fix: Single parenthesis error

fixes https://github.com/dmtrKovalenko/fff.nvim/issues/351

.github/workflows/external-tests.yml

@@ -9,6 +9,9 @@ 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:
@@ -25,7 +28,7 @@ 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
@@ -34,11 +37,11 @@ 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 || '' }}
 
@@ -87,6 +90,27 @@ jobs:
         shell: bash
         run: make test-lua
 
+      - name: Run version resolution tests
+        shell: bash
+        run: make test-version
+
       - 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

.github/workflows/lua.yml

@@ -11,7 +11,7 @@ jobs:
     name: lua-language-server type check
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
 
       - name: Install Neovim
         run: |
@@ -38,7 +38,7 @@ jobs:
     name: luacheck lint
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
 
       - name: Install luacheck
         run: |

.github/workflows/nix.yml

@@ -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

.github/workflows/panvimdoc.yaml

@@ -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

.github/workflows/release.yaml

@@ -2,9 +2,14 @@ name: Prebuild
 
 on:
   push:
-    branches: [main, feat/interchangable-ffi, feat/termux]
+    branches: [main, fix/download-version]
+    tags:
+      - "v*"
   pull_request:
 
+env:
+  FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true
+
 jobs:
   build-nvim:
     name: Build Neovim ${{ matrix.target }}
@@ -62,7 +67,7 @@ jobs:
             ext: dll
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
         with:
           persist-credentials: false
 
@@ -134,26 +139,26 @@ 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
 
@@ -168,13 +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
 
@@ -182,18 +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
 
@@ -265,16 +270,105 @@ jobs:
           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: mlugg/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.repository == 'dmtrKovalenko/fff.nvim'
+    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: Install Lua
+        uses: leafo/gh-actions-lua@v12
 
       - name: Download artifacts
         uses: actions/download-artifact@v4
@@ -311,6 +405,19 @@ 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: |
@@ -326,25 +433,22 @@ 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: lua scripts/determine-version.lua
 
       - 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
@@ -352,39 +456,67 @@ 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/fix/download-version' || startsWith(github.ref, 'refs/tags/v')))
+      || (github.event_name == 'pull_request' && github.head_ref == 'main')
+
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install Lua
+        uses: leafo/gh-actions-lua@v12
+
+      - name: Install Rust
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Install cargo-edit
+        run: cargo install cargo-edit
+
+      - name: Determine version
+        id: version
+        run: lua scripts/determine-version.lua
+
+      - 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/interchangable-ffi'))
-      || (github.event_name == 'pull_request' && (github.head_ref == 'main' || github.head_ref == 'feat/interchangable-ffi'))
+      (github.event_name == 'push' && (github.ref == 'refs/heads/main' || github.ref == 'refs/heads/fix/download-version' || startsWith(github.ref, 'refs/tags/v')))
+      || (github.event_name == 'pull_request' && github.head_ref == 'main')
     permissions:
       contents: read
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
+
+      - name: Install Lua
+        uses: leafo/gh-actions-lua@v12
 
       - 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: lua scripts/determine-version.lua
 
       - name: Download npm package artifacts
         uses: actions/download-artifact@v4
@@ -397,94 +529,45 @@ 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
-    # comments doesn't work on forks
-    if: github.event_name == 'pull_request' && github.repository == 'dmtrKovalenko/fff.nvim'
-    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 ${{ github.sha }}*
+          cd packages/fff-node
+          npm install
+          npm run build
+          npm publish --tag "$TAG" --access public || echo "Failed to publish @ff-labs/fff-node (may already exist)"

.github/workflows/rust.yml

@@ -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
@@ -37,15 +37,13 @@ jobs:
           components: rustfmt, clippy
 
       - name: Run tests
-        run: |
-          cargo test --verbose -p fff-core -p fff-query-parser -p fff-c --features zlob
-          cargo test --verbose -p grep-searcher
+        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:
@@ -58,11 +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
 
@@ -73,6 +71,4 @@ jobs:
           components: clippy
           
       - name: Run clippy
-        run: |
-          cargo clippy -p fff-core -p fff-query-parser -p fff-nvim -p fff-c --features zlob -- -D warnings
-          cargo clippy -p grep-searcher -- -D warnings
+        run: cargo clippy -- -D warnings

.github/workflows/spelling.yaml

@@ -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
 

.github/workflows/stylua.yaml

@@ -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
 

.gitignore

@@ -1,4 +1,5 @@
 doc/tags
+big-repo
 target/
 .archive.lua
 _*.lua
@@ -10,6 +11,15 @@ result
 .repro/
 .wrangler/
 *.so
-big-repo/
+*.dylib
 # all the perf like utility files
 *.data
+node_modules/
+
+dist/
+scripts/benchmark-results/
+
+# Native binaries (downloaded at install)
+*.dylib
+*.so
+*.dll

.mcp.json

@@ -0,0 +1,9 @@
+{
+  "mcpServers": {
+    "fff": {
+      "type": "stdio",
+      "command": "./target/release/fff-mcp",
+      "args": []
+    }
+  }
+}

Cargo.toml

@@ -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 = "0.5.1", path = "crates/fff-grep" }
+fff-query-parser = { version = "0.5.1", path = "crates/fff-query-parser", default-features = false }
+
 # Shared dependencies
 ahash = "0.8"
 bindet = "0.3"
@@ -23,8 +27,6 @@ git2 = { version = "0.20.2", default-features = false, features = [
 ] }
 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"
@@ -32,7 +34,7 @@ mimalloc = "0.1.47"
 zlob = "1.3.0"
 
 mlua = { version = "0.11.1", features = ["module", "luajit"] }
-neo_frizbee = "0.8.1"
+neo_frizbee = { version = "0.9.1", features = ["match_end_col"] }
 notify = "8.1.0"
 notify-debouncer-full = "0.7"
 once_cell = "1.20.2"

Makefile

@@ -1,10 +1,15 @@
 PLENARY_DIR ?= ../plenary.nvim
 
-.PHONY: build test test-rust test-lua test-bun test-setup prepare-bun
+.PHONY: build test test-rust test-lua test-version test-bun test-node prepare-bun prepare-node set-npm-version header
+
+all: format test lint
 
 build:
 	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 \
 		echo "Cloning plenary.nvim..."; \
@@ -12,12 +17,16 @@ test-setup:
 	fi
 
 test-rust:
-	cargo test --workspace --features zlob
+	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-version: test-setup
+	nvim --headless -u tests/minimal_init.lua \
+		-c "PlenaryBustedFile tests/version_spec.lua" 2>&1
+
 prepare-bun: build
 	mkdir -p packages/fff-bun/bin
 	cp target/release/libfff_c.dylib packages/fff-bun/bin/ 2>/dev/null; \
@@ -25,14 +34,65 @@ prepare-bun: build
 	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: test-rust test-lua test-bun
+test-node: prepare-node
+	cd packages/fff-node && npm run build && node test/e2e.mjs
+
+test: test-rust test-lua test-version 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

README.md

@@ -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
@@ -253,9 +266,8 @@ require('fff').setup({
       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,
@@ -282,18 +294,19 @@ require('fff').setup({
 #### Available Methods
 
 ```lua
-require('fff').find_files()                         -- Find files in current repositro
+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
@@ -301,10 +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:
@@ -365,6 +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:
@@ -458,41 +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

biome.json

@@ -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"
+      }
+    }
+  }
+}

bun.lock

@@ -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=="],
+  }
+}

crates/fff-c/Cargo.toml

@@ -1,8 +1,8 @@
 [package]
 name = "fff-c"
-version = "0.1.0"
+version = "0.5.2"
 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]
@@ -10,14 +10,14 @@ crate-type = ["cdylib"]
 
 [features]
 default = []
-zlob = ["fff-core/zlob"]
+zlob = ["fff/zlob"]
 
 [dependencies]
 mimalloc.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 = "0.5.1" }
+fff-query-parser = { path = "../fff-query-parser" , version = "0.5.2" }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"

crates/fff-c/cbindgen.toml

@@ -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"

crates/fff-c/include/fff.h

@@ -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 */

crates/fff-c/src/ffi_types.rs

@@ -1,148 +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, 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,
-    /// Opaque handle pointer (used by fff_create to return the instance)
-    pub handle: *mut c_void,
-}
-
-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(),
-            handle: 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(),
-            handle: ptr::null_mut(),
-        }))
-    }
-
-    /// Create a successful result carrying an opaque instance handle.
-    pub fn ok_handle(handle: *mut c_void) -> *mut Self {
-        Box::into_raw(Box::new(FffResult {
-            success: true,
-            data: ptr::null_mut(),
-            error: ptr::null_mut(),
-            handle,
-        }))
-    }
-
-    /// 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(),
-            handle: ptr::null_mut(),
-        }))
-    }
-}
-
-/// 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_str()),
+            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,
+            access_frecency_score: item.access_frecency_score as i64,
+            modification_frecency_score: item.modification_frecency_score as i64,
+            total_frecency_score: item.total_frecency_score() as i64,
+            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,
@@ -152,12 +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,
@@ -167,186 +141,397 @@ 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_str()),
+            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,
+            total_frecency_score: file.total_frecency_score() as i64,
+            access_frecency_score: file.access_frecency_score as i64,
+            modification_frecency_score: file.modification_frecency_score as i64,
             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.matches.len(),
-            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 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,
+    pub is_watcher_ready: bool,
+    pub is_warmup_complete: bool,
+}
+
+impl From<fff::file_picker::ScanProgress> for FffScanProgress {
+    fn from(p: fff::file_picker::ScanProgress) -> Self {
+        Self {
+            scanned_files_count: p.scanned_files_count as u64,
+            is_scanning: p.is_scanning,
+            is_watcher_ready: p.is_watcher_ready,
+            is_warmup_complete: p.is_warmup_complete,
         }
     }
 }

crates/fff-core/Cargo.toml

@@ -1,9 +1,10 @@
 [package]
-name = "fff-core"
-version = "0.1.0"
+name = "fff-search"
+version = "0.5.2"
 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"
@@ -13,31 +14,32 @@ crate-type = ["rlib", "staticlib", "cdylib"]
 default = []
 # Enable C FFI exports
 ffi = []
+# Call mi_collect(true) after large allocator churn (bigram build).
+# Requires mimalloc to be the global allocator (linked by fff-nvim).
+mimalloc-collect = ["dep:libmimalloc-sys"]
 # 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", default-features = false }
+fff-query-parser = { workspace = true  , version = "0.5.2" }
 
 # External dependencies
 bindet = { workspace = true }
 blake3 = { workspace = true }
 chrono = { workspace = true }
 dirs = { workspace = true }
+libc = "0.2"
 git2 = { workspace = true }
 glidesort = { workspace = true }
 globset = { workspace = true }
-grep-matcher = { workspace = true }
-grep-searcher = { workspace = true }
+fff-grep = { workspace = true  , version = "0.5.2" }
+aho-corasick = "1"
 memchr = "2"
 heed = { workspace = true }
 ignore = { workspace = true }
@@ -50,10 +52,13 @@ 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, optional = true }
+libmimalloc-sys = { version = "0.1", optional = true, features = ["extended"] }
 # Platform-specific: dunce for Windows to avoid \\?\ extended path prefix
 [target.'cfg(windows)'.dependencies]
 dunce = { workspace = true }
@@ -62,3 +67,15 @@ dunce = { workspace = true }
 criterion = { version = "0.5", features = ["html_reports"] }
 rand = { version = "0.8", features = ["small_rng"] }
 tempfile = "3.8"
+
+[[bench]]
+name = "parse_bench"
+harness = false
+
+[[bench]]
+name = "bigram_bench"
+harness = false
+
+[[bench]]
+name = "memmem_bench"
+harness = false

crates/fff-core/README.md

@@ -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

crates/fff-core/benches/bigram_bench.rs

@@ -0,0 +1,129 @@
+use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
+use fff_search::bigram_filter::{BigramFilter, BigramIndexBuilder};
+
+/// Build a realistic bigram index for benchmarking.
+/// Simulates a large repo by generating varied content per file.
+fn build_test_index(file_count: usize) -> BigramFilter {
+    let builder = BigramIndexBuilder::new(file_count);
+    let skip_builder = BigramIndexBuilder::new(file_count);
+
+    for i in 0..file_count {
+        // Generate varied content so we get a mix of sparse and dense columns
+        let content = format!(
+            "struct File{i} {{ fn process() {{ let controller = read(path); }} }} // module {i}"
+        );
+        builder.add_file_content(&skip_builder, i, content.as_bytes());
+    }
+
+    let mut index = builder.compress(None);
+    let skip_index = skip_builder.compress(Some(12));
+    index.set_skip_index(skip_index);
+    index
+}
+
+fn bench_bigram_query(c: &mut Criterion) {
+    let file_counts = [10_000, 100_000, 500_000];
+
+    for &file_count in &file_counts {
+        let index = build_test_index(file_count);
+        eprintln!(
+            "Index ({} files): {} columns",
+            file_count,
+            index.columns_used(),
+        );
+
+        let mut group = c.benchmark_group(format!("bigram_query_{file_count}"));
+        group.sample_size(500);
+
+        let queries: &[(&str, &[u8])] = &[
+            ("short_2char", b"st"),
+            ("medium_6char", b"struct"),
+            ("long_14char", b"let controller"),
+            ("multi_word", b"fn process"),
+        ];
+
+        for (name, query) in queries {
+            group.bench_with_input(BenchmarkId::from_parameter(name), query, |b, q| {
+                b.iter(|| {
+                    let result = index.query(black_box(q));
+                    black_box(&result);
+                });
+            });
+        }
+
+        group.finish();
+    }
+}
+
+fn bench_bigram_is_candidate(c: &mut Criterion) {
+    let index = build_test_index(500_000);
+    let candidates = match index.query(b"struct") {
+        Some(c) => c,
+        None => {
+            // All bigrams ubiquitous at this size — skip candidate benches
+            eprintln!("Skipping is_candidate bench: query returned None (all bigrams ubiquitous)");
+            return;
+        }
+    };
+
+    c.bench_function("is_candidate_500k", |b| {
+        b.iter(|| {
+            let mut count = 0u32;
+            for i in 0..500_000 {
+                if BigramFilter::is_candidate(black_box(&candidates), i) {
+                    count += 1;
+                }
+            }
+            black_box(count)
+        });
+    });
+
+    c.bench_function("count_candidates_500k", |b| {
+        b.iter(|| BigramFilter::count_candidates(black_box(&candidates)));
+    });
+}
+
+fn bench_bigram_build(c: &mut Criterion) {
+    let mut group = c.benchmark_group("bigram_build");
+    group.sample_size(10);
+
+    let file_counts = [10_000, 100_000];
+
+    for &file_count in &file_counts {
+        // Pre-generate content so we only measure index building
+        let contents: Vec<String> = (0..file_count)
+            .map(|i| {
+                format!(
+                    "struct File{i} {{ fn process() {{ let controller = read(path); }} }} // mod {i}"
+                )
+            })
+            .collect();
+
+        group.bench_with_input(
+            BenchmarkId::new("build_and_compress", file_count),
+            &file_count,
+            |b, &fc| {
+                b.iter(|| {
+                    let builder = BigramIndexBuilder::new(fc);
+                    let skip_builder = BigramIndexBuilder::new(fc);
+                    for (i, content) in contents.iter().enumerate() {
+                        builder.add_file_content(&skip_builder, i, content.as_bytes());
+                    }
+                    let index = builder.compress(None);
+                    black_box(index.columns_used())
+                });
+            },
+        );
+    }
+
+    group.finish();
+}
+
+criterion_group!(
+    benches,
+    bench_bigram_query,
+    bench_bigram_is_candidate,
+    bench_bigram_build,
+);
+
+criterion_main!(benches);

crates/fff-core/benches/memmem_bench.rs

@@ -0,0 +1,101 @@
+use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
+use fff_search::case_insensitive_memmem;
+use std::path::Path;
+
+/// Load real source files from the repository as benchmark haystacks.
+/// Falls back to concatenating all .rs files under crates/ if specific files are missing.
+fn load_real_files() -> Vec<(&'static str, Vec<u8>)> {
+    let manifest_dir = env!("CARGO_MANIFEST_DIR"); // crates/fff-core
+    let repo_root = Path::new(manifest_dir).parent().unwrap().parent().unwrap();
+
+    let files: &[(&str, &str)] = &[
+        ("grep.rs/80KB", "crates/fff-core/src/grep.rs"),
+        ("file_picker.rs/53KB", "crates/fff-core/src/file_picker.rs"),
+        ("picker_ui.lua/96KB", "lua/fff/picker_ui.lua"),
+    ];
+
+    let mut result = Vec::new();
+    for &(label, rel_path) in files {
+        let full_path = repo_root.join(rel_path);
+        if let Ok(data) = std::fs::read(&full_path) {
+            result.push((label, data));
+        }
+    }
+
+    // Also create a large synthetic file by concatenating all three
+    if result.len() == 3 {
+        let mut combined = Vec::new();
+        for (_, data) in &result {
+            combined.extend_from_slice(data);
+        }
+        // Repeat to get ~1MB
+        let base = combined.clone();
+        while combined.len() < 1024 * 1024 {
+            combined.extend_from_slice(&base);
+        }
+        combined.truncate(1024 * 1024);
+        result.push(("combined/1MB", combined));
+    }
+
+    result
+}
+
+fn bench_memmem(c: &mut Criterion) {
+    let mut group = c.benchmark_group("case_insensitive_memmem");
+
+    let files = load_real_files();
+    assert!(!files.is_empty(), "No source files found for benchmarking");
+
+    // Needles chosen to exercise different false-positive rates:
+    //
+    // "hit" needles: strings that actually appear in these source files.
+    // "miss" needles: strings with common first-bytes (lots of false positives
+    //   for memchr2) but that don't exist in any of the files.
+    let needles: &[(&str, &[u8])] = &[
+        // Hits — real identifiers from the codebase
+        ("short/hit/fn", b"fn"),
+        ("short/hit/self", b"self"),
+        ("medium/hit", b"search_file"),
+        ("long/hit", b"content_cache_budget"),
+        // Misses — common first-bytes, guaranteed not in source
+        ("short/miss", b"zqxjv"),
+        ("medium/miss", b"fluxcapacitor"),
+        ("long/miss", b"quantum_entanglement_resolver"),
+    ];
+
+    for (file_label, haystack) in &files {
+        for &(needle_label, needle) in needles {
+            let needle_lower: Vec<u8> = needle.iter().map(|b| b.to_ascii_lowercase()).collect();
+            let id = format!("{file_label}/{needle_label}");
+
+            group.bench_with_input(
+                BenchmarkId::new("packed_pair", &id),
+                &(haystack, &needle_lower),
+                |b, &(h, n)| {
+                    b.iter(|| black_box(case_insensitive_memmem::search_packed_pair(h, n)));
+                },
+            );
+
+            group.bench_with_input(
+                BenchmarkId::new("memchr2_search", &id),
+                &(haystack, &needle_lower),
+                |b, &(h, n)| {
+                    b.iter(|| black_box(case_insensitive_memmem::search(h, n)));
+                },
+            );
+
+            group.bench_with_input(
+                BenchmarkId::new("scalar_baseline", &id),
+                &(haystack, &needle_lower),
+                |b, &(h, n)| {
+                    b.iter(|| black_box(case_insensitive_memmem::search_scalar(h, n)));
+                },
+            );
+        }
+    }
+
+    group.finish();
+}
+
+criterion_group!(benches, bench_memmem);
+criterion_main!(benches);

crates/fff-core/benches/parse_bench.rs

@@ -0,0 +1,180 @@
+use criterion::{BenchmarkId, Criterion, Throughput, black_box, criterion_group, criterion_main};
+use fff_query_parser::*;
+
+fn bench_parse_simple(c: &mut Criterion) {
+    let parser = QueryParser::default();
+
+    c.bench_function("parse_simple_text", |b| {
+        b.iter(|| parser.parse(black_box("hello world")));
+    });
+
+    c.bench_function("parse_extension", |b| {
+        b.iter(|| parser.parse(black_box("*.rs")));
+    });
+
+    c.bench_function("parse_text_with_extension", |b| {
+        b.iter(|| parser.parse(black_box("name *.rs")));
+    });
+}
+
+fn bench_parse_complex(c: &mut Criterion) {
+    let parser = QueryParser::default();
+
+    c.bench_function("parse_complex_mixed", |b| {
+        b.iter(|| parser.parse(black_box("src name *.rs !test /lib/ status:modified")));
+    });
+
+    c.bench_function("parse_glob", |b| {
+        b.iter(|| parser.parse(black_box("**/*.rs")));
+    });
+
+    c.bench_function("parse_multiple_constraints", |b| {
+        b.iter(|| parser.parse(black_box("*.rs *.toml *.md !test !node_modules /src/")));
+    });
+}
+
+fn bench_parse_realistic_queries(c: &mut Criterion) {
+    let parser = QueryParser::default();
+
+    let queries = vec![
+        "file",
+        "test",
+        "mod.rs",
+        "src/*.rs",
+        "lib test",
+        "*.rs !test",
+        "src/lib/*.rs",
+        "/src/ name",
+        "status:modified *.rs",
+        "type:rust test !node_modules",
+    ];
+
+    let mut group = c.benchmark_group("realistic_queries");
+    for query in queries.iter() {
+        group.throughput(Throughput::Bytes(query.len() as u64));
+        group.bench_with_input(BenchmarkId::from_parameter(query), query, |b, q| {
+            b.iter(|| parser.parse(black_box(q)));
+        });
+    }
+    group.finish();
+}
+
+fn bench_parse_various_lengths(c: &mut Criterion) {
+    let parser = QueryParser::default();
+
+    let short = "*.rs";
+    let medium = "src name *.rs !test";
+    let long = "src lib test name *.rs *.toml !node_modules !test /src/ /lib/ status:modified";
+    let very_long =
+        "a b c d e f g h i j k l m n o p q r s t u v w x y z *.rs *.toml *.md *.txt *.js";
+
+    let mut group = c.benchmark_group("query_lengths");
+
+    group.throughput(Throughput::Bytes(short.len() as u64));
+    group.bench_with_input(BenchmarkId::new("short", short.len()), &short, |b, q| {
+        b.iter(|| parser.parse(black_box(q)));
+    });
+
+    group.throughput(Throughput::Bytes(medium.len() as u64));
+    group.bench_with_input(BenchmarkId::new("medium", medium.len()), &medium, |b, q| {
+        b.iter(|| parser.parse(black_box(q)));
+    });
+
+    group.throughput(Throughput::Bytes(long.len() as u64));
+    group.bench_with_input(BenchmarkId::new("long", long.len()), &long, |b, q| {
+        b.iter(|| parser.parse(black_box(q)));
+    });
+
+    group.throughput(Throughput::Bytes(very_long.len() as u64));
+    group.bench_with_input(
+        BenchmarkId::new("very_long", very_long.len()),
+        &very_long,
+        |b, q| {
+            b.iter(|| parser.parse(black_box(q)));
+        },
+    );
+
+    group.finish();
+}
+
+fn bench_config_comparison(c: &mut Criterion) {
+    let file_picker = QueryParser::new(FileSearchConfig);
+    let grep = QueryParser::new(GrepConfig);
+
+    let query = "src name *.rs !test";
+
+    let mut group = c.benchmark_group("config_comparison");
+
+    group.bench_function("file_picker_config", |b| {
+        b.iter(|| file_picker.parse(black_box(query)));
+    });
+
+    group.bench_function("grep_config", |b| {
+        b.iter(|| grep.parse(black_box(query)));
+    });
+
+    group.finish();
+}
+
+fn bench_constraint_types(c: &mut Criterion) {
+    let parser = QueryParser::default();
+
+    let mut group = c.benchmark_group("constraint_types");
+
+    group.bench_function("extension", |b| {
+        b.iter(|| parser.parse(black_box("*.rs")));
+    });
+
+    group.bench_function("glob", |b| {
+        b.iter(|| parser.parse(black_box("**/*.rs")));
+    });
+
+    group.bench_function("exclude", |b| {
+        b.iter(|| parser.parse(black_box("!test")));
+    });
+
+    group.bench_function("path_segment", |b| {
+        b.iter(|| parser.parse(black_box("/src/")));
+    });
+
+    group.bench_function("git_status", |b| {
+        b.iter(|| parser.parse(black_box("status:modified")));
+    });
+
+    group.bench_function("file_type", |b| {
+        b.iter(|| parser.parse(black_box("type:rust")));
+    });
+
+    group.finish();
+}
+
+fn bench_worst_case(c: &mut Criterion) {
+    let parser = QueryParser::default();
+
+    // Worst case: many constraints that all need to be checked
+    let worst_case = "a b c d e f g h i j k l m n o p q r s t u v w x y z";
+
+    c.bench_function("worst_case_many_text_tokens", |b| {
+        b.iter(|| parser.parse(black_box(worst_case)));
+    });
+
+    // Many constraints
+    let many_constraints = "*.rs *.toml *.md *.txt *.js *.ts *.jsx *.tsx *.vue *.svelte";
+
+    c.bench_function("worst_case_many_constraints", |b| {
+        b.iter(|| parser.parse(black_box(many_constraints)));
+    });
+}
+
+criterion_group!(
+    benches,
+    bench_parse_simple,
+    bench_parse_complex,
+    bench_parse_realistic_queries,
+    bench_parse_various_lengths,
+    bench_config_comparison,
+    bench_constraint_types,
+    bench_worst_case,
+);
+
+criterion_main!(benches);

crates/fff-core/src/background_watcher.rs

@@ -1,26 +1,45 @@
 use crate::error::Error;
-use crate::file_picker::FilePicker;
+use crate::file_picker::{FFFMode, FilePicker};
 use crate::git::GitStatusCache;
+use crate::shared::{SharedFrecency, SharedPicker};
 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, 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, debug, error, info, warn};
 
 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(
@@ -28,18 +47,45 @@ impl BackgroundWatcher {
         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, shared_picker, shared_frecency)?;
+            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),
         })
     }
 
@@ -48,6 +94,7 @@ impl BackgroundWatcher {
         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
@@ -66,6 +113,7 @@ impl BackgroundWatcher {
                             &git_workdir_for_handler,
                             &shared_picker,
                             &shared_frecency,
+                            mode,
                         );
                     }
                     Err(errors) => {
@@ -86,7 +134,7 @@ impl BackgroundWatcher {
         // directories like `target/` in rust causes buffer overflow, which drops real source file
         // events. Instead we watch the root non-recursively (for top-level file changes
         // and new directory detection) and each non-ignored subdirectory recursively.
-        let watch_dirs = collect_non_ignored_dirs(&base_path);
+        let watch_dirs = collect_non_ignored_dirs(&base_path, git_workdir.is_some());
 
         if watch_dirs.len() > MAX_SELECTIVE_WATCH_DIRS {
             tracing::warn!(
@@ -123,25 +171,23 @@ 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();
     }
 }
 
@@ -151,6 +197,7 @@ fn handle_debounced_events(
     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());
@@ -286,9 +333,9 @@ fn handle_debounced_events(
                             debug!(
                                 "on_create_or_modify({:?}) -> Some({})",
                                 path,
-                                file.path.display()
+                                file.path_str()
                             );
-                            files_to_update.push(file.path.clone());
+                            files_to_update.push(PathBuf::from(file.path_str()));
                         }
                         None => {
                             error!("on_create_or_modify({:?}) -> None (file not added!)", path);
@@ -316,6 +363,50 @@ fn handle_debounced_events(
             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 {
         debug!("No git repo available, skipping git status updates");
@@ -325,7 +416,7 @@ fn handle_debounced_events(
     if need_full_git_rescan {
         info!("Triggering full git rescan");
 
-        let result = FilePicker::refresh_git_status(shared_picker, shared_frecency);
+        let result = shared_picker.refresh_git_status(shared_frecency);
         if let Err(e) = result {
             error!("Failed to refresh git status: {:?}", e);
         }
@@ -388,14 +479,21 @@ fn should_include_file(path: &Path, repo: &Option<Repository>) -> bool {
         return 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,
+        None => {
+            // No git repo — apply basic sanity filters.
+            // Hidden directories are skipped by the watcher setup (hidden(true)),
+            // but events can still arrive for files in known non-code directories.
+            !is_non_code_directory(path)
+        }
     }
 }
 
+fn is_non_code_directory(path: &Path) -> bool {
+    crate::ignore::is_non_code_directory(path)
+}
+
 #[inline]
 fn is_git_file(path: &Path) -> bool {
     path.components()
@@ -481,18 +579,25 @@ fn watch_git_status_paths(debouncer: &mut Debouncer, git_workdir: Option<&PathBu
 /// 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,
 /// preventing gitignored directories like `target/` from flooding the OS event buffer.
-fn collect_non_ignored_dirs(base_path: &Path) -> Vec<PathBuf> {
+fn collect_non_ignored_dirs(base_path: &Path, has_git_repo: bool) -> Vec<PathBuf> {
+    use crate::ignore::non_git_repo_overrides;
     use ignore::WalkBuilder;
 
-    let walker = WalkBuilder::new(base_path)
-        .hidden(false)
+    let mut walk_builder = WalkBuilder::new(base_path);
+    walk_builder
+        .hidden(!has_git_repo)
         .git_ignore(true)
         .git_exclude(true)
         .git_global(true)
         .ignore(true)
         .follow_links(false)
-        .max_depth(Some(1))
-        .build();
+        .max_depth(Some(1));
+
+    if !has_git_repo && let Some(overrides) = non_git_repo_overrides(base_path) {
+        walk_builder.overrides(overrides);
+    }
+
+    let walker = walk_builder.build();
 
     let mut dirs = Vec::new();
     for entry in walker {

crates/fff-core/src/bigram_filter.rs

@@ -0,0 +1,521 @@
+use std::sync::atomic::{AtomicU16, AtomicU64, AtomicUsize, Ordering};
+
+use ahash::AHashMap;
+
+/// Maximum number of distinct bigrams tracked in the inverted index.
+/// 95 printable ASCII chars (32..=126) after lowercasing → ~70 distinct → 4900 possible.
+/// We cap at 5000 to cover all printable bigrams with margin.
+/// 5000 columns × 62.5KB (500k files) = 305MB. For 50k files: 30MB.
+const MAX_BIGRAM_COLUMNS: usize = 5000;
+
+/// Sentinel value: bigram has no allocated column.
+const NO_COLUMN: u16 = u16::MAX;
+
+/// Temporary sync dense builder for the bigram index.
+/// Builds from the many threads reading file contents in parallel
+pub struct BigramIndexBuilder {
+    // we use lookup as atomics only in the builder because it is filled by the rayon threads
+    // the actual index uses pure u16 for the allocations
+    lookup: Vec<AtomicU16>,
+    /// Per-column bitset data, lazily allocated via OnceLock.
+    col_data: Vec<AtomicU64>,
+    next_column: AtomicU16,
+    words: usize,
+    file_count: usize,
+    populated: AtomicUsize,
+}
+
+impl BigramIndexBuilder {
+    pub fn new(file_count: usize) -> Self {
+        let words = file_count.div_ceil(64);
+        let mut lookup = Vec::with_capacity(65536);
+        lookup.resize_with(65536, || AtomicU16::new(NO_COLUMN));
+        let mut col_data = Vec::with_capacity(MAX_BIGRAM_COLUMNS * words);
+        col_data.resize_with(MAX_BIGRAM_COLUMNS * words, || AtomicU64::new(0));
+        Self {
+            lookup,
+            col_data,
+            next_column: AtomicU16::new(0),
+            words,
+            file_count,
+            populated: AtomicUsize::new(0),
+        }
+    }
+
+    #[inline]
+    fn get_or_alloc_column(&self, key: u16) -> u16 {
+        let current = self.lookup[key as usize].load(Ordering::Relaxed);
+        if current != NO_COLUMN {
+            return current;
+        }
+        let new_col = self.next_column.fetch_add(1, Ordering::Relaxed);
+        if new_col >= MAX_BIGRAM_COLUMNS as u16 {
+            return NO_COLUMN;
+        }
+
+        match self.lookup[key as usize].compare_exchange(
+            NO_COLUMN,
+            new_col,
+            Ordering::Relaxed,
+            Ordering::Relaxed,
+        ) {
+            Ok(_) => new_col,
+            Err(existing) => existing,
+        }
+    }
+
+    #[inline]
+    fn column_bitset(&self, col: u16) -> &[AtomicU64] {
+        let start = col as usize * self.words;
+        &self.col_data[start..start + self.words]
+    }
+
+    pub(crate) fn add_file_content(&self, skip_builder: &Self, file_idx: usize, content: &[u8]) {
+        if content.len() < 2 {
+            return;
+        }
+
+        debug_assert!(file_idx < self.file_count);
+        let word_idx = file_idx / 64;
+        let bit_mask = 1u64 << (file_idx % 64);
+
+        // Stack-local dedup bitsets: 1024 × u64 = 8 KB each, covers all 65536 bigrams with margin
+        // have to fit in L1 cache
+        let mut seen_consec = [0u64; 1024];
+        let mut seen_skip = [0u64; 1024];
+
+        let bytes = content;
+        let len = bytes.len();
+
+        // First consecutive pair (no skip bigram possible yet).
+        let (a, b) = (bytes[0], bytes[1]);
+        if (32..=126).contains(&a) && (32..=126).contains(&b) {
+            let key = (a.to_ascii_lowercase() as u16) << 8 | b.to_ascii_lowercase() as u16;
+            let w = key as usize >> 6;
+            let bit = 1u64 << (key as usize & 63);
+            seen_consec[w] |= bit;
+            let col = self.get_or_alloc_column(key);
+            if col != NO_COLUMN {
+                self.column_bitset(col)[word_idx].fetch_or(bit_mask, Ordering::Relaxed);
+            }
+        }
+
+        // Main loop: consecutive (i-1, i) and skip-1 (i-2, i)
+        for i in 2..len {
+            let cur = bytes[i];
+
+            // Consecutive bigram: (bytes[i-1], bytes[i])
+            let prev = bytes[i - 1];
+            if (32..=126).contains(&prev) && (32..=126).contains(&cur) {
+                let key = (prev.to_ascii_lowercase() as u16) << 8 | cur.to_ascii_lowercase() as u16;
+                let w = key as usize >> 6;
+                let bit = 1u64 << (key as usize & 63);
+                if seen_consec[w] & bit == 0 {
+                    seen_consec[w] |= bit;
+                    let col = self.get_or_alloc_column(key);
+                    if col != NO_COLUMN {
+                        self.column_bitset(col)[word_idx].fetch_or(bit_mask, Ordering::Relaxed);
+                    }
+                }
+            }
+
+            // Skip-1 bigram: (bytes[i-2], bytes[i])
+            let skip_prev = bytes[i - 2];
+            if (32..=126).contains(&skip_prev) && (32..=126).contains(&cur) {
+                let key =
+                    (skip_prev.to_ascii_lowercase() as u16) << 8 | cur.to_ascii_lowercase() as u16;
+                let w = key as usize >> 6;
+                let bit = 1u64 << (key as usize & 63);
+                if seen_skip[w] & bit == 0 {
+                    seen_skip[w] |= bit;
+                    let col = skip_builder.get_or_alloc_column(key);
+                    if col != NO_COLUMN {
+                        skip_builder.column_bitset(col)[word_idx]
+                            .fetch_or(bit_mask, Ordering::Relaxed);
+                    }
+                }
+            }
+        }
+
+        self.populated.fetch_add(1, Ordering::Relaxed);
+        skip_builder.populated.fetch_add(1, Ordering::Relaxed);
+    }
+
+    pub fn is_ready(&self) -> bool {
+        self.populated.load(Ordering::Relaxed) > 0
+    }
+
+    pub fn columns_used(&self) -> u16 {
+        self.next_column
+            .load(Ordering::Relaxed)
+            .min(MAX_BIGRAM_COLUMNS as u16)
+    }
+
+    /// Compress the dense builder into a compact `BigramFilter`.
+    ///
+    /// Retains columns where the bigram appears in ≥`min_density_pct`% (or
+    /// the default ~3.1% heuristic when `None`) and <90% of indexed files.
+    /// Sparse columns carry too little data to justify their memory;
+    /// ubiquitous columns (≥90%) are nearly all-ones and barely filter.
+    pub fn compress(self, min_density_pct: Option<u32>) -> BigramFilter {
+        let cols = self.columns_used() as usize;
+        let words = self.words;
+        let file_count = self.file_count;
+        let populated = self.populated.load(Ordering::Relaxed);
+        let dense_bytes = words * 8; // cost of one dense column
+
+        let old_lookup = self.lookup;
+        let col_data = self.col_data;
+
+        let mut lookup: Vec<u16> = vec![NO_COLUMN; 65536];
+        let mut dense_data: Vec<u64> = Vec::with_capacity(cols * words);
+        let mut dense_count: usize = 0;
+
+        for key in 0..65536usize {
+            let old_col = old_lookup[key].load(Ordering::Relaxed);
+            if old_col == NO_COLUMN || old_col as usize >= cols {
+                continue;
+            }
+
+            let col_start = old_col as usize * words;
+            let bitset = &col_data[col_start..col_start + words];
+
+            // count set bits to decide if this column is worth keeping.
+            let mut popcount = 0u32;
+            for column in bitset.iter().take(words) {
+                popcount += column.load(Ordering::Relaxed).count_ones();
+            }
+
+            // drop bigrams appearing in too few files
+            let not_to_rare = if let Some(min_pct) = min_density_pct {
+                // Percentage-based: require ≥ min_pct% of populated files.
+                populated > 0 && (popcount as usize) * 100 >= populated * min_pct as usize
+            } else {
+                // Default: popcount ≥ words × 2 (~3.1% of files).
+                (popcount as usize * 4) >= dense_bytes
+            };
+
+            if !not_to_rare {
+                continue;
+            }
+
+            // Drop ubiquitous bigrams — columns ≥90% ones carry almost no
+            // filtering power and just waste memory + AND cycles.
+            if populated > 0 && (popcount as usize) * 10 >= populated * 9 {
+                continue;
+            }
+
+            let dense_idx = dense_count as u16;
+            lookup[key] = dense_idx;
+            dense_count += 1;
+
+            for column in bitset.iter().take(words) {
+                dense_data.push(column.load(Ordering::Relaxed));
+            }
+        }
+
+        // col_data + old_lookup dropped here — single deallocation each,
+        // no fragmentation.
+
+        BigramFilter {
+            lookup,
+            dense_data,
+            dense_count,
+            words,
+            file_count,
+            populated,
+            skip_index: None,
+        }
+    }
+}
+
+unsafe impl Send for BigramIndexBuilder {}
+unsafe impl Sync for BigramIndexBuilder {}
+
+/// Inverted bigram index with optional "skip-1" extension
+/// Copmressed into bitset for minimal usage, the layout of this struct actually matters
+#[derive(Debug)]
+pub struct BigramFilter {
+    lookup: Vec<u16>,
+    /// Flat buffer of all dense column data laid out at fixed stride `words`.
+    /// Column `i` starts at `i * words`.
+    dense_data: Vec<u64>, // do not try to change this to u8 it has to be wordsize
+    dense_count: usize,
+    words: usize,
+    file_count: usize,
+    populated: usize,
+    /// Optional skip-1 bigram index (stride 2). Built from character pairs
+    /// at distance 2, e.g. "ABCDE" → (A,C),(B,D),(C,E). ANDead with the
+    /// consecutive bigram candidates during query to dramatically reduce
+    /// false positives.
+    skip_index: Option<Box<BigramFilter>>,
+}
+
+/// SIMD-friendly bitwise AND of two equal-length bitsets.
+// Auto vectorized (don't touch)
+#[inline]
+fn bitset_and(result: &mut [u64], bitset: &[u64]) {
+    result
+        .iter_mut()
+        .zip(bitset.iter())
+        .for_each(|(r, b)| *r &= *b);
+}
+
+impl BigramFilter {
+    /// AND the posting lists for all query bigrams (consecutive + skip).
+    /// Returns None if no query bigrams are tracked.
+    pub fn query(&self, pattern: &[u8]) -> Option<Vec<u64>> {
+        if pattern.len() < 2 {
+            return None;
+        }
+
+        let mut result = vec![u64::MAX; self.words];
+        if !self.file_count.is_multiple_of(64) {
+            let last = self.words - 1;
+            result[last] = (1u64 << (self.file_count % 64)) - 1;
+        }
+
+        let words = self.words;
+        let mut has_filter = false;
+
+        let mut prev = pattern[0];
+        for &b in &pattern[1..] {
+            if (32..=126).contains(&prev) && (32..=126).contains(&b) {
+                let key = (prev.to_ascii_lowercase() as u16) << 8 | b.to_ascii_lowercase() as u16;
+                let col = self.lookup[key as usize];
+                if col != NO_COLUMN {
+                    let offset = col as usize * words;
+                    // SAFETY: compress() guarantees offset + words <= dense_data.len()
+                    let slice = unsafe { self.dense_data.get_unchecked(offset..offset + words) };
+                    bitset_and(&mut result, slice);
+                    has_filter = true;
+                }
+            }
+            prev = b;
+        }
+
+        // strid-1 bigrams
+        if let Some(skip) = &self.skip_index
+            && pattern.len() >= 3
+            && let Some(skip_candidates) = skip.query_skip(pattern)
+        {
+            bitset_and(&mut result, &skip_candidates);
+            has_filter = true;
+        }
+
+        has_filter.then_some(result)
+    }
+
+    /// Query using stride-2 bigrams from the pattern.
+    /// For "ABCDE" queries with keys (A,C), (B,D), (C,E).
+    fn query_skip(&self, pattern: &[u8]) -> Option<Vec<u64>> {
+        let mut result = vec![u64::MAX; self.words];
+        if !self.file_count.is_multiple_of(64) {
+            let last = self.words - 1;
+            result[last] = (1u64 << (self.file_count % 64)) - 1;
+        }
+
+        let words = self.words;
+        let mut has_filter = false;
+
+        for i in 0..pattern.len().saturating_sub(2) {
+            let a = pattern[i];
+            let b = pattern[i + 2];
+            if (32..=126).contains(&a) && (32..=126).contains(&b) {
+                let key = (a.to_ascii_lowercase() as u16) << 8 | b.to_ascii_lowercase() as u16;
+                let col = self.lookup[key as usize];
+                if col != NO_COLUMN {
+                    let offset = col as usize * words;
+                    let slice = unsafe { self.dense_data.get_unchecked(offset..offset + words) };
+                    bitset_and(&mut result, slice);
+                    has_filter = true;
+                }
+            }
+        }
+
+        has_filter.then_some(result)
+    }
+
+    /// Attach a skip-1 bigram index for tighter candidate filtering.
+    pub fn set_skip_index(&mut self, skip: BigramFilter) {
+        self.skip_index = Some(Box::new(skip));
+    }
+
+    #[inline]
+    pub fn is_candidate(candidates: &[u64], file_idx: usize) -> bool {
+        let word = file_idx / 64;
+        let bit = file_idx % 64;
+        word < candidates.len() && candidates[word] & (1u64 << bit) != 0
+    }
+
+    pub fn count_candidates(candidates: &[u64]) -> usize {
+        candidates.iter().map(|w| w.count_ones() as usize).sum()
+    }
+
+    pub fn is_ready(&self) -> bool {
+        self.populated > 0
+    }
+
+    pub fn file_count(&self) -> usize {
+        self.file_count
+    }
+
+    pub fn columns_used(&self) -> usize {
+        self.dense_count
+    }
+
+    /// Total heap bytes used by this index (lookup + dense data + skip).
+    pub fn heap_bytes(&self) -> usize {
+        let lookup_bytes = self.lookup.len() * std::mem::size_of::<u16>();
+        let dense_bytes = self.dense_data.len() * std::mem::size_of::<u64>();
+        let skip_bytes = self.skip_index.as_ref().map_or(0, |s| s.heap_bytes());
+        lookup_bytes + dense_bytes + skip_bytes
+    }
+
+    /// Check whether a bigram key is present in this index.
+    pub fn has_key(&self, key: u16) -> bool {
+        self.lookup[key as usize] != NO_COLUMN
+    }
+
+    /// Raw lookup table (65536 entries mapping bigram key → column index).
+    pub fn lookup(&self) -> &[u16] {
+        &self.lookup
+    }
+
+    /// Flat dense bitset data at fixed stride `words`.
+    pub fn dense_data(&self) -> &[u64] {
+        &self.dense_data
+    }
+
+    /// Number of u64 words per column (= ceil(file_count / 64)).
+    pub fn words(&self) -> usize {
+        self.words
+    }
+
+    /// Number of dense columns retained after compression.
+    pub fn dense_count(&self) -> usize {
+        self.dense_count
+    }
+
+    /// Number of files that contributed content to the index.
+    pub fn populated(&self) -> usize {
+        self.populated
+    }
+
+    /// Reference to the optional skip-1 bigram sub-index.
+    pub fn skip_index(&self) -> Option<&BigramFilter> {
+        self.skip_index.as_deref()
+    }
+
+    /// Create a new bigram filter from the internal data
+    pub fn new(
+        lookup: Vec<u16>,
+        dense_data: Vec<u64>,
+        dense_count: usize,
+        words: usize,
+        file_count: usize,
+        populated: usize,
+    ) -> Self {
+        Self {
+            lookup,
+            dense_data,
+            dense_count,
+            words,
+            file_count,
+            populated,
+            skip_index: None,
+        }
+    }
+}
+
+pub fn extract_bigrams(content: &[u8]) -> Vec<u16> {
+    if content.len() < 2 {
+        return Vec::new();
+    }
+    // Use a flat bitset (65536 bits = 8 KB) for dedup — faster than HashSet.
+    let mut seen = vec![0u64; 1024]; // 1024 * 64 = 65536 bits
+    let mut bigrams = Vec::new();
+
+    let mut prev = content[0];
+    for &b in &content[1..] {
+        if (32..=126).contains(&prev) && (32..=126).contains(&b) {
+            let key = (prev.to_ascii_lowercase() as u16) << 8 | b.to_ascii_lowercase() as u16;
+            let word = key as usize / 64;
+            let bit = 1u64 << (key as usize % 64);
+            if seen[word] & bit == 0 {
+                seen[word] |= bit;
+                bigrams.push(key);
+            }
+        }
+        prev = b;
+    }
+    bigrams
+}
+
+/// Modified and added files store their own bigram sets. Deleted files are
+/// tombstoned in a bitset so they can be excluded from base query results.
+/// This overlay is updated by the background watcher on every file event
+/// and cleared when the base index is rebuilt.
+#[derive(Debug)]
+pub struct BigramOverlay {
+    /// Per-file bigram sets for files modified since the base was built.
+    /// Key = file index in the base `Vec<FileItem>`.
+    modified: AHashMap<usize, Vec<u16>>,
+
+    /// Tombstone bitset — one bit per base file. Set bits are excluded
+    /// from base query results.
+    tombstones: Vec<u64>,
+
+    /// Original files count this overlay was created for.
+    base_file_count: usize,
+}
+
+impl BigramOverlay {
+    pub(crate) fn new(base_file_count: usize) -> Self {
+        let words = base_file_count.div_ceil(64);
+        Self {
+            modified: AHashMap::new(),
+            tombstones: vec![0u64; words],
+            base_file_count,
+        }
+    }
+
+    pub(crate) fn modify_file(&mut self, file_idx: usize, content: &[u8]) {
+        self.modified.insert(file_idx, extract_bigrams(content));
+    }
+
+    pub(crate) fn delete_file(&mut self, file_idx: usize) {
+        if file_idx < self.base_file_count {
+            let word = file_idx / 64;
+            self.tombstones[word] |= 1u64 << (file_idx % 64);
+        }
+        self.modified.remove(&file_idx);
+    }
+
+    /// Return base file indices of modified files whose bigrams match ALL
+    /// of the given `pattern_bigrams`.
+    pub(crate) fn query_modified(&self, pattern_bigrams: &[u16]) -> Vec<usize> {
+        if pattern_bigrams.is_empty() {
+            return self.modified.keys().copied().collect();
+        }
+        self.modified
+            .iter()
+            .filter_map(|(&file_idx, bigrams)| {
+                pattern_bigrams
+                    .iter()
+                    .all(|pb| bigrams.contains(pb))
+                    .then_some(file_idx)
+            })
+            .collect()
+    }
+
+    /// Number of base files this overlay was created for.
+    pub(crate) fn base_file_count(&self) -> usize {
+        self.base_file_count
+    }
+
+    /// Get the tombstone bitset for clearing base candidates.
+    pub(crate) fn tombstones(&self) -> &[u64] {
+        &self.tombstones
+    }
+}

crates/fff-core/src/case_insensitive_memmem.rs

@@ -0,0 +1,662 @@
+//! SIMD-accelerated case-insensitive substring search.
+//!
+//! Implementations (fastest → simplest):
+//! - `search_packed_pair`: AVX2 packed-pair scan (two rare bytes at known offsets)
+//! - `search`:             memchr2 first-byte scan + verify
+//!
+//! The packed-pair approach mirrors what `memchr::memmem` does internally for
+//! case-sensitive search — pick two rare bytes from the needle, SIMD-scan for
+//! both simultaneously, verify candidates.  This gives quadratic selectivity
+//! over the single-byte memchr2 approach.
+
+// this is stolen from the memchr2 crate
+const BYTE_FREQUENCIES: [u8; 256] = [
+    55, 52, 51, 50, 49, 48, 47, 46, 45, 103, 242, 66, 67, 229, 44, 43, // 0x00
+    42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 56, 32, 31, 30, 29, 28, // 0x10
+    255, 148, 164, 149, 136, 160, 155, 173, 221, 222, 134, 122, 232, 202, 215, 224, // 0x20
+    208, 220, 204, 187, 183, 179, 177, 168, 178, 200, 226, 195, 154, 184, 174, 126, // 0x30
+    120, 191, 157, 194, 170, 189, 162, 161, 150, 193, 142, 137, 171, 176, 185,
+    167, // 0x40 A-O
+    186, 112, 175, 192, 188, 156, 140, 143, 123, 133, 128, 147, 138, 146, 114,
+    223, // 0x50 P-_
+    151, 249, 216, 238, 236, 253, 227, 218, 230, 247, 135, 180, 241, 233, 246,
+    244, // 0x60 a-o
+    231, 139, 245, 243, 251, 235, 201, 196, 240, 214, 152, 182, 205, 181, 127,
+    27, // 0x70 p-DEL
+    212, 211, 210, 213, 228, 197, 169, 159, 131, 172, 105, 80, 98, 96, 97, 81, // 0x80
+    207, 145, 116, 115, 144, 130, 153, 121, 107, 132, 109, 110, 124, 111, 82, 108, // 0x90
+    118, 141, 113, 129, 119, 125, 165, 117, 92, 106, 83, 72, 99, 93, 65, 79, // 0xa0
+    166, 237, 163, 199, 190, 225, 209, 203, 198, 217, 219, 206, 234, 248, 158, 239, // 0xb0
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, // 0xc0
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, // 0xd0
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, // 0xe0
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, // 0xf0
+];
+
+#[inline]
+fn ascii_fold_byte(b: u8) -> u8 {
+    if b.is_ascii_uppercase() { b | 0x20 } else { b }
+}
+
+/// Toggle ASCII letter case by flipping bit 5.
+/// `'n' → 'N'`, `'N' → 'n'`.
+#[inline]
+fn ascii_swap_case(b: u8) -> u8 {
+    b ^ 0x20
+}
+
+/// Effective frequency rank for a case-insensitive byte position.
+/// Takes the max of lower/upper ranks because we must scan for both.
+#[inline]
+fn case_insensitive_rank(lower: u8) -> u8 {
+    if lower.is_ascii_lowercase() {
+        let upper = ascii_swap_case(lower);
+        BYTE_FREQUENCIES[lower as usize].max(BYTE_FREQUENCIES[upper as usize])
+    } else {
+        BYTE_FREQUENCIES[lower as usize]
+    }
+}
+
+/// Pick two needle positions with the rarest bytes (case-insensitive).
+/// Returns (index1, index2) where index1 <= index2.
+fn select_rare_pair(needle_lower: &[u8]) -> (usize, usize) {
+    debug_assert!(needle_lower.len() >= 2);
+
+    let mut best1 = (u8::MAX, 0usize); // (rank, position)
+    let mut best2 = (u8::MAX, 1usize);
+
+    for (i, &b) in needle_lower.iter().enumerate() {
+        let r = case_insensitive_rank(b);
+        if r < best1.0 {
+            best2 = best1;
+            best1 = (r, i);
+        } else if r < best2.0 && i != best1.1 {
+            best2 = (r, i);
+        }
+    }
+
+    let i1 = best1.1.min(best2.1);
+    let i2 = best1.1.max(best2.1);
+    (i1, i2)
+}
+
+#[inline]
+fn verify_scalar(h: *const u8, needle_lower: &[u8]) -> bool {
+    for (i, _) in needle_lower.iter().enumerate() {
+        if ascii_fold_byte(unsafe { *h.add(i) }) != needle_lower[i] {
+            return false;
+        }
+    }
+    true
+}
+
+/// AVX2 case-insensitive verify: checks whether `needle_lower` matches
+/// the haystack bytes starting at `h`, treating ASCII uppercase as lowercase.
+///
+/// Processes 32 bytes at a time using a SIMD trick: AVX2 only has a
+/// **signed** byte compare (`cmpgt`), but we need an **unsigned** range
+/// check (`'A' <= byte <= 'Z'`).  The trick is to XOR every byte with
+/// `0x80`, which maps the unsigned range `[0, 255]` into the signed range
+/// `[-128, 127]` while preserving order.  After the flip, signed `cmpgt`
+/// gives correct unsigned comparisons.
+///
+/// Once we know which bytes are uppercase, we set bit 5 (`0x20`) on them
+/// — this converts `'A'..'Z'` to `'a'..'z'` — then compare against the
+/// pre-lowered needle.
+#[cfg(target_arch = "x86_64")]
+#[target_feature(enable = "avx2")]
+unsafe fn verify_avx2(h: *const u8, needle_lower: &[u8]) -> bool {
+    use core::arch::x86_64::*;
+
+    let len = needle_lower.len();
+    let mut i = 0usize;
+
+    // Broadcast constants used every iteration:
+    //
+    //   flip      = 0x80 in every lane — XOR converts unsigned→signed domain
+    //   a_minus_1 = ('A' - 1) ^ 0x80  — lower bound for the range check (signed)
+    //   z_plus_1  = ('Z' + 1) ^ 0x80  — upper bound for the range check (signed)
+    //   bit20     = 0x20 in every lane — OR this onto uppercase bytes to lowercase them
+    let flip = _mm256_set1_epi8(0x80u8 as i8);
+    let a_minus_1 = _mm256_set1_epi8((b'A' - 1) as i8 ^ 0x80u8 as i8);
+    let z_plus_1 = _mm256_set1_epi8((b'Z' + 1) as i8 ^ 0x80u8 as i8);
+    let bit20 = _mm256_set1_epi8(0x20u8 as i8);
+
+    while i + 32 <= len {
+        // Load 32 bytes from the haystack candidate position.
+        let hv = unsafe { _mm256_loadu_si256(h.add(i) as *const __m256i) };
+        // Load 32 bytes from the pre-lowercased needle.
+        let nv = unsafe { _mm256_loadu_si256(needle_lower.as_ptr().add(i) as *const __m256i) };
+
+        // Flip into signed domain: x = hv ^ 0x80.
+        // After this, unsigned ordering is preserved under signed compare.
+        let x = _mm256_xor_si256(hv, flip);
+
+        // ge_a[lane] = 0xFF if x[lane] > a_minus_1, i.e. hv[lane] >= 'A' (unsigned).
+        let ge_a = _mm256_cmpgt_epi8(x, a_minus_1);
+        // le_z[lane] = 0xFF if z_plus_1 > x[lane], i.e. hv[lane] <= 'Z' (unsigned).
+        let le_z = _mm256_cmpgt_epi8(z_plus_1, x);
+        // upper[lane] = 0xFF only for bytes in the range 'A'..='Z'.
+        let upper = _mm256_and_si256(ge_a, le_z);
+
+        // Case-fold: set bit 5 on uppercase bytes → converts 'A'..'Z' to 'a'..'z'.
+        // Non-letter bytes are untouched because their `upper` lane is 0x00.
+        let folded = _mm256_or_si256(hv, _mm256_and_si256(upper, bit20));
+
+        // Compare the folded haystack against the lowercase needle.
+        let eq = _mm256_cmpeq_epi8(folded, nv);
+        // movemask extracts the high bit of each lane into a 32-bit mask.
+        // All-equal → all high bits set → mask == 0xFFFFFFFF == -1i32.
+        if _mm256_movemask_epi8(eq) != -1i32 {
+            return false;
+        }
+
+        i += 32;
+    }
+
+    // Scalar tail: handle remaining bytes that don't fill a full 32-byte vector.
+    while i < len {
+        if ascii_fold_byte(unsafe { *h.add(i) }) != needle_lower[i] {
+            return false;
+        }
+        i += 1;
+    }
+    true
+}
+
+// ======== NEON + dotprod (aarch64) ===========================================
+
+/// Extract a 16-bit bitmask from a NEON comparison result (each byte 0x00 or 0xFF).
+/// Bit *i* of the result corresponds to byte *i* of the input vector.
+#[cfg(target_arch = "aarch64")]
+#[target_feature(enable = "neon")]
+#[inline]
+unsafe fn neon_movemask(v: core::arch::aarch64::uint8x16_t) -> u16 {
+    use core::arch::aarch64::*;
+
+    // AND each byte with its bit-position mask, then horizontally sum each half.
+    // Max possible sum per half = 1+2+4+8+16+32+64+128 = 255, fits in u8.
+    static BITS: [u8; 16] = [1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128];
+    let bit_mask = unsafe { vld1q_u8(BITS.as_ptr()) };
+    let masked = vandq_u8(v, bit_mask);
+    let lo = vaddv_u8(vget_low_u8(masked));
+    let hi = vaddv_u8(vget_high_u8(masked));
+    (lo as u16) | ((hi as u16) << 8)
+}
+
+/// NEON + dotprod case-insensitive verify.
+///
+/// Uses unsigned range checks (NEON has `vcge`/`vcle` for unsigned bytes
+/// no XOR-0x80 trick needed unlike AVX2) to detect uppercase ASCII, folds
+/// to lowercase, then checks equality via UDOT: XOR the folded haystack
+/// with the pre-lowered needle and dot-product the difference with itself.
+/// Any non-zero byte produces a non-zero u32 lane.
+///
+/// The UDOT instruction is emitted via inline asm because the `vdotq_u32`
+/// intrinsic is still behind an unstable feature gate on stable Rust.
+#[cfg(target_arch = "aarch64")]
+#[target_feature(enable = "neon,dotprod")]
+unsafe fn verify_neon_dotprod(h: *const u8, needle_lower: &[u8]) -> bool {
+    use core::arch::aarch64::*;
+
+    let len = needle_lower.len();
+    let mut i = 0usize;
+
+    let a_val = vdupq_n_u8(b'A');
+    let z_val = vdupq_n_u8(b'Z');
+    let bit20 = vdupq_n_u8(0x20);
+
+    while i + 16 <= len {
+        let hv = unsafe { vld1q_u8(h.add(i)) };
+        let nv = unsafe { vld1q_u8(needle_lower.as_ptr().add(i)) };
+
+        // Unsigned range check: 'A' <= byte <= 'Z'
+        let upper = vandq_u8(vcgeq_u8(hv, a_val), vcleq_u8(hv, z_val));
+        // Case-fold: set bit 5 on uppercase bytes → 'A'..'Z' → 'a'..'z'
+        let folded = vorrq_u8(hv, vandq_u8(upper, bit20));
+
+        // XOR with needle — all-zero iff every byte matches.
+        let xored = veorq_u8(folded, nv);
+
+        // UDOT: dot(xored, xored) sums squares of 4 consecutive byte
+        // differences into each of the 4 u32 lanes (accumulates into zero).
+        // Any non-zero byte produces a positive u32 contribution.
+        let dots: uint32x4_t;
+        let zero = vdupq_n_u32(0);
+        unsafe {
+            core::arch::asm!(
+                "udot {d:v}.4s, {a:v}.16b, {b:v}.16b",
+                d = inlateout(vreg) zero => dots,
+                a = in(vreg) xored,
+                b = in(vreg) xored,
+            );
+        }
+
+        if vmaxvq_u32(dots) != 0 {
+            return false;
+        }
+
+        i += 16;
+    }
+
+    // Scalar tail
+    while i < len {
+        if ascii_fold_byte(unsafe { *h.add(i) }) != needle_lower[i] {
+            return false;
+        }
+        i += 1;
+    }
+    true
+}
+
+/// NEON packed-pair kernel: scan 16 haystack positions per iteration,
+/// checking two rare bytes (case-insensitive) simultaneously.
+/// Same algorithm as the AVX2 version but with 128-bit vectors.
+#[cfg(target_arch = "aarch64")]
+#[target_feature(enable = "neon")]
+unsafe fn search_packed_pair_neon(
+    haystack: &[u8],
+    needle_lower: &[u8],
+    i1: usize,
+    i2: usize,
+) -> bool {
+    use core::arch::aarch64::*;
+
+    let n = needle_lower.len();
+    let hlen = haystack.len();
+    let ptr = haystack.as_ptr();
+    let last_start = hlen - n;
+
+    let b1 = needle_lower[i1];
+    let b1_alt = if b1.is_ascii_lowercase() {
+        ascii_swap_case(b1)
+    } else {
+        b1
+    };
+    let b2 = needle_lower[i2];
+    let b2_alt = if b2.is_ascii_lowercase() {
+        ascii_swap_case(b2)
+    } else {
+        b2
+    };
+
+    let v1_lo = vdupq_n_u8(b1);
+    let v1_hi = vdupq_n_u8(b1_alt);
+    let v2_lo = vdupq_n_u8(b2);
+    let v2_hi = vdupq_n_u8(b2_alt);
+
+    let max_idx = i1.max(i2);
+    let max_offset = hlen.saturating_sub(max_idx + 16);
+    let mut offset = 0usize;
+
+    while offset <= max_offset {
+        let chunk1 = unsafe { vld1q_u8(ptr.add(offset + i1)) };
+        let chunk2 = unsafe { vld1q_u8(ptr.add(offset + i2)) };
+
+        // Case-insensitive match: OR both case variants, then AND the two positions.
+        let eq1 = vorrq_u8(vceqq_u8(chunk1, v1_lo), vceqq_u8(chunk1, v1_hi));
+        let eq2 = vorrq_u8(vceqq_u8(chunk2, v2_lo), vceqq_u8(chunk2, v2_hi));
+
+        let mut mask = unsafe { neon_movemask(vandq_u8(eq1, eq2)) };
+
+        while mask != 0 {
+            let bit = mask.trailing_zeros() as usize;
+            let candidate = offset + bit;
+            if candidate > last_start {
+                return false;
+            }
+            if unsafe { verify_dispatch(ptr.add(candidate), needle_lower) } {
+                return true;
+            }
+            mask &= mask - 1;
+        }
+
+        offset += 16;
+    }
+
+    // Tail: remaining positions that couldn't fill a full vector.
+    if offset <= last_start {
+        let rare_pos =
+            if case_insensitive_rank(needle_lower[i1]) <= case_insensitive_rank(needle_lower[i2]) {
+                i1
+            } else {
+                i2
+            };
+        let rare_byte = needle_lower[rare_pos];
+        let tail_start = offset + rare_pos;
+        let tail_end = last_start + rare_pos + 1;
+        if tail_start < tail_end {
+            let tail_space = &haystack[tail_start..tail_end];
+            if rare_byte.is_ascii_lowercase() {
+                for pos in memchr::memchr2_iter(rare_byte, ascii_swap_case(rare_byte), tail_space) {
+                    let candidate = offset + pos;
+                    if unsafe { verify_dispatch(ptr.add(candidate), needle_lower) } {
+                        return true;
+                    }
+                }
+            } else {
+                for pos in memchr::memchr_iter(rare_byte, tail_space) {
+                    let candidate = offset + pos;
+                    if unsafe { verify_dispatch(ptr.add(candidate), needle_lower) } {
+                        return true;
+                    }
+                }
+            }
+        }
+    }
+
+    false
+}
+
+#[inline]
+unsafe fn verify_dispatch(h: *const u8, needle_lower: &[u8]) -> bool {
+    #[cfg(target_arch = "x86_64")]
+    {
+        if needle_lower.len() >= 32 && std::is_x86_feature_detected!("avx2") {
+            return unsafe { verify_avx2(h, needle_lower) };
+        }
+    }
+    #[cfg(target_arch = "aarch64")]
+    {
+        if needle_lower.len() >= 16 && std::arch::is_aarch64_feature_detected!("dotprod") {
+            return unsafe { verify_neon_dotprod(h, needle_lower) };
+        }
+    }
+
+    verify_scalar(h, needle_lower)
+}
+
+// ── Packed-pair search (AVX2) ───────────────────────────────────────────
+
+/// AVX2 packed-pair kernel: scan 32 haystack positions per iteration,
+/// checking two rare bytes (case-insensitive) simultaneously.
+/// 4 cmpeq + 2 or + 1 and + 1 movemask per 32 bytes — same memory
+/// bandwidth as memchr2 but quadratic selectivity.
+#[cfg(target_arch = "x86_64")]
+#[target_feature(enable = "avx2")]
+unsafe fn search_packed_pair_avx2(
+    haystack: &[u8],
+    needle_lower: &[u8],
+    i1: usize,
+    i2: usize,
+) -> bool {
+    use core::arch::x86_64::*;
+
+    let n = needle_lower.len();
+    let hlen = haystack.len();
+    let ptr = haystack.as_ptr();
+    let last_start = hlen - n; // last valid match-start position
+
+    let b1 = needle_lower[i1];
+    let b1_alt = if b1.is_ascii_lowercase() {
+        ascii_swap_case(b1)
+    } else {
+        b1
+    };
+    let b2 = needle_lower[i2];
+    let b2_alt = if b2.is_ascii_lowercase() {
+        ascii_swap_case(b2)
+    } else {
+        b2
+    };
+
+    let v1_lo = _mm256_set1_epi8(b1 as i8);
+    let v1_hi = _mm256_set1_epi8(b1_alt as i8);
+    let v2_lo = _mm256_set1_epi8(b2 as i8);
+    let v2_hi = _mm256_set1_epi8(b2_alt as i8);
+
+    // Main loop: process 32 candidate positions per iteration.
+    // We load from ptr+offset+i1 and ptr+offset+i2, so we need
+    // offset + max(i1,i2) + 31 < hlen.
+    let max_idx = i1.max(i2);
+    let max_offset = hlen.saturating_sub(max_idx + 32);
+    let mut offset = 0usize;
+
+    while offset <= max_offset {
+        let chunk1 = unsafe { _mm256_loadu_si256(ptr.add(offset + i1) as *const __m256i) };
+        let chunk2 = unsafe { _mm256_loadu_si256(ptr.add(offset + i2) as *const __m256i) };
+
+        // Case-insensitive match: OR both case variants, then AND the two positions.
+        let eq1 = _mm256_or_si256(
+            _mm256_cmpeq_epi8(chunk1, v1_lo),
+            _mm256_cmpeq_epi8(chunk1, v1_hi),
+        );
+        let eq2 = _mm256_or_si256(
+            _mm256_cmpeq_epi8(chunk2, v2_lo),
+            _mm256_cmpeq_epi8(chunk2, v2_hi),
+        );
+
+        let mut mask = _mm256_movemask_epi8(_mm256_and_si256(eq1, eq2)) as u32;
+
+        while mask != 0 {
+            let bit = mask.trailing_zeros() as usize;
+            let candidate = offset + bit;
+            if candidate > last_start {
+                // Past the end — no more valid positions in this or future chunks.
+                return false;
+            }
+            if unsafe { verify_dispatch(ptr.add(candidate), needle_lower) } {
+                return true;
+            }
+            mask &= mask - 1;
+        }
+
+        offset += 32;
+    }
+
+    // Tail: remaining positions that couldn't fill a full vector.
+    // Use memchr2 on the rarest byte for these last few positions.
+    if offset <= last_start {
+        let rare_pos =
+            if case_insensitive_rank(needle_lower[i1]) <= case_insensitive_rank(needle_lower[i2]) {
+                i1
+            } else {
+                i2
+            };
+        let rare_byte = needle_lower[rare_pos];
+        let tail_start = offset + rare_pos;
+        let tail_end = last_start + rare_pos + 1;
+        if tail_start < tail_end {
+            let tail_space = &haystack[tail_start..tail_end];
+            if rare_byte.is_ascii_lowercase() {
+                for pos in memchr::memchr2_iter(rare_byte, ascii_swap_case(rare_byte), tail_space) {
+                    let candidate = offset + pos;
+                    if unsafe { verify_dispatch(ptr.add(candidate), needle_lower) } {
+                        return true;
+                    }
+                }
+            } else {
+                for pos in memchr::memchr_iter(rare_byte, tail_space) {
+                    let candidate = offset + pos;
+                    if unsafe { verify_dispatch(ptr.add(candidate), needle_lower) } {
+                        return true;
+                    }
+                }
+            }
+        }
+    }
+
+    false
+}
+
+/// Packed-pair case-insensitive substring search.
+///
+/// Selects the two rarest bytes from the needle (using the memchr byte
+/// frequency heuristic), then SIMD-scans for both at their known offsets
+/// simultaneously.  Falls back to `search` for needles shorter than 2 bytes.
+pub fn search_packed_pair(haystack: &[u8], needle_lower: &[u8]) -> bool {
+    let n = needle_lower.len();
+    if n == 0 {
+        return true;
+    }
+    if n < 2 {
+        return search(haystack, needle_lower);
+    }
+    if n > haystack.len() {
+        return false;
+    }
+
+    let (i1, i2) = select_rare_pair(needle_lower);
+
+    #[cfg(target_arch = "x86_64")]
+    {
+        if std::is_x86_feature_detected!("avx2") {
+            // Need enough haystack for at least one vector load.
+            let max_idx = i1.max(i2);
+            if haystack.len() >= max_idx + 32 {
+                return unsafe { search_packed_pair_avx2(haystack, needle_lower, i1, i2) };
+            }
+        }
+    }
+
+    #[cfg(target_arch = "aarch64")]
+    {
+        // The NEON packed-pair scan checks 16 bytes/iteration with ~7 ops,
+        // while memchr's optimized loop processes more bytes with fewer ops.
+        // Packed-pair wins when the first byte is common (lots of false
+        // positives for memchr2 that we avoid). But when the first byte is
+        // rare (z, q, x, ...) memchr2 has no false positives and its raw
+        // throughput dominates. Threshold 200 on the frequency table splits
+        // common letters (s=243, e=253, f=227) from rare ones (z=152, q=139).
+        let first_byte_rank = case_insensitive_rank(needle_lower[0]);
+        let max_idx = i1.max(i2);
+        if first_byte_rank >= 200 && haystack.len() >= max_idx + 16 {
+            return unsafe { search_packed_pair_neon(haystack, needle_lower, i1, i2) };
+        }
+    }
+
+    // Fallback for short haystacks or non-SIMD platforms.
+    search(haystack, needle_lower)
+}
+
+// ── Original memchr2 first-byte search ──────────────────────────────────
+
+/// Case-insensitive search using memchr2 on the first byte.
+pub fn search(haystack: &[u8], needle_lower: &[u8]) -> bool {
+    let n = needle_lower.len();
+    if n == 0 {
+        return true;
+    }
+    if n > haystack.len() {
+        return false;
+    }
+
+    let search_space = &haystack[..=haystack.len() - n];
+    let first = needle_lower[0];
+
+    if first.is_ascii_lowercase() {
+        let alt = ascii_swap_case(first);
+        for pos in memchr::memchr2_iter(first, alt, search_space) {
+            if unsafe { verify_dispatch(haystack.as_ptr().add(pos), needle_lower) } {
+                return true;
+            }
+        }
+    } else {
+        for pos in memchr::memchr_iter(first, search_space) {
+            if unsafe { verify_dispatch(haystack.as_ptr().add(pos), needle_lower) } {
+                return true;
+            }
+        }
+    }
+    false
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn basic_case_insensitive() {
+        assert!(search_packed_pair(b"Hello World", b"hello"));
+        assert!(search_packed_pair(b"Hello World", b"world"));
+        assert!(search_packed_pair(b"NOMORE bugs", b"nomore"));
+        assert!(!search_packed_pair(b"Hello World", b"xyz"));
+    }
+
+    #[test]
+    fn edge_cases() {
+        assert!(search_packed_pair(b"ab", b"ab"));
+        assert!(search_packed_pair(b"AB", b"ab"));
+        assert!(!search_packed_pair(b"a", b"ab"));
+        assert!(search_packed_pair(b"anything", b""));
+        assert!(!search_packed_pair(b"", b"x"));
+    }
+
+    #[test]
+    fn packed_pair_matches_search() {
+        let haystacks: &[&[u8]] = &[
+            b"The quick brown fox jumps over the lazy dog",
+            b"int mutex_lock(struct mutex *lock) { return 0; }",
+            b"#define NOMORE_RETRIES 5\nif (nomore) return;",
+            b"abcdefghijklmnopqrstuvwxyz",
+            b"short",
+        ];
+        let needles: &[&[u8]] = &[b"fox", b"mutex", b"nomore", b"xyz", b"the", b"short", b"qr"];
+        for h in haystacks {
+            for n in needles {
+                let lower: Vec<u8> = n.iter().map(|b| b.to_ascii_lowercase()).collect();
+                assert_eq!(
+                    search_packed_pair(h, &lower),
+                    search(h, &lower),
+                    "mismatch for haystack={:?} needle={:?}",
+                    std::str::from_utf8(h),
+                    std::str::from_utf8(n),
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn long_haystack_neon_path() {
+        // Haystack > 16 bytes exercises NEON packed-pair search loop
+        let haystack =
+            b"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaTHIS_IS_A_LONG_NEEDLE_TESTbbbbbbbbbbbbbbbbbb";
+        assert!(search_packed_pair(haystack, b"this_is_a_long_needle_test"));
+        assert!(!search_packed_pair(
+            haystack,
+            b"this_is_a_long_needle_testz"
+        ));
+
+        // Needle >= 16 bytes exercises NEON dotprod verify
+        let long_needle = b"struct mutex *lock";
+        let haystack2 = b"int STRUCT MUTEX *LOCK(struct mutex *lock) { return 0; }";
+        assert!(search_packed_pair(haystack2, long_needle));
+
+        // All uppercase haystack, lowercase needle
+        let upper_hay = b"ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ";
+        assert!(search_packed_pair(upper_hay, b"qrstuvwxyz0123456789a"));
+        assert!(!search_packed_pair(upper_hay, b"qrstuvwxyz01234567899"));
+
+        // Needle at very end
+        let end_hay = b"xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxfind_me";
+        assert!(search_packed_pair(end_hay, b"find_me"));
+
+        // Needle at very start
+        assert!(search_packed_pair(end_hay, b"xx"));
+
+        // 1KB haystack with needle near the end
+        let mut big = vec![b'z'; 1024];
+        big[1000..1010].copy_from_slice(b"hElLo_WoRl");
+        assert!(search_packed_pair(&big, b"hello_wo"));
+        assert!(!search_packed_pair(&big, b"hello_world"));
+    }
+
+    #[test]
+    fn rare_pair_selection() {
+        // For "nomore": n=246, o=244, m=233, o=244, r=245, e=253
+        // Rarest positions should include 'm' (pos 2, rank 233)
+        let (i1, i2) = select_rare_pair(b"nomore");
+        let ranks: Vec<u8> = b"nomore"
+            .iter()
+            .map(|&b| case_insensitive_rank(b))
+            .collect();
+        let r1 = ranks[i1];
+        let r2 = ranks[i2];
+        // Both selected ranks should be <= all other ranks
+        for (i, &r) in ranks.iter().enumerate() {
+            if i != i1 && i != i2 {
+                assert!(r1 <= r || r2 <= r, "pair ({i1},{i2}) not optimal");
+            }
+        }
+    }
+}

crates/fff-core/src/constraints.rs

@@ -12,6 +12,32 @@ use smallvec::SmallVec;
 
 use crate::git::is_modified_status;
 
+/// Case-insensitive ASCII substring search without allocation.
+/// `needle` must already be lowercase.
+#[inline]
+fn contains_ascii_ci(haystack: &str, needle: &str) -> bool {
+    let h = haystack.as_bytes();
+    let n = needle.as_bytes();
+    if n.len() > h.len() {
+        return false;
+    }
+    if n.is_empty() {
+        return true;
+    }
+    let first = n[0];
+    for i in 0..=(h.len() - n.len()) {
+        if h[i].to_ascii_lowercase() == first
+            && h[i..i + n.len()]
+                .iter()
+                .zip(n)
+                .all(|(a, b)| a.to_ascii_lowercase() == *b)
+        {
+            return true;
+        }
+    }
+    false
+}
+
 /// Minimum item count before switching to parallel iteration with rayon.
 /// Below this threshold, the overhead of thread pool dispatch outweighs the benefit.
 const PAR_THRESHOLD: usize = 10_000;
@@ -22,9 +48,6 @@ pub trait Constrainable {
     /// The file's relative path (e.g. "src/main.rs")
     fn relative_path(&self) -> &str;
 
-    /// The file's lowercased relative path for case-insensitive matching
-    fn relative_path_lower(&self) -> &str;
-
     /// The file name component (e.g. "main.rs")
     fn file_name(&self) -> &str;
 
@@ -32,6 +55,28 @@ 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 {
@@ -44,12 +89,14 @@ 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)
@@ -101,6 +148,7 @@ 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),
@@ -127,7 +175,7 @@ fn item_matches_constraint_at_index<T: Constrainable>(
         }
 
         // only works with negation
-        Constraint::Text(text) => item.relative_path_lower().contains(text),
+        Constraint::Text(text) => contains_ascii_ci(item.relative_path(), text),
 
         // Parts and Exclude are handled at a higher level
         Constraint::Parts(_) | Constraint::Exclude(_) | Constraint::FileType(_) => true,
@@ -347,8 +395,78 @@ 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"));
+    }
 }

crates/fff-core/src/error.rs

@@ -7,6 +7,10 @@ pub enum Error {
     ThreadPanic,
     #[error("Invalid path {0}")]
     InvalidPath(std::path::PathBuf),
+    #[error(
+        "Can not start fff at the file system root {0} — pass a project or at least home directory instead"
+    )]
+    FilesystemRoot(std::path::PathBuf),
     #[error("File picker not initialized")]
     FilePickerMissing,
     #[error("Failed to acquire lock for frecency")]
@@ -21,6 +25,8 @@ pub enum Error {
     EnvOpen(#[source] heed::Error),
     #[error("Failed to create frecency database: {0}")]
     DbCreate(#[source] heed::Error),
+    #[error("Failed to open frecency database: {0}")]
+    DbOpen(#[source] heed::Error),
     #[error("Failed to clear stale readers for frecency database: {0}")]
     DbClearStaleReaders(#[source] heed::Error),
 

crates/fff-core/src/frecency.rs

@@ -1,11 +1,15 @@
 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::git::is_modified_status;
+use crate::shared::SharedFrecency;
 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 +17,10 @@ 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 +35,21 @@ 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 +58,13 @@ 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);
             }
@@ -53,11 +73,28 @@ impl FrecencyTracker {
         env.clear_stale_readers()
             .map_err(Error::DbClearStaleReaders)?;
 
-        // we will open the default unnamed database
-        let mut wtxn = env.write_txn().map_err(Error::DbStartWriteTxn)?;
-        let db = env
-            .create_database(&mut wtxn, None)
-            .map_err(Error::DbCreate)?;
+        // Try read-only open first — avoids blocking on the LMDB write lock
+        // when another process (Neovim, another fff-mcp) already has it.
+        // Only fall back to create_database (which needs a write txn) if the
+        // database doesn't exist yet.
+        let rtxn = env.read_txn().map_err(Error::DbStartReadTxn)?;
+        let maybe_db: Option<Database<Bytes, SerdeBincode<VecDeque<u64>>>> =
+            env.open_database(&rtxn, None).map_err(Error::DbOpen)?;
+
+        drop(rtxn);
+
+        let db = match maybe_db {
+            Some(db) => db,
+            None => {
+                // First time: create the database (requires write lock).
+                let mut wtxn = env.write_txn().map_err(Error::DbStartWriteTxn)?;
+                let db = env
+                    .create_database(&mut wtxn, None)
+                    .map_err(Error::DbCreate)?;
+                wtxn.commit().map_err(Error::DbCommit)?;
+                db
+            }
+        };
 
         Ok(FrecencyTracker {
             db,
@@ -65,7 +102,209 @@ 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 +318,7 @@ 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 +326,15 @@ 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 +362,7 @@ 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 +373,21 @@ 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 +395,7 @@ 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 +413,31 @@ 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 +457,7 @@ 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 +535,35 @@ 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);

crates/fff-core/src/git.rs

@@ -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 => "clean",
+        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")
+}

crates/fff-core/src/ignore.rs

@@ -0,0 +1,61 @@
+use std::path::Path;
+
+pub(crate) const NON_GIT_IGNORED_DIRS: &[&str] = &[
+    "node_modules",
+    "__pycache__",
+    "venv",
+    ".venv",
+    // Rust (these are glob-only patterns for non_git_repo_overrides,
+    // is_non_code_directory matches the "target" component separately)
+    "target/debug",
+    "target/release",
+    "target/rust-analyzer",
+    "target/criterion",
+];
+
+#[cfg(target_os = "macos")]
+pub(crate) const PLATFORM_IGNORED_DIRS: &[&str] =
+    &["Library/Application Support", "Library/Caches"];
+
+#[cfg(target_os = "windows")]
+pub(crate) const PLATFORM_IGNORED_DIRS: &[&str] = &[
+    "bin/Debug",
+    "bin/Release",
+    "Program Files",
+    "Program Files (x86)",
+    "AppData/Local",
+    "AppData/Roaming",
+];
+
+#[cfg(not(any(target_os = "macos", target_os = "windows")))]
+pub(crate) const PLATFORM_IGNORED_DIRS: &[&str] = &[];
+
+pub(crate) fn non_git_repo_overrides(base_path: &Path) -> Option<ignore::overrides::Override> {
+    use ignore::overrides::OverrideBuilder;
+
+    let mut builder = OverrideBuilder::new(base_path);
+    for dir in NON_GIT_IGNORED_DIRS.iter().chain(PLATFORM_IGNORED_DIRS) {
+        let pattern = format!("!**/{dir}/");
+        if let Err(e) = builder.add(&pattern) {
+            tracing::warn!("failed to add ignore pattern {pattern}: {e}");
+        }
+    }
+
+    builder.build().ok()
+}
+
+pub(crate) fn is_non_code_directory(path: &Path) -> bool {
+    let path_str = path.as_os_str().to_str().unwrap_or("");
+    NON_GIT_IGNORED_DIRS
+        .iter()
+        .chain(PLATFORM_IGNORED_DIRS)
+        .any(|&dir| {
+            #[cfg(target_os = "windows")]
+            let dir = dir.replace('/', std::path::MAIN_SEPARATOR_STR);
+            #[cfg(target_os = "windows")]
+            return path_str.contains(dir.as_str());
+
+            #[cfg(not(target_os = "windows"))]
+            path_str.contains(dir)
+        })
+}

crates/fff-core/src/lib.rs

@@ -1,42 +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.
+//! 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.
 //!
-//! # State management
+//! ## Architecture
 //!
-//! All state is instance-based. Callers create their own `SharedPicker` /
-//! `SharedFrecency` / `SharedQueryTracker` and pass them into
-//! `FilePicker::new_with_shared_state`. Multiple independent instances can
-//! coexist in the same process.
+//! - [`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
+//! newtype wrappers around `Arc<RwLock<Option<T>>>` for thread-safe shared
+//! access. They provide `read()` / `write()` methods with built-in error
+//! conversion and convenience helpers like `wait_for_scan()`.
+//!
+//! ## Quick Start
+//!
+//! ```
+//! use fff_search::file_picker::FilePicker;
+//! use fff_search::frecency::FrecencyTracker;
+//! use fff_search::query_tracker::QueryTracker;
+//! use fff_search::{
+//!     FFFMode, FilePickerOptions, FuzzySearchOptions, PaginationArgs, QueryParser,
+//!     SharedFrecency, SharedPicker, SharedQueryTracker,
+//! };
+//!
+//! let shared_picker = SharedPicker::default();
+//! let shared_frecency = SharedFrecency::default();
+//! let shared_query_tracker = SharedQueryTracker::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.init(frecency)?;
+//!
+//! let query_tracker = QueryTracker::new(tmp.join("queries"), false)?;
+//! shared_query_tracker.init(query_tracker)?;
+//!
+//! // 2. Init the file picker (spawns background scan + watcher)
+//! FilePicker::new_with_shared_state(
+//!     shared_picker.clone(),
+//!     shared_frecency.clone(),
+//!     FilePickerOptions {
+//!         base_path: ".".into(),
+//!         mode: FFFMode::Ai,
+//!         ..Default::default()
+//!     },
+//! )?;
+//!
+//! // 3. Wait for scan
+//! shared_picker.wait_for_scan(std::time::Duration::from_secs(10));
+//!
+//! // 4. Search: lock the picker and query tracker
+//! let picker_guard = shared_picker.read()?;
+//! let picker = picker_guard.as_ref().unwrap();
+//! let qt_guard = shared_query_tracker.read()?;
+//!
+//! // 5. Parse the query and perform fuzzy search
+//! let parser = QueryParser::default();
+//! let query = parser.parse("lib.rs");
+//!
+//! let results = FilePicker::fuzzy_search(
+//!     picker.get_files(),
+//!     &query,
+//!     qt_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().as_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 bigram_filter;
+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;
+// this is pub only for benchmarks
+pub mod case_insensitive_memmem;
+
+/// 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 query_tracker::QueryTracker;
-use std::sync::{Arc, RwLock};
-
-pub type SharedPicker = Arc<RwLock<Option<FilePicker>>>;
-pub type SharedFrecency = Arc<RwLock<Option<FrecencyTracker>>>;
-pub type SharedQueryTracker = Arc<RwLock<Option<QueryTracker>>>;
+mod ignore;
+/// Thread-safe shared handles for [`FilePicker`], [`FrecencyTracker`],
+/// and [`QueryTracker`].
+pub mod shared;
 
+pub use bigram_filter::*;
 pub use db_healthcheck::{DbHealth, DbHealthChecker};
 pub use error::{Error, Result};
-pub use fff_query_parser::{
-    Constraint, FFFQuery, FuzzyQuery, Location, QueryParser, location::parse_location,
-};
-pub use file_picker::{FuzzySearchOptions, ScanProgress};
-pub use grep::{GrepMatch, GrepMode, GrepResult, GrepSearchOptions};
-pub use types::{FileItem, PaginationArgs, Score, ScoringContext, SearchResult};
+pub use fff_query_parser::*;
+pub use file_picker::*;
+pub use frecency::*;
+pub use grep::*;
+pub use query_tracker::*;
+pub use shared::*;
+pub use types::*;

crates/fff-core/src/log.rs

@@ -0,0 +1,150 @@
+//! Shared logging utilities for FFF crates.
+//!
+//! Provides file-based tracing initialization and crash handlers (panic hook
+//! + SIGSEGV signal handler) that write diagnostics to both stderr and the
+//! configured log file.
+
+use std::io;
+use std::path::{Path, PathBuf};
+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 CRASH_HANDLERS_INSTALLED: std::sync::OnceLock<()> = std::sync::OnceLock::new();
+
+/// The log file path set by `init_tracing`. Crash handlers append to this file.
+static LOG_FILE_PATH: std::sync::OnceLock<PathBuf> = std::sync::OnceLock::new();
+
+fn write_crash_report(header: &str, body: &str) {
+    let msg = format!(
+        "\n=== CRASH {} ===\n{}\n=== CRASH END {} ===\n",
+        header, body, header
+    );
+
+    let _ = std::io::Write::write_all(&mut std::io::stderr(), msg.as_bytes());
+
+    if let Some(path) = LOG_FILE_PATH.get() {
+        let _ = std::fs::OpenOptions::new()
+            .create(true)
+            .append(true)
+            .open(path)
+            .and_then(|mut f| std::io::Write::write_all(&mut f, msg.as_bytes()));
+    }
+}
+
+extern "C" fn sigsegv_handler(sig: libc::c_int) {
+    let bt = std::backtrace::Backtrace::force_capture();
+    write_crash_report("SIGSEGV", &format!("signal {}\n{}", sig, bt));
+
+    unsafe {
+        libc::signal(sig, libc::SIG_DFL);
+        libc::raise(sig);
+    }
+}
+
+/// Install both the panic hook and the SIGSEGV signal handler.
+pub fn install_panic_hook() {
+    CRASH_HANDLERS_INSTALLED.get_or_init(|| {
+        let default_panic = std::panic::take_hook();
+        std::panic::set_hook(Box::new(move |panic_info| {
+            let message = if let Some(s) = panic_info.payload().downcast_ref::<&str>() {
+                s.to_string()
+            } else if let Some(s) = panic_info.payload().downcast_ref::<String>() {
+                s.clone()
+            } else {
+                "Unknown panic payload".to_string()
+            };
+
+            let location = panic_info
+                .location()
+                .map(|l| format!("{}:{}:{}", l.file(), l.line(), l.column()))
+                .unwrap_or_else(|| "unknown location".to_string());
+
+            tracing::error!(
+                panic.message = %message,
+                panic.location = %location,
+                "PANIC occurred in FFF"
+            );
+
+            write_crash_report(
+                "RUST PANIC",
+                &format!("Message: {}\nLocation: {}", message, location),
+            );
+            default_panic(panic_info);
+        }));
+
+        unsafe {
+            libc::signal(
+                libc::SIGSEGV,
+                sigsegv_handler as *const () as libc::sighandler_t,
+            );
+        }
+    });
+}
+
+/// Parse a log level string into a `tracing::Level`.
+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.
+pub fn init_tracing(log_file_path: &str, log_level: Option<&str>) -> Result<String, io::Error> {
+    let log_path = Path::new(log_file_path);
+    if let Some(parent) = log_path.parent() {
+        std::fs::create_dir_all(parent)?;
+    }
+
+    let _ = LOG_FILE_PATH.set(log_path.to_path_buf());
+    install_panic_hook();
+
+    let file_appender = std::fs::OpenOptions::new()
+        .create(true)
+        .write(true)
+        .truncate(true) // truncates a file on restart (instead of appending)
+        .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_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())
+}

crates/fff-core/src/path_utils.rs

@@ -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 {

crates/fff-core/src/query_tracker.rs

@@ -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);
@@ -243,7 +246,6 @@ impl QueryTracker {
         min_combo_count: u32,
     ) -> Result<Option<QueryMatchEntry>, Error> {
         let query_key = Self::create_query_key(project_path, query)?;
-        tracing::debug!(?query_key, "HASH");
         let rtxn = self.env.read_txn().map_err(Error::DbStartReadTxn)?;
 
         let last_match = self

crates/fff-core/src/score.rs

@@ -34,24 +34,6 @@ impl<'a> FileItems<'a> {
         self.len() == 0
     }
 
-    #[inline]
-    fn get(&self, index: usize) -> Option<&'a FileItem> {
-        match self {
-            FileItems::All(s) => s.get(index),
-            FileItems::Filtered(v) => v.get(index).copied(),
-        }
-    }
-
-    /// Build the haystack of relative paths (original casing) for fuzzy matching.
-    /// neo_frizbee lowercases internally for comparison but preserves original casing
-    /// for capitalization_bonus and matching_case_bonus scoring.
-    fn relative_paths(&self) -> Vec<&'a str> {
-        match self {
-            FileItems::All(s) => s.iter().map(|f| f.relative_path.as_str()).collect(),
-            FileItems::Filtered(v) => v.iter().map(|f| f.relative_path.as_str()).collect(),
-        }
-    }
-
     /// Index into the file list. Panics if out of bounds (like slice indexing).
     #[inline]
     fn index(&self, index: usize) -> &'a FileItem {
@@ -66,17 +48,17 @@ impl<'a> FileItems<'a> {
 /// Single part: use optimized batch matching.
 /// Multiple parts: each part must match, scores are summed (Nucleo-style).
 /// Parts with less than 2 characters are skipped.
+///
+/// Files are passed directly to frizbee via the `Matchable` trait —
+/// deleted files return `None` from `match_str()` and are skipped
+/// without any intermediate allocation.
+#[inline]
 fn match_fuzzy_parts(
     fuzzy_parts: &[&str],
     working_files: &FileItems<'_>,
     options: &neo_frizbee::Config,
+    max_threads: usize,
 ) -> Vec<neo_frizbee::Match> {
-    if fuzzy_parts.is_empty() {
-        return vec![];
-    }
-
-    let haystack: Vec<&str> = working_files.relative_paths();
-
     // Filter out parts that are too short (< 2 chars)
     let valid_parts: Vec<&str> = fuzzy_parts
         .iter()
@@ -89,14 +71,21 @@ fn match_fuzzy_parts(
         return vec![];
     }
 
+    let first_part_matches = match working_files {
+        FileItems::All(files) => {
+            neo_frizbee::match_list_parallel(valid_parts[0], files, options, max_threads)
+        }
+        FileItems::Filtered(files) => {
+            neo_frizbee::match_list_parallel(valid_parts[0], files, options, max_threads)
+        }
+    };
+
     if valid_parts.len() == 1 {
-        let matches = neo_frizbee::match_list(valid_parts[0], &haystack, options);
-        return matches;
+        return first_part_matches;
     }
 
     // Multiple parts - match first part, then filter by remaining parts
-    // TODO figure out if we can move this logic to my frizbee fork at least
-    let mut matches = neo_frizbee::match_list(valid_parts[0], &haystack, options);
+    let mut matches = first_part_matches;
     for part in valid_parts[1..].iter() {
         let mut part_options = *options;
         part_options.max_typos = options.max_typos.map(|t| t.min(part.len() as u16));
@@ -104,8 +93,9 @@ fn match_fuzzy_parts(
         matches = matches
             .into_iter()
             .filter_map(|mut m| {
-                let path = haystack.get(m.index as usize)?;
-                let part_matches = neo_frizbee::match_list(part, &[*path], &part_options);
+                let file = working_files.index(m.index as usize);
+                let path = file.relative_path();
+                let part_matches = neo_frizbee::match_list(part, &[path], &part_options);
                 let part_match = part_matches.first()?;
 
                 // Sum scores
@@ -131,40 +121,27 @@ 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 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");
-            if query_trimmed.len() < 2 {
-                return score_filtered_by_frecency(&working_files, context);
-            }
-            &single_part_storage
+    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(),
+        _ => {
+            return score_filtered_by_frecency(&working_files, context);
         }
-        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);
-            }
-        },
     };
+    debug_assert!(!fuzzy_parts.is_empty());
 
     let has_uppercase = fuzzy_parts
         .iter()
@@ -172,7 +149,6 @@ 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 {
@@ -182,104 +158,130 @@ pub fn match_and_score_files<'a>(
         },
     };
 
-    let path_matches = match_fuzzy_parts(fuzzy_parts, &working_files, &options);
-    let primary_text = fuzzy_parts[0]; // Use first part for filename matching
-    let haystack_of_filenames: Vec<&str> = path_matches
-        .iter()
-        .filter_map(|m| {
-            working_files
-                .get(m.index as usize)
-                .map(|f| f.file_name.as_str())
-        })
-        .collect();
+    let path_matches =
+        match_fuzzy_parts(fuzzy_parts, &working_files, &options, context.max_threads);
 
-    // if there is a / in the query we don't even match filenames
-    let filename_matches = if query_contains_path_separator {
+    let main_needle = fuzzy_parts[0].as_bytes(); // safe
+    let main_needle_len = main_needle.len() as u16;
+
+    // Filename match detection: two tiers, cursor-based (no intermediate bitset/Vec<bool>).
+    // 1) Collect filenames only where match_end_col didn't land in the filename region.
+    // 2) Batch SIMD on that subset, remap indices, sort for cursor walk in the scoring loop.
+    let mut fallback_indices: Vec<u32> = Vec::new();
+    let filename_fallback_matches = if query_contains_path_separator || path_matches.len() > 15_000
+    {
         vec![]
     } else {
-        // Use parallel matching only if we have enough filenames to justify overhead
-        // Sequential matching is faster for small result sets (< 1000 matches)
-        let mut list = if haystack_of_filenames.len() > 1000 {
-            neo_frizbee::match_list_parallel(
-                primary_text,
-                &haystack_of_filenames,
-                &options,
-                context.max_threads,
-            )
-        } else {
-            neo_frizbee::match_list(primary_text, &haystack_of_filenames, &options)
-        };
+        let mut fallback_filenames: Vec<&str> = Vec::new();
 
-        // Sequential sort is faster for small lists
-        if list.len() > 1000 {
-            list.par_sort_unstable_by_key(|m| m.index);
-        } else {
-            sort_by_key_with_buffer(&mut list, |m| m.index);
+        for (i, path_match) in path_matches.iter().enumerate() {
+            let file = working_files.index(path_match.index as usize);
+            let filename_start = file.filename_offset_in_relative() as u16;
+            let match_start_approx = path_match.end_col.saturating_sub(main_needle_len - 1);
+
+            if match_start_approx < filename_start {
+                fallback_indices.push(i as u32);
+                fallback_filenames.push(file.file_name());
+            }
         }
 
-        list
+        if fallback_filenames.is_empty() {
+            vec![]
+        } else {
+            let mut matches = neo_frizbee::match_list_parallel(
+                fuzzy_parts[0],
+                &fallback_filenames,
+                &options,
+                if path_matches.len() > 10_000 {
+                    context.max_threads
+                } else {
+                    1
+                },
+            );
+
+            sort_by_key_with_buffer(&mut matches, |m| fallback_indices[m.index as usize]);
+            matches
+        }
     };
 
-    let mut next_filename_match_index = 0;
+    let mut next_filename_match_cursor = 0;
     let results: Vec<_> = path_matches
         .into_iter()
         .enumerate()
-        .map(|(index, path_match)| {
+        .map(|(match_idx, path_match)| {
             let file_idx = path_match.index as usize;
             let file = working_files.index(file_idx);
 
-            let mut base_score = path_match.score as i32;
-            let frecency_boost = base_score.saturating_mul(file.total_frecency_score as i32) / 100;
-            let distance_penalty =
-                calculate_distance_penalty(context.current_file, &file.relative_path);
+            let base_score = path_match.score as i32;
+            let frecency_boost = base_score.saturating_mul(file.total_frecency_score()) / 100;
 
-            let filename_match = filename_matches
-                .get(next_filename_match_index)
-                .and_then(|m| {
-                    if m.index == index as u32 {
-                        next_filename_match_index += 1;
-                        Some(m)
-                    } else {
-                        None
-                    }
-                });
+            // 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());
+
+            let filename_start = file.filename_offset_in_relative() as u16;
+            let match_start_approx = path_match.end_col.saturating_sub(main_needle_len - 1);
+
+            let end_col_filename_match = match_start_approx >= filename_start;
+            let simd_filename_match = if !end_col_filename_match {
+                filename_fallback_matches
+                    .get(next_filename_match_cursor)
+                    .and_then(|m| {
+                        if fallback_indices[m.index as usize] == match_idx as u32 {
+                            next_filename_match_cursor += 1;
+                            Some(m)
+                        } else {
+                            None
+                        }
+                    })
+            } else {
+                None
+            };
+
+            let is_filename_match = end_col_filename_match || simd_filename_match.is_some();
+            let is_exact_filename = simd_filename_match.is_some_and(|m| m.exact)
+                || (end_col_filename_match
+                    && main_needle_len as usize == file.file_name().len()
+                    && main_needle.eq_ignore_ascii_case(file.file_name().as_bytes()));
 
             let mut has_special_filename_bonus = false;
-            let filename_bonus = match filename_match {
-                Some(filename_match) if filename_match.exact => {
-                    filename_match.score as i32 / 5 * 2 // 40% bonus for exact filename match
+            let filename_bonus = if is_exact_filename {
+                base_score / 5 * 2 // 40% bonus for exact filename match
+            } else if is_filename_match {
+                // 16% bonus for fuzzy filename match that landed in the filename region.
+                // For fallback matches (where the path match landed in a directory segment),
+                // scale the bonus by the quality of the filename match — a contiguous match
+                // like "rename" in "rename.ts" gets the full bonus, while a scattered
+                // subsequence like r-e-n-a-m-e in "generateSessionName.ts" gets much less.
+                let max_bonus = (base_score / 6).min(30);
+                if let Some(fm) = simd_filename_match {
+                    let max_possible = main_needle_len as i32 * 16;
+                    let quality = (fm.score as i32).min(max_possible);
+                    max_bonus * quality / max_possible
+                } else {
+                    max_bonus
                 }
-                // 16% bonus for fuzzy filename match but only if the score of matched path is
-                // equal or greater than the score of matched filename, thus we are not allowing
-                // typoed filename to score higher than the path match
-                Some(filename_match)
-                    if filename_match.score >= path_match.score
-                        && !query_contains_path_separator =>
-                {
-                    base_score = filename_match.score as i32;
-
-                    (base_score / 6)
-                        // for large queries around ~300 score the bonus is too big
-                        // it might lead to situations when much more fitting path with a larger
-                        // base score getting filtered out by combination of score + filename bonus
-                        // so we cap it at 10% of the roughly largest score you can get
-                        .min(30)
-                }
-                // 5% bonus for special file but not as much as file name to avoid sitatuions
+            } else if !is_filename_match && is_special_entry_point_file(file.file_name()) {
+                // 5% bonus for special file but not as much as file name to avoid situations
                 // when you have /user_service/server.rs and /user_service/server/mod.rs
-                None if is_special_entry_point_file(&file.file_name) => {
-                    has_special_filename_bonus = true;
-                    base_score * 5 / 100
-                }
-                _ => 0,
+                has_special_filename_bonus = true;
+                base_score * 5 / 100
+            } else {
+                0
             };
 
             let current_file_penalty = calculate_current_file_penalty(file, base_score, context);
-
             let combo_match_boost = {
                 let last_same_query_match = context
                     .last_same_query_match
-                    .filter(|m| m.file_path.as_os_str() == file.path.as_os_str());
+                    .as_ref()
+                    .filter(|m| m.file_path.as_os_str() == file.as_path().as_os_str());
 
                 match last_same_query_match {
                     // if we request a combo match without a boost we have to render it anyway
@@ -296,6 +298,7 @@ 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)
@@ -312,13 +315,18 @@ 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 => "fuzzy_path",
+                exact_match: is_exact_filename || path_match.exact,
+                match_type: if is_exact_filename {
+                    "exact_filename"
+                } else if is_filename_match {
+                    "fuzzy_filename"
+                } else if path_match.exact {
+                    "exact_path"
+                } else {
+                    "fuzzy_path"
                 },
             };
 
@@ -363,9 +371,18 @@ 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 +393,7 @@ 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",
         };
@@ -384,8 +402,16 @@ pub(crate) fn score_filtered_by_frecency<'a>(
     };
 
     let results: Vec<_> = match files {
-        FileItems::All(s) => s.par_iter().map(&score_file).collect(),
-        FileItems::Filtered(v) => v.iter().map(|&file| score_file(file)).collect(),
+        FileItems::All(s) => s
+            .par_iter()
+            .filter(|f| !f.is_deleted())
+            .map(&score_file)
+            .collect(),
+        FileItems::Filtered(v) => v
+            .iter()
+            .filter(|f| !f.is_deleted())
+            .map(|&file| score_file(file))
+            .collect(),
     };
 
     sort_and_paginate(results, context)
@@ -400,7 +426,7 @@ fn calculate_current_file_penalty(
     let mut penalty = 0i32;
 
     if let Some(current) = context.current_file
-        && file.relative_path.as_str() == current
+        && file.relative_path() == current
     {
         penalty -= match file.git_status {
             Some(status) if is_modified_status(status) => base_score / 2,
@@ -482,14 +508,14 @@ fn sort_and_paginate<'a>(
 mod tests {
     use super::*;
     use crate::types::PaginationArgs;
-    use std::path::PathBuf;
+    use fff_query_parser::QueryParser;
 
     fn create_test_file(path: &str, score: i32, modified: u64) -> (FileItem, Score) {
-        let file_name = path.split('/').last().unwrap_or(path).to_string();
+        let filename_start = path.rfind('/').map(|i| i + 1).unwrap_or(0) as u16;
         let file = FileItem::new_raw(
-            PathBuf::from(path),
             path.to_string(),
-            file_name,
+            0,
+            filename_start,
             0,
             modified,
             None,
@@ -503,6 +529,7 @@ 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 +559,11 @@ 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,
@@ -560,15 +589,15 @@ mod tests {
         assert_eq!(scores[2].total, 200, "Third should be third highest");
 
         // Verify the files match
-        assert_eq!(items[0].relative_path, "file4.rs");
-        assert_eq!(items[1].relative_path, "file6.rs");
-        assert_eq!(items[2].relative_path, "file2.rs");
+        assert_eq!(items[0].relative_path(), "file4.rs");
+        assert_eq!(items[1].relative_path(), "file6.rs");
+        assert_eq!(items[2].relative_path(), "file2.rs");
     }
 
     #[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 +610,11 @@ 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 +648,7 @@ 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 +659,11 @@ 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,
@@ -652,9 +685,127 @@ mod tests {
         assert_eq!(scores[0].total, 200);
         assert_eq!(scores[1].total, 100);
         assert_eq!(scores[2].total, 50);
-        assert_eq!(items[0].relative_path, "file2.rs");
-        assert_eq!(items[1].relative_path, "file1.rs");
-        assert_eq!(items[2].relative_path, "file3.rs");
+        assert_eq!(items[0].relative_path(), "file2.rs");
+        assert_eq!(items[1].relative_path(), "file1.rs");
+        assert_eq!(items[2].relative_path(), "file3.rs");
+    }
+}
+
+#[cfg(test)]
+mod filename_bonus_tests {
+    use super::*;
+    use crate::types::PaginationArgs;
+    use fff_query_parser::QueryParser;
+
+    fn make_file(path: &str) -> FileItem {
+        let filename_start = path.rfind('/').map(|i| i + 1).unwrap_or(0) as u16;
+        FileItem::new_raw(path.to_string(), 0, filename_start, 0, 0, None, false)
+    }
+
+    fn search(files: &[FileItem], query: &str) -> Vec<(String, Score)> {
+        let parser = QueryParser::default();
+        let parsed = parser.parse(query);
+        let ctx = ScoringContext {
+            query: &parsed,
+            max_threads: 1,
+            max_typos: 2,
+            current_file: None,
+            last_same_query_match: None,
+            project_path: None,
+            combo_boost_score_multiplier: 100,
+            min_combo_count: 3,
+            pagination: PaginationArgs {
+                offset: 0,
+                limit: 100,
+            },
+        };
+        let (items, scores, _) = match_and_score_files(files, &ctx);
+        items
+            .iter()
+            .zip(scores.iter())
+            .map(|(f, s)| (f.relative_path().to_string(), s.clone()))
+            .collect()
+    }
+
+    #[test]
+    fn test_filename_match_ranks_above_path_only_match() {
+        let files = vec![
+            make_file("src/username/handler.rs"),
+            make_file("src/username/username.rs"),
+        ];
+
+        let results = search(&files, "usrnmea");
+
+        assert!(
+            results.len() >= 2,
+            "both files should match, got {}",
+            results.len()
+        );
+        assert_eq!(
+            results[0].0, "src/username/username.rs",
+            "filename match should rank first"
+        );
+        assert!(
+            results[0].1.filename_bonus > 0,
+            "username.rs should have filename bonus"
+        );
+        assert_eq!(
+            results[1].1.filename_bonus, 0,
+            "handler.rs should have no filename bonus"
+        );
+    }
+
+    #[test]
+    fn test_exact_filename_beats_fuzzy_filename() {
+        // "username.rs" exactly matches "username.rs" → exact filename
+        // "username.rs" is a fuzzy match of "user_name_handler.rs" → fuzzy bonus only
+        let files = vec![
+            make_file("src/user_name_handler.rs"),
+            make_file("src/username.rs"),
+        ];
+
+        let results = search(&files, "username.rs");
+
+        assert!(results.len() >= 2);
+        assert_eq!(
+            results[0].0, "src/username.rs",
+            "exact filename should rank first"
+        );
+        assert_eq!(results[0].1.match_type, "exact_filename");
+        assert!(results[0].1.filename_bonus > results[1].1.filename_bonus);
+    }
+
+    #[test]
+    fn test_same_length_filename_no_false_exact() {
+        // "item.rs" exactly matches "item.rs" → exact_filename
+        // "item.rs" should NOT get exact_filename on "file.rs" even though stem lengths match
+        let files = vec![
+            make_file("src/item_sync/file.rs"),
+            make_file("src/models/item.rs"),
+        ];
+
+        let results = search(&files, "item.rs");
+
+        assert!(results.len() >= 2);
+        assert_eq!(results[0].0, "src/models/item.rs");
+        assert_eq!(results[0].1.match_type, "exact_filename");
+        assert_ne!(
+            results[1].1.match_type, "exact_filename",
+            "file.rs should not get exact_filename"
+        );
+    }
+
+    #[test]
+    fn test_path_separator_disables_filename_bonus() {
+        let files = vec![make_file("src/controllers/user.rs")];
+
+        let results = search(&files, "src/user");
+
+        assert!(!results.is_empty());
+        assert_eq!(
+            results[0].1.filename_bonus, 0,
+            "path-like query should not get filename bonus"
+        );
     }
 }
 
@@ -666,7 +817,6 @@ 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 +848,6 @@ 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()

crates/fff-core/src/shared.rs

@@ -0,0 +1,180 @@
+use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
+use std::time::Duration;
+
+use crate::error::Error;
+use crate::file_picker::FilePicker;
+use crate::frecency::FrecencyTracker;
+use crate::git::GitStatusCache;
+use crate::query_tracker::QueryTracker;
+
+/// Thread-safe shared handle to the [`FilePicker`] instance.
+///
+/// Uses `parking_lot::RwLock` which is reader-fair — new readers are not
+/// blocked when a writer is waiting, preventing search query stalls during
+/// background bigram builds or watcher writes.
+///
+/// `Clone` gives a new handle to the same picker (Arc clone).
+/// `Default` creates an empty handle suitable for `Lazy::new(SharedPicker::default)`.
+#[derive(Clone, Default)]
+pub struct SharedPicker(pub(crate) Arc<parking_lot::RwLock<Option<FilePicker>>>);
+
+impl std::fmt::Debug for SharedPicker {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_tuple("SharedPicker").field(&"..").finish()
+    }
+}
+
+impl SharedPicker {
+    pub fn read(&self) -> Result<parking_lot::RwLockReadGuard<'_, Option<FilePicker>>, Error> {
+        Ok(self.0.read())
+    }
+
+    pub fn write(&self) -> Result<parking_lot::RwLockWriteGuard<'_, Option<FilePicker>>, Error> {
+        Ok(self.0.write())
+    }
+
+    /// Block until the background filesystem scan finishes.
+    /// Returns `true` if scan completed, `false` on timeout.
+    pub fn wait_for_scan(&self, timeout: Duration) -> bool {
+        let signal = {
+            let guard = self.0.read();
+            match &*guard {
+                Some(picker) => picker.scan_signal(),
+                None => return true,
+            }
+        };
+
+        let start = std::time::Instant::now();
+        while signal.load(std::sync::atomic::Ordering::Acquire) {
+            if start.elapsed() >= timeout {
+                return false;
+            }
+            std::thread::sleep(Duration::from_millis(10));
+        }
+        true
+    }
+
+    /// Block until the background file watcher is ready.
+    /// Returns `true` if watcher ready, `false` on timeout.
+    pub fn wait_for_watcher(&self, timeout: Duration) -> bool {
+        let signal = {
+            let guard = self.0.read();
+            match &*guard {
+                Some(picker) => picker.watcher_signal(),
+                None => return true,
+            }
+        };
+
+        let start = std::time::Instant::now();
+        while !signal.load(std::sync::atomic::Ordering::Acquire) {
+            if start.elapsed() >= timeout {
+                return false;
+            }
+            std::thread::sleep(Duration::from_millis(10));
+        }
+        true
+    }
+
+    /// Refresh git statuses for all indexed files.
+    pub fn refresh_git_status(&self, shared_frecency: &SharedFrecency) -> Result<usize, Error> {
+        use git2::StatusOptions;
+        use tracing::debug;
+
+        let git_status = {
+            let guard = self.read()?;
+            let Some(ref picker) = *guard else {
+                return Err(Error::FilePickerMissing);
+            };
+
+            debug!(
+                "Refreshing git statuses for picker: {:?}",
+                picker.git_root()
+            );
+
+            GitStatusCache::read_git_status(
+                picker.git_root(),
+                StatusOptions::new()
+                    .include_untracked(true)
+                    .recurse_untracked_dirs(true)
+                    .include_unmodified(true)
+                    .exclude_submodules(true),
+            )
+        };
+
+        let mut guard = self.write()?;
+        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, shared_frecency)?;
+            count
+        } else {
+            0
+        };
+
+        Ok(statuses_count)
+    }
+}
+
+/// Thread-safe shared handle to the [`FrecencyTracker`] instance.
+#[derive(Clone, Default)]
+pub struct SharedFrecency(pub(crate) Arc<RwLock<Option<FrecencyTracker>>>);
+
+impl std::fmt::Debug for SharedFrecency {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_tuple("SharedFrecency").field(&"..").finish()
+    }
+}
+
+impl SharedFrecency {
+    pub fn read(&self) -> Result<RwLockReadGuard<'_, Option<FrecencyTracker>>, Error> {
+        self.0.read().map_err(|_| Error::AcquireFrecencyLock)
+    }
+
+    pub fn write(&self) -> Result<RwLockWriteGuard<'_, Option<FrecencyTracker>>, Error> {
+        self.0.write().map_err(|_| Error::AcquireFrecencyLock)
+    }
+
+    /// Initialize the frecency tracker, replacing any existing one.
+    pub fn init(&self, tracker: FrecencyTracker) -> Result<(), Error> {
+        let mut guard = self.write()?;
+        *guard = Some(tracker);
+        Ok(())
+    }
+
+    /// Spawn a background GC thread for this frecency tracker.
+    pub fn spawn_gc(
+        &self,
+        db_path: String,
+        use_unsafe_no_lock: bool,
+    ) -> crate::Result<std::thread::JoinHandle<()>> {
+        FrecencyTracker::spawn_gc(self.clone(), db_path, use_unsafe_no_lock)
+    }
+}
+
+/// Thread-safe shared handle to the [`QueryTracker`] instance.
+#[derive(Clone, Default)]
+pub struct SharedQueryTracker(pub(crate) Arc<RwLock<Option<QueryTracker>>>);
+
+impl std::fmt::Debug for SharedQueryTracker {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_tuple("SharedQueryTracker").field(&"..").finish()
+    }
+}
+
+impl SharedQueryTracker {
+    pub fn read(&self) -> Result<RwLockReadGuard<'_, Option<QueryTracker>>, Error> {
+        self.0.read().map_err(|_| Error::AcquireFrecencyLock)
+    }
+
+    pub fn write(&self) -> Result<RwLockWriteGuard<'_, Option<QueryTracker>>, Error> {
+        self.0.write().map_err(|_| Error::AcquireFrecencyLock)
+    }
+
+    /// Initialize the query tracker, replacing any existing one.
+    pub fn init(&self, tracker: QueryTracker) -> Result<(), Error> {
+        let mut guard = self.write()?;
+        *guard = Some(tracker);
+        Ok(())
+    }
+}

crates/fff-core/src/sort_buffer.rs

@@ -82,9 +82,9 @@ mod tests {
 
     #[test]
     fn test_sort_by_key_with_buffer() {
-        let mut data = vec![(2, "b"), (1, "a"), (3, "c")];
-        sort_by_key_with_buffer(&mut data, |item| item.0);
-        assert_eq!(data, vec![(1, "a"), (2, "b"), (3, "c")]);
+        let mut data = vec![(1, 50), (2, 20), (3, 80), (4, 10), (5, 90)];
+        sort_by_key_with_buffer(&mut data, |a| a.1);
+        assert_eq!(data, vec![(4, 10), (2, 20), (1, 50), (3, 80), (5, 90)]);
     }
 
     #[test]

crates/fff-core/src/types.rs

@@ -1,145 +1,368 @@
-use std::path::{Path, PathBuf};
+use std::path::Path;
 use std::sync::OnceLock;
-
-use memmap2::Mmap;
+use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
 
 use crate::constraints::Constrainable;
 use crate::query_tracker::QueryMatchEntry;
 use fff_query_parser::{FFFQuery, FuzzyQuery, Location};
+use neo_frizbee::Matchable;
 
-/// 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.
+///
+/// The `Buffer` variant is also used on Unix for temporary (uncached) reads
+/// where the mmap/munmap syscall overhead exceeds the cost of a heap copy.
+#[derive(Debug)]
+#[allow(dead_code)] // variants are conditionally used per platform
+pub enum FileContent {
+    #[cfg(not(target_os = "windows"))]
+    Mmap(memmap2::Mmap),
+    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,
+            FileContent::Buffer(b) => b,
+        }
+    }
+}
+
+pub struct FileItemFlags;
+
+impl FileItemFlags {
+    pub const BINARY: u8 = 1 << 0;
+    /// Tombstone — file was deleted but index slot is preserved so
+    /// bigram indices for other files stay valid.
+    pub const DELETED: u8 = 1 << 1;
+}
+
+/// A single indexed file with metadata, frecency scores, and lazy content cache.
+/// Occupies ~100 bytes + file path per file
+///
+/// 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`.
 #[derive(Debug)]
 pub struct FileItem {
-    pub path: PathBuf,
-    pub relative_path: String,
-    pub relative_path_lower: String,
-    pub file_name: String,
-    pub file_name_lower: String,
+    /// File size in bytes
     pub size: u64,
+    /// Modification time in UNIX timestamp
     pub modified: u64,
-    pub access_frecency_score: i64,
-    pub modification_frecency_score: i64,
-    pub total_frecency_score: i64,
+    /// Frecency access score
+    pub access_frecency_score: i16,
+    /// Frecency modification score
+    pub modification_frecency_score: i16,
+    /// The file's git status
     pub git_status: Option<git2::Status>,
-    pub is_binary: bool,
-    /// Lazily-initialized memory-mapped file contents for grep.
+
+    /// Absolute path stored as a plain String. We never use path components —
+    /// only slicing, comparison, and passing to fs/DB APIs via `as_path()`.
+    path: String,
+    /// Byte offset where the relative path begins (after base_path + separator).
+    relative_start: u16,
+    /// Byte offset where the filename begins (after last separator).
+    filename_start: u16,
+    /// Packed boolean flags — see `FileItemFlags`.
+    flags: u8,
+    /// 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 {
     fn clone(&self) -> Self {
         Self {
             path: self.path.clone(),
-            relative_path: self.relative_path.clone(),
-            relative_path_lower: self.relative_path_lower.clone(),
-            file_name: self.file_name.clone(),
-            file_name_lower: self.file_name_lower.clone(),
+            relative_start: self.relative_start,
+            filename_start: self.filename_start,
             size: self.size,
             modified: self.modified,
             access_frecency_score: self.access_frecency_score,
             modification_frecency_score: self.modification_frecency_score,
-            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(),
+            flags: self.flags,
+            // Don't clone the content — the clone lazily re-creates it on demand
+            content: OnceLock::new(),
         }
     }
 }
 
-const MAX_MMAP_FILE_SIZE: u64 = 10 * 1024 * 1024;
+/// File content that is either borrowed from the persistent cache or owned
+/// from a temporary mmap. Dereferences to `&[u8]` so callers can use it
+/// transparently.
+///
+/// On Unix the uncached variant holds a temporary `memmap2::Mmap` that is
+/// backed by the kernel page cache — same zero-copy benefit as the cached
+/// path, but the mapping is released (munmap) as soon as this value is
+/// dropped instead of being retained for the lifetime of the `FileItem`.
+pub enum FileContentRef<'a> {
+    /// Content is stored in the `FileItem`'s `OnceLock` cache (fast path).
+    Cached(&'a [u8]),
+    /// Temporary mmap (Unix) / heap buffer (Windows) created because the
+    /// persistent cache budget was exceeded. Unmapped on drop.
+    Temp(FileContent),
+}
+
+impl std::ops::Deref for FileContentRef<'_> {
+    type Target = [u8];
+    fn deref(&self) -> &[u8] {
+        match self {
+            FileContentRef::Cached(s) => s,
+            FileContentRef::Temp(c) => c,
+        }
+    }
+}
 
 impl FileItem {
     /// Create a new `FileItem` with all fields specified and an empty (not yet loaded) mmap.
     pub fn new_raw(
-        path: PathBuf,
-        relative_path: String,
-        file_name: String,
+        path: String,
+        relative_start: u16,
+        filename_start: u16,
         size: u64,
         modified: u64,
         git_status: Option<git2::Status>,
         is_binary: bool,
     ) -> Self {
+        let mut flags = 0u8;
+        if is_binary {
+            flags |= FileItemFlags::BINARY;
+        }
+
         Self {
-            relative_path_lower: relative_path.to_lowercase(),
-            file_name_lower: file_name.to_lowercase(),
             path,
-            relative_path,
-            file_name,
+            relative_start,
+            filename_start,
             size,
             modified,
             access_frecency_score: 0,
             modification_frecency_score: 0,
-            total_frecency_score: 0,
             git_status,
-            is_binary,
-            mmap: OnceLock::new(),
+            flags,
+            content: OnceLock::new(),
         }
     }
 
-    /// Invalidate the cached mmap so the next `get_mmap()` call creates a fresh one.
+    /// The full absolute path as a string slice.
+    #[inline]
+    pub fn path_str(&self) -> &str {
+        &self.path
+    }
+
+    /// The full absolute path as a `&Path` (zero-cost on Unix).
+    #[inline]
+    pub fn as_path(&self) -> &Path {
+        Path::new(&self.path)
+    }
+
+    /// The relative path (from the base directory).
+    #[inline]
+    pub fn relative_path(&self) -> &str {
+        &self.path[self.relative_start as usize..]
+    }
+
+    /// Just the filename component.
+    #[inline]
+    pub fn file_name(&self) -> &str {
+        &self.path[self.filename_start as usize..]
+    }
+
+    /// Byte offset of the filename within the relative path.
+    /// Equivalent to `relative_path().len() - file_name().len()`.
+    #[inline]
+    pub fn filename_offset_in_relative(&self) -> usize {
+        (self.filename_start - self.relative_start) as usize
+    }
+
+    #[inline]
+    pub fn total_frecency_score(&self) -> i32 {
+        self.access_frecency_score as i32 + self.modification_frecency_score as i32
+    }
+
+    #[inline]
+    pub fn is_binary(&self) -> bool {
+        self.flags & FileItemFlags::BINARY != 0
+    }
+
+    #[inline]
+    pub fn set_binary(&mut self, val: bool) {
+        if val {
+            self.flags |= FileItemFlags::BINARY;
+        } else {
+            self.flags &= !FileItemFlags::BINARY;
+        }
+    }
+
+    #[inline]
+    pub fn is_deleted(&self) -> bool {
+        self.flags & FileItemFlags::DELETED != 0
+    }
+
+    #[inline]
+    pub fn set_deleted(&mut self, val: bool) {
+        if val {
+            self.flags |= FileItemFlags::DELETED;
+        } else {
+            self.flags &= !FileItemFlags::DELETED;
+        }
+    }
+}
+
+impl Matchable for FileItem {
+    #[inline]
+    fn match_str(&self) -> Option<&str> {
+        (!self.is_deleted()).then(|| self.relative_path())
+    }
+}
+
+impl Matchable for &FileItem {
+    #[inline]
+    fn match_str(&self) -> Option<&str> {
+        (!self.is_deleted()).then(|| self.relative_path())
+    }
+}
+
+impl FileItem {
+    /// 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.
-    pub fn invalidate_mmap(&mut self) {
-        self.mmap = 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.
-    ///
-    /// 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.
-    #[inline]
-    pub fn get_mmap(&self) -> Option<&Mmap> {
-        if let Some(mmap) = self.mmap.get() {
-            return Some(mmap);
+    /// 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, budget: &ContentCacheBudget) {
+        if self.content.get().is_some() {
+            budget.cached_count.fetch_sub(1, Ordering::Relaxed);
+            budget.cached_bytes.fetch_sub(self.size, Ordering::Relaxed);
         }
 
-        if self.size == 0 || self.size > MAX_MMAP_FILE_SIZE {
+        self.content = OnceLock::new();
+    }
+
+    /// Get the cached file contents or lazily load and cache them.
+    ///
+    /// Returns `None` if the file is too large, empty, can't be opened, **or
+    /// the cache budget is exhausted**. Callers that need content regardless
+    /// of the budget should use [`get_content_for_search`].
+    ///
+    /// After the first call, this is lock-free (just an atomic load + pointer deref).
+    pub fn get_content(&self, budget: &ContentCacheBudget) -> Option<&[u8]> {
+        if let Some(content) = self.content.get() {
+            return Some(content);
+        }
+
+        let max_file_size = budget.max_file_size;
+        if self.size == 0 || self.size > max_file_size {
             return None;
         }
 
-        let file = std::fs::File::open(&self.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()?;
+        // Check cache budget before creating a new persistent cache entry.
+        let count = budget.cached_count.load(Ordering::Relaxed);
+        let bytes = budget.cached_bytes.load(Ordering::Relaxed);
+        let max_files = budget.max_files;
+        let max_bytes = budget.max_bytes;
+        if count >= max_files || bytes + self.size > max_bytes {
+            return None;
+        }
 
-        // 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))
+        let content = load_file_content(self.as_path(), self.size)?;
+        let result = self.content.get_or_init(|| content);
+
+        // Bump counters. Slight over-count under races is fine — the budget
+        // is a soft limit and the overshoot is bounded by rayon thread count.
+        budget.cached_count.fetch_add(1, Ordering::Relaxed);
+        budget.cached_bytes.fetch_add(self.size, Ordering::Relaxed);
+
+        Some(result)
+    }
+
+    /// Get file content for searching — **always returns content** for eligible
+    /// files, even when the persistent cache budget is exhausted.
+    ///
+    /// Tries the `OnceLock` cache first (fast path). If the cache is full,
+    /// falls back to a temporary mmap that is unmapped when the returned
+    /// [`FileContentRef`] is dropped — no persistent kernel resources retained.
+    #[inline]
+    pub fn get_content_for_search<'a>(
+        &'a self,
+        budget: &ContentCacheBudget,
+    ) -> Option<FileContentRef<'a>> {
+        if let Some(cached) = self.get_content(budget) {
+            return Some(FileContentRef::Cached(cached));
+        }
+
+        // get_content returned None — either ineligible or over budget.
+        let max_file_size = budget.max_file_size;
+        if self.is_binary() || self.size == 0 || self.size > max_file_size {
+            return None;
+        }
+
+        // Over budget: create a temporary mmap that is unmapped on drop.
+        let content = load_file_content(self.as_path(), self.size)?;
+        Some(FileContentRef::Temp(content))
+    }
+}
+
+/// Page size on Apple Silicon is 16KB; on x86-64 it's 4KB.
+/// Files smaller than one page waste the remainder when mmapped.
+/// Reading them into a heap buffer avoids this overhead.
+#[cfg(target_arch = "aarch64")]
+const MMAP_THRESHOLD: u64 = 16 * 1024;
+#[cfg(not(target_arch = "aarch64"))]
+const MMAP_THRESHOLD: u64 = 4 * 1024;
+
+/// Load file contents: small files are read into a heap buffer to avoid
+/// mmap page alignment waste; large files use mmap for zero-copy access.
+/// On Windows, always uses heap buffer (mmap holds the file handle open).
+fn load_file_content(path: &Path, size: u64) -> Option<FileContent> {
+    #[cfg(not(target_os = "windows"))]
+    {
+        if size < MMAP_THRESHOLD {
+            let data = std::fs::read(path).ok()?;
+            Some(FileContent::Buffer(data))
+        } else {
+            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 { memmap2::Mmap::map(&file) }.ok()?;
+            Some(FileContent::Mmap(mmap))
+        }
+    }
+
+    #[cfg(target_os = "windows")]
+    {
+        let _ = size;
+        let data = std::fs::read(path).ok()?;
+        Some(FileContent::Buffer(data))
+    }
+}
+
+impl AsRef<Path> for FileItem {
+    #[inline]
+    fn as_ref(&self) -> &Path {
+        Path::new(&self.path)
     }
 }
 
 impl Constrainable for FileItem {
     #[inline]
     fn relative_path(&self) -> &str {
-        &self.relative_path
-    }
-
-    #[inline]
-    fn relative_path_lower(&self) -> &str {
-        &self.relative_path_lower
+        FileItem::relative_path(self)
     }
 
     #[inline]
     fn file_name(&self) -> &str {
-        &self.file_name
+        FileItem::file_name(self)
     }
 
     #[inline]
@@ -148,13 +371,14 @@ 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 +392,42 @@ 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(),
         }
     }
 }
@@ -212,3 +440,91 @@ pub struct SearchResult<'a> {
     pub total_files: usize,
     pub location: Option<Location>,
 }
+
+const MAX_MMAP_FILE_SIZE: u64 = 10 * 1024 * 1024;
+
+// Limits the total number of files (and bytes) whose content is kept in
+// memory via the `OnceLock<FileContent>` cache. On Unix every cached file
+// holds a live `mmap`, which consumes a kernel `vm_map_entry`. On a 500k-file
+// monorepo, caching everything exhausts macOS/Linux kernel resources and
+// crashes the machine (see issue #294).
+//
+// Each `FilePicker` owns its own `ContentCacheBudget`. The budget is passed
+// to `grep_search` and `warmup_mmaps` so that multiple pickers can coexist
+// without interfering with each other's counters.
+
+const MAX_CACHED_CONTENT_BYTES: u64 = 512 * 1024 * 1024;
+
+/// Per-picker budget controlling how many files may have their content
+/// persistently cached (mmap on Unix, heap buffer on Windows).
+#[derive(Debug)]
+pub struct ContentCacheBudget {
+    pub max_files: usize,
+    pub max_bytes: u64,
+    pub max_file_size: u64,
+    pub cached_count: AtomicUsize,
+    pub cached_bytes: AtomicU64,
+}
+
+impl ContentCacheBudget {
+    /// No limits — every eligible file is cached. Useful for tests and
+    /// short-lived tools that don't need resource protection.
+    pub fn unlimited() -> Self {
+        Self {
+            max_files: usize::MAX,
+            max_bytes: u64::MAX,
+            max_file_size: MAX_MMAP_FILE_SIZE,
+            cached_count: AtomicUsize::new(0),
+            cached_bytes: AtomicU64::new(0),
+        }
+    }
+
+    pub fn zero() -> Self {
+        Self {
+            max_files: 0,
+            max_bytes: 0,
+            max_file_size: 0,
+            cached_count: AtomicUsize::new(0),
+            cached_bytes: AtomicU64::new(0),
+        }
+    }
+
+    pub fn new_for_repo(file_count: usize) -> Self {
+        let max_files = if file_count > 50_000 {
+            5_000
+        } else if file_count > 10_000 {
+            10_000
+        } else {
+            30_000 // effectively unlimited for small repos
+        };
+
+        let max_bytes = if file_count > 50_000 {
+            128 * 1024 * 1024 // 128 MB
+        } else if file_count > 10_000 {
+            256 * 1024 * 1024 // 256 MB
+        } else {
+            MAX_CACHED_CONTENT_BYTES // 512 MB
+        };
+
+        Self {
+            max_files,
+            max_bytes,
+            max_file_size: MAX_MMAP_FILE_SIZE,
+            cached_count: AtomicUsize::new(0),
+            cached_bytes: AtomicU64::new(0),
+        }
+    }
+
+    /// Reset the counters. Called when the file index is rebuilt (rescan /
+    /// directory change) and all old `FileItem`s are dropped.
+    pub fn reset(&self) {
+        self.cached_count.store(0, Ordering::Relaxed);
+        self.cached_bytes.store(0, Ordering::Relaxed);
+    }
+}
+
+impl Default for ContentCacheBudget {
+    fn default() -> Self {
+        Self::new_for_repo(30_000)
+    }
+}

crates/fff-core/tests/bigram_overlay_integration.rs

@@ -0,0 +1,334 @@
+//! Integration test: verify that modifying a file after the bigram index is built
+//! still makes the new content findable via grep (through the overlay layer).
+
+use std::fs;
+use std::time::Duration;
+use tempfile::TempDir;
+
+use fff_search::file_picker::{FFFMode, FilePicker};
+use fff_search::grep::{GrepMode, GrepSearchOptions, parse_grep_query};
+use fff_search::{FilePickerOptions, SharedFrecency, SharedPicker};
+
+/// Create a temp directory with some initial files, run the full picker lifecycle,
+/// then modify a file and verify grep finds the new content.
+#[test]
+fn modified_file_findable_via_overlay() {
+    let tmp = TempDir::new().unwrap();
+    let base = tmp.path();
+
+    // Create initial files with known content.
+    fs::write(base.join("alpha.txt"), "hello world\nfoo bar\n").unwrap();
+    fs::write(
+        base.join("beta.txt"),
+        "some other content\nnothing special\n",
+    )
+    .unwrap();
+    fs::write(base.join("gamma.txt"), "yet another file\nmore lines\n").unwrap();
+
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
+
+    FilePicker::new_with_shared_state(
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        FilePickerOptions {
+            base_path: base.to_string_lossy().to_string(),
+            warmup_mmap_cache: true,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
+    )
+    .expect("Failed to create FilePicker");
+
+    // Wait for scan + bigram build to complete.
+    let deadline = std::time::Instant::now() + Duration::from_secs(30);
+    loop {
+        std::thread::sleep(Duration::from_millis(50));
+
+        let ready = shared_picker
+            .read()
+            .ok()
+            .map(|guard| {
+                guard
+                    .as_ref()
+                    .map_or(false, |p| !p.is_scan_active() && p.bigram_index().is_some())
+            })
+            .unwrap_or(false);
+
+        if ready {
+            break;
+        }
+        assert!(
+            std::time::Instant::now() < deadline,
+            "Timed out waiting for scan + bigram build"
+        );
+    }
+
+    // Sanity check: the 3 files are indexed.
+    {
+        let guard = shared_picker.read().unwrap();
+        let picker = guard.as_ref().unwrap();
+        assert_eq!(picker.get_files().len(), 3, "Expected 3 files after scan");
+        assert!(
+            picker.bigram_index().is_some(),
+            "Bigram index should be built"
+        );
+        assert!(
+            picker.bigram_overlay().is_some(),
+            "Overlay should be initialized"
+        );
+    }
+
+    // "UNIQUE_NEEDLE" should NOT exist in any file yet.
+    {
+        let guard = shared_picker.read().unwrap();
+        let picker = guard.as_ref().unwrap();
+        let parsed = parse_grep_query("UNIQUE_NEEDLE");
+        let opts = grep_opts();
+        let result = picker.grep(&parsed, &opts);
+        assert_eq!(
+            result.matches.len(),
+            0,
+            "UNIQUE_NEEDLE should not exist before modification"
+        );
+    }
+
+    // Sleep so the filesystem mtime (seconds granularity) advances past the
+    // value recorded during scan. Without this, on_create_or_modify skips
+    // mmap invalidation and grep reads stale cached content.
+    std::thread::sleep(Duration::from_millis(1100));
+
+    // Write new content containing the needle.
+    let modified_path = base.join("beta.txt");
+    fs::write(
+        &modified_path,
+        "some other content\nUNIQUE_NEEDLE is here\nnothing special\n",
+    )
+    .unwrap();
+
+    // Simulate watcher event: call on_create_or_modify.
+    // This updates the overlay's bigrams and invalidates the mmap cache.
+    {
+        let mut guard = shared_picker.write().unwrap();
+        let picker = guard.as_mut().unwrap();
+        let result = picker.on_create_or_modify(&modified_path);
+        assert!(
+            result.is_some(),
+            "on_create_or_modify should return the file"
+        );
+    }
+
+    // The bigram index was built BEFORE the modification, so without the
+    // overlay, beta.txt would be filtered out (its old bigrams don't contain
+    // "UNIQUE_NEEDLE"). The overlay should fix that.
+    {
+        let guard = shared_picker.read().unwrap();
+        let picker = guard.as_ref().unwrap();
+        let parsed = parse_grep_query("UNIQUE_NEEDLE");
+        let opts = grep_opts();
+        let result = picker.grep(&parsed, &opts);
+        assert!(
+            !result.matches.is_empty(),
+            "UNIQUE_NEEDLE should be findable after modification (overlay adds the candidate back)"
+        );
+        assert_eq!(result.matches.len(), 1);
+        assert!(result.matches[0].line_content.contains("UNIQUE_NEEDLE"));
+    }
+
+    // Prove the overlay is actually doing something: without it, the bigram
+    // index would filter out beta.txt and the search would miss the needle.
+    {
+        let guard = shared_picker.read().unwrap();
+        let picker = guard.as_ref().unwrap();
+        let parsed = parse_grep_query("UNIQUE_NEEDLE");
+        let opts = grep_opts();
+        let result = picker.grep_without_overlay(&parsed, &opts);
+        assert_eq!(
+            result.matches.len(),
+            0,
+            "Without overlay, bigram prefiltering should exclude the modified file"
+        );
+    }
+
+    // Cleanup: stop background watcher.
+    if let Ok(mut guard) = shared_picker.write() {
+        if let Some(ref mut picker) = *guard {
+            picker.stop_background_monitor();
+        }
+    }
+}
+
+/// Verify that deleting a file makes its content un-findable via grep.
+#[test]
+fn deleted_file_excluded_via_overlay() {
+    let tmp = TempDir::new().unwrap();
+    let base = tmp.path();
+
+    fs::write(base.join("keep.txt"), "keep this content\n").unwrap();
+    fs::write(base.join("remove.txt"), "DELETEME_TOKEN is here\n").unwrap();
+
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
+
+    FilePicker::new_with_shared_state(
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        FilePickerOptions {
+            base_path: base.to_string_lossy().to_string(),
+            warmup_mmap_cache: true,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
+    )
+    .unwrap();
+
+    wait_for_bigram(&shared_picker);
+
+    // Sanity: DELETEME_TOKEN is findable.
+    {
+        let guard = shared_picker.read().unwrap();
+        let picker = guard.as_ref().unwrap();
+        let result = grep_for(picker, "DELETEME_TOKEN");
+        assert_eq!(
+            result.matches.len(),
+            1,
+            "Token should be found before delete"
+        );
+    }
+
+    // Delete the file on disk and via picker.
+    let remove_path = base.join("remove.txt");
+    fs::remove_file(&remove_path).unwrap();
+    {
+        let mut guard = shared_picker.write().unwrap();
+        let picker = guard.as_mut().unwrap();
+        assert!(
+            picker.remove_file_by_path(&remove_path),
+            "remove should succeed"
+        );
+    }
+
+    // Token should no longer be found (tombstone in overlay clears the candidate).
+    {
+        let guard = shared_picker.read().unwrap();
+        let picker = guard.as_ref().unwrap();
+        let result = grep_for(picker, "DELETEME_TOKEN");
+        assert_eq!(
+            result.matches.len(),
+            0,
+            "DELETEME_TOKEN should not be found after deletion (tombstone in overlay)"
+        );
+    }
+
+    if let Ok(mut guard) = shared_picker.write() {
+        if let Some(ref mut picker) = *guard {
+            picker.stop_background_monitor();
+        }
+    }
+}
+
+/// Verify that a newly added file (in overflow) is findable via grep.
+#[test]
+fn new_file_findable_after_add() {
+    let tmp = TempDir::new().unwrap();
+    let base = tmp.path();
+
+    fs::write(base.join("existing.txt"), "original content\n").unwrap();
+
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
+
+    FilePicker::new_with_shared_state(
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        FilePickerOptions {
+            base_path: base.to_string_lossy().to_string(),
+            warmup_mmap_cache: true,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
+    )
+    .unwrap();
+
+    wait_for_bigram(&shared_picker);
+
+    // Create a new file on disk after the index was built.
+    let new_path = base.join("newcomer.txt");
+    fs::write(&new_path, "BRAND_NEW_TOKEN lives here\n").unwrap();
+
+    // Simulate watcher detecting the new file.
+    {
+        let mut guard = shared_picker.write().unwrap();
+        let picker = guard.as_mut().unwrap();
+        let result = picker.on_create_or_modify(&new_path);
+        assert!(
+            result.is_some(),
+            "on_create_or_modify should return the new file"
+        );
+    }
+
+    // The new file is in overflow, not in the base files slice.
+    // grep_search currently only searches base files, so we need to verify
+    // the overflow file is accessible.
+    {
+        let guard = shared_picker.read().unwrap();
+        let picker = guard.as_ref().unwrap();
+        let overflow = picker.get_overflow_files();
+        assert_eq!(overflow.len(), 1, "Should have 1 overflow file");
+        assert!(
+            overflow[0].as_path().ends_with("newcomer.txt"),
+            "Overflow file should be newcomer.txt"
+        );
+    }
+
+    if let Ok(mut guard) = shared_picker.write() {
+        if let Some(ref mut picker) = *guard {
+            picker.stop_background_monitor();
+        }
+    }
+}
+
+// ── Helpers ─────────────────────────────────────────────────────────────
+
+fn grep_opts() -> GrepSearchOptions {
+    GrepSearchOptions {
+        max_file_size: 10 * 1024 * 1024,
+        max_matches_per_file: 200,
+        smart_case: true,
+        file_offset: 0,
+        page_limit: 200,
+        mode: GrepMode::PlainText,
+        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
+    }
+}
+
+fn grep_for<'a>(picker: &'a FilePicker, query: &str) -> fff_search::grep::GrepResult<'a> {
+    let parsed = parse_grep_query(query);
+    picker.grep(&parsed, &grep_opts())
+}
+
+fn wait_for_bigram(shared_picker: &SharedPicker) {
+    let deadline = std::time::Instant::now() + Duration::from_secs(30);
+    loop {
+        std::thread::sleep(Duration::from_millis(50));
+        let ready = shared_picker
+            .read()
+            .ok()
+            .map(|guard| {
+                guard
+                    .as_ref()
+                    .map_or(false, |p| !p.is_scan_active() && p.bigram_index().is_some())
+            })
+            .unwrap_or(false);
+        if ready {
+            break;
+        }
+        assert!(
+            std::time::Instant::now() < deadline,
+            "Timed out waiting for bigram build"
+        );
+    }
+}

crates/fff-grep/Cargo.toml

@@ -0,0 +1,11 @@
+[package]
+name = "fff-grep"
+description = "File grepping logic for fff"
+license = "MIT"
+authors = ["Dmitriy Kovalenko <dmtr.kovalenko@outlok.com>"]
+version = "0.5.2"
+edition = "2024"
+
+[dependencies]
+bstr = { version = "1.6.2", default-features = false, features = ["std"] }
+memchr = "2.6.3"

crates/fff-grep/src/lib.rs

@@ -8,10 +8,12 @@ Only `search_slice` is supported -- no file/reader/mmap search.
 #![deny(missing_docs)]
 
 pub use crate::{
+    matcher::{LineTerminator, Match, Matcher, NoError},
     searcher::{Searcher, SearcherBuilder},
     sink::{Sink, SinkError, SinkFinish, SinkMatch},
 };
 
 pub mod lines;
+pub mod matcher;
 mod searcher;
 mod sink;

crates/fff-grep/src/lines.rs

@@ -2,10 +2,9 @@
 A collection of routines for performing operations on lines.
 */
 
-use {
-    bstr::ByteSlice,
-    grep_matcher::{LineTerminator, Match},
-};
+use bstr::ByteSlice;
+
+use crate::matcher::{LineTerminator, Match};
 
 /// An explicit iterator over lines in a particular slice of bytes.
 ///

crates/fff-grep/src/matcher.rs

@@ -0,0 +1,175 @@
+//! Matcher trait inspired by ripgrep's `Matcher` just simpler
+
+/// A byte range representing a match.
+#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
+pub struct Match {
+    start: usize,
+    end: usize,
+}
+
+impl Match {
+    /// Create a new match from start/end byte offsets.
+    #[inline]
+    pub fn new(start: usize, end: usize) -> Match {
+        debug_assert!(start <= end);
+        Match { start, end }
+    }
+
+    /// Create a zero-width match at `offset`.
+    #[inline]
+    pub fn zero(offset: usize) -> Match {
+        Match {
+            start: offset,
+            end: offset,
+        }
+    }
+
+    /// Start byte offset.
+    #[inline]
+    pub fn start(&self) -> usize {
+        self.start
+    }
+
+    /// End byte offset (exclusive).
+    #[inline]
+    pub fn end(&self) -> usize {
+        self.end
+    }
+
+    /// Return a copy with a different end offset.
+    #[inline]
+    pub fn with_end(&self, end: usize) -> Match {
+        debug_assert!(self.start <= end);
+        Match { end, ..*self }
+    }
+
+    /// Shift both offsets forward by `amount`.
+    #[inline]
+    pub fn offset(&self, amount: usize) -> Match {
+        Match {
+            start: self.start + amount,
+            end: self.end + amount,
+        }
+    }
+
+    /// Byte length of the match.
+    #[inline]
+    pub fn len(&self) -> usize {
+        self.end - self.start
+    }
+
+    /// True if this is a zero-width match.
+    #[inline]
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+}
+
+impl std::ops::Index<Match> for [u8] {
+    type Output = [u8];
+
+    #[inline]
+    fn index(&self, index: Match) -> &[u8] {
+        &self[index.start..index.end]
+    }
+}
+
+impl std::ops::IndexMut<Match> for [u8] {
+    #[inline]
+    fn index_mut(&mut self, index: Match) -> &mut [u8] {
+        &mut self[index.start..index.end]
+    }
+}
+
+impl std::ops::Index<Match> for str {
+    type Output = str;
+
+    #[inline]
+    fn index(&self, index: Match) -> &str {
+        &self[index.start..index.end]
+    }
+}
+
+/// A line terminator (always a single byte for fff — no CRLF support needed).
+#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
+pub struct LineTerminator(u8);
+
+impl LineTerminator {
+    /// Create a line terminator from a single byte.
+    #[inline]
+    pub fn byte(byte: u8) -> LineTerminator {
+        LineTerminator(byte)
+    }
+
+    /// Return the terminator byte.
+    #[inline]
+    pub fn as_byte(&self) -> u8 {
+        self.0
+    }
+
+    /// Return the terminator as a single-element byte slice.
+    #[inline]
+    pub fn as_bytes(&self) -> &[u8] {
+        std::slice::from_ref(&self.0)
+    }
+}
+
+impl Default for LineTerminator {
+    #[inline]
+    fn default() -> LineTerminator {
+        LineTerminator(b'\n')
+    }
+}
+
+/// An error type for matchers that never produce errors.
+#[derive(Debug, Eq, PartialEq)]
+pub struct NoError(());
+
+impl std::error::Error for NoError {}
+
+impl std::fmt::Display for NoError {
+    fn fmt(&self, _: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        unreachable!("NoError should never be instantiated")
+    }
+}
+
+/// A matcher finds byte-level matches in a haystack.
+pub trait Matcher {
+    /// The error type (use [`NoError`] for infallible matchers).
+    type Error: std::fmt::Display;
+
+    /// Find the first match at or after `at` in `haystack`.
+    fn find_at(&self, haystack: &[u8], at: usize) -> Result<Option<Match>, Self::Error>;
+
+    /// Find the first match in `haystack`.
+    #[inline]
+    fn find(&self, haystack: &[u8]) -> Result<Option<Match>, Self::Error> {
+        self.find_at(haystack, 0)
+    }
+
+    /// The line terminator this matcher guarantees will never appear in a match.
+    /// Return `None` if the matcher can match across lines.
+    #[inline]
+    fn line_terminator(&self) -> Option<LineTerminator> {
+        None
+    }
+}
+
+impl<M: Matcher> Matcher for &M {
+    type Error = M::Error;
+
+    #[inline]
+    fn find_at(&self, haystack: &[u8], at: usize) -> Result<Option<Match>, Self::Error> {
+        (*self).find_at(haystack, at)
+    }
+
+    #[inline]
+    fn find(&self, haystack: &[u8]) -> Result<Option<Match>, Self::Error> {
+        (*self).find(haystack)
+    }
+
+    #[inline]
+    fn line_terminator(&self) -> Option<LineTerminator> {
+        (*self).line_terminator()
+    }
+}

crates/fff-grep/src/searcher/core.rs

@@ -0,0 +1,139 @@
+use crate::{
+    lines,
+    matcher::Matcher,
+    searcher::{Config, Range, Searcher},
+    sink::{Sink, SinkError, SinkFinish, SinkMatch},
+};
+
+#[derive(Debug)]
+pub(crate) struct Core<'s, M: 's, S> {
+    config: &'s Config,
+    matcher: M,
+    searcher: &'s Searcher,
+    sink: S,
+    pos: usize,
+    absolute_byte_offset: u64,
+    line_number: Option<u64>,
+    last_line_counted: usize,
+    last_line_visited: usize,
+}
+
+impl<'s, M: Matcher, S: Sink> Core<'s, M, S> {
+    pub(crate) fn new(searcher: &'s Searcher, matcher: M, sink: S) -> Core<'s, M, S> {
+        let line_number = if searcher.config.line_number {
+            Some(1)
+        } else {
+            None
+        };
+        Core {
+            config: &searcher.config,
+            matcher,
+            searcher,
+            sink,
+            pos: 0,
+            absolute_byte_offset: 0,
+            line_number,
+            last_line_counted: 0,
+            last_line_visited: 0,
+        }
+    }
+
+    pub(crate) fn pos(&self) -> usize {
+        self.pos
+    }
+
+    pub(crate) fn set_pos(&mut self, pos: usize) {
+        self.pos = pos;
+    }
+
+    pub(crate) fn matched(&mut self, buf: &[u8], range: &Range) -> Result<bool, S::Error> {
+        self.sink_matched(buf, range)
+    }
+
+    pub(crate) fn find(&mut self, slice: &[u8]) -> Result<Option<Range>, S::Error> {
+        self.matcher.find(slice).map_err(S::Error::error_message)
+    }
+
+    pub(crate) fn begin(&mut self) -> Result<bool, S::Error> {
+        self.sink.begin(self.searcher)
+    }
+
+    pub(crate) fn finish(&mut self, byte_count: u64) -> Result<(), S::Error> {
+        self.sink.finish(self.searcher, &SinkFinish { byte_count })
+    }
+
+    pub(crate) fn match_by_line(&mut self, buf: &[u8]) -> Result<bool, S::Error> {
+        while !buf[self.pos()..].is_empty() {
+            if let Some(line) = self.find_by_line(buf)? {
+                self.set_pos(line.end());
+                if !self.sink_matched(buf, &line)? {
+                    return Ok(false);
+                }
+            } else {
+                break;
+            }
+        }
+        self.set_pos(buf.len());
+        Ok(true)
+    }
+
+    #[inline(always)]
+    fn find_by_line(&mut self, buf: &[u8]) -> Result<Option<Range>, S::Error> {
+        let mut pos = self.pos();
+        while !buf[pos..].is_empty() {
+            let mat = match self
+                .matcher
+                .find(&buf[pos..])
+                .map_err(S::Error::error_message)?
+            {
+                None => return Ok(None),
+                Some(m) => m,
+            };
+            let line = lines::locate(
+                buf,
+                self.config.line_term.as_byte(),
+                Range::zero(mat.start()).offset(pos),
+            );
+            if line.start() == buf.len() {
+                pos = buf.len();
+                continue;
+            }
+            return Ok(Some(line));
+        }
+        Ok(None)
+    }
+
+    #[inline(always)]
+    fn sink_matched(&mut self, buf: &[u8], range: &Range) -> Result<bool, S::Error> {
+        self.count_lines(buf, range.start());
+        let offset = self.absolute_byte_offset + range.start() as u64;
+        let linebuf = &buf[*range];
+        let keepgoing = self.sink.matched(
+            self.searcher,
+            &SinkMatch {
+                bytes: linebuf,
+                absolute_byte_offset: offset,
+                line_number: self.line_number,
+                buffer: buf,
+                bytes_range_in_buffer: range.start()..range.end(),
+            },
+        )?;
+        if !keepgoing {
+            return Ok(false);
+        }
+        self.last_line_visited = range.end();
+        Ok(true)
+    }
+
+    fn count_lines(&mut self, buf: &[u8], upto: usize) {
+        if let Some(ref mut line_number) = self.line_number {
+            if self.last_line_counted >= upto {
+                return;
+            }
+            let slice = &buf[self.last_line_counted..upto];
+            let count = lines::count(slice, self.config.line_term.as_byte());
+            *line_number += count;
+            self.last_line_counted = upto;
+        }
+    }
+}

crates/fff-grep/src/searcher/glue.rs

@@ -1,7 +1,6 @@
-use grep_matcher::Matcher;
-
 use crate::{
     lines,
+    matcher::Matcher,
     searcher::{Config, Range, Searcher, core::Core},
     sink::Sink,
 };

crates/fff-grep/src/searcher/mod.rs

@@ -1,6 +1,5 @@
-use grep_matcher::{LineTerminator, Match, Matcher};
-
 use crate::{
+    matcher::{LineTerminator, Match, Matcher},
     searcher::glue::{MultiLine, SliceByLine},
     sink::{Sink, SinkError},
 };
@@ -190,11 +189,6 @@ impl Searcher {
         {
             return false;
         }
-        if let Some(non_matching) = matcher.non_matching_bytes()
-            && non_matching.contains(self.line_terminator().as_byte())
-        {
-            return false;
-        }
         true
     }
 }

crates/fff-mcp/Cargo.toml

@@ -0,0 +1,27 @@
+[package]
+name = "fff-mcp"
+version = "0.5.2"
+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 = "0.5.1" }
+fff-query-parser = { path = "../fff-query-parser", default-features = false , version = "0.5.2" }
+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", "env"] }

crates/fff-mcp/build.rs

@@ -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/");
+}

crates/fff-mcp/src/cursor.rs

@@ -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()
+    }
+}

crates/fff-mcp/src/healthcheck.rs

@@ -0,0 +1,117 @@
+use crate::Args;
+use git2::Repository;
+
+fn check(label: &str, ok: bool, detail: &str) -> bool {
+    let marker = if ok { "+" } else { "x" };
+    println!("  [{marker}] {label}: {detail}");
+    ok
+}
+
+fn warn(label: &str, detail: &str) {
+    println!("  [!] {label}: {detail}");
+}
+
+pub fn run_healthcheck(args: &Args) -> Result<(), Box<dyn std::error::Error>> {
+    let version = concat!(env!("CARGO_PKG_VERSION"), " (", env!("FFF_GIT_HASH"), ")");
+    println!("fff-mcp {version}\n");
+
+    let mut all_ok = true;
+
+    // 1. Base path
+    let base_path = args.base_path.clone().unwrap_or_else(|| {
+        std::env::current_dir()
+            .unwrap_or_default()
+            .to_string_lossy()
+            .to_string()
+    });
+
+    let path_exists = std::path::Path::new(&base_path).is_dir();
+    all_ok &= check(
+        "Base path",
+        path_exists,
+        if path_exists {
+            &base_path
+        } else {
+            "directory does not exist"
+        },
+    );
+
+    // 2. Git repository
+    match Repository::discover(&base_path) {
+        Ok(repo) => {
+            if let Some(workdir) = repo.workdir() {
+                all_ok &= check("Git repository", true, &format!("{}", workdir.display()));
+            } else {
+                all_ok &= check("Git repository", true, "bare repository");
+            }
+        }
+        Err(_) => {
+            // Not fatal — fff-mcp works without git, but worth flagging.
+            warn(
+                "Git repository",
+                "not found (fff-mcp will still work, but git-status features are disabled)",
+            );
+        }
+    }
+
+    // 3. Frecency database
+    if let Some(ref db_path) = args.frecency_db_path {
+        let parent_ok = std::path::Path::new(db_path)
+            .parent()
+            .is_some_and(|p| p.is_dir());
+        all_ok &= check(
+            "Frecency DB",
+            parent_ok,
+            if parent_ok {
+                db_path
+            } else {
+                "parent directory does not exist"
+            },
+        );
+    } else {
+        check("Frecency DB", false, "path not resolved");
+    }
+
+    // 4. Query history database
+    if let Some(ref db_path) = args.history_db_path {
+        let parent_ok = std::path::Path::new(db_path)
+            .parent()
+            .is_some_and(|p| p.is_dir());
+        all_ok &= check(
+            "History DB",
+            parent_ok,
+            if parent_ok {
+                db_path
+            } else {
+                "parent directory does not exist"
+            },
+        );
+    } else {
+        check("History DB", false, "path not resolved");
+    }
+
+    // 5. Log file
+    if let Some(ref log_path) = args.log_file {
+        let parent_ok = std::path::Path::new(log_path)
+            .parent()
+            .is_some_and(|p| p.is_dir());
+        all_ok &= check(
+            "Log file",
+            parent_ok,
+            if parent_ok {
+                log_path
+            } else {
+                "parent directory does not exist"
+            },
+        );
+    } else {
+        check("Log file", false, "path not resolved");
+    }
+
+    if all_ok {
+        println!("All checks passed.");
+        Ok(())
+    } else {
+        Err("Some checks failed — review the items marked [x] above.".into())
+    }
+}

crates/fff-mcp/src/main.rs

@@ -0,0 +1,334 @@
+//! 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 healthcheck;
+mod output;
+mod server;
+mod update_check;
+
+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"), ")"))]
+pub(crate) 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,
+
+    /// Disable eager mmap warmup after the initial scan. Grep results will
+    /// still work (files are mmap'd lazily on first access), but the first
+    /// search may be slightly slower. Useful on very large repos where the
+    /// warmup would consume too many kernel resources.
+    #[arg(long = "no-warmup")]
+    no_warmup: bool,
+
+    /// Maximum number of files whose content is kept persistently in memory.
+    /// Files beyond this limit are still searchable via temporary mmaps that
+    /// are released after each grep. Defaults to 30 000.
+    /// Also settable via the FFF_MAX_CACHED_FILES environment variable.
+    #[arg(long = "max-cached-files", env = "FFF_MAX_CACHED_FILES")]
+    max_cached_files: Option<usize>,
+
+    /// Run a health check and print diagnostic information, then exit.
+    #[arg(long = "healthcheck")]
+    pub(crate) healthcheck: 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);
+
+    if args.healthcheck {
+        return healthcheck::run_healthcheck(&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()
+    });
+
+    let base_path = match Repository::discover(&base_path) {
+        Ok(repo) => {
+            if let Some(workdir) = repo.workdir() {
+                let git_root = workdir.to_string_lossy().to_string();
+                tracing::info!("Discovered git root: {}", git_root);
+                git_root
+            } else {
+                tracing::info!("Git repository is bare, using base path: {}", base_path);
+                base_path
+            }
+        }
+        Err(_) => {
+            tracing::info!(
+                "No git repository found, indexing from base path: {}",
+                base_path
+            );
+            base_path
+        }
+    };
+
+    let frecency_db_path = args.frecency_db_path.unwrap_or_default();
+
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
+    match FrecencyTracker::new(&frecency_db_path, false) {
+        Ok(tracker) => {
+            let _ = shared_frecency.init(tracker);
+            let _ = shared_frecency.spawn_gc(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(
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        fff::FilePickerOptions {
+            base_path,
+            warmup_mmap_cache: !args.no_warmup,
+            mode: FFFMode::Ai,
+            cache_budget: args
+                .max_cached_files
+                .map(fff::ContentCacheBudget::new_for_repo),
+            ..Default::default()
+        },
+    )
+    .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 = shared_picker.clone();
+    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(())
+}

crates/fff-mcp/src/output.rs

@@ -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: i32) -> 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: i32) -> 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() != 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())
+            {
+                let expand_limit = if def_expanded_files.is_empty() {
+                    MAX_DEF_EXPAND_FIRST
+                } else {
+                    MAX_DEF_EXPAND
+                };
+                def_expanded_files.insert(file.relative_path());
+                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();
+        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
+    }
+}

crates/fff-mcp/src/server.rs

@@ -0,0 +1,684 @@
+//! 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 crate::cursor::CursorStore;
+use crate::output::{GrepFormatter, OutputMode, file_suffix};
+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};
+
+/// 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")]
+    // this has to be float because llms are stupid
+    pub max_results: Option<f64>,
+    /// 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<f64>, // this has to be float because llms are stupid
+    /// 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<f64>,
+    /// 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<f64>,
+}
+
+#[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 parser = QueryParser::new(AiGrepConfig);
+        let parsed = parser.parse(query);
+        let result = picker.grep(&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 = picker.grep(&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 = picker.grep(&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() != 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(picker.get_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(20.0).round() as usize; // safe
+        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(20.0) as usize;
+        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(20.0).round() as usize;
+        let context = params.context.map(|v| v.round() as usize);
+        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, 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 budget = picker.cache_budget();
+        let result =
+            grep::multi_grep_search(files, &patterns_refs, constraints, &options, budget, None);
+        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, 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, budget, None, None, None);
+
+                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)
+    }
+}

crates/fff-mcp/src/update_check.rs

@@ -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)
+}

crates/fff-nvim/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "fff-nvim"
-version = "0.1.0"
+version = "0.5.2"
 edition = "2024"
 
 [lib]
@@ -9,7 +9,7 @@ crate-type = ["cdylib", "rlib"]
 
 [features]
 default = []
-zlob = ["fff-core/zlob"]
+zlob = ["fff/zlob"]
 
 [[bin]]
 name = "test_watcher"
@@ -35,6 +35,11 @@ path = "src/bin/grep_profiler.rs"
 name = "grep_vs_rg"
 path = "src/bin/grep_vs_rg.rs"
 
+[[bin]]
+name = "bench_grep_query"
+ppath = "src/bin/bench_grep_query.rs"
+path = "src/bin/bench_grep_query.rs"
+
 [dependencies]
 # Workspace dependencies
 ahash = { workspace = true }
@@ -44,8 +49,8 @@ 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 = "0.5.1", features = ["mimalloc-collect"] }
+fff-query-parser = { path = "../fff-query-parser" , version = "0.5.2" }
 
 # External dependencies
 blake3 = "1.8.2"
@@ -65,8 +70,6 @@ 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"] }
 
 [dev-dependencies]
 criterion = { version = "0.5", features = ["html_reports"] }
@@ -80,4 +83,3 @@ harness = false
 [[bench]]
 name = "query_tracker_bench"
 harness = false
-

crates/fff-nvim/benches/indexing_and_search.rs

@@ -1,9 +1,11 @@
 use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
-use fff_core::file_picker::FilePicker;
-use fff_core::types::{FileItem, PaginationArgs};
-use fff_core::{FuzzySearchOptions, SharedFrecency, SharedPicker};
+use fff::file_picker::{FFFMode, FilePicker};
+use fff::types::{ContentCacheBudget, FileItem, PaginationArgs};
+use fff::{
+    FilePickerOptions, FuzzySearchOptions, GrepMode, GrepSearchOptions, QueryParser,
+    SharedFrecency, SharedPicker, build_bigram_index, grep,
+};
 use std::path::PathBuf;
-use std::sync::{Arc, RwLock};
 use std::time::Duration;
 
 /// Initialize tracing to output to console
@@ -27,10 +29,14 @@ fn init_file_picker_internal(
     shared_frecency: &SharedFrecency,
 ) -> Result<(), String> {
     FilePicker::new_with_shared_state(
-        path.to_string(),
-        false,
-        Arc::clone(shared_picker),
-        Arc::clone(shared_frecency),
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        FilePickerOptions {
+            base_path: path.to_string(),
+            warmup_mmap_cache: false,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
     )
     .map_err(|e| format!("Failed to create FilePicker: {:?}", e))
 }
@@ -128,12 +134,12 @@ fn setup_once() -> Result<(Vec<FileItem>, SharedPicker, SharedFrecency), 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 shared_picker: SharedPicker = Arc::new(RwLock::new(None));
-    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
 
     init_file_picker_internal(
         &canonical_path.to_string_lossy(),
@@ -164,7 +170,7 @@ 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);
@@ -178,8 +184,8 @@ fn bench_indexing(c: &mut Criterion) {
 
     group.bench_function("index_big_repo", |b| {
         b.iter(|| {
-            let sp: SharedPicker = Arc::new(RwLock::new(None));
-            let sf: SharedFrecency = Arc::new(RwLock::new(None));
+            let sp = SharedPicker::default();
+            let sf = SharedFrecency::default();
 
             let start = std::time::Instant::now();
             init_file_picker_internal(black_box(&canonical_path.to_string_lossy()), &sp, &sf)
@@ -225,19 +231,21 @@ 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),
                     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 {
@@ -268,6 +276,8 @@ 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 {
@@ -278,14 +288,13 @@ fn bench_search_thread_scaling(c: &mut Criterion) {
                 b.iter(|| {
                     let results = FilePicker::fuzzy_search(
                         black_box(&files),
-                        black_box(query),
+                        black_box(&parsed),
                         None,
                         FuzzySearchOptions {
                             max_threads: threads,
                             current_file: None,
-
                             project_path: None,
-                            last_same_query_match: None,
+
                             combo_boost_score_multiplier: 100,
                             min_combo_count: 3,
                             pagination: PaginationArgs {
@@ -317,6 +326,8 @@ 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 {
@@ -324,14 +335,13 @@ fn bench_search_result_limits(c: &mut Criterion) {
             b.iter(|| {
                 let results = FilePicker::fuzzy_search(
                     black_box(&files),
-                    black_box(query),
+                    black_box(&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 {
@@ -350,8 +360,8 @@ 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;
@@ -370,6 +380,8 @@ 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 {
@@ -382,14 +394,13 @@ fn bench_search_scalability(c: &mut Criterion) {
             b.iter(|| {
                 let results = FilePicker::fuzzy_search(
                     black_box(subset),
-                    black_box(query),
+                    black_box(&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 {
@@ -419,21 +430,22 @@ 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),
                 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 {
@@ -451,14 +463,13 @@ fn bench_search_ordering(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box(query),
+                black_box(&parsed_controller),
                 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 {
@@ -476,14 +487,13 @@ fn bench_search_ordering(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box("mod"),
+                black_box(&parsed_mod),
                 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 {
@@ -500,14 +510,13 @@ fn bench_search_ordering(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box("mod"),
+                black_box(&parsed_mod),
                 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 {
@@ -525,14 +534,13 @@ fn bench_search_ordering(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box("controller"),
+                black_box(&parsed_controller),
                 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 {
@@ -549,14 +557,13 @@ fn bench_search_ordering(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box("controller"),
+                black_box(&parsed_controller),
                 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 {
@@ -586,6 +593,8 @@ 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)
@@ -593,14 +602,13 @@ fn bench_pagination_performance(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box(query),
+                black_box(&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 {
@@ -618,14 +626,13 @@ fn bench_pagination_performance(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box(query),
+                black_box(&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 {
@@ -643,14 +650,13 @@ fn bench_pagination_performance(c: &mut Criterion) {
         b.iter(|| {
             let results = FilePicker::fuzzy_search(
                 black_box(&files),
-                black_box(query),
+                black_box(&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 {
@@ -666,6 +672,90 @@ fn bench_pagination_performance(c: &mut Criterion) {
     group.finish();
 }
 
+/// Benchmark grep search with bigram index prefiltering
+fn bench_grep_search(c: &mut Criterion) {
+    let (files, _sp, _sf) = match setup_once() {
+        Ok(result) => result,
+        Err(e) => {
+            eprintln!("Skipping grep benchmarks: {}", e);
+            return;
+        }
+    };
+
+    let budget = ContentCacheBudget::new_for_repo(files.len());
+
+    eprintln!("  Building bigram index for {} files...", files.len());
+    let start = std::time::Instant::now();
+    let (bigram_filter, _overflow_indices) = build_bigram_index(&files, &budget);
+    eprintln!(
+        "  Bigram index built in {:.2}s ({} columns)",
+        start.elapsed().as_secs_f64(),
+        bigram_filter.columns_used(),
+    );
+
+    let mut group = c.benchmark_group("grep");
+    group.sample_size(50);
+
+    let options = GrepSearchOptions {
+        max_file_size: 10 * 1024 * 1024,
+        max_matches_per_file: 0,
+        smart_case: true,
+        file_offset: 0,
+        page_limit: 100,
+        mode: GrepMode::PlainText,
+        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
+    };
+
+    let test_queries = vec![
+        ("common", "struct"),
+        ("specific", "DEFINE_MUTEX"),
+        ("path_filter", "*.h mutex"),
+    ];
+
+    let grep_parser = fff::QueryParser::new(fff::GrepConfig);
+
+    for (name, query) in &test_queries {
+        let parsed = grep_parser.parse(query);
+
+        // With bigram index
+        group.bench_with_input(BenchmarkId::new("with_bigram", name), query, |b, _| {
+            b.iter(|| {
+                let result = grep::grep_search(
+                    black_box(&files),
+                    black_box(&parsed),
+                    black_box(&options),
+                    &budget,
+                    Some(&bigram_filter),
+                    None,
+                    None,
+                );
+                result.matches.len()
+            });
+        });
+
+        // Without bigram index
+        group.bench_with_input(BenchmarkId::new("without_bigram", name), query, |b, _| {
+            b.iter(|| {
+                let result = grep::grep_search(
+                    black_box(&files),
+                    black_box(&parsed),
+                    black_box(&options),
+                    &budget,
+                    None,
+                    None,
+                    None,
+                );
+                result.matches.len()
+            });
+        });
+    }
+
+    group.finish();
+}
+
 criterion_group!(
     benches,
     bench_indexing,
@@ -675,6 +765,7 @@ criterion_group!(
     bench_search_scalability,
     bench_search_ordering,
     bench_pagination_performance,
+    bench_grep_search,
 );
 
 criterion_main!(benches);

crates/fff-nvim/benches/query_tracker_bench.rs

@@ -1,5 +1,5 @@
 use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
-use fff_core::query_tracker::QueryTracker;
+use fff::query_tracker::QueryTracker;
 use rand::distributions::Alphanumeric;
 use rand::prelude::*;
 use std::path::PathBuf;

crates/fff-nvim/src/bin/bench_ci_memmem.rs

@@ -0,0 +1,191 @@
+/// Benchmark: AVX2 vs scalar case-insensitive memmem prefilter.
+///
+/// Loads all non-binary file contents from a repo, then times both
+/// implementations scanning every file for the query.
+///
+/// Usage:
+///   cargo build --release --bin bench_ci_memmem
+///   ./target/release/bench_ci_memmem --path ./big-repo --query "nomore" --iters 5
+use fff::case_insensitive_memmem;
+use std::io::Read;
+use std::path::Path;
+use std::time::Instant;
+
+fn fmt_dur(us: u128) -> String {
+    if us > 1_000_000 {
+        format!("{:.2}s", us as f64 / 1_000_000.0)
+    } else if us > 1000 {
+        format!("{:.2}ms", us as f64 / 1000.0)
+    } else {
+        format!("{}µs", us)
+    }
+}
+
+fn stats(times_us: &mut [u128]) -> (u128, u128, u128, u128) {
+    times_us.sort();
+    let sum: u128 = times_us.iter().sum();
+    let mean = sum / times_us.len() as u128;
+    let median = times_us[times_us.len() / 2];
+    (mean, median, times_us[0], times_us[times_us.len() - 1])
+}
+
+fn detect_binary(path: &Path, size: u64) -> bool {
+    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)
+}
+
+fn load_file_contents(base_path: &Path) -> Vec<Vec<u8>> {
+    use ignore::WalkBuilder;
+
+    let mut contents = Vec::new();
+    let max_size = 10 * 1024 * 1024u64;
+
+    WalkBuilder::new(base_path)
+        .hidden(false)
+        .git_ignore(true)
+        .git_exclude(true)
+        .git_global(true)
+        .ignore(true)
+        .follow_links(false)
+        .build()
+        .filter_map(|e| e.ok())
+        .filter(|e| e.file_type().is_some_and(|ft| ft.is_file()))
+        .for_each(|entry| {
+            let path = entry.path();
+            let size = entry.metadata().ok().map_or(0, |m| m.len());
+            if size == 0 || size > max_size || detect_binary(path, size) {
+                return;
+            }
+            if let Ok(data) = std::fs::read(path) {
+                contents.push(data);
+            }
+        });
+
+    contents
+}
+
+fn bench_impl(
+    label: &str,
+    contents: &[Vec<u8>],
+    needle_lower: &[u8],
+    total_bytes: u64,
+    iters: usize,
+    search_fn: fn(&[u8], &[u8]) -> bool,
+) {
+    eprintln!("\n  [{}]", label);
+    let mut times = Vec::with_capacity(iters);
+    let mut hit_count = 0u32;
+
+    for i in 0..iters {
+        let t = Instant::now();
+        let mut hits = 0u32;
+        for content in contents {
+            if search_fn(content, needle_lower) {
+                hits += 1;
+            }
+        }
+        let us = t.elapsed().as_micros();
+        times.push(us);
+        hit_count = hits;
+        let tp = total_bytes as f64 / (us as f64 / 1_000_000.0) / (1024.0 * 1024.0 * 1024.0);
+        eprintln!(
+            "    iter {}: {}  ({} hits, {:.2} GB/s)",
+            i + 1,
+            fmt_dur(us),
+            hits,
+            tp
+        );
+    }
+
+    let (mean, median, min, max) = stats(&mut times);
+    let med_tp = total_bytes as f64 / (median as f64 / 1_000_000.0) / (1024.0 * 1024.0 * 1024.0);
+    eprintln!(
+        "    mean: {}  median: {} ({:.2} GB/s)  min: {}  max: {}  hits: {}",
+        fmt_dur(mean),
+        fmt_dur(median),
+        med_tp,
+        fmt_dur(min),
+        fmt_dur(max),
+        hit_count
+    );
+}
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+
+    let path = args
+        .iter()
+        .position(|a| a == "--path")
+        .and_then(|i| args.get(i + 1))
+        .map(|s| s.as_str())
+        .unwrap_or(".");
+
+    let query = args
+        .iter()
+        .position(|a| a == "--query")
+        .and_then(|i| args.get(i + 1))
+        .map(|s| s.as_str())
+        .unwrap_or("TODO");
+
+    let iters: usize = args
+        .iter()
+        .position(|a| a == "--iters")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(5);
+
+    let repo = std::path::PathBuf::from(path);
+    if !repo.exists() {
+        eprintln!("Path not found: {}", path);
+        eprintln!("Usage: bench_ci_memmem --path <dir> --query <text> [--iters N]");
+        std::process::exit(1);
+    }
+
+    let canonical = fff::path_utils::canonicalize(&repo).expect("Failed to canonicalize path");
+    let needle_lower: Vec<u8> = query.bytes().map(|b| b.to_ascii_lowercase()).collect();
+
+    eprintln!("=== bench_ci_memmem: AVX2 vs Scalar ===");
+    eprintln!("Path:   {}", canonical.display());
+    eprintln!("Query:  \"{}\"", query);
+    eprintln!("Needle: {:?}", std::str::from_utf8(&needle_lower).unwrap());
+    eprintln!("Iters:  {}", iters);
+
+    eprint!("\n[1/2] Loading files into memory... ");
+    let t = Instant::now();
+    let contents = load_file_contents(&canonical);
+    let total_bytes: u64 = contents.iter().map(|c| c.len() as u64).sum();
+    eprintln!(
+        "{} files, {:.1} MB in {:.2}s",
+        contents.len(),
+        total_bytes as f64 / (1024.0 * 1024.0),
+        t.elapsed().as_secs_f64()
+    );
+
+    eprintln!("\n[2/2] Benchmarking memmem prefilter (scanning ALL files)");
+
+    bench_impl(
+        "Packed pair: (AVX2 two-byte scan)",
+        &contents,
+        &needle_lower,
+        total_bytes,
+        iters,
+        case_insensitive_memmem::search_packed_pair,
+    );
+
+    bench_impl(
+        "scalar: memchr2 first-byte + AVX2 verify",
+        &contents,
+        &needle_lower,
+        total_bytes,
+        iters,
+        case_insensitive_memmem::search,
+    );
+}

crates/fff-nvim/src/bin/bench_grep_query.rs

@@ -0,0 +1,163 @@
+/// Single-query grep benchmark with bigram index profiling.
+///
+/// Usage:
+///   cargo build --release --bin bench_grep_query
+///   ./target/release/bench_grep_query --path ~/dev/chromium --query "MAX_FILE_SIZE" --iters 3
+///   ./target/release/bench_grep_query --path ~/dev/chromium --query "TODO" --no-bigram
+use fff::grep::{GrepMode, GrepSearchOptions, grep_search, parse_grep_query};
+use fff::types::ContentCacheBudget;
+use std::time::Instant;
+
+fn fmt_dur(us: u128) -> String {
+    if us > 1_000_000 {
+        format!("{:.2}s", us as f64 / 1_000_000.0)
+    } else if us > 1000 {
+        format!("{:.2}ms", us as f64 / 1000.0)
+    } else {
+        format!("{}µs", us)
+    }
+}
+
+fn run_grep(files: &[fff::FileItem], index: Option<&fff::BigramFilter>, query: &str, iters: usize) {
+    let options = GrepSearchOptions {
+        max_file_size: 10 * 1024 * 1024,
+        max_matches_per_file: 200,
+        smart_case: true,
+        file_offset: 0,
+        page_limit: usize::MAX,
+        mode: GrepMode::PlainText,
+        time_budget_ms: 0,
+        before_context: 0,
+        after_context: 0,
+        classify_definitions: false,
+    };
+
+    let parsed = parse_grep_query(query);
+    let budget = ContentCacheBudget::default();
+    let mut times_us = Vec::with_capacity(iters);
+
+    for i in 0..iters {
+        let t = Instant::now();
+        let result = grep_search(files, &parsed, &options, &budget, index, None, None);
+        let us = t.elapsed().as_micros();
+        times_us.push(us);
+
+        eprintln!(
+            "  iter {}: {} ({} matches in {} files, {}/{} searched)",
+            i + 1,
+            fmt_dur(us),
+            result.matches.len(),
+            result.files_with_matches,
+            result.total_files_searched,
+            result.total_files,
+        );
+    }
+
+    if times_us.len() > 1 {
+        times_us.sort();
+        let sum: u128 = times_us.iter().sum();
+        let mean = sum / times_us.len() as u128;
+        let median = times_us[times_us.len() / 2];
+        let min = times_us[0];
+        let max = times_us[times_us.len() - 1];
+        eprintln!(
+            "  mean: {}  median: {}  min: {}  max: {}",
+            fmt_dur(mean),
+            fmt_dur(median),
+            fmt_dur(min),
+            fmt_dur(max)
+        );
+    }
+}
+
+fn build_bigram(files: &mut [fff::FileItem]) -> fff::BigramFilter {
+    let budget = ContentCacheBudget::default();
+    let (index, binary_indices) = fff::build_bigram_index(files, &budget);
+
+    for &i in &binary_indices {
+        files[i].set_binary(true);
+    }
+
+    index
+}
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+
+    let path = args
+        .iter()
+        .position(|a| a == "--path")
+        .and_then(|i| args.get(i + 1))
+        .map(|s| s.as_str())
+        .unwrap_or(".");
+
+    let query = args
+        .iter()
+        .position(|a| a == "--query")
+        .and_then(|i| args.get(i + 1))
+        .map(|s| s.as_str())
+        .unwrap_or("TODO");
+
+    let iters: usize = args
+        .iter()
+        .position(|a| a == "--iters")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(5);
+
+    let no_bigram = args.iter().any(|a| a == "--no-bigram");
+
+    let repo = std::path::PathBuf::from(path);
+    if !repo.exists() {
+        eprintln!("Path not found: {}", path);
+        eprintln!("Usage: bench_grep_query --path <dir> --query <text> [--iters N] [--no-bigram]");
+        std::process::exit(1);
+    }
+
+    let canonical = fff::path_utils::canonicalize(&repo).expect("Failed to canonicalize path");
+    eprintln!("=== bench_grep_query ===");
+    eprintln!("Path:  {}", canonical.display());
+    eprintln!("Query: \"{}\"", query);
+    eprintln!("Iters: {}", iters);
+    eprintln!();
+
+    // ── 1. Scan files ──────────────────────────────────────────────────
+    eprint!("[1/3] Scanning files... ");
+    let t = Instant::now();
+    let mut files = fff::scan_files(&canonical);
+    let non_binary = files.iter().filter(|f| !f.is_binary()).count();
+    eprintln!(
+        "{} files in {:.2}s ({} non-binary)",
+        files.len(),
+        t.elapsed().as_secs_f64(),
+        non_binary,
+    );
+
+    if no_bigram {
+        eprintln!("[2/3] Bigram index skipped (--no-bigram)");
+        eprintln!(
+            "\n[3/3] Running grep \"{}\" x {} iterations\n",
+            query, iters
+        );
+        run_grep(&files, None, query, iters);
+        return;
+    }
+
+    // ── 2. Build bigram index ──────────────────────────────────────────
+    eprint!("[2/3] Bigram index... ");
+    let t = Instant::now();
+    let index = build_bigram(&mut files);
+    eprintln!(
+        "done in {:.2}s  ({} cols, {:.1} MB)",
+        t.elapsed().as_secs_f64(),
+        index.columns_used(),
+        index.heap_bytes() as f64 / (1024.0 * 1024.0),
+    );
+
+    // ── 3. Grep ───────────────────────────────────────────────────────
+    eprintln!(
+        "\n[3/3] Running grep \"{}\" x {} iterations\n",
+        query, iters
+    );
+    run_grep(&files, Some(&index), query, iters);
+}

crates/fff-nvim/src/bin/bench_search_only.rs

@@ -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);
 
@@ -35,12 +35,17 @@ fn main() {
                     pathdiff::diff_paths(&path, &canonical_path).unwrap_or_else(|| path.clone());
 
                 let relative_path = relative.to_string_lossy().into_owned();
-                let file_name = entry.file_name().to_string_lossy().into_owned();
 
+                let path_string = path.to_string_lossy().into_owned();
+                let relative_start = (path_string.len() - relative_path.len()) as u16;
+                let filename_start = path_string
+                    .rfind('/')
+                    .map(|i| i + 1)
+                    .unwrap_or(relative_start as usize) as u16;
                 files.push(FileItem::new_raw(
-                    path,
-                    relative_path,
-                    file_name,
+                    path_string,
+                    relative_start,
+                    filename_start,
                     entry.metadata().ok().map_or(0, |m| m.len()),
                     0,
                     None,
@@ -64,7 +69,6 @@ fn main() {
         ("long_rare", "user_authentication", 100),
         ("typo_resistant", "contrlr", 200),
         ("path_like", "src/lib", 150),
-        ("single_char", "a", 300),
         ("two_char", "st", 300),
         ("partial_word", "test", 200),
         ("deep_path", "drivers/net", 100),
@@ -87,13 +91,12 @@ 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 {

crates/fff-nvim/src/bin/fuzzy_grep_test.rs

@@ -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;
@@ -31,14 +31,19 @@ fn load_files(base_path: &Path) -> Vec<FileItem> {
             let path = entry.path().to_path_buf();
             let relative = pathdiff::diff_paths(&path, base_path).unwrap_or_else(|| path.clone());
             let relative_path = relative.to_string_lossy().into_owned();
-            let file_name = entry.file_name().to_string_lossy().into_owned();
             let size = entry.metadata().ok().map_or(0, |m| m.len());
             let is_binary = detect_binary(&path, size);
 
+            let path_string = path.to_string_lossy().into_owned();
+            let relative_start = (path_string.len() - relative_path.len()) as u16;
+            let filename_start = path_string
+                .rfind('/')
+                .map(|i| i + 1)
+                .unwrap_or(relative_start as usize) as u16;
             files.push(FileItem::new_raw(
-                path,
-                relative_path,
-                file_name,
+                path_string,
+                relative_start,
+                filename_start,
                 size,
                 0,
                 None,
@@ -71,11 +76,22 @@ 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,
+        &fff::ContentCacheBudget::zero(),
+        None,
+        None,
+        None,
+    );
     let elapsed = start.elapsed();
 
     eprintln!("══════════════════════════════════════════════════════════════");
@@ -99,7 +115,7 @@ fn run_fuzzy_query(files: &[FileItem], query: &str, label: &str) {
         if m.file_index != current_file_idx {
             current_file_idx = m.file_index;
             let file = &result.files[m.file_index];
-            eprintln!("\n  ┌─ {}", file.relative_path);
+            eprintln!("\n  ┌─ {}", file.relative_path());
         }
 
         // Truncate long lines for display
@@ -163,15 +179,14 @@ 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);
 
     eprintln!("Loading files...");
     let load_start = Instant::now();
     let files = load_files(&canonical);
-    let non_binary = files.iter().filter(|f| !f.is_binary).count();
+    let non_binary = files.iter().filter(|f| !f.is_binary()).count();
     eprintln!(
         "Loaded {} files ({} non-binary) in {:.2}s\n",
         files.len(),

crates/fff-nvim/src/bin/grep_profiler.rs

@@ -1,4 +1,3 @@
-use fff_core::FileItem;
 /// Live grep benchmark profiler for fff.nvim
 ///
 /// Benchmarks the full grep pipeline against a large repository (Linux kernel).
@@ -10,7 +9,11 @@ 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::{
+    BigramFilter, FileItem,
+    grep::{GrepMode, GrepSearchOptions, grep_search, parse_grep_query},
+    types::ContentCacheBudget,
+};
 use std::io::Read;
 use std::path::Path;
 use std::time::{Duration, Instant};
@@ -34,14 +37,19 @@ fn load_files(base_path: &Path) -> Vec<FileItem> {
             let path = entry.path().to_path_buf();
             let relative = pathdiff::diff_paths(&path, base_path).unwrap_or_else(|| path.clone());
             let relative_path = relative.to_string_lossy().into_owned();
-            let file_name = entry.file_name().to_string_lossy().into_owned();
             let size = entry.metadata().ok().map_or(0, |m| m.len());
             let is_binary = detect_binary(&path, size);
 
+            let path_string = path.to_string_lossy().into_owned();
+            let relative_start = (path_string.len() - relative_path.len()) as u16;
+            let filename_start = path_string
+                .rfind('/')
+                .map(|i| i + 1)
+                .unwrap_or(relative_start as usize) as u16;
             files.push(FileItem::new_raw(
-                path,
-                relative_path,
-                file_name,
+                path_string,
+                relative_start,
+                filename_start,
                 size,
                 0,
                 None,
@@ -115,29 +123,51 @@ impl BenchStats {
 struct GrepBench<'a> {
     files: &'a [FileItem],
     options: GrepSearchOptions,
+    bigram_index: Option<&'a BigramFilter>,
 }
 
 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,
+            bigram_index: None,
             options: GrepSearchOptions {
                 max_file_size: 10 * 1024 * 1024,
                 max_matches_per_file: 200,
                 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,
             },
         }
     }
 
+    fn with_bigram(mut self, index: &'a BigramFilter) -> Self {
+        self.bigram_index = Some(index);
+        self
+    }
+
     /// Run a single grep search, return (duration, match_count, files_searched)
     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,
+            &ContentCacheBudget::default(),
+            self.bigram_index,
+            None,
+            None,
+        );
         let elapsed = start.elapsed();
         (elapsed, result.matches.len(), result.total_files_searched)
     }
@@ -159,6 +189,17 @@ impl<'a> GrepBench<'a> {
     }
 }
 
+fn build_bigram(files: &mut [FileItem]) -> BigramFilter {
+    let budget = ContentCacheBudget::default();
+    let (index, binary_indices) = fff::build_bigram_index(files, &budget);
+
+    for &i in &binary_indices {
+        files[i].set_binary(true);
+    }
+
+    index
+}
+
 fn fmt_dur(d: Duration) -> String {
     let us = d.as_micros();
     if us > 1_000_000 {
@@ -215,16 +256,16 @@ 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 mut files = load_files(&canonical);
     let load_time = load_start.elapsed();
-    let non_binary = files.iter().filter(|f| !f.is_binary).count();
+    let non_binary = files.iter().filter(|f| !f.is_binary()).count();
     let large_files = files.iter().filter(|f| f.size > 10 * 1024 * 1024).count();
     eprintln!(
         "  Loaded {} files in {:.2}s ({} non-binary, {} >10MB skipped)\n",
@@ -236,7 +277,7 @@ 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();
 
@@ -259,7 +300,7 @@ 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();
 
@@ -293,9 +334,107 @@ fn main() {
         print_row(name, &stats, matches, files_searched, *iters);
     }
 
-    eprintln!("\n[4/5] Incremental typing simulation");
+    eprintln!("\n[3b/7] Building bigram index...");
+    let bigram_start = Instant::now();
+    let bigram_index = build_bigram(&mut files);
+    eprintln!(
+        "  Built in {:.2}s ({} columns, {:.1} MB)\n",
+        bigram_start.elapsed().as_secs_f64(),
+        bigram_index.file_count(),
+        bigram_index.heap_bytes() as f64 / (1024.0 * 1024.0),
+    );
+
+    eprintln!("[3c/7] Bigram-accelerated warm benchmarks (same queries, with bigram prefilter)");
+    print_header();
+
+    let bigram_bench = GrepBench::new(&files).with_bigram(&bigram_index);
+    for (name, query, iters) in &warm_queries {
+        let bigram_name = format!("bg_{}", name.strip_prefix("warm_").unwrap_or(name));
+        let (stats, matches, files_searched) = bigram_bench.bench_query(query, *iters);
+        print_row(&bigram_name, &stats, matches, files_searched, *iters);
+    }
+
+    // ── 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 bench = GrepBench::new(&files);
     let typing_sequences: Vec<(&str, Vec<&str>)> = vec![
         (
             "mutex_lock",
@@ -338,7 +477,7 @@ 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";
@@ -363,9 +502,20 @@ 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,
+            &fff::ContentCacheBudget::unlimited(),
+            None,
+            None,
+            None,
+        );
         let elapsed = start.elapsed();
         eprintln!(
             "    {:>6} | {:>12} | {:>8} | {:>6} | {:>12}",
@@ -384,7 +534,13 @@ fn main() {
     }
 
     eprintln!("\n=== Summary ===");
-    let mmap_count = files.iter().filter(|f| f.get_mmap().is_some()).count();
+    let mmap_count = files
+        .iter()
+        .filter(|f| {
+            f.get_content_for_search(&fff::ContentCacheBudget::unlimited())
+                .is_some()
+        })
+        .count();
     eprintln!("  Files with cached mmap: {}", mmap_count);
     eprintln!("  Total indexed files: {}", files.len());
     eprintln!("  Non-binary files: {}", non_binary);

crates/fff-nvim/src/bin/grep_vs_rg.rs

@@ -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;
@@ -47,14 +48,19 @@ fn load_files(base_path: &Path) -> Vec<FileItem> {
             let path = entry.path().to_path_buf();
             let relative = pathdiff::diff_paths(&path, base_path).unwrap_or_else(|| path.clone());
             let relative_path = relative.to_string_lossy().into_owned();
-            let file_name = entry.file_name().to_string_lossy().into_owned();
             let size = entry.metadata().ok().map_or(0, |m| m.len());
             let is_binary = detect_binary(&path, size);
 
+            let path_string = path.to_string_lossy().into_owned();
+            let relative_start = (path_string.len() - relative_path.len()) as u16;
+            let filename_start = path_string
+                .rfind('/')
+                .map(|i| i + 1)
+                .unwrap_or(relative_start as usize) as u16;
             files.push(FileItem::new_raw(
-                path,
-                relative_path,
-                file_name,
+                path_string,
+                relative_start,
+                filename_start,
                 size,
                 0,
                 None,
@@ -204,9 +210,48 @@ 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,
+        &fff::ContentCacheBudget::zero(),
+        None,
+        None,
+        None,
+    );
+    let elapsed = start.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,
+        &fff::ContentCacheBudget::unlimited(),
+        None,
+        None,
+        None,
+    );
     let elapsed = start.elapsed();
     (result.matches.len(), elapsed)
 }
@@ -222,9 +267,20 @@ 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,
+        &fff::ContentCacheBudget::unlimited(),
+        None,
+        None,
+        None,
+    );
     let elapsed = start.elapsed();
     (result.matches.len(), elapsed)
 }
@@ -292,7 +348,7 @@ 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")
@@ -314,7 +370,7 @@ fn main() {
 
     eprintln!("[1/5] Indexing files...");
     let files = load_files(&canonical);
-    let non_binary = files.iter().filter(|f| !f.is_binary).count();
+    let non_binary = files.iter().filter(|f| !f.is_binary()).count();
     eprintln!("  {} files ({} searchable)\n", files.len(), non_binary);
 
     eprintln!("[2/5] Warming caches (fff mmap + OS page cache)...");

crates/fff-nvim/src/bin/jemalloc_profile.rs

@@ -1,7 +1,6 @@
-use fff_core::file_picker::FilePicker;
-use fff_core::{FuzzySearchOptions, PaginationArgs, QueryParser, SharedFrecency, SharedPicker};
+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;
 
@@ -90,13 +89,12 @@ fn test_search_memory_pattern(
                 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 {
@@ -180,16 +178,20 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
     println!();
 
     // Create shared state
-    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
-    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
 
     // Initialize FilePicker
     println!("Initializing FilePicker...");
     FilePicker::new_with_shared_state(
-        base_path.clone(),
-        false,
-        Arc::clone(&shared_picker),
-        Arc::clone(&shared_frecency),
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        fff::FilePickerOptions {
+            base_path: base_path.clone(),
+            warmup_mmap_cache: false,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
     )?;
 
     // Wait for initial scan

crates/fff-nvim/src/bin/search_profiler.rs

@@ -1,50 +1,23 @@
-use fff_core::file_picker::FilePicker;
-use fff_core::{
+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(shared_picker: &SharedPicker, timeout_secs: u64) -> Result<usize, String> {
-    let start = Instant::now();
     let timeout = Duration::from_secs(timeout_secs);
-    let mut iteration = 0;
+    if !shared_picker.wait_for_scan(timeout) {
+        return Err(format!("Scan timed out after {} seconds", timeout_secs));
+    }
 
-    loop {
-        iteration += 1;
-
-        let picker_guard = shared_picker
-            .read()
-            .map_err(|_| "Failed to acquire read lock")?;
-        if let Some(ref picker) = *picker_guard {
-            let is_scanning = picker.is_scan_active();
-            let file_count = picker.get_files().len();
-
-            if iteration % 20 == 0 {
-                eprintln!(
-                    "  [{:.1}s] Scanning: {}, Files: {}",
-                    start.elapsed().as_secs_f64(),
-                    is_scanning,
-                    file_count
-                );
-            }
-
-            if !is_scanning && file_count > 0 {
-                return Ok(file_count);
-            }
-        } else if iteration % 20 == 0 {
-            eprintln!(
-                "  [{:.1}s] FilePicker is None",
-                start.elapsed().as_secs_f64()
-            );
-        }
-
-        if start.elapsed() > timeout {
-            return Err(format!("Scan timed out after {} seconds", timeout_secs));
-        }
-
-        std::thread::sleep(Duration::from_millis(100));
+    let picker_guard = shared_picker
+        .read()
+        .map_err(|e| format!("Failed to acquire read lock: {}", e))?;
+    if let Some(ref picker) = *picker_guard {
+        Ok(picker.get_files().len())
+    } else {
+        Err("FilePicker not initialized".to_string())
     }
 }
 
@@ -52,7 +25,7 @@ fn wait_for_scan(shared_picker: &SharedPicker, timeout_secs: u64) -> Result<usiz
 fn get_files(shared_picker: &SharedPicker) -> Result<Vec<FileItem>, String> {
     let picker_guard = shared_picker
         .read()
-        .map_err(|_| "Failed to acquire read lock")?;
+        .map_err(|e| format!("Failed to acquire read lock: {}", e))?;
     if let Some(ref picker) = *picker_guard {
         Ok(picker.get_files().to_vec())
     } else {
@@ -71,20 +44,24 @@ 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));
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
 
     eprintln!("Initializing FilePicker for: {:?}", canonical_path);
     FilePicker::new_with_shared_state(
-        canonical_path.to_string_lossy().to_string(),
-        false,
-        Arc::clone(&shared_picker),
-        Arc::clone(&shared_frecency),
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        fff::FilePickerOptions {
+            base_path: canonical_path.to_string_lossy().to_string(),
+            warmup_mmap_cache: false,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
     )
-    .expect("Failed to init FilePicker");
+    .expect("Failed to init FilePicker with shared state");
 
     // Give background thread time to start
     std::thread::sleep(Duration::from_millis(200));
@@ -97,16 +74,16 @@ fn main() {
 
     // Test queries representing different search patterns
     let test_queries = vec![
-        ("short_common", "mod", 5000),
-        ("medium_specific", "controller", 2000),
-        ("long_rare", "user_authentication", 1000),
-        ("typo_resistant", "contrlr", 2000),
-        ("path_like", "src/lib", 1500),
-        ("single_char", "a", 3000),
-        ("two_char", "st", 3000),
-        ("partial_word", "test", 2000),
-        ("deep_path", "drivers/net", 1000),
-        ("extension", ".rs", 2000),
+        ("short_common", "mod", 100),
+        ("medium_specific", "controller", 100),
+        ("long_rare", "user_authentication", 100),
+        ("typo_resistant", "contrlr", 100),
+        ("path_like", "src/lib", 100),
+        ("single_char", "a", 100),
+        ("two_char", "st", 100),
+        ("partial_word", "test", 100),
+        ("deep_path", "drivers/net", 100),
+        ("extension", ".rs", 100),
     ];
 
     eprintln!("Running search profiler...");
@@ -125,13 +102,12 @@ 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 {

crates/fff-nvim/src/bin/test_memory_leak.rs

@@ -1,8 +1,7 @@
-use fff_core::file_picker::FilePicker;
-use fff_core::{FuzzySearchOptions, PaginationArgs, QueryParser, SharedFrecency, SharedPicker};
+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};
 
@@ -79,16 +78,20 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
     println!();
 
     // Create shared state
-    let shared_picker: SharedPicker = Arc::new(RwLock::new(None));
-    let shared_frecency: SharedFrecency = Arc::new(RwLock::new(None));
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
 
     // Initialize the file picker
     println!("📁 Initializing FilePicker...");
     FilePicker::new_with_shared_state(
-        base_path.clone(),
-        false,
-        Arc::clone(&shared_picker),
-        Arc::clone(&shared_frecency),
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        fff::FilePickerOptions {
+            base_path: base_path.clone(),
+            warmup_mmap_cache: false,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
     )?;
 
     // Wait for initial scan to complete
@@ -141,7 +144,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
             if !files.is_empty() {
                 println!("Sample files:");
                 for (i, file) in files.iter().take(5).enumerate() {
-                    println!("  {}. {}", i + 1, file.relative_path);
+                    println!("  {}. {}", i + 1, file.relative_path());
                 }
             }
             files.len()
@@ -199,13 +202,12 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                 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 {

crates/fff-nvim/src/bin/test_watcher.rs

@@ -2,13 +2,13 @@
 #![allow(dead_code)]
 #![allow(clippy::enum_variant_names)]
 
-use fff_core::file_picker::FilePicker;
-use fff_core::git::format_git_status;
-use fff_core::{FuzzySearchOptions, PaginationArgs, QueryParser, SharedFrecency, SharedPicker};
+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;
 
@@ -25,11 +25,11 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
     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));
+    let shared_picker = SharedPicker::default();
+    let shared_frecency = SharedFrecency::default();
 
     // Clone for signal handler
-    let picker_for_cleanup = Arc::clone(&shared_picker);
+    let picker_for_cleanup = shared_picker.clone();
     ctrlc::set_handler(move || {
         println!("\n🛑 Received interrupt signal, shutting down...");
         if let Ok(mut guard) = picker_for_cleanup.write() {
@@ -46,10 +46,14 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     // Initialize the file picker using shared state
     FilePicker::new_with_shared_state(
-        base_path.clone(),
-        false,
-        Arc::clone(&shared_picker),
-        Arc::clone(&shared_frecency),
+        shared_picker.clone(),
+        shared_frecency.clone(),
+        fff::FilePickerOptions {
+            base_path: base_path.clone(),
+            warmup_mmap_cache: false,
+            mode: FFFMode::default(),
+            ..Default::default()
+        },
     )?;
 
     // Get initial file count from shared state
@@ -64,7 +68,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                 println!(
                     "  {}. {} ({})",
                     i + 1,
-                    file.relative_path,
+                    file.relative_path(),
                     format_git_status(file.git_status)
                 );
             }
@@ -106,7 +110,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                 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);
+                    println!("   ➕ {}", file.relative_path());
                 }
             } else {
                 let removed = last_count - current_count;
@@ -152,13 +156,12 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
             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 {
@@ -183,7 +186,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                 println!(
                     "   {}. {} (score: {})",
                     i + 1,
-                    file.relative_path,
+                    file.relative_path(),
                     score.total
                 );
             }
@@ -198,5 +201,6 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
             picker.stop_background_monitor();
         }
     }
+
     Ok(())
 }

crates/fff-nvim/src/error.rs

@@ -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>;
 }
@@ -24,14 +24,3 @@ impl<T> IntoLuaResult<T> for Result<T, CoreError> {
         self.map_err(to_lua_error)
     }
 }
-
-/// Extension trait for Result<T, PoisonError> to convert to Result<T, CoreError>
-pub trait IntoCoreError<T> {
-    fn with_lock_error(self, err: CoreError) -> Result<T, CoreError>;
-}
-
-impl<T, G> IntoCoreError<T> for Result<T, std::sync::PoisonError<G>> {
-    fn with_lock_error(self, err: CoreError) -> Result<T, CoreError> {
-        self.map_err(|_| err)
-    }
-}

crates/fff-nvim/src/hex_dump.rs

@@ -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))
+}

crates/fff-nvim/src/lib.rs

@@ -1,21 +1,22 @@
 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, SharedFrecency,
-    SharedPicker, SharedQueryTracker,
+use error::IntoLuaResult;
+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;
@@ -25,74 +26,67 @@ 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 static FILE_PICKER: Lazy<SharedPicker> = Lazy::new(SharedPicker::default);
+pub static FRECENCY: Lazy<SharedFrecency> = Lazy::new(SharedFrecency::default);
+pub static QUERY_TRACKER: Lazy<SharedQueryTracker> = Lazy::new(SharedQueryTracker::default);
 
 pub fn init_db(
     _: &Lua,
     (frecency_db_path, history_db_path, use_unsafe_no_lock): (String, String, bool),
 ) -> LuaResult<bool> {
-    let mut frecency = FRECENCY
-        .write()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let mut frecency = FRECENCY.write().into_lua_result()?;
     if frecency.is_some() {
         *frecency = None;
     }
     *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);
 
-    let mut query_tracker = QUERY_TRACKER
-        .write()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    // Spawn background GC to purge stale entries without blocking startup
+    let _ = FRECENCY.spawn_gc(frecency_db_path, use_unsafe_no_lock);
+
+    let mut query_tracker = QUERY_TRACKER.write().into_lua_result()?;
     if query_tracker.is_some() {
         *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)
 }
 
 pub fn destroy_frecency_db(_: &Lua, _: ()) -> LuaResult<bool> {
-    let mut frecency = FRECENCY
-        .write()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let mut frecency = FRECENCY.write().into_lua_result()?;
     *frecency = None;
     Ok(true)
 }
 
 pub fn destroy_query_db(_: &Lua, _: ()) -> LuaResult<bool> {
-    let mut query_tracker = QUERY_TRACKER
-        .write()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let mut query_tracker = QUERY_TRACKER.write().into_lua_result()?;
     *query_tracker = None;
     Ok(true)
 }
 
 pub fn init_file_picker(_: &Lua, base_path: String) -> LuaResult<bool> {
     {
-        let guard = FILE_PICKER
-            .read()
-            .with_lock_error(Error::AcquireItemLock)
-            .into_lua_result()?;
+        let guard = FILE_PICKER.read().into_lua_result()?;
         if guard.is_some() {
             return Ok(false);
         }
     }
 
     FilePicker::new_with_shared_state(
-        base_path,
-        false,
-        Arc::clone(&FILE_PICKER),
-        Arc::clone(&FRECENCY),
+        FILE_PICKER.clone(),
+        FRECENCY.clone(),
+        fff::FilePickerOptions {
+            base_path,
+            warmup_mmap_cache: true,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
     )
     .into_lua_result()?;
 
@@ -104,9 +98,7 @@ fn reinit_file_picker_internal(path: &Path) -> Result<(), Error> {
     // 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)?;
+        let mut guard = FILE_PICKER.write()?;
         if let Some(ref mut picker) = *guard {
             // Signal cancellation BEFORE stopping — this tells any orphaned
             // scan threads from this picker to discard their results.
@@ -119,10 +111,14 @@ fn reinit_file_picker_internal(path: &Path) -> Result<(), Error> {
 
     // Create new picker — this atomically replaces the old one via write lock
     FilePicker::new_with_shared_state(
-        path.to_string_lossy().to_string(),
-        false,
-        Arc::clone(&FILE_PICKER),
-        Arc::clone(&FRECENCY),
+        FILE_PICKER.clone(),
+        FRECENCY.clone(),
+        fff::FilePickerOptions {
+            base_path: path.to_string_lossy().to_string(),
+            warmup_mmap_cache: true,
+            mode: FFFMode::Neovim,
+            ..Default::default()
+        },
     )?;
 
     Ok(())
@@ -137,7 +133,7 @@ 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))
     })?;
 
@@ -164,10 +160,7 @@ pub fn restart_index_in_path(_: &Lua, new_path: String) -> LuaResult<()> {
 }
 
 pub fn scan_files(_: &Lua, _: ()) -> LuaResult<()> {
-    let mut file_picker = FILE_PICKER
-        .write()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let mut file_picker = FILE_PICKER.write().into_lua_result()?;
     let picker = file_picker
         .as_mut()
         .ok_or(Error::FilePickerMissing)
@@ -199,10 +192,7 @@ pub fn fuzzy_search_files(
         Option<usize>,
     ),
 ) -> LuaResult<LuaValue> {
-    let file_picker_guard = FILE_PICKER
-        .read()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let file_picker_guard = FILE_PICKER.read().into_lua_result()?;
     let Some(ref picker) = *file_picker_guard else {
         return Err(error::to_lua_error(Error::FilePickerMissing));
     };
@@ -210,26 +200,13 @@ 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().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,
@@ -238,19 +215,18 @@ 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 {
@@ -260,6 +236,34 @@ 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.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)
 }
 
@@ -286,23 +290,19 @@ pub fn live_grep(
         Option<u64>,
     ),
 ) -> LuaResult<LuaValue> {
-    let file_picker_guard = FILE_PICKER
-        .read()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let file_picker_guard = FILE_PICKER.read().into_lua_result()?;
     let Some(ref picker) = *file_picker_guard else {
         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),
@@ -310,22 +310,71 @@ 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 = picker.grep(&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", 0i32)?;
+    item.set("modification_frecency_score", 0i32)?;
+    item.set("total_frecency_score", 0i32)?;
+    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);
 
     // Track access in frecency DB (expensive LMDB write, ~100-200ms)
     // Do this WITHOUT holding FILE_PICKER lock to avoid blocking searches
-    let frecency_guard = FRECENCY
-        .read()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let frecency_guard = FRECENCY.read().into_lua_result()?;
     let Some(ref frecency) = *frecency_guard else {
         return Ok(false);
     };
@@ -335,18 +384,12 @@ pub fn track_access(_: &Lua, file_path: String) -> LuaResult<bool> {
     drop(frecency_guard);
 
     // Quick lock to update single file's frecency score in picker
-    let mut file_picker = FILE_PICKER
-        .write()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let mut file_picker = FILE_PICKER.write().into_lua_result()?;
     let Some(ref mut picker) = *file_picker else {
         return Err(error::to_lua_error(Error::FilePickerMissing));
     };
 
-    let frecency_guard = FRECENCY
-        .read()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let frecency_guard = FRECENCY.read().into_lua_result()?;
     let Some(ref frecency) = *frecency_guard else {
         return Ok(false);
     };
@@ -358,10 +401,7 @@ pub fn track_access(_: &Lua, file_path: String) -> LuaResult<bool> {
 }
 
 pub fn get_scan_progress(lua: &Lua, _: ()) -> LuaResult<LuaValue> {
-    let file_picker = FILE_PICKER
-        .read()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let file_picker = FILE_PICKER.read().into_lua_result()?;
     let picker = file_picker
         .as_ref()
         .ok_or(Error::FilePickerMissing)
@@ -375,10 +415,7 @@ pub fn get_scan_progress(lua: &Lua, _: ()) -> LuaResult<LuaValue> {
 }
 
 pub fn is_scanning(_: &Lua, _: ()) -> LuaResult<bool> {
-    let file_picker = FILE_PICKER
-        .read()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let file_picker = FILE_PICKER.read().into_lua_result()?;
     let picker = file_picker
         .as_ref()
         .ok_or(Error::FilePickerMissing)
@@ -387,10 +424,7 @@ pub fn is_scanning(_: &Lua, _: ()) -> LuaResult<bool> {
 }
 
 pub fn get_git_root(_: &Lua, _: ()) -> LuaResult<Option<String>> {
-    let file_picker = FILE_PICKER
-        .read()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let file_picker = FILE_PICKER.read().into_lua_result()?;
     let Some(ref picker) = *file_picker else {
         return Ok(None);
     };
@@ -399,22 +433,16 @@ pub fn get_git_root(_: &Lua, _: ()) -> LuaResult<Option<String>> {
 }
 
 pub fn refresh_git_status(_: &Lua, _: ()) -> LuaResult<usize> {
-    FilePicker::refresh_git_status(&FILE_PICKER, &FRECENCY).into_lua_result()
+    FILE_PICKER.refresh_git_status(&FRECENCY).into_lua_result()
 }
 
 pub fn update_single_file_frecency(_: &Lua, file_path: String) -> LuaResult<bool> {
-    let frecency_guard = FRECENCY
-        .read()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let frecency_guard = FRECENCY.read().into_lua_result()?;
     let Some(ref frecency) = *frecency_guard else {
         return Ok(false);
     };
 
-    let mut file_picker = FILE_PICKER
-        .write()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let mut file_picker = FILE_PICKER.write().into_lua_result()?;
     let Some(ref mut picker) = *file_picker else {
         return Err(error::to_lua_error(Error::FilePickerMissing));
     };
@@ -426,10 +454,7 @@ pub fn update_single_file_frecency(_: &Lua, file_path: String) -> LuaResult<bool
 }
 
 pub fn stop_background_monitor(_: &Lua, _: ()) -> LuaResult<bool> {
-    let mut file_picker = FILE_PICKER
-        .write()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let mut file_picker = FILE_PICKER.write().into_lua_result()?;
     let Some(ref mut picker) = *file_picker else {
         return Err(error::to_lua_error(Error::FilePickerMissing));
     };
@@ -440,10 +465,7 @@ pub fn stop_background_monitor(_: &Lua, _: ()) -> LuaResult<bool> {
 }
 
 pub fn cleanup_file_picker(_: &Lua, _: ()) -> LuaResult<bool> {
-    let mut file_picker = FILE_PICKER
-        .write()
-        .with_lock_error(Error::AcquireItemLock)
-        .into_lua_result()?;
+    let mut file_picker = FILE_PICKER.write().into_lua_result()?;
     if let Some(picker) = file_picker.take() {
         drop(picker);
         ::tracing::info!("FilePicker cleanup completed");
@@ -461,10 +483,7 @@ pub fn cancel_scan(_: &Lua, _: ()) -> LuaResult<bool> {
 pub fn track_query_completion(_: &Lua, (query, file_path): (String, String)) -> LuaResult<bool> {
     // Get the project path before spawning thread
     let project_path = {
-        let file_picker = FILE_PICKER
-            .read()
-            .with_lock_error(Error::AcquireItemLock)
-            .into_lua_result()?;
+        let file_picker = FILE_PICKER.read().into_lua_result()?;
         let Some(ref picker) = *file_picker else {
             return Ok(false);
         };
@@ -472,7 +491,7 @@ 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");
@@ -481,9 +500,10 @@ 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);
+    let query_tracker = QUERY_TRACKER.clone();
     std::thread::spawn(move || {
-        if let Ok(Some(tracker)) = query_tracker.write().as_deref_mut()
+        if let Ok(mut guard) = query_tracker.write()
+            && let Some(tracker) = guard.as_mut()
             && let Err(e) = tracker.track_query_completion(&query, &project_path, &file_path)
         {
             tracing::error!(
@@ -500,20 +520,14 @@ pub fn track_query_completion(_: &Lua, (query, file_path): (String, String)) ->
 
 pub fn get_historical_query(_: &Lua, offset: usize) -> LuaResult<Option<String>> {
     let project_path = {
-        let file_picker = FILE_PICKER
-            .read()
-            .with_lock_error(Error::AcquireItemLock)
-            .into_lua_result()?;
+        let file_picker = FILE_PICKER.read().into_lua_result()?;
         let Some(ref picker) = *file_picker else {
             return Ok(None);
         };
         picker.base_path().to_path_buf()
     };
 
-    let query_tracker = QUERY_TRACKER
-        .read()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let query_tracker = QUERY_TRACKER.read().into_lua_result()?;
     let Some(ref tracker) = *query_tracker else {
         return Ok(None);
     };
@@ -525,19 +539,17 @@ pub fn get_historical_query(_: &Lua, offset: usize) -> LuaResult<Option<String>>
 
 pub fn track_grep_query(_: &Lua, query: String) -> LuaResult<bool> {
     let project_path = {
-        let file_picker = FILE_PICKER
-            .read()
-            .with_lock_error(Error::AcquireItemLock)
-            .into_lua_result()?;
+        let file_picker = FILE_PICKER.read().into_lua_result()?;
         let Some(ref picker) = *file_picker else {
             return Ok(false);
         };
         picker.base_path().to_path_buf()
     };
 
-    let query_tracker = Arc::clone(&QUERY_TRACKER);
+    let query_tracker = QUERY_TRACKER.clone();
     std::thread::spawn(move || {
-        if let Ok(Some(tracker)) = query_tracker.write().as_deref_mut()
+        if let Ok(mut guard) = query_tracker.write()
+            && let Some(ref mut tracker) = *guard
             && let Err(e) = tracker.track_grep_query(&query, &project_path)
         {
             tracing::error!(
@@ -553,20 +565,14 @@ pub fn track_grep_query(_: &Lua, query: String) -> LuaResult<bool> {
 
 pub fn get_historical_grep_query(_: &Lua, offset: usize) -> LuaResult<Option<String>> {
     let project_path = {
-        let file_picker = FILE_PICKER
-            .read()
-            .with_lock_error(Error::AcquireItemLock)
-            .into_lua_result()?;
+        let file_picker = FILE_PICKER.read().into_lua_result()?;
         let Some(ref picker) = *file_picker else {
             return Ok(None);
         };
         picker.base_path().to_path_buf()
     };
 
-    let query_tracker = QUERY_TRACKER
-        .read()
-        .with_lock_error(Error::AcquireFrecencyLock)
-        .into_lua_result()?;
+    let query_tracker = QUERY_TRACKER.read().into_lua_result()?;
     let Some(ref tracker) = *query_tracker else {
         return Ok(None);
     };
@@ -582,10 +588,7 @@ pub fn wait_for_initial_scan(_: &Lua, timeout_ms: Option<u64>) -> LuaResult<bool
     // 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 file_picker = FILE_PICKER.read().into_lua_result()?;
         let picker = file_picker
             .as_ref()
             .ok_or(Error::FilePickerMissing)
@@ -818,6 +821,7 @@ 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)
 }

crates/fff-nvim/src/log.rs

@@ -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};

crates/fff-nvim/src/lua_types.rs

@@ -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>,
 }
@@ -41,9 +35,9 @@ impl IntoLua for LuaPosition {
 
 fn file_item_into_lua(item: &FileItem, lua: &Lua) -> LuaResult<LuaValue> {
     let table = lua.create_table()?;
-    table.set("path", item.path.to_string_lossy().to_string())?;
-    table.set("relative_path", item.relative_path.clone())?;
-    table.set("name", item.file_name.clone())?;
+    table.set("path", item.path_str())?;
+    table.set("relative_path", item.relative_path())?;
+    table.set("name", item.file_name())?;
     table.set("size", item.size)?;
     table.set("modified", item.modified)?;
     table.set("access_frecency_score", item.access_frecency_score)?;
@@ -51,9 +45,9 @@ fn file_item_into_lua(item: &FileItem, lua: &Lua) -> LuaResult<LuaValue> {
         "modification_frecency_score",
         item.modification_frecency_score,
     )?;
-    table.set("total_frecency_score", item.total_frecency_score)?;
+    table.set("total_frecency_score", item.total_frecency_score())?;
     table.set("git_status", format_git_status(item.git_status))?;
-    table.set("is_binary", item.is_binary)?;
+    table.set("is_binary", item.is_binary())?;
     Ok(LuaValue::Table(table))
 }
 
@@ -128,14 +122,14 @@ impl IntoLua for GrepResultLua<'_> {
 
             // File metadata from the deduplicated files vec
             let file = self.inner.files[m.file_index];
-            item.set("path", file.path.to_string_lossy().to_string())?;
-            item.set("relative_path", file.relative_path.as_str())?;
-            item.set("name", file.file_name.as_str())?;
-            item.set("is_binary", file.is_binary)?;
+            item.set("path", file.path_str())?;
+            item.set("relative_path", file.relative_path())?;
+            item.set("name", file.file_name())?;
+            item.set("is_binary", file.is_binary())?;
             item.set("git_status", format_git_status(file.git_status))?;
             item.set("size", file.size)?;
             item.set("modified", file.modified)?;
-            item.set("total_frecency_score", file.total_frecency_score)?;
+            item.set("total_frecency_score", file.total_frecency_score())?;
             item.set("access_frecency_score", file.access_frecency_score)?;
             item.set(
                 "modification_frecency_score",

crates/fff-nvim/src/path_shortening.rs

@@ -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());
         }
     }

crates/fff-query-parser/Cargo.toml

@@ -1,7 +1,10 @@
 [package]
 name = "fff-query-parser"
-version = "0.1.0"
+version = "0.5.2"
 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"
@@ -11,7 +14,6 @@ default = []
 zlob = ["dep:zlob"]
 
 [dependencies]
-smallvec = { workspace = true }
 zlob = { workspace = true, optional = true }
 
 [dev-dependencies]

crates/fff-query-parser/benches/parse_bench.rs

@@ -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";

crates/fff-query-parser/src/config.rs

@@ -1,6 +1,47 @@
 use crate::constraints::Constraint;
 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 {
     fn enable_glob(&self) -> bool {
@@ -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
@@ -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
@@ -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
+        }
+    }
+}

crates/fff-query-parser/src/constraints.rs

@@ -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>;

crates/fff-query-parser/src/lib.rs

@@ -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")));
@@ -45,16 +45,12 @@ 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 {
@@ -64,32 +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) => {
@@ -107,9 +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")));
@@ -118,9 +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] {
@@ -132,7 +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) => {
@@ -145,7 +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],
@@ -156,9 +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],
@@ -169,9 +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],
@@ -182,9 +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 {
@@ -224,21 +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) => {

crates/fff-query-parser/src/parser.rs

@@ -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();
@@ -44,33 +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,
+            };
         }
 
         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);
                 }
@@ -82,7 +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() {
@@ -114,11 +158,12 @@ impl<C: ParserConfig> QueryParser<C> {
             }
         };
 
-        Some(FFFQuery {
+        FFFQuery {
+            raw_query,
             constraints,
             fuzzy_query,
             location,
-        })
+        }
     }
 }
 
@@ -145,20 +190,27 @@ impl<'a> FFFQuery<'a> {
     }
 }
 
-/// Strip the leading `\` from a backslash-escaped token, returning the rest.
-/// For all other tokens returns the input unchanged.
+/// 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.starts_with('\\') && token.len() > 1 {
-        &token[1..]
-    } else {
-        token
+    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::FilePickerConfig> {
+impl Default for QueryParser<crate::FileSearchConfig> {
     fn default() -> Self {
-        Self::new(crate::FilePickerConfig)
+        Self::new(crate::FileSearchConfig)
     }
 }
 
@@ -341,11 +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
@@ -384,7 +437,7 @@ fn parse_git_status(value: &str) -> Option<Constraint<'_>> {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::{FilePickerConfig, GrepConfig};
+    use crate::{FileSearchConfig, GrepConfig};
 
     #[test]
     fn test_parse_extension() {
@@ -399,7 +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);
@@ -428,18 +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],
@@ -474,11 +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) => {
@@ -490,10 +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) => {
@@ -505,10 +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) => {
@@ -520,10 +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) => {
@@ -538,10 +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
@@ -557,10 +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) => {
@@ -573,118 +619,91 @@ 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 =")
-            .expect("should parse");
+        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")
-            .expect("should parse");
+        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 =")
-            .expect("should parse");
+        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/")
-            .expect("should parse");
+        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")
-            .expect("should parse");
+        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")
-            .expect("should parse");
+        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")
-            .expect("should parse");
+        let q = QueryParser::new(GrepConfig).parse("\\*.rs foo");
         assert_eq!(q.grep_text(), "*.rs foo");
 
-        let q = QueryParser::new(GrepConfig)
-            .parse("\\/src/ foo")
-            .expect("should parse");
+        let q = QueryParser::new(GrepConfig).parse("\\/src/ foo");
         assert_eq!(q.grep_text(), "/src/ foo");
 
-        let q = QueryParser::new(GrepConfig)
-            .parse("\\!test foo")
-            .expect("should parse");
+        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")
-            .expect("should parse");
+        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")
-            .expect("should parse");
+        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")
-            .expect("should parse");
+        let q = QueryParser::new(GrepConfig).parse("pattern src/**/*.rs");
         assert_eq!(q.grep_text(), "pattern");
     }
 
     #[test]
     fn test_grep_question_mark_is_text() {
         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() {
         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);
     }
@@ -692,9 +711,7 @@ mod tests {
     #[test]
     fn test_grep_path_glob_is_constraint() {
         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!(
@@ -706,9 +723,7 @@ mod tests {
     #[test]
     fn test_grep_brace_is_constraint() {
         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],
@@ -716,12 +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() {
         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,
@@ -732,9 +786,7 @@ mod tests {
     #[test]
     fn test_grep_negated_text() {
         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) => {
@@ -751,9 +803,7 @@ mod tests {
     #[test]
     fn test_grep_negated_path_segment() {
         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) => {
@@ -770,9 +820,7 @@ mod tests {
     #[test]
     fn test_grep_negated_extension() {
         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) => {
@@ -785,4 +833,341 @@ mod tests {
             other => panic!("Expected Not constraint, got {:?}", other),
         }
     }
+
+    #[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
+        );
+    }
+
+    #[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
+        );
+    }
 }

crates/fff-searcher/Cargo.toml

@@ -1,9 +0,0 @@
-[package]
-name = "grep-searcher"
-version = "0.1.16"
-edition = "2024"
-
-[dependencies]
-bstr = { version = "1.6.2", default-features = false, features = ["std"] }
-grep-matcher = { workspace = true }
-memchr = "2.6.3"

crates/fff-searcher/src/searcher/core.rs

@@ -1,212 +0,0 @@
-use grep_matcher::{LineMatchKind, Matcher};
-
-use crate::{
-    lines::{self, LineStep},
-    searcher::{Config, Range, Searcher},
-    sink::{Sink, SinkError, SinkFinish, SinkMatch},
-};
-
-#[derive(Debug)]
-pub(crate) struct Core<'s, M: 's, S> {
-    config: &'s Config,
-    matcher: M,
-    searcher: &'s Searcher,
-    sink: S,
-    pos: usize,
-    absolute_byte_offset: u64,
-    line_number: Option<u64>,
-    last_line_counted: usize,
-    last_line_visited: usize,
-}
-
-impl<'s, M: Matcher, S: Sink> Core<'s, M, S> {
-    pub(crate) fn new(searcher: &'s Searcher, matcher: M, sink: S) -> Core<'s, M, S> {
-        let line_number = if searcher.config.line_number {
-            Some(1)
-        } else {
-            None
-        };
-        Core {
-            config: &searcher.config,
-            matcher,
-            searcher,
-            sink,
-            pos: 0,
-            absolute_byte_offset: 0,
-            line_number,
-            last_line_counted: 0,
-            last_line_visited: 0,
-        }
-    }
-
-    pub(crate) fn pos(&self) -> usize {
-        self.pos
-    }
-
-    pub(crate) fn set_pos(&mut self, pos: usize) {
-        self.pos = pos;
-    }
-
-    pub(crate) fn matched(&mut self, buf: &[u8], range: &Range) -> Result<bool, S::Error> {
-        self.sink_matched(buf, range)
-    }
-
-    pub(crate) fn find(&mut self, slice: &[u8]) -> Result<Option<Range>, S::Error> {
-        match self.matcher.find(slice) {
-            Err(err) => Err(S::Error::error_message(err)),
-            Ok(m) => Ok(m),
-        }
-    }
-
-    fn shortest_match(&mut self, slice: &[u8]) -> Result<Option<usize>, S::Error> {
-        match self.matcher.shortest_match(slice) {
-            Err(err) => Err(S::Error::error_message(err)),
-            Ok(m) => Ok(m),
-        }
-    }
-
-    pub(crate) fn begin(&mut self) -> Result<bool, S::Error> {
-        self.sink.begin(self.searcher)
-    }
-
-    pub(crate) fn finish(&mut self, byte_count: u64) -> Result<(), S::Error> {
-        self.sink.finish(self.searcher, &SinkFinish { byte_count })
-    }
-
-    pub(crate) fn match_by_line(&mut self, buf: &[u8]) -> Result<bool, S::Error> {
-        if self.is_line_by_line_fast() {
-            self.match_by_line_fast(buf)
-        } else {
-            self.match_by_line_slow(buf)
-        }
-    }
-
-    fn match_by_line_slow(&mut self, buf: &[u8]) -> Result<bool, S::Error> {
-        debug_assert!(!self.searcher.multi_line_with_matcher(&self.matcher));
-
-        let range = Range::new(self.pos(), buf.len());
-        let mut stepper =
-            LineStep::new(self.config.line_term.as_byte(), range.start(), range.end());
-        while let Some(line) = stepper.next_match(buf) {
-            let matched = {
-                let slice = lines::without_terminator(&buf[line], self.config.line_term);
-                self.shortest_match(slice)?.is_some()
-            };
-            self.set_pos(line.end());
-            if matched && !self.sink_matched(buf, &line)? {
-                return Ok(false);
-            }
-        }
-        Ok(true)
-    }
-
-    fn match_by_line_fast(&mut self, buf: &[u8]) -> Result<bool, S::Error> {
-        while !buf[self.pos()..].is_empty() {
-            if let Some(line) = self.find_by_line_fast(buf)? {
-                self.set_pos(line.end());
-                if !self.sink_matched(buf, &line)? {
-                    return Ok(false);
-                }
-            } else {
-                break;
-            }
-        }
-        self.set_pos(buf.len());
-        Ok(true)
-    }
-
-    #[inline(always)]
-    fn find_by_line_fast(&mut self, buf: &[u8]) -> Result<Option<Range>, S::Error> {
-        debug_assert!(!self.searcher.multi_line_with_matcher(&self.matcher));
-        debug_assert!(self.is_line_by_line_fast());
-
-        let mut pos = self.pos();
-        while !buf[pos..].is_empty() {
-            match self.matcher.find_candidate_line(&buf[pos..]) {
-                Err(err) => return Err(S::Error::error_message(err)),
-                Ok(None) => return Ok(None),
-                Ok(Some(LineMatchKind::Confirmed(i))) => {
-                    let line = lines::locate(
-                        buf,
-                        self.config.line_term.as_byte(),
-                        Range::zero(i).offset(pos),
-                    );
-                    if line.start() == buf.len() {
-                        pos = buf.len();
-                        continue;
-                    }
-                    return Ok(Some(line));
-                }
-                Ok(Some(LineMatchKind::Candidate(i))) => {
-                    let line = lines::locate(
-                        buf,
-                        self.config.line_term.as_byte(),
-                        Range::zero(i).offset(pos),
-                    );
-                    let slice = lines::without_terminator(&buf[line], self.config.line_term);
-                    if self
-                        .matcher
-                        .is_match(slice)
-                        .map_err(S::Error::error_message)?
-                    {
-                        return Ok(Some(line));
-                    }
-                    pos = line.end();
-                }
-            }
-        }
-        Ok(None)
-    }
-
-    #[inline(always)]
-    fn sink_matched(&mut self, buf: &[u8], range: &Range) -> Result<bool, S::Error> {
-        self.count_lines(buf, range.start());
-        let offset = self.absolute_byte_offset + range.start() as u64;
-        let linebuf = &buf[*range];
-        let keepgoing = self.sink.matched(
-            self.searcher,
-            &SinkMatch {
-                bytes: linebuf,
-                absolute_byte_offset: offset,
-                line_number: self.line_number,
-                buffer: buf,
-                bytes_range_in_buffer: range.start()..range.end(),
-            },
-        )?;
-        if !keepgoing {
-            return Ok(false);
-        }
-        self.last_line_visited = range.end();
-        Ok(true)
-    }
-
-    fn count_lines(&mut self, buf: &[u8], upto: usize) {
-        if let Some(ref mut line_number) = self.line_number {
-            if self.last_line_counted >= upto {
-                return;
-            }
-            let slice = &buf[self.last_line_counted..upto];
-            let count = lines::count(slice, self.config.line_term.as_byte());
-            *line_number += count;
-            self.last_line_counted = upto;
-        }
-    }
-
-    fn is_line_by_line_fast(&self) -> bool {
-        debug_assert!(!self.searcher.multi_line_with_matcher(&self.matcher));
-        if let Some(line_term) = self.matcher.line_terminator() {
-            if line_term.as_byte() == b'\x00' {
-                return false;
-            }
-            if line_term == self.config.line_term {
-                return true;
-            }
-        }
-        if let Some(non_matching) = self.matcher.non_matching_bytes()
-            && non_matching.contains(self.config.line_term.as_byte())
-        {
-            return true;
-        }
-        false
-    }
-}

doc/fff.nvim.txt

@@ -1,52 +1,68 @@
-*fff.nvim.txt*         For Neovim >= 0.10.0         Last change: 2026 March 06
+*fff.nvim.txt*         For Neovim >= 0.10.0         Last change: 2026 April 08
 
 ==============================================================================
 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
 
@@ -261,9 +277,8 @@ all available options:
           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,
@@ -292,19 +307,21 @@ KEY FEATURES ~
 AVAILABLE METHODS
 
 >lua
-    require('fff').find_files()                         -- Find files in current repositro
+    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
@@ -313,13 +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:
@@ -387,6 +397,34 @@ 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
@@ -496,47 +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>
 

flake.lock

@@ -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": {

install-mcp.sh

@@ -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

lua/fff/conf.lua

@@ -313,9 +313,8 @@ local function init()
       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,
@@ -356,7 +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

lua/fff/core.lua

@@ -88,11 +88,11 @@ M.ensure_initialized = function()
   local history_db_path = config.history.db_path or (vim.fn.stdpath('data') .. '/fff_history')
 
   local ok, result = pcall(fuzzy.init_db, frecency_db_path, history_db_path, true)
-  if not ok then vim.notify('Failed to databases: ' .. result, vim.log.levels.WARN) end
+  if not ok then vim.notify('Failed to databases: ' .. tostring(result), vim.log.levels.WARN) end
 
   ok, result = pcall(fuzzy.init_file_picker, config.base_path)
   if not ok then
-    vim.notify('Failed to initialize file picker: ' .. result, vim.log.levels.ERROR)
+    vim.notify('Failed to initialize file picker: ' .. tostring(result), vim.log.levels.ERROR)
     return fuzzy
   end
 

lua/fff/download.lua

@@ -1,19 +1,10 @@
 local M = {}
 local system = require('fff.utils.system')
 local fs_utils = require('fff.utils.fs')
+local fff_version = require('fff.utils.version')
 
 local GITHUB_REPO = 'dmtrKovalenko/fff.nvim'
 
-local function get_current_version(plugin_dir, callback)
-  vim.system({ 'git', 'rev-parse', '--short', 'HEAD' }, { cwd = plugin_dir }, function(result)
-    if result.code ~= 0 or not result.stdout or result.stdout == '' then
-      callback(nil)
-      return
-    end
-    callback(result.stdout:gsub('%s+', ''))
-  end)
-end
-
 local function get_binary_dir(plugin_dir) return plugin_dir .. '/../target/release' end
 
 local function get_binary_path(plugin_dir)
@@ -114,7 +105,7 @@ local function download_from_github(version, binary_path, opts, callback)
     extra_curl_args = opts.extra_curl_args,
   }, function(success, err)
     if not success then
-      vim.uv.fanoushkas_unlink(tmp_path)
+      vim.uv.fs_unlink(tmp_path)
       callback(false, err)
       return
     end
@@ -168,20 +159,38 @@ function M.ensure_downloaded(opts, callback)
     return
   end
 
-  local function on_version(target_version)
-    if not target_version then
+  local function on_release_tag(release_tag)
+    if not release_tag then
       callback(false, 'Could not determine target version')
       return
     end
 
     local binary_path = get_binary_path(plugin_dir)
-    download_from_github(target_version, binary_path, opts, callback)
+    download_from_github(release_tag, binary_path, opts, callback)
   end
 
   if opts.version then
-    on_version(opts.version)
+    on_release_tag(opts.version)
   else
-    get_current_version(plugin_dir, on_version)
+    -- plugin_dir is <repo>/lua; parent is the repo root
+    local repo_root = vim.fn.fnamemodify(plugin_dir, ':h')
+
+    -- 1. Try reading the CI-created tag on HEAD (no version computation)
+    local tag = fff_version.current_release_tag(repo_root)
+    if tag then
+      on_release_tag(tag)
+      return
+    end
+
+    -- 2. No local tag — construct the nightly version (bumps patch so
+    --    the prerelease is higher than Cargo.toml base in semver)
+    local info, err = fff_version.resolve(repo_root)
+    if info then
+      on_release_tag(info.release_tag)
+      return
+    end
+
+    callback(false, err or 'Could not determine target version')
   end
 end
 

lua/fff/file_picker/init.lua

@@ -24,7 +24,7 @@ function M.scan_files()
 
   local ok, result = pcall(fuzzy.scan_files)
   if not ok then
-    vim.notify('Failed to trigger file scan: ' .. result, vim.log.levels.ERROR)
+    vim.notify('Failed to trigger file scan: ' .. tostring(result), vim.log.levels.ERROR)
     return
   end
 
@@ -121,7 +121,7 @@ function M.track_access(file_path)
   if not M.state.initialized then return end
 
   local ok, result = pcall(fuzzy.track_access, file_path)
-  if not ok then vim.notify('Failed to record file access: ' .. result, vim.log.levels.WARN) end
+  if not ok then vim.notify('Failed to record file access: ' .. tostring(result), vim.log.levels.WARN) end
 end
 
 --- Get file content for preview
@@ -156,7 +156,7 @@ function M.get_scan_progress()
 
   local ok, result = pcall(fuzzy.get_scan_progress)
   if not ok then
-    vim.notify('Failed to get scan progress: ' .. result, vim.log.levels.WARN)
+    vim.notify('Failed to get scan progress: ' .. tostring(result), vim.log.levels.WARN)
     return { scanned_files_count = 0, is_scanning = false }
   end
 
@@ -170,7 +170,7 @@ function M.refresh_git_status()
 
   local ok, result = pcall(fuzzy.refresh_git_status)
   if not ok then
-    vim.notify('Failed to refresh git status: ' .. result, vim.log.levels.WARN)
+    vim.notify('Failed to refresh git status: ' .. tostring(result), vim.log.levels.WARN)
     return {}
   end
 
@@ -185,7 +185,7 @@ function M.stop_background_monitor()
 
   local ok, result = pcall(fuzzy.stop_background_monitor)
   if not ok then
-    vim.notify('Failed to stop background monitor: ' .. result, vim.log.levels.WARN)
+    vim.notify('Failed to stop background monitor: ' .. tostring(result), vim.log.levels.WARN)
     return false
   end
   return result
@@ -199,7 +199,7 @@ function M.wait_for_initial_scan(timeout_ms)
 
   local ok, result = pcall(fuzzy.wait_for_initial_scan, timeout_ms)
   if not ok then
-    vim.notify('Failed to wait for initial scan: ' .. result, vim.log.levels.WARN)
+    vim.notify('Failed to wait for initial scan: ' .. tostring(result), vim.log.levels.WARN)
     return false
   end
   return result

lua/fff/file_picker/preview.lua

@@ -2,9 +2,32 @@ 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 = {}
 
+-- Preview buffers are scratch buffers. Detect the file's language and attach
+-- highlighting directly, but keep buffer filetype empty to avoid ftplugin and
+-- LSP side effects that are meant for real editing buffers.
+local function attach_preview_highlighter(bufnr, filetype)
+  if not bufnr or not vim.api.nvim_buf_is_valid(bufnr) then return end
+
+  pcall(vim.treesitter.stop, bufnr)
+  vim.api.nvim_set_option_value('filetype', '', { buf = bufnr })
+  vim.api.nvim_set_option_value('syntax', '', { buf = bufnr })
+
+  if not filetype or filetype == '' then return end
+
+  local lang_ok, lang = pcall(vim.treesitter.language.get_lang, filetype)
+  if not lang_ok or not lang then lang = filetype end
+
+  if pcall(vim.treesitter.language.add, lang) then
+    pcall(vim.treesitter.start, bufnr, lang)
+  else
+    vim.api.nvim_set_option_value('syntax', filetype, { buf = bufnr })
+  end
+end
+
 local function set_buffer_lines(bufnr, lines)
   if not bufnr or not vim.api.nvim_buf_is_valid(bufnr) then return end
 
@@ -273,7 +296,7 @@ local function link_buffer_content(source_bufnr, target_bufnr)
   set_buffer_lines(target_bufnr, lines)
 
   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
+  if source_ft ~= '' then attach_preview_highlighter(target_bufnr, source_ft) end
 
   M.state.has_more_content = false
   M.state.total_file_lines = #lines
@@ -301,6 +324,8 @@ M.state = {
   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
@@ -529,7 +554,7 @@ function M.preview_file(file_path, bufnr)
       set_buffer_lines(bufnr, content)
 
       local file_config = M.get_file_config(file_path)
-      vim.api.nvim_set_option_value('filetype', info.filetype, { buf = bufnr })
+      attach_preview_highlighter(bufnr, info.filetype)
       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 })
@@ -554,7 +579,69 @@ 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
@@ -562,49 +649,40 @@ function M.preview_binary_file(file_path, bufnr)
   local info = M.get_file_info(file_path)
   local lines = {}
 
+  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 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
+
+  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
+
   set_buffer_lines(bufnr, lines)
-  vim.api.nvim_set_option_value('filetype', 'text', { buf = bufnr })
+  attach_preview_highlighter(bufnr, 'text')
   vim.api.nvim_set_option_value('modifiable', false, { buf = bufnr })
   vim.api.nvim_set_option_value('readonly', true, { buf = bufnr })
 
-  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
+  M.state.content_height = #lines
+  M.state.loaded_lines = #lines
 
-        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
-
-          if vim.fn.executable('xxd') == 1 then
-            table.insert(lines, '')
-            set_buffer_lines(bufnr, lines)
-
-            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
-
-                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
@@ -641,6 +719,9 @@ 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
@@ -672,15 +753,19 @@ 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)
@@ -732,7 +817,12 @@ function M.update_file_info_buffer(file, bufnr, file_index)
   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 })
-  vim.api.nvim_set_option_value('wrap', false, { 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
@@ -770,11 +860,8 @@ function M.clear_buffer(bufnr)
   cleanup_file_operation()
   M.clear_preview_visual_state(bufnr)
 
-  pcall(vim.treesitter.stop, bufnr)
-
   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 })
+  attach_preview_highlighter(bufnr, '')
   vim.api.nvim_set_option_value('buftype', 'nofile', { buf = bufnr })
 
   set_buffer_lines(bufnr, {})
@@ -797,6 +884,8 @@ 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

lua/fff/file_renderer.lua

@@ -225,7 +225,7 @@ function M.apply_highlights(item, ctx, item_idx, buf, ns_id, line_idx, line_cont
 
   -- 9. Query match
   if ctx.query and ctx.query ~= '' then
-    local match_start, match_end = string.find(line_content, ctx.query, 1)
+    local match_start, match_end = string.find(line_content, ctx.query, 1, true)
     if match_start and match_end then
       vim.api.nvim_buf_set_extmark(
         buf,