnr, location, is_binary)
  M.state.total_file_lines = nil
  M.state.has_more_content = true
  M.state.is_loading = false
+  M.state.hex_byte_offset = 0
+
+  M.state.is_binary_preview = false

  M.state.current_file = file_path
  M.state.bufnr = bufnr
@@ -663,10 +710,7 @@ function M.preview(file_path, bufnr, location, is_binary)

    if not M.state.winid or not vim.api.nvim_win_is_valid(M.state.winid) then return false end

-    local win_width = vim.api.nvim_win_get_width(M.state.winid) - 2
-    local win_height = vim.api.nvim_win_get_height(M.state.winid) - 2
-
-    return image.display_image(file_path, bufnr, win_width, win_height)
+    return image.display_image(file_path, bufnr)
  elseif is_binary then
    return M.preview_binary_file(file_path, bufnr)
  else
@@ -687,15 +731,19 @@ function M.scroll(lines)
  -- If scrolling down and approaching end of loaded content, try to load more
  if lines > 0 and not M.state.is_loading then
    local target_line = new_offset + win_height
-    local buffer_needed = target_line + 20 -- Load a bit ahead
+    local buffer_needed = target_line + 20

    if current_buffer_lines < buffer_needed and M.state.has_more_content then
-      -- Load more content asynchronously but don't wait for it
-      ensure_content_loaded_async(target_line)
+      if M.state.is_binary_preview then
+        load_more_hex_content()
+        -- Re-read line count after loading more
+        current_buffer_lines = vim.api.nvim_buf_line_count(M.state.bufnr)
+      else
+        ensure_content_loaded_async(target_line)
+      end
    end
  end

-  -- Use actual buffer line count for scroll calculations
  local content_height = current_buffer_lines
  local half_screen = math.floor(win_height / 2)
  local max_scroll = math.max(0, content_height + half_screen - win_height)
@@ -744,10 +792,15 @@ function M.update_file_info_buffer(file, bufnr, file_index)
  end
  set_buffer_lines(bufnr, file_info_lines)

-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', false)
-  vim.api.nvim_buf_set_option(bufnr, 'readonly', true)
-  vim.api.nvim_buf_set_option(bufnr, 'buftype', 'nofile')
-  vim.api.nvim_buf_set_option(bufnr, 'wrap', false)
+  vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
+  vim.api.nvim_set_option_value('readonly', true, { buf = bufnr })
+  vim.api.nvim_set_option_value('buftype', 'nofile', { buf = bufnr })
+
+  -- Set wrap on the window (wrap is window-local, not buffer-local)
+  local wins = vim.fn.win_findbuf(bufnr)
+  for _, win in ipairs(wins) do
+    if vim.api.nvim_win_is_valid(win) then vim.api.nvim_set_option_value('wrap', false, { win = win }) end
+  end

  return true
 end
@@ -787,10 +840,10 @@ function M.clear_buffer(bufnr)

  pcall(vim.treesitter.stop, bufnr)

-  vim.api.nvim_buf_set_option(bufnr, 'modifiable', true)
-  vim.api.nvim_buf_set_option(bufnr, 'filetype', '')
-  vim.api.nvim_buf_set_option(bufnr, 'syntax', '')
-  vim.api.nvim_buf_set_option(bufnr, 'buftype', 'nofile')
+  vim.api.nvim_set_option_value('modifiable', true, { buf = bufnr })
+  vim.api.nvim_set_option_value('filetype', '', { buf = bufnr })
+  vim.api.nvim_set_option_value('syntax', '', { buf = bufnr })
+  vim.api.nvim_set_option_value('buftype', 'nofile', { buf = bufnr })

  set_buffer_lines(bufnr, {})
 end
@@ -812,6 +865,8 @@ function M.clear()
  M.state.scroll_offset = 0
  M.state.content_height = 0
  M.state.location = nil
+  M.state.is_binary_preview = false
+  M.state.hex_byte_offset = 0
 end

 --- Apply location highlighting to the preview buffer
@@ -829,7 +884,6 @@ function M.apply_location_highlighting(bufnr)

  if not M.state.location then return end

-  -- Apply highlighting
  location_utils.highlight_location(bufnr, M.state.location, M.state.location_namespace)

  if M.state.winid and vim.api.nvim_win_is_valid(M.state.winid) then
@@ -2,30 +2,6 @@
 --- Simple renderer for file items with 2 functions: render_line and apply_highlights
 local M = {}

--- Render Context passed to renderer functions
--- @class RenderContext
--- @field config table User configuration from conf.get()
--- @field items table[] Array of file items being rendered
--- @field cursor number Current cursor position (1-based index into items)
--- @field win_height number Window height in lines
--- @field win_width number Window width in columns
--- @field max_path_width number Maximum width for file paths
--- @field debug_enabled boolean Whether debug mode is enabled (shows frecency scores)
--- @field prompt_position string Prompt position: 'top' or 'bottom'
--- @field has_combo boolean Whether combo boost is active
--- @field combo_header_line string Formatted combo header line (if has_combo)
--- @field combo_header_text_len number Length of combo header text (if has_combo)
--- @field combo_item_index number Index of item with combo (usually 1)
--- @field display_start number Start index for displayed items
--- @field display_end number End index for displayed items
--- @field iter_start number Iteration start (may differ from display_start for bottom prompt)
--- @field iter_end number Iteration end (may differ from display_end for bottom prompt)
--- @field iter_step number Iteration step (1 for top prompt, -1 for bottom prompt)
--- @field format_file_display fun(item: table, max_width: number): string, string Helper function to format filename and dir path
--- @field selected_files table<string, boolean> Map of selected file paths
--- @field query string Current search query
--- @field renderer table|nil Custom renderer (if provided via opts)
-
 --- File Item structure from Rust
 --- @class FileItem
 --- @field path string Absolute file path
@@ -37,28 +13,23 @@ local M = {}
 --- @field total_frecency_score number Total frecency score
 --- @field access_frecency_score number Access-based frecency score
 --- @field modification_frecency_score number Modification-based frecency score
--- @field git_status number|nil Git status enum (if file is in git repo)
-
--- Renderer Interface:
--- @field render_line fun(item: FileItem, ctx: RenderContext, item_idx: number): string[] Returns array of line strings
--- @field apply_highlights fun(item: FileItem, ctx: RenderContext, item_idx: number, buf: number, ns_id: number, line_idx: number, line_content: string): nil Applies highlights to the rendered line
+--- @field git_status string|nil Git status string (e.g. 'modified', 'untracked') if file is in git repo
+--- internal:
+--- @field _has_group_header boolean Internal flag for render_line to indicate if this item has a combo header line (not from Rust)

 --- Render a file item line
 --- @param item FileItem File item from Rust
--- @param ctx RenderContext Render context with all state
+--- @param ctx ListRenderContext Render context with all state
 --- @param item_idx number Item index (1-based)
 --- @return string[] Array of line strings (1 or 2 lines if combo)
 function M.render_line(item, ctx, item_idx)
  local icons = require('fff.file_picker.icons')
  local lines = {}

-  -- Check if this should have combo header (first item with combo boost)
  local has_combo = item_idx == 1 and ctx.has_combo and ctx.combo_header_line
-
  if has_combo then table.insert(lines, ctx.combo_header_line) end

-  -- Get icon
-  local icon, icon_hl_group = icons.get_icon(item.name, item.extension, false)
+  local icon, _ = icons.get_icon(item.name, item.extension, false)

  -- Build frecency indicator (debug mode only)
  local frecency = ''
@@ -100,7 +71,7 @@ end

 --- Apply highlights to a rendered line
 --- @param item FileItem File item from Rust
--- @param ctx RenderContext Render context with all state
+--- @param ctx ListRenderContext Render context with all state
 --- @param item_idx number Item index (1-based)
 --- @param buf number Buffer handle
 --- @param ns_id number Namespace ID
@@ -126,7 +97,7 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
    vim.api.nvim_buf_set_extmark(buf, ns_id, line_idx - 1, 0, {
      end_col = 0,
      end_row = line_idx,
-      hl_group = ctx.config.hl.active_file,
+      hl_group = ctx.config.hl.cursor,
      hl_eol = true,
      priority = 100,
    })
@@ -135,7 +106,13 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
  -- 2. Icon
  if icon and icon_hl_group and vim.fn.strdisplaywidth(icon) > 0 then
    local icon_hl = is_current_file and 'Comment' or icon_hl_group
-    vim.api.nvim_buf_add_highlight(buf, ns_id, icon_hl, line_idx - 1, 0, vim.fn.strdisplaywidth(icon))
+    vim.api.nvim_buf_set_extmark(
+      buf,
+      ns_id,
+      line_idx - 1,
+      0,
+      { end_col = vim.fn.strdisplaywidth(icon), hl_group = icon_hl }
+    )
  end

  -- 3. Git text color (filename)
@@ -143,15 +120,27 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
    local git_text_hl = item.git_status and git_utils.get_text_highlight(item.git_status) or nil
    if git_text_hl and git_text_hl ~= '' and not is_current_file then
      local filename_start = #icon + 1
-      vim.api.nvim_buf_add_highlight(buf, ns_id, git_text_hl, line_idx - 1, filename_start, filename_start + #filename)
+      vim.api.nvim_buf_set_extmark(
+        buf,
+        ns_id,
+        line_idx - 1,
+        filename_start,
+        { end_col = filename_start + #filename, hl_group = git_text_hl }
+      )
    end
  end

  -- 4. Frecency indicator
  if ctx.debug_enabled then
    local start_pos, end_pos = line_content:find('[⭐️🔥✨•]%d+')
-    if start_pos then
-      vim.api.nvim_buf_add_highlight(buf, ns_id, ctx.config.hl.frecency, line_idx - 1, start_pos - 1, end_pos)
+    if start_pos and end_pos then
+      vim.api.nvim_buf_set_extmark(
+        buf,
+        ns_id,
+        line_idx - 1,
+        start_pos - 1,
+        { end_col = end_pos, hl_group = ctx.config.hl.frecency }
+      )
    end
  end

@@ -161,22 +150,26 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
    if icon then
      prefix_len = prefix_len + #icon + 1 -- if icon add icon bytes + space
    end
-    vim.api.nvim_buf_add_highlight(
+    vim.api.nvim_buf_set_extmark(
      buf,
      ns_id,
-      ctx.config.hl.directory_path or 'Comment',
      line_idx - 1,
      prefix_len,
-      prefix_len + #dir_path
+      { end_col = prefix_len + #dir_path, hl_group = ctx.config.hl.directory_path }
    )
  end

  -- 6. Current file
  if is_current_file then
-    if not is_cursor then vim.api.nvim_buf_add_highlight(buf, ns_id, 'Comment', line_idx - 1, 0, -1) end
-    local virt_text_hl = is_cursor and ctx.config.hl.active_file or 'Comment'
+    local hl
+    if is_cursor then
+      hl = ctx.config.hl.cursor
+    else
+      hl = 'Comment'
+    end
+
    vim.api.nvim_buf_set_extmark(buf, ns_id, line_idx - 1, 0, {
-      virt_text = { { ' (current)', virt_text_hl } },
+      virt_text = { { ' ' .. ctx.config.file_picker.current_file_label, hl } },
      virt_text_pos = 'right_align',
    })
  end
@@ -187,10 +180,10 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
    local border_hl

    if is_cursor then
-      local base_hl = git_utils.get_border_highlight(item.git_status)
+      local base_hl = git_utils.get_border_highlight_selected(item.git_status)
      if base_hl and base_hl ~= '' then
        local border_fg = vim.fn.synIDattr(vim.fn.synIDtrans(vim.fn.hlID(base_hl)), 'fg')
-        local cursor_bg = vim.fn.synIDattr(vim.fn.synIDtrans(vim.fn.hlID(ctx.config.hl.active_file)), 'bg')
+        local cursor_bg = vim.fn.synIDattr(vim.fn.synIDtrans(vim.fn.hlID(ctx.config.hl.cursor)), 'bg')
        local temp_hl_name = 'FFFGitBorderSelected_' .. item_idx
        if border_fg ~= '' and cursor_bg ~= '' then
          vim.api.nvim_set_hl(0, temp_hl_name, { fg = border_fg, bg = cursor_bg })
@@ -199,7 +192,7 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
          border_hl = git_utils.get_border_highlight_selected(item.git_status)
        end
      else
-        border_hl = ctx.config.hl.active_file
+        border_hl = ctx.config.hl.cursor
      end
    else
      border_hl = git_utils.get_border_highlight(item.git_status)
@@ -215,7 +208,7 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
  elseif is_cursor then
    vim.api.nvim_buf_set_extmark(buf, ns_id, line_idx - 1, 0, {
      sign_text = ' ',
-      sign_hl_group = ctx.config.hl.active_file,
+      sign_hl_group = ctx.config.hl.cursor,
      priority = 1000,
    })
  end
@@ -234,13 +227,12 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
  if ctx.query and ctx.query ~= '' then
    local match_start, match_end = string.find(line_content, ctx.query, 1)
    if match_start and match_end then
-      vim.api.nvim_buf_add_highlight(
+      vim.api.nvim_buf_set_extmark(
        buf,
        ns_id,
-        ctx.config.hl.matched or 'IncSearch',
        line_idx - 1,
        match_start - 1,
-        match_end
+        { end_col = match_end, hl_group = ctx.config.hl.matched or 'IncSearch' }
      )
    end
  end
@@ -109,20 +109,20 @@ function M.setup_highlights()
  vim.cmd([[
    " Symbol highlights
    highlight default FFFGitStaged guifg=#10B981 ctermfg=2
-    highlight default FFFGitModified guifg=#F59E0B ctermfg=3  
+    highlight default FFFGitModified guifg=#F59E0B ctermfg=3
    highlight default FFFGitDeleted guifg=#EF4444 ctermfg=1
    highlight default FFFGitRenamed guifg=#8B5CF6 ctermfg=5
    highlight default FFFGitUntracked guifg=#10B981 ctermfg=2
    highlight default FFFGitIgnored guifg=#4B5563 ctermfg=8
-    
-    " Thin border highlights 
+
+    " Thin border highlights
    highlight default FFFGitSignStaged guifg=#10B981 ctermfg=2
-    highlight default FFFGitSignModified guifg=#F59E0B ctermfg=3  
+    highlight default FFFGitSignModified guifg=#F59E0B ctermfg=3
    highlight default FFFGitSignDeleted guifg=#EF4444 ctermfg=1
    highlight default FFFGitSignRenamed guifg=#8B5CF6 ctermfg=5
    highlight default FFFGitSignUntracked guifg=#10B981 ctermfg=2
    highlight default FFFGitSignIgnored guifg=#4B5563 ctermfg=8
-    
+
    " Fallback to GitSigns highlights if they exist
    highlight default link FFFGitSignStaged GitSignsAdd
    highlight default link FFFGitSignModified GitSignsChange
@@ -6,27 +6,26 @@
 local M = {}

 local file_renderer = require('fff.file_renderer')
+local tresitter_highlight = require('fff.treesitter_hl')

 --- Build the file group header line using the same layout as file_renderer.
 --- Delegates to file_renderer.render_line (with combo disabled).
---@param item table Grep match item (used for file metadata)
+---@param item FileItem Grep match
 ---@param ctx table Render context
 ---@return string The header line string
 local function build_group_header(item, ctx)
-  -- file_renderer.render_line checks (item_idx == 1 and ctx.has_combo) for combo header.
-  -- We pass item_idx=0 and disable has_combo to suppress combo logic entirely.
-  local saved_has_combo = ctx.has_combo
  ctx.has_combo = false
+  ---@diagnostic disable-next-line: param-type-mismatch
  local lines = file_renderer.render_line(item, ctx, 0)
-  ctx.has_combo = saved_has_combo
+  ctx.has_combo = false -- never has a combo in grep
  return lines[1]
 end

 --- Apply highlights for a file group header line using file_renderer.
 --- Delegates to file_renderer.apply_highlights so all highlight groups
 --- (icon, filename, git text color, directory path, git sign) match exactly.
---@param item table Grep match item
---@param ctx table Render context
+---@param item FileItem Grep match item
+---@param ctx ListRenderContext Render context
 ---@param buf number Buffer handle
 ---@param ns_id number Namespace id
 ---@param row number 0-based row in buffer (header line)
@@ -48,17 +47,17 @@ end
 local function render_match_line(item, ctx)
  local location = string.format(':%d:%d', item.line_number or 0, (item.col or 0) + 1)
  local separator = '  '
-  local raw_content = item.line_content or ''
-  local leading_ws = #raw_content - #raw_content:match('^%s*(.*)')
-  local content = vim.trim(raw_content)
+  -- vim.json.decode may return Blobs for strings with NUL bytes; coerce to string.
+  local raw_content = item.line_content
+  if type(raw_content) ~= 'string' then raw_content = raw_content and tostring(raw_content) or '' end
+  local content = raw_content

  -- Indent + location + separator + content
  local indent = ' '
-  -- Prefix is always ASCII so byte length == display width
  local prefix_display_w = #indent + #location + #separator
  local available = ctx.win_width - prefix_display_w - 2
-  local was_truncated = false
  local content_display_w = vim.fn.strdisplaywidth(content)
+
  if content_display_w > available and available > 3 then
    -- UTF-8 aware truncation: binary search for the character count that
    -- fits within the available display width (handles multi-byte and wide chars)
@@ -73,15 +72,11 @@ local function render_match_line(item, ctx)
      end
    end
    content = vim.fn.strcharpart(content, 0, lo) .. '…'
-    was_truncated = true
  end

  local line = indent .. location .. separator .. content
  local padding = math.max(0, ctx.win_width - vim.fn.strdisplaywidth(line) + 5)

-  -- Store transient data on item for highlight pass
-  item._leading_ws = leading_ws
-  item._was_truncated = was_truncated
  item._match_indent = #indent
  item._content_offset = prefix_display_w -- byte offset where content starts in the line
  item._trimmed_content = content -- trimmed content string for treesitter parsing
@@ -102,9 +97,6 @@ local function apply_match_highlights(item, ctx, item_idx, buf, ns_id, row, line
  local is_cursor = item_idx == ctx.cursor
  local indent = item._match_indent or 1

-  -- 1. Cursor line highlight — use hl_group + hl_eol instead of line_hl_group
-  -- so that higher-priority inline extmarks (IncSearch match ranges at 200)
-  -- cleanly override both fg and bg on the cursor line.
  if is_cursor then
    vim.api.nvim_buf_set_extmark(buf, ns_id, row, 0, {
      end_col = 0,
@@ -143,16 +135,16 @@ local function apply_match_highlights(item, ctx, item_idx, buf, ns_id, row, line
  -- below IncSearch match ranges (200) so search matches take precedence.
  local content_start = sep_end
  if item._trimmed_content and item.name then
-    local ts_hl = require('fff.treesitter_hl')
    -- Resolve language once per file group (cache on the render context)
    ctx._ts_lang_cache = ctx._ts_lang_cache or {}
    local lang = ctx._ts_lang_cache[item.name]
    if lang == nil then
-      lang = ts_hl.lang_from_filename(item.name) or false
+      lang = tresitter_highlight.lang_from_filename(item.name) or false
      ctx._ts_lang_cache[item.name] = lang
    end
+
    if lang then
-      local highlights = ts_hl.get_line_highlights(item._trimmed_content, lang)
+      local highlights = tresitter_highlight.get_line_highlights(item._trimmed_content, lang)
      for _, hl in ipairs(highlights) do
        local hl_start = content_start + hl.col
        local hl_end = content_start + hl.end_col
@@ -171,16 +163,14 @@ local function apply_match_highlights(item, ctx, item_idx, buf, ns_id, row, line
  -- Use extmarks with priority > cursor line (100) so IncSearch renders
  -- properly on the selected line instead of being overridden by CursorLine.
  if item.match_ranges then
-    local leading_ws = item._leading_ws or 0
    for _, range in ipairs(item.match_ranges) do
      local raw_start = range[1] or 0
      local raw_end = range[2] or 0
-      local adj_start = raw_start - leading_ws
-      local adj_end = raw_end - leading_ws
-      if adj_end > 0 then
-        adj_start = math.max(0, adj_start)
-        local hl_start = content_start + adj_start
-        local hl_end = content_start + adj_end
+
+      if raw_end > 0 then
+        raw_start = math.max(0, raw_start)
+        local hl_start = content_start + raw_start
+        local hl_end = content_start + raw_end
        if hl_start < #line_content and hl_end <= #line_content then
          pcall(vim.api.nvim_buf_set_extmark, buf, ns_id, row, hl_start, {
            end_col = hl_end,
@@ -208,11 +198,10 @@ end
 --- Render a single item's lines (called by list_renderer's generate_item_lines).
 --- Returns 2 lines [header, match] for the first match of a file group,
 --- or 1 line [match] for subsequent matches in the same file.
---@param item table Grep match item
+---@param item FileItem Grep match item
 ---@param ctx table Render context
---@param item_idx number 1-based item index
 ---@return string[]
-function M.render_line(item, ctx, item_idx)
+function M.render_line(item, ctx)
  -- Track file grouping across the render pass via ctx
  -- ctx._grep_last_file is reset each render (ctx is fresh per render_list call)
  local is_new_group = (item.path ~= ctx._grep_last_file)
@@ -221,8 +210,8 @@ function M.render_line(item, ctx, item_idx)
  local match_line = render_match_line(item, ctx)

  if is_new_group then
-    local header_line = build_group_header(item, ctx)
    item._has_group_header = true
+    local header_line = build_group_header(item, ctx)
    return { header_line, match_line }
  else
    item._has_group_header = false
@@ -233,8 +222,8 @@ end
 --- Apply highlights for rendered lines (called by list_renderer's apply_all_highlights).
 --- line_idx is the 1-based index of the item's LAST line (the match line).
 --- If the item has a group header, it's at line_idx - 1.
---@param item table Grep match item
---@param ctx table Render context
+---@param item FileItem Grep match item
+---@param ctx ListRenderContext Render context
 ---@param item_idx number 1-based item index
 ---@param buf number Buffer handle
 ---@param ns_id number Namespace id
@@ -12,6 +12,7 @@ local fuzzy = require('fff.fuzzy')
 ---@field total_files number Total indexed files
 ---@field filtered_file_count number Total searchable files after filtering
 ---@field next_file_offset number File offset to pass for the next page (0 = no more results)
+---@field regex_fallback_error string|nil Error message if regex compilation failed and search fell back to literal

 local last_result = nil

@@ -15,7 +15,7 @@
 local M = {}

 --- @class ListRenderContext
--- @field config table User configuration
+--- @field config FffConfig User configuration
 --- @field items table[] Array of data items to render
 --- @field cursor number Current cursor position (1-based index into items)
 --- @field win_height number Window height in lines
@@ -37,6 +37,7 @@ local M = {}
 --- @field selected_files table<string, boolean> Selected file paths set
 --- @field mode string|nil Current mode (nil or 'grep')
 --- @field format_file_display function Helper for formatting file display
+--- @field suggestion_source string|nil Active cross-mode suggestion source ('grep' or 'files')

 --- @class ItemLineMapping
 --- @field first number First buffer line (1-based) this item occupies
@@ -54,7 +55,7 @@ local M = {}
 --- When cross-mode suggestions are active, a suggestion banner is prepended
 --- (for top prompt) or appended (for bottom prompt) so it always appears
 --- above the suggestion items visually.
--- @param ctx ListRenderContext
+--- @param ctx table
 --- @return string[] lines Array of line strings
 --- @return table<number, ItemLineMapping> item_to_lines
 local function generate_item_lines(ctx)
@@ -68,8 +69,16 @@ local function generate_item_lines(ctx)
  local has_suggestion_header = ctx.suggestion_source ~= nil and #ctx.items > 0
  if has_suggestion_header then
    table.insert(suggestion_header_lines, '')
-    local mode_label = ctx.suggestion_source == 'grep' and 'content matches' or 'file name matches'
-    table.insert(suggestion_header_lines, '  No results found. Suggested ' .. mode_label .. ':')
+    if ctx.mode == 'grep' and ctx.suggestion_source == 'files' then
+      -- Grep mode with no results — hint about mode cycling to fuzzy search
+      local config = require('fff.conf').get()
+      local keybind = config.keymaps.cycle_grep_modes
+      if type(keybind) == 'table' then keybind = keybind[1] or '<S-Tab>' end
+      table.insert(suggestion_header_lines, '  No results, try ' .. keybind .. ' to fuzzy search')
+    else
+      local mode_label = ctx.suggestion_source == 'grep' and 'content matches' or 'file name matches'
+      table.insert(suggestion_header_lines, '  No results found. Suggested ' .. mode_label .. ':')
+    end
    table.insert(suggestion_header_lines, '')
  end

@@ -90,11 +99,9 @@ local function generate_item_lines(ctx)
    -- Renderer returns 1+ lines: virtual headers first, content line last.
    -- This contract is shared by file_renderer (combo header) and
    -- grep_renderer (file group header).
+    ---@diagnostic disable-next-line: param-type-mismatch
    local item_lines = renderer.render_line(item, ctx, i)
-
-    for _, line in ipairs(item_lines) do
-      table.insert(lines, line)
-    end
+    vim.list_extend(lines, item_lines)

    local item_end_line = #lines
    local virtual_count = item_end_line - item_start_line -- 0 if single line, 1 if header + content
@@ -120,7 +127,7 @@ end
 --- Adjusts all line indices in item_to_lines accordingly.
 --- @param lines string[] Lines array (mutated)
 --- @param item_to_lines table<number, ItemLineMapping> Mapping (mutated)
--- @param ctx ListRenderContext
+--- @param ctx table
 --- @return number padding_offset Number of empty lines prepended
 local function apply_bottom_padding(lines, item_to_lines, ctx)
  if ctx.prompt_position ~= 'bottom' then return 0 end
@@ -151,7 +158,7 @@ end
 --- never a virtual header line.
 --- @param lines string[]
 --- @param item_to_lines table<number, ItemLineMapping>
--- @param ctx ListRenderContext
+--- @param ctx table
 --- @param list_buf number Buffer handle
 --- @param list_win number Window handle
 --- @param ns_id number Namespace id
@@ -163,9 +170,9 @@ local function update_buffer_and_cursor(lines, item_to_lines, ctx, list_buf, lis
    if cursor_item then cursor_line = cursor_item.last end
  end

-  vim.api.nvim_buf_set_option(list_buf, 'modifiable', true)
+  vim.api.nvim_set_option_value('modifiable', true, { buf = list_buf })
  vim.api.nvim_buf_set_lines(list_buf, 0, -1, false, lines)
-  vim.api.nvim_buf_set_option(list_buf, 'modifiable', false)
+  vim.api.nvim_set_option_value('modifiable', false, { buf = list_buf })

  vim.api.nvim_buf_clear_namespace(list_buf, ns_id, 0, -1)

@@ -180,7 +187,7 @@ end
 --- header highlights internally via the item._has_group_header flag.
 --- @param lines string[]
 --- @param item_to_lines table<number, ItemLineMapping>
--- @param ctx ListRenderContext
+--- @param ctx table
 --- @param list_buf number
 --- @param ns_id number
 local function apply_all_highlights(lines, item_to_lines, ctx, list_buf, ns_id)
@@ -190,23 +197,24 @@ local function apply_all_highlights(lines, item_to_lines, ctx, list_buf, ns_id)
  for i = ctx.display_start, ctx.display_end do
    local item = ctx.items[i]
    local item_lines = item_to_lines[i]
-    if not item_lines then goto continue end

-    -- The content line is always the last line in the mapping
-    local line_idx = item_lines.last
-    local line_content = lines[line_idx]
+    if item_lines then
+      -- The content line is always the last line in the mapping
+      local line_idx = item_lines.last
+      local line_content = lines[line_idx]

-    if not line_content then goto continue end
-
-    renderer.apply_highlights(item, ctx, i, list_buf, ns_id, line_idx, line_content)
-    ::continue::
+      if line_content then
+        ---@diagnostic disable-next-line: param-type-mismatch
+        renderer.apply_highlights(item, ctx, i, list_buf, ns_id, line_idx, line_content)
+      end
+    end
  end
 end

 --- Render the full item list into the buffer.
 --- This is the main entry point — replaces the inline rendering in picker_ui.
 ---
--- @param ctx ListRenderContext Render context built by picker_ui
+--- @param ctx table Render context built by picker_ui
 --- @param list_buf number List buffer handle
 --- @param list_win number List window handle
 --- @param ns_id number Highlight namespace
@@ -224,8 +232,15 @@ function M.render(ctx, list_buf, list_win, ns_id)
    local suggestion_hl = ctx.config.hl.suggestion_header or 'WarningMsg'
    for i = 0, #lines - 1 do
      local line = lines[i + 1]
-      if line and line:match('^%s+No results found') then
-        pcall(vim.api.nvim_buf_add_highlight, list_buf, ns_id, suggestion_hl, i, 0, -1)
+      if line and (line:match('^%s+No results found') or line:match('^%s+No results,')) then
+        pcall(
+          vim.api.nvim_buf_set_extmark,
+          list_buf,
+          ns_id,
+          i,
+          0,
+          { end_row = i + 1, end_col = 0, hl_group = suggestion_hl }
+        )
      end
    end
  end
@@ -18,14 +18,7 @@ function M.find_files(opts)
 end

 --- Live grep: search file contents in the current directory
--- @param opts? table Optional configuration overrides
--- @param opts.cwd? string Custom working directory
--- @param opts.title? string Window title (default: "Live Grep")
--- @param opts.prompt? string Input prompt text (default: "grep> ")
--- @param opts.layout? table Layout overrides
--- @param opts.grep? table Grep-specific overrides {max_file_size, smart_case, max_matches_per_file, modes}
--- @param opts.grep.modes? table Available search modes and their cycling order (default: {'plain', 'regex', 'fuzzy'})
--- @param opts.query? string Initial search query to pre-fill
+--- @param opts? {cwd?: string, title?: string, prompt?: string, layout?: table, grep?: {max_file_size?: number, smart_case?: boolean, max_matches_per_file?: number, modes?: string[]}, query?: string} Optional configuration overrides
 function M.live_grep(opts)
  local picker_ok, picker_ui = pcall(require, 'fff.picker_ui')
  if not picker_ok then
@@ -38,17 +31,19 @@ function M.live_grep(opts)

  local grep_config = vim.tbl_deep_extend('force', config.grep or {}, (opts and opts.grep) or {})

-  local picker_opts = vim.tbl_deep_extend('force', opts or {}, {
-    title = (opts and opts.title) or 'Live Grep',
+  local picker_opts = vim.tbl_deep_extend('force', {
+    title = 'Live Grep',
    mode = 'grep',
    renderer = grep_renderer,
    grep_config = grep_config,
-    query = opts and opts.query or nil,
-  })
+    query = '',
+  }, opts or {})

  picker_ui.open(picker_opts)
 end

+--- Changes the directory indexed by the file picker to the git root and opens the file picker
+--- @deprecated Use `find_files` instead
 function M.find_in_git_root()
  local fuzzy = require('fff.core').ensure_initialized()
  local ok, git_root = pcall(fuzzy.get_git_root)
@@ -170,11 +165,12 @@ function M.find_files_in_dir(directory)
    return
  end

-  M.change_indexing_directory(directory)
-
  local picker_ok, picker_ui = pcall(require, 'fff.picker_ui')
  if picker_ok then
-    picker_ui.open({ title = 'Files in ' .. vim.fn.fnamemodify(directory, ':t') })
+    picker_ui.open({
+      title = 'Files in ' .. vim.fn.fnamemodify(directory, ':t'),
+      cwd = directory,
+    })
  else
    vim.notify('Failed to load picker UI', vim.log.levels.ERROR)
  end
@@ -184,35 +180,15 @@ end
 --- @param new_path string New directory path to use as base
 --- @return boolean `true` if successful, `false` otherwise
 function M.change_indexing_directory(new_path)
-  if not new_path or new_path == '' then
-    vim.notify('Directory path is required', vim.log.levels.ERROR)
-    return false
-  end
-
-  local expanded_path = vim.fn.expand(new_path)
-
-  if vim.fn.isdirectory(expanded_path) ~= 1 then
-    vim.notify('Directory does not exist: ' .. expanded_path, vim.log.levels.ERROR)
-    return false
-  end
-
-  local fuzzy = require('fff.core').ensure_initialized()
-  local ok, result = pcall(fuzzy.restart_index_in_path, expanded_path)
-  if not ok then
-    vim.notify('Failed to change directory: ' .. result, vim.log.levels.ERROR)
-    return false
-  end
-
-  local config = require('fff.conf').get()
-  config.base_path = expanded_path
-  return true
+  local picker_ok, picker_ui = pcall(require, 'fff.picker_ui')
+  if picker_ok then return picker_ui.change_indexing_directory(new_path) end
+  return false
 end

 --- Opens the file under the cursor with an optional callback if the only file
 --- is found and we are about to inline open it
 --- @param open_cb function|nil Optional callback function to execute after opening the file
 function M.open_file_under_cursor(open_cb)
-  local filename = vim.fn.expand('<cfile>')
  local full_path_with_suffix = vim.fn.expand('<cWORD>')

  local picker_ok, picker_ui = pcall(require, 'fff.picker_ui')
@@ -221,7 +197,7 @@ function M.open_file_under_cursor(open_cb)
    return
  end

-  picker_ui.open_with_callback(full_path_with_suffix, function(files, metadata, location, get_file_score)
+  picker_ui.open_with_callback(full_path_with_suffix, function(files, _, location)
    if #files == 1 or require('fff.file_picker').get_file_score(1).exact_match then
      if open_cb and type(open_cb) == 'function' then open_cb(files[1].path) end
      vim.api.nvim_command(string.format('e %s', vim.fn.fnameescape(files[1].path)))
@@ -38,7 +38,7 @@ function M.render(layout, config, list_win, pagination, prompt_position)
  -- rendering in a separate buffer to overflow the border
  if not scrollbar_exists then
    scrollbar_state.buf = vim.api.nvim_create_buf(false, true)
-    vim.api.nvim_buf_set_option(scrollbar_state.buf, 'bufhidden', 'wipe')
+    vim.api.nvim_set_option_value('bufhidden', 'wipe', { buf = scrollbar_state.buf })

    scrollbar_state.win = vim.api.nvim_open_win(scrollbar_state.buf, false, {
      relative = 'editor',
@@ -52,7 +52,7 @@ function M.render(layout, config, list_win, pagination, prompt_position)
    })

    local scrollbar_hl = string.format('Normal:%s', config.hl.border)
-    vim.api.nvim_win_set_option(scrollbar_state.win, 'winhighlight', scrollbar_hl)
+    vim.api.nvim_set_option_value('winhighlight', scrollbar_hl, { win = scrollbar_state.win })

    scrollbar_state.ever_shown = true
  end
@@ -83,9 +83,9 @@ function M.render(layout, config, list_win, pagination, prompt_position)
    end
  end

-  pcall(vim.api.nvim_buf_set_option, scrollbar_state.buf, 'modifiable', true)
+  pcall(vim.api.nvim_set_option_value, 'modifiable', true, { buf = scrollbar_state.buf })
  pcall(vim.api.nvim_buf_set_lines, scrollbar_state.buf, 0, -1, false, lines)
-  pcall(vim.api.nvim_buf_set_option, scrollbar_state.buf, 'modifiable', false)
+  pcall(vim.api.nvim_set_option_value, 'modifiable', false, { buf = scrollbar_state.buf })

  pcall(vim.api.nvim_buf_clear_namespace, scrollbar_state.buf, ns_id, 0, -1)
  if thumb_size > 0 then
@@ -1,7 +1,5 @@
--- Treesitter Highlight Extraction
--- Extracts syntax highlights from a code string using treesitter.
--- Uses a per-language scratch buffer pool to avoid repeated buffer creation.
--- Results are returned as extmark-style tables { col, end_col, hl_group }.
+local utils = require('fff.utils')
+
 local M = {}

 --- Per-language scratch buffer cache
@@ -33,11 +31,7 @@ end
 function M.lang_from_filename(filename)
  if not filename or filename == '' then return nil end

-  -- Use vim.filetype.match to get the filetype from the filename
-  local ok, ft = pcall(vim.filetype.match, { filename = filename })
-  if not ok or not ft then return nil end
-
-  -- Convert filetype to treesitter language
+  local ft = utils.detect_filetype(filename) or 'text'
  local lang_ok, lang = pcall(vim.treesitter.language.get_lang, ft)
  if not lang_ok or not lang then lang = ft end

@@ -83,7 +77,7 @@ function M.get_line_highlights(text, lang)
    local query_ok, query = pcall(vim.treesitter.query.get, tree_lang, 'highlights')
    if not query_ok or not query then return end

-    for capture, node, metadata in query:iter_captures(root, buf, 0, 1) do
+    for capture, node, _ in query:iter_captures(root, buf, 0, 1) do
      local name = query.captures[capture]
      if name and name ~= 'spell' and name ~= 'conceal' then
        local start_row, start_col, end_row, end_col = node:range()
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Dmitriy Kovalenko	398f93759c	feat: Migrate to stable rust docs / docs (push) Has been cancelled Details	2026-03-18 15:42:33 -07:00
Dmitriy Kovalenko	2b6ace888d	feat: SDKs for Nodejs FFI, Rust, and C (#295 ) * feat: Nodejs FFI package * chore: Update docs for - feat: Nodejs FFI package	2026-03-18 15:39:43 -07:00
Dmitriy Kovalenko	4b83987ea1	feat: Colorized hex dump for binary file previews (#291 ) * feat: Colorized hex dump for binary file previews * chore: Update docs for - feat: Colorized hex dump for binary file previews	2026-03-17 12:27:48 -07:00
Dmitriy Kovalenko	c6cb66b597	fix: Windows dir locking (#293 ) * fix: Windows dir locking * chore: Update docs for - fix: Windows dir locking	2026-03-17 10:40:09 -07:00
George	64861f8142	chore(readme) Fix install-fff.sh link to install-mcp.sh (#283 ) * install-fff.sh -> install.mcp.sh * install-fff.sh -> install.mcp.sh	2026-03-15 08:00:18 -07:00
Mohamed Ibraheem	eb5f2b3648	refactor: move change_indexing_directory to picker_ui module (#286 ) * refactor: move change_indexing_directory to picker_ui module closes #237 * chore: format * chore: format	2026-03-15 07:59:38 -07:00
Dmitriy Kovalenko	fcdf4a9172	feat(mcp): Make install script understand upgrade use case	2026-03-13 16:45:21 -07:00
Dmitriy Kovalenko	1001eb8b5e	feat(mcp): Add file path fallback for grep (#281 ) * feat(mcp): Add file path fallback for grep * chore: Update docs for - feat(mcp): Add file path fallback for grep	2026-03-13 16:29:24 -07:00
Dmitriy Kovalenko	f0ce2dd50d	chore: fix format	2026-03-13 09:58:27 -07:00
Dmitriy Kovalenko	1c2a1c1204	fix(grep): file path single file constraint parsing (#280 )	2026-03-12 17:10:28 -07:00
Salomon Popp	66bdfff454	fix: correct typo in `vim.uv.fs_unlink()` (#279 )	2026-03-12 16:51:20 -07:00
Dmitriy Kovalenko	1e50f8df80	fix: Prevent mcp from running root large folders (#278 )	2026-03-12 08:35:27 -07:00
Dmitriy Kovalenko	736c41ecd6	feat: MCP (#272 ) * feat: MCP * chore: Update docs for - feat: MCP * feat: Perf optimizations * chore: Update docs for - feat: Perf optimizations	2026-03-11 21:27:11 -07:00
Dmitriy Kovalenko	ac8df4c9e4	chore: Update docs for - feat: Show the actual matches for typo resistant fuzzy search	2026-03-06 13:31:12 -08:00
Dmitriy Kovalenko	e3e534f4ad	feat: Show the actual matches for typo resistant fuzzy search	2026-03-06 13:31:12 -08:00
Dmitriy Kovalenko	764e3ecf18	fix: Race when openning many sequenital dirs closes https://github.com/dmtrKovalenko/fff.nvim/issues/269 closes https://github.com/dmtrKovalenko/fff.nvim/issues/256	2026-03-06 13:31:12 -08:00
Nick Kadutskyi	10a27f9678	fix: proper git hl group for sign on selected file (#268 )	2026-03-03 20:27:03 -08:00
Nico Domino	66e712e066	fix: prompt_position=bottom fix scrolling of viewport to top (#267 )	2026-03-03 09:18:56 -08:00
Dmitriy Kovalenko	c5a3f89c5e	feat: Make zig optional for building (#266 ) * feat: Make zig optional for building * chore: Update docs for - feat: Make zig optional for building	2026-02-28 14:13:07 -08:00
bansheerubber	29f81a141f	feat: add cursorlineopt config option for styling livegrep file preview (#262 ) * feat: add cursorlineopt config option for styling livegrep file preview * fix: better validation for junk `cursorlineopt` values	2026-02-27 21:00:30 -08:00
Dmitriy Kovalenko	d2e3993398	feat: cross-compile for android (! for termux users) (#264 ) * feat: cross-compile for android (! for termux users) * chore: Update docs for - feat: cross-compile for android (! for termux users)	2026-02-27 20:57:35 -08:00
bansheerubber	a411100fa4	chore(docs): replace flex = nil with flex = false in README.md (#263 )	2026-02-27 17:57:26 -08:00
Italo Silva	0a993d7a30	feat!: Make layout flex on small screens & responsive (#255 ) * feat: Add `flex` option for preview * chore: Update docs for - feat: Add `flex` option for preview * feat: Recalculate layout on resize --------- Co-authored-by: Dmitriy Kovalenko <dmtr.kovalenko@outlook.com>	2026-02-25 18:36:20 -08:00
Dmitriy Kovalenko	7c2d46f1e2	chore: Upgrade every used neovim deprecated lua apis (#261 ) * chore: Do not use any neovim deprecated apis * chore: Update docs for - chore: Do not use any neovim deprecated apis	2026-02-25 17:44:13 -08:00
Dmitriy Kovalenko	c3ed9fb17b	fix: Binary download integrity issues (#260 ) * feat: cleanup hl groups resolution (#259) * feat: cleanup hl groups resolution * chore: Update docs for - feat: cleanup hl groups resolution * chore: Update docs for - feat: cleanup hl groups resolution (#259) * feat: cleanup hl groups resolution (#259) * feat: cleanup hl groups resolution * chore: Update docs for - feat: cleanup hl groups resolution	2026-02-24 17:32:10 -08:00
Dmitriy Kovalenko	0f40c66eb7	feat: cleanup hl groups resolution (#259 ) * feat: cleanup hl groups resolution * chore: Update docs for - feat: cleanup hl groups resolution	2026-02-24 09:54:24 -08:00
Nicolò Francesco Maria Spingola	7298978bcb	fix(download): correctly parse certutil output on Windows (#258 )	2026-02-23 06:57:07 -08:00
Federico	8f69f987a4	docs: update readme example keybindings (#253 ) * docs: update readme example keybindings * update indenntation Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>	2026-02-20 07:41:24 -08:00
Dmitriy Kovalenko	3e9b8655b7	fix: Atomic download checks for binaries + codesign for macos (#252 )	2026-02-19 22:16:22 -08:00
Dmitriy Kovalenko	727935ede6	fix: Macos background watcher memory usage (#251 ) * fix: Macos background watcher memory usage closes https://github.com/dmtrKovalenko/fff.nvim/issues/239 Fixes a memory leak of the backgorund watcher on macos when the internal fsevents buffer is droping certain elements * chore: Update docs for - fix: Macos background watcher memory usage	2026-02-19 18:10:59 -08:00
Dmitriy Kovalenko	9a2612b1b5	feat: Grep perfromance and stability improvements (#235 ) * feat: Allow many fff consumers over ffi * parallize grep * chore: Update docs for - parallize grep	2026-02-19 15:59:23 -08:00