.

Co-authored-by: Yuge Zhang <scottyugochang@gmail.com>

.dockerignore

@@ -0,0 +1,14 @@
+.venv
+**/.venv
+__pycache__
+.git
+.gitignore
+**/node_modules
+dist
+build
+.env
+docker
+.pytest_cache
+.vscode
+**/*.log
+examples/**/data

.github/workflows/badge-azure.yml

@@ -0,0 +1,29 @@
+name: Badge - Azure
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Azure
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-azure.yml', label: 'azure', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });

.github/workflows/badge-claude-code.yml

@@ -0,0 +1,29 @@
+name: Badge - Claude Code
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Claude Code
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-claude-code.yml', label: 'claude-code', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });

.github/workflows/badge-compat.yml

@@ -0,0 +1,29 @@
+name: Badge - Compatibility
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Backward Compatibility
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-compat.yml', label: 'examples-compat', variants: ['legacy', 'stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });

.github/workflows/badge-examples.yml

@@ -7,6 +7,9 @@ on:
       - Examples - Spider
       - Examples - APO
       - Examples - Unsloth
+      - Examples - Tinker
+      - Examples - Azure
+      - Examples - Claude Code
     types: [completed]
 
   workflow_dispatch:
@@ -31,5 +34,8 @@ jobs:
               { workflow: 'examples-spider.yml', label: 'examples-spider.stable', variants: ['stable'] },
               { workflow: 'examples-apo.yml', label: 'examples-apo.stable', variants: ['stable'] },
               { workflow: 'examples-unsloth.yml', label: 'examples-unsloth.stable', variants: ['stable'] },
+              { workflow: 'examples-tinker.yml', label: 'examples-tinker.stable', variants: ['stable'] },
+              { workflow: 'examples-azure.yml', label: 'examples-azure.stable', variants: ['stable'] },
+              { workflow: 'examples-claude-code.yml', label: 'examples-claude-code.stable', variants: ['stable'] },
             ];
             await badgeAggregation({ github, context, core, dependencies });

.github/workflows/badge-tinker.yml

@@ -0,0 +1,29 @@
+name: Badge - Tinker
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Tinker
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-tinker.yml', label: 'tinker', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });

.github/workflows/badge-unit.yml

@@ -0,0 +1,31 @@
+name: Badge - Unit Test
+
+on:
+  workflow_run:
+    workflows:
+      - CPU Test
+      - GPU Test
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'tests-full.yml', label: 'tests-full', variants: ['legacy', 'stable'] },
+              { workflow: 'tests.yml', label: 'tests', variants: ['legacy', 'stable', 'Lint', 'documentation', 'JavaScript'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });

.github/workflows/benchmark.yml

@@ -0,0 +1,187 @@
+name: Benchmark
+permissions:
+  contents: read
+on:
+  workflow_dispatch:
+
+jobs:
+  benchmark:
+    name: Benchmark (${{ matrix.backend.id }}, ${{ matrix.scenario.display }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-cpu]
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        backend:
+          - id: memory
+            compose_file: compose.prometheus-memory-store.yml
+          - id: mongo
+            compose_file: compose.prometheus-mongo-store.yml
+        scenario:
+          - id: minimal-production
+            display: Minimal production scale
+            store_workers: 4
+            args: >-
+              --mode batch
+              --total-tasks 4096
+              --batch-size 256
+              --n-runners 32
+              --max-rounds 6
+              --sleep-seconds 0.5
+          - id: medium-production
+            display: Medium production scale
+            store_workers: 16
+            args: >-
+              --mode batch
+              --total-tasks 10000
+              --batch-size 1000
+              --n-runners 100
+              --max-rounds 10
+              --sleep-seconds 0.1
+          - id: large-batch
+            display: Large batch waves
+            store_workers: 32
+            args: >-
+              --mode batch
+              --total-tasks 100000
+              --batch-size 8192
+              --n-runners 256
+              --max-rounds 6
+              --sleep-seconds 0.1
+          - id: long-queues
+            display: Long rollout queues
+            store_workers: 32
+            args: >-
+              --mode batch_partial
+              --total-tasks 100000
+              --batch-size 1024
+              --n-runners 256
+              --remaining-tasks 4096
+              --max-rounds 4
+              --sleep-seconds 0.1
+          - id: high-concurrency
+            display: High-throughput concurrent requests
+            store_workers: 32
+            args: >-
+              --mode single
+              --total-tasks 100000
+              --concurrency 2048
+              --n-runners 256
+              --max-rounds 2
+              --sleep-seconds 0.1
+          - id: heavy-traces
+            display: Heavy rollouts with deep traces
+            store_workers: 64
+            args: >-
+              --mode batch_partial
+              --total-tasks 10000
+              --batch-size 1024
+              --remaining-tasks 256
+              --n-runners 512
+              --max-rounds 20
+              --sleep-seconds 1.0
+    env:
+      STORE_URL: http://localhost:4747
+      STORE_API_URL: http://localhost:4747/v1/agl
+      PROM_URL: http://localhost:9090
+      SCENARIO_ID: ${{ matrix.scenario.id }}
+      BACKEND_ID: ${{ matrix.backend.id }}
+      ARTIFACT_DIR: artifacts/${{ matrix.scenario.id }}-${{ matrix.backend.id }}
+      COMPOSE_FILE: ${{ matrix.backend.compose_file }}
+      AGL_STORE_N_WORKERS: ${{ matrix.scenario.store_workers }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: '3.12'
+
+      - name: Sync dependencies
+        run: uv sync --frozen --extra mongo --group core-stable --group dev
+
+      - name: Check disk space
+        run: df -h
+
+      - name: Reset benchmark data directories
+        run: |
+          set -euo pipefail
+          cd docker
+          rm -rf data
+          bash setup.sh
+
+      - name: Launch ${{ matrix.backend.id }} Prometheus stack
+        run: |
+          set -euo pipefail
+          cd docker
+          docker compose -f "$COMPOSE_FILE" down -v || true
+          docker compose -f "$COMPOSE_FILE" up -d --quiet-pull
+
+      - name: Wait for store readiness
+        run: |
+          set -euo pipefail
+          for attempt in {1..60}; do
+            if curl -fsS "$STORE_API_URL/health" >/dev/null 2>&1; then
+              exit 0
+            fi
+            sleep 1
+          done
+          echo "Store did not become ready in time" >&2
+          exit 1
+
+      - name: Prepare artifact directory
+        run: mkdir -p "$ARTIFACT_DIR"
+
+      - name: Record benchmark start
+        run: echo "BENCHMARK_START=$(date -u +%FT%TZ)" >> "$GITHUB_ENV"
+
+      - name: Run ${{ matrix.scenario.display }} workload
+        run: |
+          set -euo pipefail
+          uv run --locked --no-sync python -m tests.benchmark.benchmark_store \
+            --store-url "$STORE_URL" \
+            ${{ matrix.scenario.args }}
+
+      - name: Record benchmark end
+        if: ${{ always() }}
+        run: echo "BENCHMARK_END=$(date -u +%FT%TZ)" >> "$GITHUB_ENV"
+
+      - name: Run benchmark analysis
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          mkdir -p "$ARTIFACT_DIR"
+          if [ -z "${BENCHMARK_START:-}" ] || [ -z "${BENCHMARK_END:-}" ]; then
+            echo "Analysis skipped: benchmark window not recorded." > "$ARTIFACT_DIR/analysis.txt"
+            exit 1
+          fi
+          uv run --locked --no-sync python -m tests.benchmark.analysis \
+            --prom-url "$PROM_URL" \
+            --store-url "$STORE_API_URL" \
+            --start "$BENCHMARK_START" \
+            --end "$BENCHMARK_END" \
+            | tee "$ARTIFACT_DIR/analysis.txt"
+
+      - name: Stop ${{ matrix.backend.id }} Prometheus stack
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          cd docker
+          docker compose -f "$COMPOSE_FILE" down -v || true
+
+      - name: Archive Prometheus metrics
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          mkdir -p "$ARTIFACT_DIR"
+          if [ -d docker/data/prometheus ]; then
+            tar -C docker/data -czf "$ARTIFACT_DIR/prometheus-${SCENARIO_ID}-${BACKEND_ID}.tar.gz" prometheus
+          fi
+
+      - name: Upload benchmark artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: benchmark-${{ matrix.scenario.id }}-${{ matrix.backend.id }}
+          path: ${{ env.ARTIFACT_DIR }}
+          if-no-files-found: error

.github/workflows/dashboard.yml

@@ -0,0 +1,33 @@
+name: Dashboard
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 5 AM UTC+8
+    - cron: '0 21 * * *'
+
+  workflow_dispatch:
+
+  push:
+    branches: [ main, stable/**/* ]
+
+jobs:
+  dashboard:
+    name: Chromatic
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Run Chromatic
+        uses: chromaui/action@v13
+        with:
+          projectToken: ${{ secrets.CHROMATIC_PROJECT_TOKEN }}
+          workingDir: dashboard
+          exitZeroOnChanges: false

.github/workflows/examples-apo.yml

@@ -14,7 +14,7 @@ on:
 run-name: >-
   ${{ github.event_name == 'repository_dispatch'
       && format(
-        'PR #{0} - Label {1} - {2}',
+        'APO - PR #{0} - {1} - {2}',
         github.event.client_payload.pull_number,
         github.event.client_payload.ci_label,
         github.event.client_payload.correlation_id

.github/workflows/examples-azure.yml

@@ -0,0 +1,98 @@
+name: Examples - Azure
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 4 AM UTC+8
+    - cron: '0 20 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-azure, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Azure - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Azure - {0}', github.event_name) }}
+
+jobs:
+  azure:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-azure' ||
+      github.event.action == 'ci-all'
+    name: Azure (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-cpu]
+    timeout-minutes: 400
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.12'
+          setup-script: 'stable'
+      fail-fast: false
+    steps:
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups \
+            --group dev --group experiment --group agents --group core-stable
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: dependencies-azure-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Azure Login
+        run: |
+          az login --identity
+        shell: bash
+
+      - name: Azure OpenAI Sanity Check
+        run: |
+          source .venv/bin/activate
+          cd examples/azure
+          python capital_agent.py
+        shell: bash
+        env:
+          AZURE_OPENAI_ENDPOINT: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          AZURE_OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+        id: azure_openai_sanity_check
+
+      - name: Azure OpenAI Supervised Fine-tuning
+        run: |
+          source .venv/bin/activate
+          cd examples/azure
+          python train_capital_agent.py --n-iterations 2 --cleanup
+        shell: bash
+        env:
+          AZURE_OPENAI_ENDPOINT: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          AZURE_OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          AZURE_SUBSCRIPTION_ID: ${{ secrets.AZURE_SUBSCRIPTION_ID }}
+          AZURE_OPENAI_API_VERSION: 2025-04-01-preview
+          AZURE_RESOURCE_GROUP: ${{ secrets.AZURE_RESOURCE_GROUP }}
+          AZURE_RESOURCE_NAME: ${{ secrets.AZURE_RESOURCE_NAME }}
+        id: azure_openai_finetune

.github/workflows/examples-calc-x.yml

@@ -14,7 +14,7 @@ on:
 run-name: >-
   ${{ github.event_name == 'repository_dispatch'
       && format(
-        'PR #{0} - Label {1} - {2}',
+        'Calc-X - PR #{0} - {1} - {2}',
         github.event.client_payload.pull_number,
         github.event.client_payload.ci_label,
         github.event.client_payload.correlation_id
@@ -22,12 +22,12 @@ run-name: >-
       || format('Calc-X - {0}', github.event_name) }}
 
 jobs:
-  calc-x:
+  calc-x-perf:
     if: >
       github.event_name != 'repository_dispatch' ||
       github.event.action == 'ci-calc-x' ||
       github.event.action == 'ci-all'
-    name: Calc-X (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    name: Calc-X Performance (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
     runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
     timeout-minutes: 90
     strategy:
@@ -74,7 +74,7 @@ jobs:
       - name: Upload dependencies artifact
         uses: actions/upload-artifact@v4
         with:
-          name: dependencies-calc-x-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          name: dependencies-calc-x-performance-${{ matrix.python-version }}-${{ matrix.setup-script }}
           path: requirements-freeze.txt
           compression-level: 0
 
@@ -116,13 +116,11 @@ jobs:
       # Don't ask why. Don't touch this.
       - name: Calc-X training
         run: |
-          set -ex
           source .venv/bin/activate
           cd examples/calc_x
           ../../scripts/restart_ray.sh
           sleep 5
-          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci
-          sleep 10
+          python train_calc_agent.py --val-file data/test_mini.parquet --ci
         shell: bash
         env:
           WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
@@ -137,14 +135,126 @@ jobs:
           WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
           WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
 
-      - name: Calc-X training LLM Proxy
+  calc-x-variants:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-calc-x' ||
+      github.event.action == 'ci-all'
+    name: Calc-X Variants (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 90
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group torch-gpu-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable & legacy)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group torch-gpu-${{ matrix.setup-script }}
+        if: matrix.setup-script != 'latest'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: dependencies-calc-x-variants-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Prepare Calc-X dataset
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run gdown --fuzzy https://drive.google.com/file/d/1FQMyKLLd6hP9dw9rfZn1EZOWNvKaDsqw/view
+          unzip calc-x-data.zip -d data
+          rm calc-x-data.zip
+
+      - name: Calc-X MCP sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run tests/test_mcp_calculator.py
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+      - name: Calc-X sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run legacy_calc_agent_debug.py
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+
+      - name: Training with local model
         run: |
           set -ex
           source .venv/bin/activate
           cd examples/calc_x
           ../../scripts/restart_ray.sh
           sleep 5
-          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci --llm-proxy
+          hf download Qwen/Qwen2.5-0.5B-Instruct --local-dir data/qwen_model
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --model $(realpath data/qwen_model)
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_local_model
+
+      - name: Validate training with local model
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_local_model.outputs.project_name }} ${{ steps.calc_x_train_local_model.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with LLM Proxy
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --llm-proxy
           sleep 10
         shell: bash
         env:
@@ -152,7 +262,15 @@ jobs:
           WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
         id: calc_x_train_llm_proxy
 
-      - name: Calc-X training with external store
+      - name: Validate training with LLM Proxy
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_llm_proxy.outputs.project_name }} ${{ steps.calc_x_train_llm_proxy.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with external store
         run: |
           set -euo pipefail
           source .venv/bin/activate
@@ -182,7 +300,15 @@ jobs:
           WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
         id: calc_x_train_external_store
 
-      - name: Calc-X training with role-based environment variables
+      - name: Validate training with external store
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_external_store.outputs.project_name }} ${{ steps.calc_x_train_external_store.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with role-based environment variables
         run: |
           set -euo pipefail
           source .venv/bin/activate
@@ -203,3 +329,12 @@ jobs:
         env:
           WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
           WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_role_based_env_var
+
+      - name: Validate training with role-based environment variables
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_role_based_env_var.outputs.project_name }} ${{ steps.calc_x_train_role_based_env_var.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}

.github/workflows/examples-claude-code.yml

@@ -0,0 +1,151 @@
+name: Examples - Claude Code
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 4 AM UTC+8
+    - cron: "0 20 * * *"
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-claude-code, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Claude Code - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Claude Code - {0}', github.event_name) }}
+
+jobs:
+  claude-code:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-claude-code' ||
+      github.event.action == 'ci-all'
+    name: Claude Code (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 60
+    strategy:
+      matrix:
+        include:
+          - python-version: "3.12"
+            setup-script: "stable"
+          - python-version: "3.13"
+            setup-script: "latest"
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups \
+            --group dev --group experiment --group agents --group torch-gpu-stable
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: dependencies-claude-code-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Download model
+        run: |
+          source .venv/bin/activate
+          python -c "from transformers import AutoModelForCausalLM; AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-Coder-30B-A3B-Instruct')"
+
+      - name: Launch vLLM server
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct \
+            --max-model-len 131072 \
+            --enable-auto-tool-choice \
+            --tool-call-parser qwen3_coder \
+            --port 45993 &
+
+          VLLM_READY=0
+          for i in {1..100}; do
+            if curl -sSf http://localhost:45993/v1/models > /dev/null 2>&1; then
+              echo "vLLM server is ready!"
+              VLLM_READY=1
+              break
+            fi
+            echo "Waiting for vLLM server to be ready... (${i})"
+            sleep 5
+          done
+          if [[ "$VLLM_READY" != "1" ]]; then
+            echo "vLLM server failed to start!"
+            exit 1
+          fi
+
+      - name: Claude Code sanity check with vLLM models
+        run: |
+          source .venv/bin/activate
+          cd examples/claude_code
+          python claude_code_agent.py vllm --backend-model-high Qwen/Qwen3-Coder-30B-A3B-Instruct --backend-model-low Qwen/Qwen3-Coder-30B-A3B-Instruct --base-url http://localhost:45993/v1 --debug
+        shell: bash
+
+      - name: Upload sanity check artifacts for vLLM
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: claude-code-sanity-check-vllm-${{ matrix.setup-script }}
+          path: |
+            examples/claude_code/data/
+            examples/claude_code/logs/
+          if-no-files-found: error
+
+      - name: Cleanup vLLM
+        run: |
+          set -euo pipefail
+          pkill -f vllm
+          for i in {1..60}; do
+            if ! pgrep -f vllm; then
+              break
+            fi
+            sleep 5
+          done
+          rm -rf examples/claude_code/data/
+          rm -rf examples/claude_code/logs/
+
+      - name: Claude Code sanity check with OpenAI models
+        run: |
+          source .venv/bin/activate
+          cd examples/claude_code
+          python claude_code_agent.py openai --backend-model-high gpt-5.1-codex-mini --backend-model-low gpt-4.1-mini --debug
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+
+      - name: Upload sanity check artifacts for OpenAI
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: claude-code-sanity-check-openai-${{ matrix.setup-script }}
+          path: |
+            examples/claude_code/data/
+            examples/claude_code/logs/
+          if-no-files-found: error

.github/workflows/examples-compat.yml

@@ -14,7 +14,7 @@ on:
 run-name: >-
   ${{ github.event_name == 'repository_dispatch'
       && format(
-        'PR #{0} - Label {1} - {2}',
+        'Backward Compatibility - PR #{0} - {1} - {2}',
         github.event.client_payload.pull_number,
         github.event.client_payload.ci_label,
         github.event.client_payload.correlation_id
@@ -54,6 +54,10 @@ jobs:
         run: |
           uv sync --frozen --no-default-groups --extra apo --extra verl \
             --group dev --group experiment --group agents --group torch-gpu-${{ matrix.setup-script }}
+      - name: Override VERL (stable)
+        run: |
+          uv pip install verl==0.5.0 vllm==0.10.2
+        if: matrix.setup-script == 'stable'
       - name: Freeze dependencies
         run: |
           set -ex

.github/workflows/examples-spider.yml

@@ -14,7 +14,7 @@ on:
 run-name: >-
   ${{ github.event_name == 'repository_dispatch'
       && format(
-        'PR #{0} - Label {1} - {2}',
+        'Spider - PR #{0} - {1} - {2}',
         github.event.client_payload.pull_number,
         github.event.client_payload.ci_label,
         github.event.client_payload.correlation_id
@@ -121,7 +121,7 @@ jobs:
       - name: Validate Spider training
         run: |
           set -ex
-          uv run scripts/validate_example_wandb.py ${{ steps.spider_train.outputs.project_name }} ${{ steps.spider_train.outputs.run_name }}
+          uv run scripts/validate_example_wandb.py ${{ steps.spider_train.outputs.project_name }} ${{ steps.spider_train.outputs.run_name }} --reward-tolerance 5
         env:
           WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
           WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}

.github/workflows/examples-tinker.yml

@@ -0,0 +1,160 @@
+name: Examples - Tinker
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 3 AM UTC+8
+    - cron: '0 19 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-tinker, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Tinker - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Tinker - {0}', github.event_name) }}
+
+jobs:
+  tinker:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-tinker' ||
+      github.event.action == 'ci-all'
+    name: Tinker (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-cpu]
+    timeout-minutes: 150
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups \
+            --group dev --group experiment --group agents --group torch-cpu --group core-stable --group tinker
+      - name: Freeze dependencies
+        run: |
+          set -euo pipefail
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: dependencies-tinker-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Tinker LLM sanity check
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          # TODO: Currently only test the client tracer implementation.
+          python -m tests.test_tinker_llm
+        shell: bash
+        env:
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Hello
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python hello.py oneclick --ci
+        shell: bash
+        env:
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Evaluate (GPT-4.1)
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          mkdir -p logs
+          python q20_evaluate.py --ci --model gpt-4.1 --output-file logs/q20_evaluate_gpt-4.1.jsonl
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Evaluate (Qwen3-30B-A3B-Instruct-2507)
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python q20_evaluate.py --ci --model Qwen/Qwen3-30B-A3B-Instruct-2507 --output-file logs/q20_evaluate_qwen3-30b-a3b.jsonl
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Training Dry Run
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python q20_train.py dryrun --model qwen4b
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Training
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          agl store --port 4747 &
+          sleep 5
+          python q20_train.py runner --n-runners 4 &
+          sleep 5
+          python q20_train.py algo --model qwen4b --ci
+          sleep 5
+
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+          pkill -f q20_train.py && echo "SIGTERM sent to q20_train.py" || echo "No q20_train.py process found"
+          while pgrep -f q20_train.py; do
+            echo "Waiting for q20_train.py to finish..."
+            sleep 5
+          done
+          echo "q20_train.py has finished."
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}

.github/workflows/examples-unsloth.yml

@@ -14,7 +14,7 @@ on:
 run-name: >-
   ${{ github.event_name == 'repository_dispatch'
       && format(
-        'PR #{0} - Label {1} - {2}',
+        'Unsloth - PR #{0} - {1} - {2}',
         github.event.client_payload.pull_number,
         github.event.client_payload.ci_label,
         github.event.client_payload.correlation_id

.github/workflows/pypi-nightly.yml

@@ -26,6 +26,14 @@ jobs:
       - name: Sync dependencies
         run: uv sync --frozen --no-default-groups --group dev
 
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
       - name: Get current version
         id: get_version
         run: |

.github/workflows/pypi-release.yml

@@ -60,6 +60,14 @@ jobs:
       - name: Sync dependencies
         run: uv sync --frozen --no-default-groups --group dev
 
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
       - name: Build package
         run: |
           uv build

.github/workflows/tests-full.yml

@@ -14,7 +14,7 @@ on:
 run-name: >-
   ${{ github.event_name == 'repository_dispatch'
       && format(
-        'PR #{0} - Label {1} - {2}',
+        'GPU Test - PR #{0} - {1} - {2}',
         github.event.client_payload.pull_number,
         github.event.client_payload.ci_label,
         github.event.client_payload.correlation_id
@@ -46,7 +46,130 @@ jobs:
         run: nvidia-smi
       - uses: actions/checkout@v4
         with:
-          ref: ${{ github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number) || github.ref }}
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: uv sync --frozen --no-default-groups --extra apo --extra mongo --group dev --group agents --group torch-gpu-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable & legacy)
+        run: uv sync --frozen --no-default-groups --extra apo --extra mongo --group dev --group agents --group torch-gpu-${{ matrix.setup-script }}
+        if: matrix.setup-script != 'latest'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: dependencies-tests-full-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
+      - name: Setup Docker environments
+        run: |
+          set -euo pipefail
+
+          cd docker
+
+          # Setup data directories
+          ./setup.sh
+
+          # Start Dockers
+          docker compose -f compose.mongo.yml up -d
+
+          SERVICE_NAME=mongo
+          TIMEOUT=60   # seconds
+          SLEEP=2
+
+          cid="$(docker compose -f compose.mongo.yml ps -q "$SERVICE_NAME")"
+          if [ -z "$cid" ]; then
+            echo "Service $SERVICE_NAME is not running"
+            exit 1
+          fi
+
+          echo "Waiting for $SERVICE_NAME to become healthy..."
+          end=$((SECONDS + TIMEOUT))
+
+          while [ "$SECONDS" -lt "$end" ]; do
+            status="$(docker inspect -f '{{.State.Health.Status}}' "$cid")"
+            echo "Current status: $status"
+
+            if [ "$status" = "healthy" ]; then
+              echo "$SERVICE_NAME is healthy ✅"
+              exit 0
+            elif [ "$status" = "unhealthy" ]; then
+              echo "$SERVICE_NAME is unhealthy ❌"
+              docker logs "$cid" || true
+              exit 1
+            fi
+
+            sleep "$SLEEP"
+          done
+
+          echo "Timed out waiting for $SERVICE_NAME to become healthy after ${TIMEOUT}s"
+          docker logs "$cid" || true
+          exit 1
+        shell: bash
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Run tests
+        run: |
+          uv run pytest -v --durations=0 tests
+        env:
+          PYTEST_ADDOPTS: "--color=yes"
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+          AGL_TEST_MONGO_URI: mongodb://localhost:27017/?replicaSet=rs0
+
+
+  minimal-examples:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-gpu' ||
+      github.event.action == 'ci-all'
+    name: Minimal Examples with Python ${{ matrix.python-version }} (${{ matrix.setup-script }})
+
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 30
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
       - uses: astral-sh/setup-uv@v7
         with:
           enable-cache: true
@@ -69,7 +192,7 @@ jobs:
       - name: Upload dependencies artifact
         uses: actions/upload-artifact@v4
         with:
-          name: dependencies-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          name: dependencies-minimal-examples-${{ matrix.python-version }}-${{ matrix.setup-script }}
           path: requirements-freeze.txt
           compression-level: 0
 
@@ -80,10 +203,120 @@ jobs:
           AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
           AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
 
-      - name: Run tests
+      - name: Write Traces via Otel Tracer
         run: |
-          uv run pytest -v --durations=0 tests
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python write_traces.py otel
+          sleep 5
+
+      - name: Write Traces via AgentOps Tracer
         env:
-          PYTEST_ADDOPTS: "--color=yes"
           OPENAI_BASE_URL: http://localhost:12306/
           OPENAI_API_KEY: dummy
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python write_traces.py agentops
+          sleep 5
+
+      - name: Write Traces via Otel Tracer with Client
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          agl store --port 45993 --log-level DEBUG &
+          sleep 5
+          python write_traces.py otel --use-client
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+
+      - name: Write Traces via AgentOps Tracer with Client
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          agl store --port 45993 --log-level DEBUG &
+          sleep 5
+          python write_traces.py agentops --use-client
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+
+      - name: vLLM Server
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python vllm_server.py Qwen/Qwen2.5-0.5B-Instruct
+
+      - name: LLM Proxy (OpenAI backend)
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+
+          python llm_proxy.py openai gpt-4.1-mini &
+
+          LLM_PROXY_READY=0
+          for attempt in $(seq 1 30); do
+            if curl -sSf http://localhost:43886/health > /dev/null 2>&1; then
+              LLM_PROXY_READY=1
+              break
+            fi
+            sleep 2
+          done
+          if [[ "$LLM_PROXY_READY" != "1" ]]; then
+            echo "LLM proxy failed to become healthy" >&2
+            exit 1
+          fi
+
+          python llm_proxy.py test gpt-4.1-mini
+
+          pkill -f llm_proxy.py && echo "SIGTERM sent to llm_proxy.py" || echo "No llm_proxy.py process found"
+          while pgrep -f llm_proxy.py; do
+            echo "Waiting for llm_proxy.py to finish..."
+            sleep 5
+          done
+
+      - name: LLM Proxy (vLLM backend)
+        if: matrix.setup-script != 'legacy'  # Skip if return_token_ids is not supported
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python llm_proxy.py vllm Qwen/Qwen2.5-0.5B-Instruct &
+
+          LLM_PROXY_READY=0
+          for attempt in $(seq 1 30); do
+            if curl -sSf http://localhost:43886/health > /dev/null 2>&1; then
+              LLM_PROXY_READY=1
+              break
+            fi
+            sleep 2
+          done
+          if [[ "$LLM_PROXY_READY" != "1" ]]; then
+            echo "LLM proxy failed to become healthy" >&2
+            exit 1
+          fi
+
+          python llm_proxy.py test Qwen/Qwen2.5-0.5B-Instruct
+
+          pkill -f llm_proxy.py && echo "SIGTERM sent to llm_proxy.py" || echo "No llm_proxy.py process found"
+          while pgrep -f llm_proxy.py; do
+            echo "Waiting for llm_proxy.py to finish..."
+            sleep 5
+          done

.github/workflows/tests.yml

@@ -37,6 +37,7 @@ jobs:
           uv sync --frozen \
             --extra apo \
             --extra verl \
+            --extra mongo \
             --group dev \
             --group torch-cpu \
             --group torch-stable \
@@ -45,10 +46,11 @@ jobs:
             --group agents \
             --no-default-groups
         if: matrix.setup == 'slow'
+      # This pre-commit skips JavaScript on purpose.
       - name: Run pre-commit
         uses: pre-commit/action@v3.0.1
       - name: Check Python headers
-        run: uv run --locked --no-sync scripts/check_python_headers.py
+        run: uv run --locked --no-sync scripts/check_headers.py
       - name: Run Black
         run: uv run --locked --no-sync black --check .
       - name: Run isort
@@ -60,6 +62,28 @@ jobs:
         run: uv run --locked --no-sync pyright -p pyrightconfig.json
         if: matrix.setup == 'slow'
 
+  lint-js:
+    name: Lint - JavaScript
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install dependencies
+        run: cd dashboard && npm ci
+      - name: Run ESLint
+        run: cd dashboard && npm run eslint
+      - name: Run Prettier
+        run: cd dashboard && npm run prettier
+      - name: Run Stylelint
+        run: cd dashboard && npm run stylelint
+      - name: Run Typecheck
+        run: cd dashboard && npm run typecheck
+      - name: Verify build
+        run: cd dashboard && npm run build
+
   docs:
     name: Build documentation
     runs-on: ubuntu-latest
@@ -132,8 +156,39 @@ jobs:
           name: dependencies-${{ matrix.python-version }}-${{ matrix.setup-script }}
           path: requirements-freeze.txt
           compression-level: 0
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
       - name: Run tests
         run: |
-          uv run pytest -v --durations=0 tests
+          uv run pytest -v --durations=0 tests -m "not mongo"
         env:
           PYTEST_ADDOPTS: "--color=yes"
+
+  test-js:
+    name: Test - JavaScript
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: '3.12'
+      - name: Sync Python dependencies
+        run: uv sync --frozen --no-default-groups --extra apo --group dev --group agents --group core-stable
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Run vitest
+        run: cd dashboard && npm run vitest

.gitignore

@@ -207,3 +207,12 @@ cython_debug/
 
 # Claude
 .claude/*.local.json
+
+# Dashboard generated files
+agentlightning/dashboard/**/*.css
+agentlightning/dashboard/**/*.js
+agentlightning/dashboard/**/*.html
+agentlightning/dashboard/**/*.svg
+
+# Docker data
+docker/data/

.pre-commit-config.yaml

@@ -24,3 +24,53 @@ repos:
         pass_filenames: false
         always_run: true
         args: ["."]
+
+  - repo: local
+    hooks:
+    - id: prettier
+      name: prettier (dashboard)
+      language: system
+      pass_filenames: false
+      always_run: true
+      entry: >
+        bash -c '
+          cd dashboard || exit 1
+          if [ -d node_modules ]; then
+            echo "✅ node_modules already exists"
+            npx prettier --cache --write "**/*.{ts,tsx,mjs,cjs}"
+          else
+            echo "⚠️ node_modules not found — npx is not reliable. Skipping."
+          fi
+        '
+
+    - id: eslint
+      name: eslint (dashboard)
+      language: system
+      pass_filenames: false
+      always_run: true
+      entry: >
+        bash -c '
+          cd dashboard || exit 1
+          if [ -d node_modules ]; then
+            echo "✅ node_modules already exists"
+            npx eslint --cache --fix .
+          else
+            echo "⚠️ node_modules not found — npx is not reliable. Skipping."
+          fi
+        '
+
+    - id: stylelint
+      name: stylelint (dashboard)
+      language: system
+      pass_filenames: false
+      always_run: true
+      entry: >
+        bash -c '
+          cd dashboard || exit 1
+          if [ -d node_modules ]; then
+            echo "✅ node_modules already exists"
+            npx stylelint --cache --fix "**/*.css"
+          else
+            echo "⚠️ node_modules not found — npx is not reliable. Skipping."
+          fi
+        '

README.md

@@ -4,10 +4,11 @@
 
 # Agent Lightning⚡
 
-[![Test](https://github.com/microsoft/agent-lightning/actions/workflows/tests-full.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/tests-full.yml)
+[![Unit Tests](https://github.com/microsoft/agent-lightning/actions/workflows/badge-unit.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-unit.yml)
 [![Documentation](https://img.shields.io/badge/GitHub%20Pages-Documentation-blue)](https://microsoft.github.io/agent-lightning/)
 [![PyPI version](https://badge.fury.io/py/agentlightning.svg)](https://badge.fury.io/py/agentlightning)
 [![License](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+[![Ask DeepWiki](https://deepwiki.com/badge.svg)](https://deepwiki.com/microsoft/agent-lightning)
 [![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/RYk7CdvDR7)
 
 **The absolute trainer to light up AI agents.**
@@ -33,12 +34,19 @@ Read more on our [documentation website](https://microsoft.github.io/agent-light
 pip install agentlightning
 ```
 
+For the latest nightly build (cutting-edge features), you can install from Test PyPI:
+
+```bash
+pip install --upgrade --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ agentlightning
+```
+
 Please refer to our [installation guide](https://microsoft.github.io/agent-lightning/stable/tutorials/installation/) for more details.
 
 To start using Agent-lightning, check out our [documentation](https://microsoft.github.io/agent-lightning/) and [examples](./examples).
 
 ## ⚡ Articles
 
+- 11/4/2025 [Tuning ANY AI agent with Tinker ✕ Agent-lightning](https://medium.com/@yugez/tuning-any-ai-agent-with-tinker-agent-lightning-part-1-1d8c9a397f0e) Medium. See also [Part 2](https://medium.com/@yugez/tuning-any-ai-agent-with-tinker-agent-lightning-part-2-332c5437f0dc).
 - 10/22/2025 [No More Retokenization Drift: Returning Token IDs via the OpenAI Compatible API Matters in Agent RL](https://blog.vllm.ai/2025/10/22/agent-lightning.html) vLLM blog. See also [Zhihu writeup](https://zhuanlan.zhihu.com/p/1965067274642785725).
 - 8/11/2025 [Training AI Agents to Write and Self-correct SQL with Reinforcement Learning](https://medium.com/@yugez/training-ai-agents-to-write-and-self-correct-sql-with-reinforcement-learning-571ed31281ad) Medium.
 - 8/5/2025 [Agent Lightning: Train ANY AI Agents with Reinforcement Learning](https://arxiv.org/abs/2508.03680) arXiv paper.
@@ -67,10 +75,11 @@ No rewrites, no lock-in, just a clear path from first rollout to steady improvem
 | Workflow | Status |
 |----------|--------|
 | CPU Tests | [![tests workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/tests.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/tests.yml) |
-| GPU Tests | [![tests-full workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/tests-full.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/tests-full.yml) |
+| Full Tests | [![tests summary workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-unit.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-unit.yml) |
+| UI Tests | [![UI Tests](https://github.com/microsoft/agent-lightning/actions/workflows/dashboard.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/dashboard.yml) |
 | Examples Integration | [![examples summary workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-examples.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-examples.yml) |
 | Latest Dependency Compatibility | [![latest summary workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-latest.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-latest.yml) |
-| Legacy Examples Compatibility | [![examples compatibility workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/examples-compat.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/examples-compat.yml) |
+| Legacy Examples Compatibility | [![compat summary workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-compat.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-compat.yml) |
 
 ## ⚡ Citation
 
@@ -90,7 +99,7 @@ If you find Agent Lightning useful in your research or projects, please cite our
 
 ## ⚡ Contributing
 
-This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
+This project welcomes contributions and suggestions. Start by reading the [Contributing Guide](docs/community/contributing.md) for recommended contribution points, environment setup, branching conventions, and pull request expectations. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
 
 When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
 

agentlightning/__init__.py

@@ -1,16 +1,19 @@
 # Copyright (c) Microsoft. All rights reserved.
 
-__version__ = "0.2.1"
+__version__ = "0.3.0"
 
 from .adapter import *
 from .algorithm import *
 from .client import AgentLightningClient, DevTaskLoader  # deprecated  # type: ignore
 from .config import *
 from .emitter import *
+from .env_var import *
 from .execution import *
 from .litagent import *
 from .llm_proxy import *
-from .logging import *
+from .logging import configure_logger  # deprecated  # type: ignore
+from .logging import setup as setup_logging  # type: ignore
+from .logging import setup_module as setup_module_logging  # type: ignore
 from .runner import *
 from .server import AgentLightningServer  # deprecated  # type: ignore
 from .store import *

agentlightning/adapter/base.py

@@ -1,6 +1,6 @@
 # Copyright (c) Microsoft. All rights reserved.
 
-from typing import Generic, List, TypeVar
+from typing import Generic, Sequence, TypeVar
 
 from opentelemetry.sdk.trace import ReadableSpan
 
@@ -66,7 +66,7 @@ class Adapter(Generic[T_from, T_to]):
         raise NotImplementedError("Adapter.adapt() is not implemented")
 
 
-class OtelTraceAdapter(Adapter[List[ReadableSpan], T_to], Generic[T_to]):
+class OtelTraceAdapter(Adapter[Sequence[ReadableSpan], T_to], Generic[T_to]):
     """Base class for adapters that convert OpenTelemetry trace spans into other formats.
 
     This specialization of [`Adapter`][agentlightning.Adapter] expects a list of
@@ -84,7 +84,7 @@ class OtelTraceAdapter(Adapter[List[ReadableSpan], T_to], Generic[T_to]):
     """
 
 
-class TraceAdapter(Adapter[List[Span], T_to], Generic[T_to]):
+class TraceAdapter(Adapter[Sequence[Span], T_to], Generic[T_to]):
     """Base class for adapters that convert trace spans into other formats.
 
     This class specializes [`Adapter`][agentlightning.Adapter] for working with

agentlightning/adapter/messages.py

@@ -4,7 +4,7 @@ from __future__ import annotations
 
 import json
 from collections import defaultdict
-from typing import TYPE_CHECKING, Any, Dict, Generator, Iterable, List, Optional, TypedDict, Union, cast
+from typing import TYPE_CHECKING, Any, Dict, Generator, Iterable, List, Optional, Sequence, TypedDict, Union, cast
 
 from pydantic import TypeAdapter
 
@@ -208,7 +208,7 @@ class TraceToMessages(TraceAdapter[List[OpenAIMessages]]):
         children of the associated completion span.
     """
 
-    def get_tool_calls(self, completion: Span, all_spans: List[Span], /) -> Iterable[Dict[str, Any]]:
+    def get_tool_calls(self, completion: Span, all_spans: Sequence[Span], /) -> Iterable[Dict[str, Any]]:
         """Yield tool call payloads for a completion span.
 
         Args:
@@ -231,7 +231,7 @@ class TraceToMessages(TraceAdapter[List[OpenAIMessages]]):
             if tool_call:
                 yield tool_call
 
-    def adapt(self, source: List[Span], /) -> List[OpenAIMessages]:
+    def adapt(self, source: Sequence[Span], /) -> List[OpenAIMessages]:
         """Transform trace spans into OpenAI chat payloads.
 
         Args:

agentlightning/adapter/triplet.py

@@ -6,12 +6,13 @@ import json
 import logging
 import re
 from enum import Enum
-from typing import Any, Dict, List, Optional, Tuple, Union, cast
+from typing import Any, Dict, List, Optional, Sequence, Tuple, Union, cast
 
 from opentelemetry.sdk.trace import ReadableSpan
 from pydantic import BaseModel
 
-from agentlightning.types import Span, SpanNames, Triplet
+from agentlightning.emitter.reward import get_reward_value
+from agentlightning.types import Span, Triplet
 
 from .base import TraceAdapter
 
@@ -313,24 +314,11 @@ class TraceTree:
         Returns:
             Dictionary containing reward metadata, or an empty dictionary when no reward is found.
         """
-        for key in [
-            "agentops.task.output",  # newer versions of agentops
-            "agentops.entity.output",
-        ]:
-            output = self.span.attributes.get(key)  # type: ignore
-            if output:
-                if isinstance(output, dict):
-                    return output
-                elif isinstance(output, str):
-                    try:
-                        return json.loads(output)
-                    except json.JSONDecodeError:
-                        return {}
-
-        # Latest emit reward format
-        if self.span.name == SpanNames.REWARD.value and self.span.attributes:
-            return {"type": "reward", "value": self.span.attributes.get("reward", None)}
-        return {}
+        reward_value = get_reward_value(self.span)
+        if reward_value is not None:
+            return {"type": "reward", "value": reward_value}
+        else:
+            return {}
 
     def is_reward_span(self) -> bool:
         """Return whether the span explicitly encodes a reward.
@@ -429,7 +417,16 @@ class TraceTree:
         If we don't, when we want to select the LLM completion span with agent as filter.
         We will never get the correct span underneath.
         """
+        # If the current node has only one child, recursively repair its hierarchy directly.
+        # This special-case handling is needed because when a trace is manually ended
+        # (via agentops.end_trace), the AgentOps provider automatically wraps all spans
+        # under an extra synthetic root node (e.g., "run_one.session").
+        if len(self.children) == 1:
+            self.children[0].repair_hierarchy()
+            return
+
         nodes_to_repair = list(self.children)
+
         for repair_node in nodes_to_repair:
             if len(self.children) == 1:
                 # If there is only one child, we don't need to repair the hierarchy.
@@ -661,7 +658,7 @@ class TracerTraceToTriplet(TraceToTripletBase):
         trace_tree.visualize(filename, interested_span_match=interested_span_match)
         return trace_tree
 
-    def adapt(self, source: Union[List[Span], List[ReadableSpan]], /) -> List[Triplet]:  # type: ignore
+    def adapt(self, source: Union[Sequence[Span], Sequence[ReadableSpan]], /) -> List[Triplet]:  # type: ignore
         """Convert tracer spans into [`Triplet`][agentlightning.Triplet] trajectories.
 
         Args:
@@ -767,31 +764,14 @@ class LlmProxyTraceToTriplet(TraceToTripletBase):
 
     def _maybe_reward_value(self, span: Span) -> Optional[float]:
         """Parse reward from typical AgentOps payloads or explicit reward spans."""
-        attrs = span.attributes or {}
-
-        # AgentOps new/old keys
-        for k in ("agentops.task.output", "agentops.entity.output"):
-            v = attrs.get(k)
-            v = self._literal_eval_maybe(v)
-            if isinstance(v, dict) and cast(Dict[str, Any], v).get("type") == "reward":
-                rv = cast(Dict[str, Any], v).get("value", None)
-                if rv is None or isinstance(rv, (int, float)):
-                    return None if rv is None else float(rv)
-
-        # Explicit reward span
-        if span.name == SpanNames.REWARD.value:
-            rv = attrs.get("reward", None)
-            if rv is None or isinstance(rv, (int, float)):
-                return None if rv is None else float(rv)
-
-        return None
+        return get_reward_value(span)
 
     def _request_id_from_attrs(self, attrs: Dict[str, Any]) -> Optional[str]:
         # Prefer OpenAI-like id if present, else proxy raw id.
         rid = attrs.get("gen_ai.response.id") or attrs.get("llm.hosted_vllm.id")
         return str(rid) if isinstance(rid, str) and rid else None
 
-    def adapt(self, source: List[Span], /) -> List[Triplet]:  # type: ignore
+    def adapt(self, source: Sequence[Span], /) -> List[Triplet]:  # type: ignore
         """Convert LLM Proxy spans into [`Triplet`][agentlightning.Triplet] trajectories.
 
         Args:

agentlightning/algorithm/apo/apo.py

@@ -20,6 +20,7 @@ from openai import AsyncOpenAI
 
 from agentlightning.adapter.messages import TraceToMessages
 from agentlightning.algorithm.base import Algorithm
+from agentlightning.algorithm.utils import batch_iter_over_dataset
 from agentlightning.reward import find_final_reward
 from agentlightning.types import Dataset, NamedResources, PromptTemplate, Rollout, RolloutMode, RolloutStatus
 
@@ -56,41 +57,6 @@ APPLY_EDIT_PROMPT_FILES = [
 ]
 
 
-def batch_iter_over_dataset(dataset: Dataset[T_task], batch_size: int) -> Iterator[Sequence[T_task]]:
-    """
-    Create an infinite iterator that yields batches from the dataset.
-
-    When batch_size >= dataset size, yields the entire shuffled dataset repeatedly.
-    When batch_size < dataset size, yields batches of the specified size, reshuffling
-    after each complete pass through the dataset.
-
-    Args:
-        dataset: The dataset to iterate over.
-        batch_size: The desired batch size.
-
-    Yields:
-        Sequences of tasks from the dataset. Each task appears at most once per epoch.
-    """
-    if batch_size >= len(dataset):
-        while True:
-            dataset_copy = [dataset[i] for i in range(len(dataset))]
-            random.shuffle(dataset_copy)
-            yield dataset_copy
-
-    else:
-        current_batch: List[int] = []
-        while True:
-            indices = list(range(len(dataset)))
-            random.shuffle(indices)
-            for index in indices:
-                if index in current_batch:
-                    continue
-                current_batch.append(index)
-                if len(current_batch) == batch_size:
-                    yield [dataset[index] for index in current_batch]
-                    current_batch = []
-
-
 class APO(Algorithm, Generic[T_task]):
     """Automatic Prompt Optimization (APO) algorithm using textual gradients and beam search.
 

agentlightning/algorithm/fast.py

@@ -143,7 +143,7 @@ class Baseline(FastAlgorithm):
         store = self.get_store()
 
         for index in train_indices + val_indices:
-            queuing_rollouts = await store.query_rollouts(status=["queuing", "requeuing"])
+            queuing_rollouts = await store.query_rollouts(status_in=["queuing", "requeuing"])
             if len(queuing_rollouts) <= 1:
                 # Only enqueue a new rollout when there is at most 1 rollout in the queue.
                 sample = dataset[index]
@@ -222,7 +222,7 @@ class Baseline(FastAlgorithm):
                     f"Processing index {index}. {len(train_indices)} train indices and {len(val_indices)} val indices in total."
                 )
                 while True:
-                    queuing_rollouts = await store.query_rollouts(status=["queuing", "requeuing"])
+                    queuing_rollouts = await store.query_rollouts(status_in=["queuing", "requeuing"])
                     if len(queuing_rollouts) <= self.max_queue_length:
                         # Only enqueue a new rollout when there is at most "max_queue_length" rollout in the queue.
                         sample = concatenated_dataset[index]

agentlightning/algorithm/utils.py

@@ -0,0 +1,43 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+import random
+from typing import Iterator, List, Sequence, TypeVar
+
+from agentlightning.types import Dataset
+
+T_task = TypeVar("T_task")
+
+
+def batch_iter_over_dataset(dataset: Dataset[T_task], batch_size: int) -> Iterator[Sequence[T_task]]:
+    """
+    Create an infinite iterator that yields batches from the dataset.
+
+    When batch_size >= dataset size, yields the entire shuffled dataset repeatedly.
+    When batch_size < dataset size, yields batches of the specified size, reshuffling
+    after each complete pass through the dataset.
+
+    Args:
+        dataset: The dataset to iterate over.
+        batch_size: The desired batch size.
+
+    Yields:
+        Sequences of tasks from the dataset. Each task appears at most once per epoch.
+    """
+    if batch_size >= len(dataset):
+        while True:
+            dataset_copy = [dataset[i] for i in range(len(dataset))]
+            random.shuffle(dataset_copy)
+            yield dataset_copy
+
+    else:
+        current_batch: List[int] = []
+        while True:
+            indices = list(range(len(dataset)))
+            random.shuffle(indices)
+            for index in indices:
+                if index in current_batch:
+                    continue
+                current_batch.append(index)
+                if len(current_batch) == batch_size:
+                    yield [dataset[index] for index in current_batch]
+                    current_batch = []

agentlightning/algorithm/verl/interface.py

@@ -99,6 +99,8 @@ class VERL(Algorithm):
 
         # Merge your dict overrides
         override_conf = OmegaConf.create(config)
+        # Allow adding new fields
+        OmegaConf.set_struct(base_cfg, False)
         self.config = OmegaConf.merge(base_cfg, override_conf)
 
     def run(

agentlightning/cli/agentops_server.py

@@ -1,30 +0,0 @@
-# Copyright (c) Microsoft. All rights reserved.
-
-from __future__ import annotations
-
-import argparse
-import time
-from typing import Iterable
-
-from agentlightning.instrumentation.agentops import AgentOpsServerManager
-
-
-def main(argv: Iterable[str] | None = None) -> int:
-    parser = argparse.ArgumentParser(description="Start AgentOps server")
-    parser.add_argument("--daemon", action="store_true", help="Run server as a daemon")
-    parser.add_argument("--port", type=int, default=8002, help="Port to run the server on")
-    args = parser.parse_args(list(argv) if argv is not None else None)
-
-    manager = AgentOpsServerManager(daemon=args.daemon, port=args.port)
-    try:
-        manager.start()
-        # Wait forever
-        while True:
-            time.sleep(1)
-    except KeyboardInterrupt:
-        manager.stop()
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())

agentlightning/cli/store.py

@@ -9,7 +9,7 @@ import asyncio
 import logging
 from typing import Iterable
 
-from agentlightning.logging import configure_logger
+from agentlightning import setup_logging
 from agentlightning.store.client_server import LightningStoreServer
 from agentlightning.store.memory import InMemoryLightningStore
 
@@ -18,13 +18,75 @@ logger = logging.getLogger(__name__)
 
 def main(argv: Iterable[str] | None = None) -> int:
     parser = argparse.ArgumentParser(description="Run a LightningStore server")
+    parser.add_argument("--host", default="0.0.0.0", help="Host to bind the server to")
     parser.add_argument("--port", type=int, default=4747, help="Port to run the server on")
+    parser.add_argument(
+        "--cors-origin",
+        dest="cors_origins",
+        action="append",
+        help="Allowed CORS origin. Repeat for multiple origins. Use '*' to allow all origins.",
+    )
+    parser.add_argument(
+        "--log-level",
+        default="INFO",
+        choices=["DEBUG", "INFO", "WARNING", "ERROR"],
+        help="Configure the logging level for the store.",
+    )
+    parser.add_argument(
+        "--prometheus",
+        action="store_true",
+        help="Enable Prometheus metrics.",
+    )
+    parser.add_argument(
+        "--n-workers",
+        default=1,
+        type=int,
+        help=(
+            "Number of workers to run in the server. When it's greater than 1, the server will be run using `mp` launch mode. "
+            "Only applicable for zero-copy stores such as MongoDB backend."
+        ),
+    )
+
+    parser.add_argument(
+        "--backend",
+        choices=["memory", "mongo"],
+        default="memory",
+        help="Backend to use for the store.",
+    )
+    parser.add_argument(
+        "--mongo-uri",
+        default="mongodb://localhost:27017/?replicaSet=rs0",
+        help="MongoDB URI to use for the store. Applicable only if --backend is 'mongo'.",
+    )
+
     args = parser.parse_args(list(argv) if argv is not None else None)
 
-    configure_logger()
+    setup_logging(args.log_level)
 
-    store = InMemoryLightningStore()
-    server = LightningStoreServer(store, host="0.0.0.0", port=args.port)
+    if args.backend == "memory":
+        store = InMemoryLightningStore(prometheus=args.prometheus)
+    elif args.backend == "mongo":
+        from agentlightning.store.mongo import MongoLightningStore
+
+        store = MongoLightningStore(client=args.mongo_uri, prometheus=args.prometheus)
+    else:
+        raise ValueError(f"Invalid backend: {args.backend}")
+
+    if args.n_workers > 1:
+        logger.info(f"Running the server using `mp` launch mode with {args.n_workers} workers.")
+        launch_mode = "mp"
+    else:
+        logger.info("Running the server using `asyncio` launch mode.")
+        launch_mode = "asyncio"
+    server = LightningStoreServer(
+        store,
+        host=args.host,
+        port=args.port,
+        cors_allow_origins=args.cors_origins,
+        launch_mode=launch_mode,
+        prometheus=args.prometheus,
+        n_workers=args.n_workers,
+    )
     try:
         asyncio.run(server.run_forever())
     except RuntimeError as exc:

agentlightning/client.py

@@ -332,7 +332,9 @@ class DevTaskLoader(AgentLightningClient):
         if isinstance(resources, ResourcesUpdate):
             self._resources_update = resources
         else:
-            self._resources_update = ResourcesUpdate(resources_id="local", resources=resources)
+            self._resources_update = ResourcesUpdate(
+                resources_id="local", resources=resources, create_time=time.time(), update_time=time.time(), version=1
+            )
 
         # Store rollouts posted back to the loader for easy debugging of local runs
         self._rollouts: List[RolloutLegacy] = []

agentlightning/emitter/__init__.py

@@ -1,25 +1,32 @@
 # Copyright (c) Microsoft. All rights reserved.
 
+from .annotation import emit_annotation, operation
 from .exception import emit_exception
-from .message import emit_message
-from .object import emit_object
+from .message import emit_message, get_message_value
+from .object import emit_object, get_object_value
 from .reward import (
     emit_reward,
     find_final_reward,
     find_reward_spans,
     get_reward_value,
+    get_rewards_from_span,
     is_reward_span,
     reward,
 )
 
 __all__ = [
     "reward",
+    "operation",
     "emit_reward",
     "get_reward_value",
+    "get_rewards_from_span",
     "is_reward_span",
     "find_reward_spans",
     "find_final_reward",
     "emit_message",
     "emit_object",
     "emit_exception",
+    "emit_annotation",
+    "get_message_value",
+    "get_object_value",
 ]

agentlightning/emitter/annotation.py

@@ -0,0 +1,364 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Helpers for emitting annotation/operation spans."""
+
+import asyncio
+import functools
+import inspect
+import json
+import logging
+from types import TracebackType
+from typing import (
+    Any,
+    Callable,
+    ContextManager,
+    Dict,
+    Optional,
+    Tuple,
+    Type,
+    TypeVar,
+    Union,
+    cast,
+    overload,
+)
+
+from opentelemetry import trace
+from opentelemetry.sdk.trace import ReadableSpan
+from opentelemetry.trace import Status, StatusCode
+
+from agentlightning.semconv import AGL_ANNOTATION, AGL_OPERATION, LightningSpanAttributes
+from agentlightning.utils.otel import flatten_attributes, get_tracer
+
+_FnType = TypeVar("_FnType", bound=Callable[..., Any])
+
+logger = logging.getLogger(__name__)
+
+
+def emit_annotation(annotation: Dict[str, Any], propagate: bool = True) -> ReadableSpan:
+    """Emit a new annotation span.
+
+    This is the underlying implementation of [`emit_reward`][agentlightning.emit_reward].
+
+    Annotation spans are used to annotate a specific event or a part of rollout.
+    See [semconv][agentlightning.semconv] for conventional annotation keys in Agent-lightning.
+
+    If annotations contain nested dicts, they will be flattened before emitting.
+    Complex objects will lead to emitting failures.
+
+    Args:
+        annotation: Dictionary containing annotation key-value pairs.
+            Representatives are rewards, tags, and metadata.
+        propagate: Whether to propagate the span to exporters automatically.
+    """
+    annotation_attributes = flatten_attributes(annotation)
+    if any(not isinstance(v, (str, int, float, bool, bytes)) for v in annotation_attributes.values()):
+        raise TypeError("All annotation attributes must be primitive types (str, int, float, bool, bytes)")
+
+    # TODO: this should use a tracer from current context rather than the singleton
+    tracer = get_tracer(use_active_span_processor=propagate)
+    span = tracer.start_span(
+        AGL_ANNOTATION,
+        attributes=annotation_attributes,
+    )
+    logger.debug("Emitting annotation span with keys %s", annotation_attributes)
+    with span:
+        pass
+    if not isinstance(span, ReadableSpan):
+        raise ValueError(f"Span is not a ReadableSpan: {span}")
+
+    return span
+
+
+def _safe_json_dump(obj: Any) -> str:
+    """Serialize an object to JSON, falling back to ``str(obj)`` if needed.
+
+    Args:
+        obj: Object to be serialized.
+
+    Returns:
+        The JSON-encoded string representation of the object, or its string
+        representation if JSON encoding fails.
+    """
+    try:
+        return json.dumps(obj, default=str, ensure_ascii=False)
+    except Exception:
+        return str(obj)
+
+
+class OperationContext:
+    """Context manager and decorator for tracing operations.
+
+    This class manages an OpenTelemetry span for a logical unit of work. It can
+    be used either:
+
+    * As a decorator, in which case inputs and outputs are inferred
+      automatically from the wrapped function's signature.
+    * As a context manager, in which case inputs and outputs can be recorded
+      explicitly via :meth:`set_input` and :meth:`set_output`.
+
+    Attributes:
+        name: Human-readable span name.
+        initial_attributes: Attributes applied when the span is created.
+        tracer: OpenTelemetry tracer used to create spans.
+        span: The currently active span, if any.
+    """
+
+    def __init__(self, name: str, attributes: Dict[str, Any], *, propagate: bool = True) -> None:
+        """Initialize a new operation context.
+
+        Args:
+            name: Human-readable name of the span.
+            attributes: Initial attributes attached to the span. Values are
+                JSON-serialized where necessary.
+            propagate: Whether the span should be sent to active exporters.
+        """
+        self.name: str = name
+        self.initial_attributes: Dict[str, Any] = attributes
+        self.propagate: bool = propagate
+        self.tracer: trace.Tracer = get_tracer(use_active_span_processor=propagate)
+        self.span: Optional[trace.Span] = None
+        self._ctx_token: Optional[ContextManager[Any]] = None
+
+    def __enter__(self) -> "OperationContext":
+        """Enter the context manager and start a new span.
+
+        Returns:
+            The current :class:`OperationContext` instance with an active span.
+        """
+        # 1. Start the span with initial attributes (JSON serialized)
+        sanitized_attrs = {
+            k: _safe_json_dump(v) if not isinstance(v, (str, int, float, bool)) else v
+            for k, v in self.initial_attributes.items()
+        }
+
+        self.span = self.tracer.start_span(self.name, attributes=sanitized_attrs)
+        self._ctx_token = trace.use_span(self.span, end_on_exit=True)
+        self._ctx_token.__enter__()
+        return self
+
+    def __exit__(
+        self,
+        exc_type: Optional[Type[BaseException]],
+        exc_val: Optional[BaseException],
+        exc_tb: Optional[TracebackType],
+    ) -> None:
+        """Exit the context manager and finish the span.
+
+        Any exception raised inside the context is recorded on the span and the
+        span status is set to error.
+
+        Args:
+            exc_type: Exception type, if an exception occurred.
+            exc_val: Exception instance, if an exception occurred.
+            exc_tb: Traceback object, if an exception occurred.
+        """
+        # 1. Record Exception if present
+        if exc_val and self.span:
+            self.span.record_exception(exc_val)
+            self.span.set_status(Status(StatusCode.ERROR, str(exc_val)))
+
+        # 2. Close span
+        if self._ctx_token:
+            self._ctx_token.__exit__(exc_type, exc_val, exc_tb)
+
+    def set_input(self, *args: Any, **kwargs: Any) -> None:
+        """Record input arguments on the current span.
+
+        Positional arguments are stored under the ``input.args`` attribute,
+        and keyword arguments are stored under ``input.<name>`` attributes.
+
+        This is intended for use inside a ``with operation(...) as op`` block.
+
+        Args:
+            *args: Positional arguments to record.
+            **kwargs: Keyword arguments to record.
+        """
+        if not self.span:
+            return
+
+        if args:
+            self.span.set_attribute("input.args", _safe_json_dump(args))
+        if kwargs:
+            for k, v in kwargs.items():
+                self.span.set_attribute(f"input.{k}", _safe_json_dump(v))
+
+    def set_output(self, output: Any) -> None:
+        """Record the output value on the current span.
+
+        This is intended for use inside a ``with operation(...) as op`` block.
+
+        Args:
+            output: The output value to record.
+        """
+        if not self.span:
+            return
+        self.span.set_attribute("output", _safe_json_dump(output))
+
+    def __call__(self, fn: _FnType) -> _FnType:
+        """Wrap a callable so its execution is traced in a span.
+
+        When used as a decorator, a new span is created for each call to
+        the wrapped function. The bound arguments are recorded as input
+        attributes, the return value is recorded as an output attribute,
+        and any exception is recorded and marks the span as an error.
+
+        Args:
+            fn: The function or coroutine function to wrap.
+
+        Returns:
+            The wrapped callable.
+        """
+        function_name = fn.__name__
+
+        sig = inspect.signature(fn)
+
+        def _record_auto_inputs(span: trace.Span, args: Tuple[Any, ...], kwargs: Dict[str, Any]) -> None:
+            """Bind arguments to signature and log them on the span.
+
+            Args:
+                span: Span on which to record attributes.
+                args: Positional arguments passed to the wrapped callable.
+                kwargs: Keyword arguments passed to the wrapped callable.
+            """
+            try:
+                bound = sig.bind(*args, **kwargs)
+                bound.apply_defaults()
+                for k, v in bound.arguments.items():
+                    span.set_attribute(
+                        f"{LightningSpanAttributes.OPERATION_INPUT.value}.{k}",
+                        _safe_json_dump(v),
+                    )
+            except Exception:
+                span.set_attribute(
+                    f"{LightningSpanAttributes.OPERATION_INPUT.value}.args",
+                    _safe_json_dump(args),
+                )
+                span.set_attribute(
+                    f"{LightningSpanAttributes.OPERATION_INPUT.value}.kwargs",
+                    _safe_json_dump(kwargs),
+                )
+
+        if asyncio.iscoroutinefunction(fn) or inspect.iscoroutinefunction(fn):
+
+            @functools.wraps(fn)
+            async def async_wrapper(*args: Any, **kwargs: Any) -> Any:
+                """Async wrapper that traces the wrapped coroutine."""
+                # Reuse __enter__ logic via 'with self' would share state incorrectly
+                # across concurrent calls. We must create a new span per call.
+                # So we manually reimplement the span logic for the wrapper here.
+
+                sanitized_attrs = {
+                    k: _safe_json_dump(v) if not isinstance(v, (str, int, float, bool)) else v
+                    for k, v in self.initial_attributes.items()
+                }
+
+                with self.tracer.start_as_current_span(self.name, attributes=sanitized_attrs) as span:
+                    span.set_attribute(LightningSpanAttributes.OPERATION_NAME.value, function_name)
+                    _record_auto_inputs(span, args, kwargs)
+                    try:
+                        result = await fn(*args, **kwargs)
+                        span.set_attribute(
+                            LightningSpanAttributes.OPERATION_OUTPUT.value,
+                            _safe_json_dump(result),
+                        )
+                        return result
+                    except Exception as e:
+                        span.record_exception(e)
+                        span.set_status(Status(StatusCode.ERROR, str(e)))
+                        raise
+
+            return cast(_FnType, async_wrapper)
+
+        else:
+
+            @functools.wraps(fn)
+            def sync_wrapper(*args: Any, **kwargs: Any) -> Any:
+                """Sync wrapper that traces the wrapped callable."""
+                sanitized_attrs = {
+                    k: _safe_json_dump(v) if not isinstance(v, (str, int, float, bool)) else v
+                    for k, v in self.initial_attributes.items()
+                }
+
+                with self.tracer.start_as_current_span(self.name, attributes=sanitized_attrs) as span:
+                    span.set_attribute(LightningSpanAttributes.OPERATION_NAME.value, function_name)
+                    _record_auto_inputs(span, args, kwargs)
+                    try:
+                        result = fn(*args, **kwargs)
+                        span.set_attribute(
+                            LightningSpanAttributes.OPERATION_OUTPUT.value,
+                            _safe_json_dump(result),
+                        )
+                        return result
+                    except Exception as e:
+                        span.record_exception(e)
+                        span.set_status(Status(StatusCode.ERROR, str(e)))
+                        raise
+
+            return cast(_FnType, sync_wrapper)
+
+
+@overload
+def operation(fn: _FnType, *, propagate: bool = True, **additional_attributes: Any) -> _FnType: ...
+
+
+@overload
+def operation(*, propagate: bool = True, **additional_attributes: Any) -> OperationContext: ...
+
+
+def operation(
+    fn: Optional[_FnType] = None,
+    *,
+    propagate: bool = True,
+    **additional_attributes: Any,
+) -> Union[_FnType, OperationContext]:
+    """Entry point for tracking operations.
+
+    This helper can be used either as a decorator or as a context manager.
+    The span name is fixed to [`AGL_OPERATION`][agentlightning.semconv.AGL_OPERATION];
+    custom span names are not supported. Any keyword arguments are recorded as span attributes.
+
+    Usage as a decorator:
+
+    ```python
+    @operation
+    def func(...):
+        ...
+
+    @operation(category="compute")
+    def func(...):
+        ...
+    ```
+
+    Usage as a context manager:
+
+    ```python
+    with operation(user_id=123) as op:
+        op.set_input(data=data)
+        # ... do work ...
+        op.set_output(result)
+    ```
+
+    Args:
+        fn: When used as `@operation`, this is the wrapped function.
+            When used as `operation(**attrs)`, this should be omitted (or
+            left as `None`) and only keyword attributes are provided.
+        propagate: Whether spans should use the active span processor. When False,
+            spans will stay local and not be exported.
+        **additional_attributes: Additional span attributes to attach at
+            creation time.
+
+    Returns:
+        Either a wrapped callable (when used as a decorator) or an
+        [`OperationContext`][agentlightning.emitter.annotation.OperationContext]
+        (when used as a context manager factory).
+    """
+    # Case 1: Used as @operation (bare decorator or with attributes)
+    if callable(fn):
+        # Create context with fixed name, then immediately wrap the function
+        return OperationContext(AGL_OPERATION, additional_attributes, propagate=propagate)(fn)
+
+    # Case 2: Used as operation(...) / with operation(...)
+    # Custom span names are intentionally not supported; use AGL_OPERATION.
+    if fn is not None:
+        raise ValueError("Custom span names are intentionally not supported when used as a context manager.")
+    return OperationContext(AGL_OPERATION, additional_attributes, propagate=propagate)

agentlightning/emitter/exception.py

@@ -2,43 +2,53 @@
 
 import logging
 import traceback
+from typing import Any, Dict, Optional
 
 from opentelemetry.semconv.attributes import exception_attributes
 
-from agentlightning.types import SpanNames
-
-from .utils import get_tracer
+from agentlightning.semconv import AGL_EXCEPTION
+from agentlightning.utils.otel import get_tracer
 
 logger = logging.getLogger(__name__)
 
 
-def emit_exception(exception: BaseException) -> None:
+def emit_exception(
+    exception: BaseException, attributes: Optional[Dict[str, Any]] = None, propagate: bool = True
+) -> None:
     """Record an exception with OpenTelemetry metadata.
 
+    Classic OpenTelemetry records exceptions in a dedicated logging service.
+    We simplify the model and use trace spans to record exceptions as well.
+
     Args:
         exception: Raised exception instance to serialize into telemetry attributes.
+        attributes: Additional attributes to attach to the exception span.
+        propagate: Whether to propagate the span to exporters automatically.
 
     !!! note
-        The helper validates its input. Non-exception values are ignored to prevent
-        noisy telemetry and indicate programming mistakes via the logger.
+
+        The helper validates its input. If a non-exception value is provided,
+        a TypeError is raised to indicate a programming mistake.
     """
     if not isinstance(exception, BaseException):  # type: ignore
-        logger.error(f"Expected an BaseException instance, got: {type(exception)}. Skip emit_exception.")
-        return
+        raise TypeError(f"Expected a BaseException instance, got: {type(exception)}.")
 
-    tracer = get_tracer()
+    tracer = get_tracer(use_active_span_processor=propagate)
     stacktrace = "".join(traceback.format_exception(type(exception), exception, exception.__traceback__))
-    attributes = {
+    span_attributes = {
         exception_attributes.EXCEPTION_TYPE: type(exception).__name__,
         exception_attributes.EXCEPTION_MESSAGE: str(exception),
         exception_attributes.EXCEPTION_ESCAPED: True,
     }
     if stacktrace.strip():
-        attributes[exception_attributes.EXCEPTION_STACKTRACE] = stacktrace
+        span_attributes[exception_attributes.EXCEPTION_STACKTRACE] = stacktrace
+
+    if attributes:
+        span_attributes.update(attributes)
 
     span = tracer.start_span(
-        SpanNames.EXCEPTION.value,
-        attributes=attributes,
+        AGL_EXCEPTION,
+        attributes=span_attributes,
     )
     logger.debug("Emitting exception span for %s", type(exception).__name__)
     with span:

agentlightning/emitter/message.py

@@ -1,33 +1,55 @@
 # Copyright (c) Microsoft. All rights reserved.
 
 import logging
+from typing import Any, Dict, Optional
 
-from agentlightning.types import SpanAttributeNames, SpanNames
-
-from .utils import get_tracer
+from agentlightning.semconv import AGL_MESSAGE, LightningSpanAttributes
+from agentlightning.types import SpanLike
+from agentlightning.utils.otel import get_tracer
 
 logger = logging.getLogger(__name__)
 
 
-def emit_message(message: str) -> None:
+def emit_message(message: str, attributes: Optional[Dict[str, Any]] = None, propagate: bool = True) -> None:
     """Emit a textual message as an OpenTelemetry span.
 
+    Commonly used for sending debugging and logging messages.
+
     Args:
         message: Human readable message to attach as a span attribute.
+        attributes: Additional attributes to attach to the message span.
+        propagate: Whether to propagate the span to exporters automatically.
 
     !!! note
         OpenTelemetry distinguishes between logs and spans. Emitting the message as a
         span keeps all Agent Lightning telemetry in a single data store for analysis.
     """
     if not isinstance(message, str):  # type: ignore
-        logger.error(f"Message must be a string, got: {type(message)}. Skip emit_message.")
-        return
+        raise TypeError(f"Message must be a string or list of strings, got: {type(message)}.")
 
-    tracer = get_tracer()
+    tracer = get_tracer(use_active_span_processor=propagate)
+    span_attributes = {LightningSpanAttributes.MESSAGE_BODY.value: message}
+    if attributes:
+        span_attributes.update(attributes)
     span = tracer.start_span(
-        SpanNames.MESSAGE.value,
-        attributes={SpanAttributeNames.MESSAGE.value: message},
+        AGL_MESSAGE,
+        attributes=span_attributes,
     )
     logger.debug("Emitting message span with message: %s", message)
     with span:
         pass
+
+
+def get_message_value(span: SpanLike) -> Optional[str]:
+    """Extract the message string from a message span.
+
+    Args:
+        span: Span-like object to extract the message from.
+    """
+    span_attributes = span.attributes or {}
+    if LightningSpanAttributes.MESSAGE_BODY.value not in span_attributes:
+        return None
+    message = span_attributes[LightningSpanAttributes.MESSAGE_BODY.value]
+    if isinstance(message, str):
+        return message
+    raise TypeError(f"Message must be a string, got: {type(message)}.")

agentlightning/emitter/object.py

@@ -1,37 +1,106 @@
 # Copyright (c) Microsoft. All rights reserved.
 
+import base64
 import json
 import logging
-from typing import Any
+from typing import Any, Dict, Optional
 
-from agentlightning.types import SpanAttributeNames, SpanNames
-
-from .utils import get_tracer
+from agentlightning.semconv import AGL_OBJECT, LightningSpanAttributes
+from agentlightning.types import SpanLike
+from agentlightning.utils.otel import full_qualified_name, get_tracer
 
 logger = logging.getLogger(__name__)
 
 
-def emit_object(object: Any) -> None:
+def emit_object(object: Any, attributes: Optional[Dict[str, Any]] = None, propagate: bool = True) -> None:
     """Emit an object's serialized representation as an OpenTelemetry span.
 
     Args:
         object: Data structure to encode as JSON and attach to the span payload.
+        attributes: Additional attributes to attach to the object span.
+        propagate: Whether to propagate the span to exporters automatically.
 
     !!! note
-        The payload must be JSON serializable. Non-serializable objects are ignored and
-        an error is logged to aid debugging.
+        The payload must be JSON serializable. Non-serializable objects will lead to a RuntimeError.
     """
-    try:
-        serialized = json.dumps(object)
-    except (TypeError, ValueError):
-        logger.error(f"Object must be JSON serializable, got: {type(object)}. Skip emit_object.")
-        return
-
-    tracer = get_tracer()
+    span_attributes = encode_object(object)
+    if attributes:
+        span_attributes.update(attributes)
+    tracer = get_tracer(use_active_span_processor=propagate)
     span = tracer.start_span(
-        SpanNames.OBJECT.value,
-        attributes={SpanAttributeNames.OBJECT.value: serialized},
+        AGL_OBJECT,
+        attributes=span_attributes,
     )
-    logger.debug("Emitting object span with payload size %d characters", len(serialized))
+    attr_length = 0
+    if LightningSpanAttributes.OBJECT_JSON.value in span_attributes:
+        attr_length = len(span_attributes[LightningSpanAttributes.OBJECT_JSON.value])
+    elif LightningSpanAttributes.OBJECT_LITERAL.value in span_attributes:
+        attr_length = len(span_attributes[LightningSpanAttributes.OBJECT_LITERAL.value])
+    logger.debug("Emitting object span with payload size %d characters", attr_length)
     with span:
         pass
+
+
+def encode_object(object: Any) -> Dict[str, Any]:
+    """Encode an object as span attributes.
+
+    Args:
+        object: Data structure to encode as JSON.
+    """
+    span_attributes = {}
+    if isinstance(object, (str, int, float, bool)):
+        span_attributes = {
+            LightningSpanAttributes.OBJECT_TYPE.value: type(object).__name__,
+            LightningSpanAttributes.OBJECT_LITERAL.value: str(object),
+        }
+    elif isinstance(object, bytes):
+        b64_encoded = base64.b64encode(object).decode("utf-8")
+        span_attributes = {
+            LightningSpanAttributes.OBJECT_TYPE.value: "bytes",
+            LightningSpanAttributes.OBJECT_LITERAL.value: b64_encoded,
+        }
+    else:
+        try:
+            serialized = json.dumps(object)
+        except (TypeError, ValueError) as exc:
+            raise RuntimeError(f"Object must be JSON serializable, got: {type(object)}.") from exc
+
+        span_attributes = {
+            LightningSpanAttributes.OBJECT_TYPE.value: full_qualified_name(type(object)),  # type: ignore
+            LightningSpanAttributes.OBJECT_JSON.value: serialized,
+        }
+
+    return span_attributes
+
+
+def get_object_value(span: SpanLike) -> Any:
+    """Extract the object payload from an object span.
+
+    Args:
+        span: Span object produced by Agent Lightning emitters.
+    """
+    attributes = span.attributes or {}
+    if LightningSpanAttributes.OBJECT_JSON.value in attributes:
+        serialized = attributes[LightningSpanAttributes.OBJECT_JSON.value]
+        try:
+            return json.loads(serialized)  # type: ignore
+        except (TypeError, ValueError) as exc:
+            raise RuntimeError("Failed to deserialize object JSON from span.") from exc
+    elif LightningSpanAttributes.OBJECT_LITERAL.value in attributes:
+        literal = attributes[LightningSpanAttributes.OBJECT_LITERAL.value]
+        obj_type = attributes.get(LightningSpanAttributes.OBJECT_TYPE.value, "str")
+        if obj_type == "str":
+            return literal
+        elif obj_type == "int":
+            # Let it raise errors if there are any
+            return int(literal)  # type: ignore
+        elif obj_type == "float":
+            return float(literal)  # type: ignore
+        elif obj_type == "bool":
+            return literal.lower() == "true"  # type: ignore
+        elif obj_type == "bytes":
+            return base64.b64decode(literal.encode("utf-8"))  # type: ignore
+        else:
+            raise RuntimeError(f"Unsupported object type for literal deserialization: {obj_type}")
+    else:
+        return None

agentlightning/emitter/reward.py

@@ -23,10 +23,13 @@ from typing import (
 import agentops
 from agentops.sdk.decorators import operation
 from opentelemetry.sdk.trace import ReadableSpan
+from pydantic import TypeAdapter
 
-from agentlightning.types import SpanLike, SpanNames
+from agentlightning.semconv import AGL_ANNOTATION, LightningSpanAttributes, RewardPydanticModel
+from agentlightning.types import SpanLike
+from agentlightning.utils.otel import filter_and_unflatten_attributes
 
-from .utils import get_tracer
+from .annotation import emit_annotation
 
 logger = logging.getLogger(__name__)
 
@@ -34,18 +37,26 @@ __all__ = [
     "reward",
     "emit_reward",
     "get_reward_value",
+    "get_rewards_from_span",
     "is_reward_span",
     "find_reward_spans",
     "find_final_reward",
 ]
 
 
-class RewardSpanData(TypedDict):
+class RewardDimension(TypedDict):
+    """Type representing a single dimension in a multi-dimensional reward."""
+
+    name: str
+    value: float
+
+
+class _RewardSpanData(TypedDict):
     type: Literal["reward"]
     value: Optional[float]
 
 
-FnType = TypeVar("FnType", bound=Callable[..., Any])
+_FnType = TypeVar("_FnType", bound=Callable[..., Any])
 
 
 def _agentops_initialized() -> bool:
@@ -53,7 +64,7 @@ def _agentops_initialized() -> bool:
     return agentops.get_client().initialized
 
 
-def reward(fn: FnType) -> FnType:
+def reward(fn: _FnType) -> _FnType:
     """Decorate a reward function so its outputs are tracked as spans.
 
     The decorator integrates with AgentOps when it is available and falls back to
@@ -70,7 +81,7 @@ def reward(fn: FnType) -> FnType:
         Wrapped callable that preserves the original signature.
     """
 
-    def wrap_result(result: Optional[float]) -> RewardSpanData:
+    def wrap_result(result: Optional[float]) -> _RewardSpanData:
         """Normalize the reward value into the span payload format."""
         if result is None:
             return {"type": "reward", "value": None}
@@ -94,7 +105,7 @@ def reward(fn: FnType) -> FnType:
             result: Optional[float] = None
 
             @operation
-            async def agentops_reward_operation() -> RewardSpanData:
+            async def agentops_reward_operation() -> _RewardSpanData:
                 # The reward function we are interested in tracing
                 # It takes zero inputs and return a formatted dict
                 nonlocal result
@@ -118,7 +129,7 @@ def reward(fn: FnType) -> FnType:
             result: Optional[float] = None
 
             @operation
-            def agentops_reward_operation() -> RewardSpanData:
+            def agentops_reward_operation() -> _RewardSpanData:
                 nonlocal result
                 result = fn(*args, **kwargs)
                 return wrap_result(result)
@@ -129,12 +140,36 @@ def reward(fn: FnType) -> FnType:
         return wrapper  # type: ignore
 
 
-def emit_reward(reward: float) -> ReadableSpan:
+def emit_reward(
+    reward: float | Dict[str, Any],
+    *,
+    primary_key: str | None = None,
+    attributes: Dict[str, Any] | None = None,
+    propagate: bool = True,
+) -> ReadableSpan:
     """Emit a reward value as an OpenTelemetry span.
 
+    Examples:
+        Emit a single-dimensional reward:
+        >>> emit_reward(1.0)
+
+        Emit multi-dimensional rewards:
+        >>> emit_reward({"task_completion": 1.0, "efficiency": 0.8}, primary_key="task_completion")
+
+        Emit a reward with additional attributes (for example linking to another response span):
+        >>> from agentlightning.utils.otel import make_link_attributes
+        >>> emit_reward(0.5, attributes=make_link_attributes({"gen_ai.response.id": "response-123"}))
+
+        Or adding tags onto the reward span:
+        >>> from agentlightning.utils.otel import make_tag_attributes
+        >>> emit_reward(0.7, attributes=make_tag_attributes(["fast", "reliable"]))
+
     Args:
         reward: Numeric reward to record. Integers and booleans are converted to
             floating point numbers for consistency.
+            Use a dictionary to represent a multi-dimensional reward.
+        attributes: Other optional span attributes.
+        propagate: Whether to propagate the span to exporters automatically.
 
     Returns:
         Readable span capturing the recorded reward.
@@ -144,20 +179,34 @@ def emit_reward(reward: float) -> ReadableSpan:
             resulting span is not a [`ReadableSpan`](https://opentelemetry.io/docs/concepts/signals/traces/) instance.
     """
     logger.debug(f"Emitting reward: {reward}")
-    if isinstance(reward, (int, bool)):
-        reward = float(reward)
-    if not isinstance(reward, float):
-        raise ValueError(f"Reward must be a number, got: {type(reward)}")
+    reward_dimensions: List[RewardDimension] = []
+    if isinstance(reward, dict):
+        reward_dict: Dict[str, float] = {}
+        for k, v in reward.items():
+            if isinstance(v, (int, bool)):
+                reward_dict[k] = float(v)
+            elif isinstance(v, float):
+                reward_dict[k] = v
+            else:
+                raise ValueError(f"Reward value must be a number, got: {type(v)} for key {k}")
+        if primary_key is None:
+            raise ValueError("When emitting a multi-dimensional reward as a dict, primary_key must be provided.")
+        if primary_key not in reward_dict:
+            raise ValueError(f"Primary key '{primary_key}' not found in reward dict keys: {list(reward_dict.keys())}")
+        reward_dimensions.append(RewardDimension(name=primary_key, value=reward_dict[primary_key]))
+        for k, v in reward_dict.items():
+            if k != primary_key:
+                reward_dimensions.append(RewardDimension(name=k, value=v))
+    else:
+        if isinstance(reward, (int, bool)):
+            reward = float(reward)
+        elif not isinstance(reward, float):  # pyright: ignore[reportUnnecessaryIsInstance]
+            raise TypeError(f"Reward must be a number, got: {type(reward)}")
+        reward_dimensions.append(RewardDimension(name="primary", value=reward))
 
-    # TODO: This should use the tracer from current context by tracer
-    tracer = get_tracer()
-    span = tracer.start_span(SpanNames.REWARD.value, attributes={"reward": reward})
-    # Do nothing; it's just a number
-    with span:
-        pass
-    if not isinstance(span, ReadableSpan):
-        raise ValueError(f"Span is not a ReadableSpan: {span}")
-    return span
+    return emit_annotation(
+        {LightningSpanAttributes.REWARD.value: reward_dimensions, **(attributes or {})}, propagate=propagate
+    )
 
 
 def get_reward_value(span: SpanLike) -> Optional[float]:
@@ -167,8 +216,14 @@ def get_reward_value(span: SpanLike) -> Optional[float]:
         span: Span object produced by AgentOps or Agent Lightning emitters.
 
     Returns:
-        The reward encoded in the span or `None` when the span does not represent a reward.
+        The primary reward encoded in the span or `None` when the span does not represent a reward.
     """
+    # v0.3+ emit reward format
+    reward_list = get_rewards_from_span(span)
+    if reward_list:
+        # Reward list is ordered and the first element is the primary reward
+        return reward_list[0].value
+
     for key in [
         "agentops.task.output",  # newer versions of agentops
         "agentops.entity.output",
@@ -191,19 +246,45 @@ def get_reward_value(span: SpanLike) -> Optional[float]:
                 return None
             if not isinstance(reward_value, float):
                 logger.error(f"Reward is not a number, got: {type(reward_value)}. This may cause undefined behaviors.")
+            logger.warning(
+                f"Extracted reward {reward_value} from AgentOps. This format is deprecated, please migrate to using `emit_reward`."
+            )
             return cast(float, reward_value)
 
-    # Latest emit reward format
-    if span.name == SpanNames.REWARD.value and span.attributes:
+    # v0.2 emit reward format
+    if span.name == AGL_ANNOTATION and span.attributes:
         reward_value = span.attributes.get("reward", None)
         if reward_value is None:
             return None
         if not isinstance(reward_value, float):
             logger.error(f"Reward is not a number, got: {type(reward_value)}. This may cause undefined behaviors.")
+        logger.warning(
+            f"Extracted reward {reward_value} from a legacy version of reward span. You might have inconsistent agent-lightning versions."
+        )
         return cast(float, reward_value)
+
     return None
 
 
+def get_rewards_from_span(span: SpanLike) -> List[RewardPydanticModel]:
+    """Extract the reward as a list from a span, if available.
+
+    Args:
+        span: Span object produced by AgentOps or Agent Lightning emitters.
+
+    Returns:
+        A list of reward dimensions encoded in the span or an empty list when the span does not represent a reward.
+    """
+    if span.attributes and any(key.startswith(LightningSpanAttributes.REWARD.value) for key in span.attributes):
+        reward_attr = filter_and_unflatten_attributes(
+            cast(Any, span.attributes or {}), LightningSpanAttributes.REWARD.value
+        )
+        recovered_rewards = TypeAdapter(List[RewardPydanticModel]).validate_python(reward_attr)
+        return recovered_rewards
+    else:
+        return []
+
+
 def is_reward_span(span: SpanLike) -> bool:
     """Return ``True`` when the provided span encodes a reward value."""
     maybe_reward = get_reward_value(span)

agentlightning/emitter/utils.py

@@ -1,22 +0,0 @@
-# Copyright (c) Microsoft. All rights reserved.
-
-"""Utilities shared across emitter implementations."""
-
-import opentelemetry.trace as trace_api
-from opentelemetry.trace import get_tracer_provider
-
-
-def get_tracer() -> trace_api.Tracer:
-    """Resolve the OpenTelemetry tracer configured for Agent Lightning.
-
-    Returns:
-        OpenTelemetry tracer tagged with the `agentlightning` instrumentation name.
-
-    Raises:
-        RuntimeError: If OpenTelemetry was not initialized before calling this helper.
-    """
-    if hasattr(trace_api, "_TRACER_PROVIDER") and trace_api._TRACER_PROVIDER is None:  # type: ignore[attr-defined]
-        raise RuntimeError("Tracer is not initialized. Cannot emit a meaningful span.")
-
-    tracer_provider = get_tracer_provider()
-    return tracer_provider.get_tracer("agentlightning")

agentlightning/env_var.py

@@ -0,0 +1,156 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Environment variable managements."""
+
+from __future__ import annotations
+
+import os
+from enum import Enum
+from typing import overload
+
+__all__ = [
+    "LightningEnvVar",
+    "resolve_bool_env_var",
+    "resolve_int_env_var",
+    "resolve_str_env_var",
+]
+
+
+class LightningEnvVar(Enum):
+    """Environment variables for Agent Lightning."""
+
+    AGL_EMITTER_DEBUG = "AGL_EMITTER_DEBUG"
+    """Enable debug logging for the emitter."""
+
+    AGL_MANAGED_STORE = "AGL_MANAGED_STORE"
+    """If yes, the [`ExecutionStrategy`][agentlightning.ExecutionStrategy]
+    constructs LightningStore wrappers automatically. When `False` the provided
+    `store` is passed directly to the bundles, allowing callers to manage
+    store wrappers manually."""
+
+    AGL_CURRENT_ROLE = "AGL_CURRENT_ROLE"
+    """Which side(s) to run in this process. Used in
+    [`ClientServerExecutionStrategy`][agentlightning.ClientServerExecutionStrategy]."""
+
+    AGL_SERVER_HOST = "AGL_SERVER_HOST"
+    """Interface the [`LightningStoreServer`][agentlightning.LightningStoreServer]
+    binds to when running the algorithm bundle locally."""
+
+    AGL_SERVER_PORT = "AGL_SERVER_PORT"
+    """Port the [`LightningStoreServer`][agentlightning.LightningStoreServer] listens to."""
+
+
+_TRUTHY_VALUES = {"1", "true", "yes", "on"}
+_FALSY_VALUES = {"0", "false", "no", "off"}
+
+
+@overload
+def resolve_bool_env_var(env_var: LightningEnvVar, override: bool, fallback: bool) -> bool: ...
+
+
+@overload
+def resolve_bool_env_var(env_var: LightningEnvVar, *, fallback: bool) -> bool: ...
+
+
+@overload
+def resolve_bool_env_var(
+    env_var: LightningEnvVar, override: bool | None = None, fallback: bool | None = None
+) -> bool | None: ...
+
+
+def resolve_bool_env_var(
+    env_var: LightningEnvVar, override: bool | None = None, fallback: bool | None = None
+) -> bool | None:
+    """Resolve a boolean environment variable.
+
+    Args:
+        env_var: The environment variable to resolve.
+        override: Optional override supplied by the caller.
+        fallback: Default value if the environment variable is not set.
+    """
+
+    if override is not None:
+        return override
+
+    env_value = os.getenv(env_var.value)
+    if env_value is None:
+        return fallback
+
+    normalized = env_value.strip().lower()
+    if normalized in _TRUTHY_VALUES:
+        return True
+    if normalized in _FALSY_VALUES:
+        return False
+
+    raise ValueError(f"{env_var.value} must be one of {_TRUTHY_VALUES} or {_FALSY_VALUES}")
+
+
+@overload
+def resolve_int_env_var(env_var: LightningEnvVar, override: int, fallback: int) -> int: ...
+
+
+@overload
+def resolve_int_env_var(env_var: LightningEnvVar, *, fallback: int) -> int: ...
+
+
+@overload
+def resolve_int_env_var(
+    env_var: LightningEnvVar, override: int | None = None, fallback: int | None = None
+) -> int | None: ...
+
+
+def resolve_int_env_var(
+    env_var: LightningEnvVar, override: int | None = None, fallback: int | None = None
+) -> int | None:
+    """Resolve an integer environment variable.
+
+    Args:
+        env_var: The environment variable to resolve.
+        override: Optional override supplied by the caller.
+        fallback: Default value if the environment variable is not set.
+    """
+    if override is not None:
+        return override
+
+    env_value = os.getenv(env_var.value)
+    if env_value is None:
+        return fallback
+
+    try:
+        return int(env_value)
+    except ValueError:
+        raise ValueError(f"{env_var.value} must be an integer")
+
+
+@overload
+def resolve_str_env_var(env_var: LightningEnvVar, override: str, fallback: str) -> str: ...
+
+
+@overload
+def resolve_str_env_var(env_var: LightningEnvVar, *, fallback: str) -> str: ...
+
+
+@overload
+def resolve_str_env_var(
+    env_var: LightningEnvVar, override: str | None = None, fallback: str | None = None
+) -> str | None: ...
+
+
+def resolve_str_env_var(
+    env_var: LightningEnvVar, override: str | None = None, fallback: str | None = None
+) -> str | None:
+    """Resolve a string environment variable.
+
+    Args:
+        env_var: The environment variable to resolve.
+        override: Optional override supplied by the caller.
+        fallback: Default value if the environment variable is not set.
+    """
+    if override is not None:
+        return override
+
+    env_value = os.getenv(env_var.value)
+    if env_value is None:
+        return fallback
+
+    return env_value

agentlightning/execution/base.py

@@ -3,7 +3,6 @@
 from __future__ import annotations
 
 import logging
-import os
 from typing import Protocol
 
 from agentlightning.store.base import LightningStore
@@ -13,47 +12,6 @@ from .events import ExecutionEvent
 logger = logging.getLogger(__name__)
 
 
-_TRUTHY_VALUES = {"1", "true", "yes", "on"}
-_FALSY_VALUES = {"0", "false", "no", "off"}
-
-
-def resolve_managed_store_flag(value: bool | None) -> bool:
-    """Determine whether execution helpers should wrap the provided store.
-
-    The helper first honours an explicit `value`. When `None` it falls back
-    to the `AGL_MANAGED_STORE` environment variable, accepting a variety
-    of truthy and falsy spellings. Missing environment configuration defaults to
-    `True` so that higher-level strategies create the appropriate client or
-    server wrappers automatically.
-
-    Args:
-        value: Optional override supplied by the caller.
-
-    Returns:
-        `True` when a managed store should be created around the provided
-        instance, otherwise `False`.
-
-    Raises:
-        ValueError: If `AGL_MANAGED_STORE` is set to an unsupported
-            value.
-    """
-
-    if value is not None:
-        return value
-
-    env_value = os.getenv("AGL_MANAGED_STORE")
-    if env_value is None:
-        return True
-
-    normalized = env_value.strip().lower()
-    if normalized in _TRUTHY_VALUES:
-        return True
-    if normalized in _FALSY_VALUES:
-        return False
-
-    raise ValueError("AGL_MANAGED_STORE must be one of 1, 0, true, false, yes, no, on, or off")
-
-
 class AlgorithmBundle(Protocol):
     """Callable bundle produced by [`Trainer`][agentlightning.Trainer].
 

agentlightning/execution/client_server.py

@@ -9,10 +9,11 @@ import time
 from multiprocessing.context import BaseContext
 from typing import Callable, Iterable, Literal, cast
 
+from agentlightning.env_var import LightningEnvVar, resolve_bool_env_var, resolve_int_env_var, resolve_str_env_var
 from agentlightning.store.base import LightningStore
 from agentlightning.store.client_server import LightningStoreClient, LightningStoreServer
 
-from .base import AlgorithmBundle, ExecutionStrategy, RunnerBundle, resolve_managed_store_flag
+from .base import AlgorithmBundle, ExecutionStrategy, RunnerBundle
 from .events import ExecutionEvent, MultiprocessingEvent
 
 logger = logging.getLogger(__name__)
@@ -67,10 +68,11 @@ class ClientServerExecutionStrategy(ExecutionStrategy):
         server_host: str | None = None,
         server_port: int | None = None,
         n_runners: int = 1,
-        graceful_timeout: float = 5.0,
-        terminate_timeout: float = 5.0,
+        graceful_timeout: float = 10.0,
+        terminate_timeout: float = 10.0,
         main_process: Literal["algorithm", "runner"] = "algorithm",
         managed_store: bool | None = None,
+        allowed_exit_codes: Iterable[int] = (0, -15),
     ) -> None:
         """Configure the strategy.
 
@@ -94,45 +96,33 @@ class ClientServerExecutionStrategy(ExecutionStrategy):
                 LightningStore client/server wrappers automatically. When
                 `False` the provided `store` is passed directly to the
                 bundles, allowing callers to manage store wrappers manually.
+            allowed_exit_codes: Allowed exit codes for subprocesses.
+                By default, runner can exit gracefully with code 0 or terminated
+                by SIGTERM (-15).
         """
-        if role is None:
-            role_env = os.getenv("AGL_CURRENT_ROLE")
-            if role_env is None:
-                # Use both if not specified via env var or argument
-                role = "both"
-            elif role_env not in ("algorithm", "runner", "both"):
-                raise ValueError("role must be one of 'algorithm', 'runner', or 'both'")
-            else:
-                role = role_env
-
-        if server_host is None:
-            server_host = os.getenv("AGL_SERVER_HOST", "localhost")
-
-        if server_port is None:
-            server_port_env = os.getenv("AGL_SERVER_PORT")
-            if server_port_env is None:
-                server_port = 4747
-            else:
-                try:
-                    server_port = int(server_port_env)
-                except ValueError as exc:
-                    raise ValueError("AGL_SERVER_PORT must be an integer") from exc
-
-        self.role = role
+        resolved_role = resolve_str_env_var(LightningEnvVar.AGL_CURRENT_ROLE, override=role, fallback="both")
+        if resolved_role not in ("algorithm", "runner", "both"):
+            raise ValueError("role must be one of 'algorithm', 'runner', or 'both'")
+        self.role: Literal["algorithm", "runner", "both"] = resolved_role
         self.n_runners = n_runners
-        self.server_host = server_host
-        self.server_port = server_port
+        self.server_host = resolve_str_env_var(
+            LightningEnvVar.AGL_SERVER_HOST, override=server_host, fallback="localhost"
+        )
+        self.server_port = resolve_int_env_var(LightningEnvVar.AGL_SERVER_PORT, override=server_port, fallback=4747)
         self.graceful_timeout = graceful_timeout
         self.terminate_timeout = terminate_timeout
         if main_process not in ("algorithm", "runner"):
             raise ValueError("main_process must be 'algorithm' or 'runner'")
         if main_process == "runner":
-            if role != "both":
+            if self.role != "both":
                 raise ValueError("main_process='runner' is only supported when role='both'")
             if n_runners != 1:
                 raise ValueError("main_process='runner' requires n_runners to be 1")
         self.main_process = main_process
-        self.managed_store = resolve_managed_store_flag(managed_store)
+        self.managed_store = resolve_bool_env_var(
+            LightningEnvVar.AGL_MANAGED_STORE, override=managed_store, fallback=True
+        )
+        self.allowed_exit_codes = tuple(allowed_exit_codes)
 
     async def _execute_algorithm(
         self, algorithm: AlgorithmBundle, store: LightningStore, stop_evt: ExecutionEvent
@@ -338,10 +328,10 @@ class ClientServerExecutionStrategy(ExecutionStrategy):
 
     def _check_process_exitcodes(self, processes: Iterable[multiprocessing.Process]) -> None:
         """Raise an error if any managed process exited with a non-zero status."""
-        failed = [p for p in processes if p.exitcode not in (0, None)]
+        failed = [p for p in processes if p.exitcode not in self.allowed_exit_codes + (None,)]
         if failed:
             formatted = ", ".join(f"{p.name or p.pid} (exitcode={p.exitcode})" for p in failed)
-            raise RuntimeError(f"Subprocesses failed: {formatted}")
+            raise RuntimeError(f"Subprocesses failed with unexpected exit codes: {formatted}")
 
     def execute(self, algorithm: AlgorithmBundle, runner: RunnerBundle, store: LightningStore) -> None:
         logger.info(

agentlightning/execution/shared_memory.py

@@ -7,10 +7,11 @@ from contextlib import suppress
 from queue import SimpleQueue
 from typing import Any, Awaitable, Callable, List, Literal, Optional, Tuple
 
+from agentlightning.env_var import LightningEnvVar, resolve_bool_env_var
 from agentlightning.store.base import LightningStore
 from agentlightning.store.threading import LightningStoreThreaded
 
-from .base import AlgorithmBundle, ExecutionStrategy, RunnerBundle, resolve_managed_store_flag
+from .base import AlgorithmBundle, ExecutionStrategy, RunnerBundle
 from .events import ExecutionEvent, ThreadingEvent
 
 logger = logging.getLogger(__name__)
@@ -62,7 +63,9 @@ class SharedMemoryExecutionStrategy(ExecutionStrategy):
         self.join_timeout = join_timeout
         self.graceful_delay = graceful_delay
         self.poll_interval = poll_interval
-        self.managed_store = resolve_managed_store_flag(managed_store)
+        self.managed_store = resolve_bool_env_var(
+            LightningEnvVar.AGL_MANAGED_STORE, override=managed_store, fallback=True
+        )
 
     async def _run_until_completed_or_canceled(self, coro: Awaitable[Any], stop_evt: ExecutionEvent) -> Any:
         """Run `coro` until it finishes or a cooperative stop is requested.

agentlightning/instrumentation/agentops.py

@@ -4,28 +4,76 @@ from __future__ import annotations
 
 import json
 import logging
-import multiprocessing
-import signal
-import socket
-import time
 from typing import Any, Callable, no_type_check
 
-import flask
 import requests
-import setproctitle
+from agentops.client.api import V3Client, V4Client
+from agentops.client.api.types import AuthTokenResponse
+from agentops.sdk.exporters import AuthenticatedOTLPExporter
+from opentelemetry.exporter.otlp.proto.http.metric_exporter import OTLPMetricExporter
+from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+from opentelemetry.sdk.metrics.export import MetricExportResult
+
+from agentlightning.utils.otlp import LightningStoreOTLPExporter
 
 logger = logging.getLogger(__name__)
 
 __all__ = [
     "instrument_agentops",
     "uninstrument_agentops",
-    "agentops_local_server",
-    "AgentOpsServerManager",
 ]
 
 # Module-level storage for originals
 _original_handle_chat_attributes: Callable[..., Any] | None = None
 _original_handle_response: Callable[..., Any] | None = None
+_agentops_service_enabled = False
+
+
+def enable_agentops_service(enabled: bool = True) -> None:
+    """
+    Enable or disable communication with the AgentOps service.
+
+    By default, AgentOps exporters and clients will run in local mode
+    and will NOT attempt to communicate with the remote AgentOps service.
+
+    Args:
+        enabled: If True, enable all AgentOps exporters and clients.
+            All exporters and clients will operate in normal mode and send data
+            to the [AgentOps service](https://www.agentops.ai).
+    """
+    global _agentops_service_enabled
+    _agentops_service_enabled = enabled
+    logger.info(f"AgentOps service enabled is set to {enabled}.")
+
+
+def _patch_exporters():
+    import agentops.client.api
+    import agentops.sdk.core
+
+    agentops.sdk.core.AuthenticatedOTLPExporter = BypassableAuthenticatedOTLPExporter  # type: ignore
+    agentops.sdk.core.OTLPMetricExporter = BypassableOTLPMetricExporter
+    if hasattr(agentops.sdk.core, "OTLPSpanExporter"):
+        agentops.sdk.core.OTLPSpanExporter = BypassableOTLPSpanExporter  # type: ignore
+    agentops.client.api.V3Client = BypassableV3Client
+    agentops.client.api.V4Client = BypassableV4Client
+
+
+def _unpatch_exporters():
+    import agentops.client.api
+    import agentops.sdk.core
+
+    agentops.sdk.core.AuthenticatedOTLPExporter = AuthenticatedOTLPExporter  # type: ignore
+    agentops.sdk.core.OTLPMetricExporter = OTLPMetricExporter
+    if hasattr(agentops.sdk.core, "OTLPSpanExporter"):
+        agentops.sdk.core.OTLPSpanExporter = OTLPSpanExporter  # type: ignore
+    agentops.client.api.V3Client = V3Client
+    agentops.client.api.V4Client = V4Client
+
+
+def _unwrap_legacy_response(response: Any) -> Any:
+    if hasattr(response, "parse") and callable(response.parse):
+        return response.parse()
+    return response
 
 
 def _patch_new_agentops():
@@ -44,6 +92,11 @@ def _patch_new_agentops():
     @no_type_check
     def _handle_chat_attributes_with_tokens(args=None, kwargs=None, return_value=None, **kws):  # type: ignore
         attributes = _original_handle_chat_attributes(args=args, kwargs=kwargs, return_value=return_value, **kws)
+
+        # In some cases, response is a openai._legacy_response.LegacyAPIResponse (e.g., LiteLLM, or LangChain),
+        # This is created by client.with_raw_response.create()
+        return_value = _unwrap_legacy_response(return_value)
+
         if (
             return_value is not None
             and hasattr(return_value, "prompt_token_ids")
@@ -89,20 +142,6 @@ def _patch_new_agentops():
                         [logprob.model_dump() for logprob in first_choice.logprobs.refusal]
                     )
 
-        # For LiteLLM, response is a openai._legacy_response.LegacyAPIResponse
-        if (
-            return_value is not None
-            and hasattr(return_value, "http_response")
-            and return_value.http_response is not None
-            and hasattr(return_value.http_response, "json")
-        ):
-            json_data = return_value.http_response.json()
-            if isinstance(json_data, dict):
-                if json_data.get("prompt_token_ids") is not None:
-                    attributes["prompt_token_ids"] = list(json_data["prompt_token_ids"])
-                if json_data.get("response_token_ids") is not None:
-                    attributes["response_token_ids"] = list(json_data["response_token_ids"][0])
-
         return attributes
 
     agentops.instrumentation.providers.openai.wrappers.chat.handle_chat_attributes = _handle_chat_attributes_with_tokens
@@ -173,6 +212,8 @@ def instrument_agentops():
     Instrument agentops to capture token IDs.
     Automatically detects and uses the appropriate patching method based on the installed agentops version.
     """
+    _patch_exporters()
+
     # Try newest version first (tested for 0.4.16)
     try:
         return _patch_new_agentops()
@@ -192,6 +233,8 @@ def instrument_agentops():
 
 def uninstrument_agentops():
     """Uninstrument agentops to stop capturing token IDs."""
+    _unpatch_exporters()
+
     try:
         _unpatch_new_agentops()
     except Exception:
@@ -202,114 +245,70 @@ def uninstrument_agentops():
         pass
 
 
-def agentops_local_server():
+class BypassableAuthenticatedOTLPExporter(LightningStoreOTLPExporter, AuthenticatedOTLPExporter):
     """
-    Returns a Flask app that can be used to test agentops integration.
-    This server provides endpoints for token fetching and a catch-all endpoint.
+    AuthenticatedOTLPExporter with switchable service control.
+
+    When `_agentops_service_enabled` is False, skip export and return success.
     """
-    app = flask.Flask(__name__)
 
-    @app.route("/v3/auth/token", methods=["POST"])
-    def fetch_token():  # type: ignore
-        return {"token": "dummy", "project_id": "dummy"}
-
-    @app.route("/", defaults={"path": ""}, methods=["GET", "POST"])
-    @app.route("/<path:path>", methods=["GET", "POST"])
-    def catch_all(path: str):  # type: ignore
-        return {"path": path}
-
-    return app
+    def should_bypass(self) -> bool:
+        return not _agentops_service_enabled
 
 
-def _run_server(**kwargs: Any):  # type: ignore
+class BypassableOTLPMetricExporter(OTLPMetricExporter):
     """
-    Internal function to run the Flask server.
-    This is used to avoid issues with multiprocessing and Flask's reloader.
+    OTLPMetricExporter with switchable service control.
+    When `_agentops_service_enabled` is False, skip export and return success.
     """
-    signal.signal(signal.SIGINT, signal.SIG_IGN)  # Ignore SIGINT in worker processes
-    setproctitle.setproctitle(multiprocessing.current_process().name)
-    app = agentops_local_server()
-    app.run(**kwargs)
 
-
-class AgentOpsServerManager:
-    """Manages a AgentOps local server to bypass the online service of AgentOps."""
-
-    def __init__(self, daemon: bool = True, port: int | None = None):
-        self.server_process: multiprocessing.Process | None = None
-        self.server_port = port
-        self.daemon = daemon
-        logger.info("AgentOpsServerManager initialized.")
-
-    def _find_available_port(self) -> int:
-        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
-            s.bind(("", 0))
-            return s.getsockname()[1]
-
-    def start(self):
-        if self.server_process and self.server_process.is_alive():
-            logger.warning("AgentOps server process appears to be already running.")
-            return
-
-        if self.server_port is None:
-            self.server_port = self._find_available_port()
-
-        logger.info(f"Starting AgentOps local server on port {self.server_port}...")
-
-        self.server_process = multiprocessing.Process(
-            target=_run_server,
-            kwargs={"host": "127.0.0.1", "port": self.server_port, "use_reloader": False, "debug": False},
-            daemon=self.daemon,
-            name="AgentLightning-AgentOpsServer",
-        )
-        self.server_process.start()
-        logger.info(
-            f"AgentOps local server process (PID: {self.server_process.pid}) started, targeting port {self.server_port}."
-        )
-        for attempt in range(20):  # 10 seconds total
-            time.sleep(0.5)  # Brief wait for server to start up
-            try:
-                result = requests.get(f"http://127.0.0.1:{self.server_port}/")
-                if result.status_code == 200:
-                    break
-            except Exception as e:
-                logger.debug(f"Error checking AgentOps server: {e}")
-            logger.warning(f"AgentOps still not ready after {attempt} attempts. Retrying...")
+    def export(self, *args: Any, **kwargs: Any) -> MetricExportResult:
+        if _agentops_service_enabled:
+            return super().export(*args, **kwargs)  # type: ignore[reportUnknownMemberType]
         else:
-            logger.error(f"AgentOps local server failed to start or exited prematurely.")
-            return
+            logger.debug("SwitchableOTLPMetricExporter is switched off, skipping export.")
+            return MetricExportResult.SUCCESS
 
-        if not self.server_process.is_alive():
-            logger.error(f"AgentOps local server failed to start or exited prematurely.")
 
-    def is_alive(self) -> bool:
-        if self.server_process and self.server_process.is_alive():
-            return True
-        return False
+class BypassableOTLPSpanExporter(LightningStoreOTLPExporter):
+    """
+    OTLPSpanExporter with switchable service control.
+    When `_agentops_service_enabled` is False, skip export and return success.
 
-    def stop(self):
-        if self.server_process is not None and self.server_process.is_alive():
-            logger.info(f"Stopping AgentOps local server (PID: {self.server_process.pid})...")
-            self.server_process.terminate()  # Send SIGTERM
-            self.server_process.join(timeout=5)  # Wait for clean exit
-            if self.server_process.is_alive():
-                logger.warning(
-                    f"AgentOps server (PID: {self.server_process.pid}) did not terminate gracefully, killing..."
-                )
-                self.server_process.kill()  # Force kill
-                self.server_process.join(timeout=10)  # Wait for kill
-            self.server_process = None
-            logger.info(f"AgentOps local server stopped.")
+    This is used instead of BypassableAuthenticatedOTLPExporter on legacy AgentOps versions.
+    """
+
+    def should_bypass(self) -> bool:
+        return not _agentops_service_enabled
+
+
+class BypassableV3Client(V3Client):
+    """
+    V3Client with toggleable authentication calls.
+    Returns dummy auth response when `_agentops_service_enabled` is False.
+    """
+
+    # Temporary synchronous override of fetch_auth_token for mock purposes.
+    def fetch_auth_token(self, *args: Any, **kwargs: Any) -> AuthTokenResponse:  # type: ignore[override]
+        if _agentops_service_enabled:
+            return super().fetch_auth_token(*args, **kwargs)  # type: ignore[override]
         else:
-            logger.info("AgentOps local server was not running or already stopped.")
+            logger.debug("SwitchableV3Client is switched off, skipping fetch_auth_token request.")
+            return AuthTokenResponse(token="dummy", project_id="dummy")
 
-    def get_port(self) -> int | None:
-        # Check liveness again in case it died since start()
-        if self.is_alive() and self.server_port is not None:
-            return self.server_port
-        # If called after server stopped or failed, port might be stale or None
-        if self.server_port is not None and (self.server_process is None or not self.server_process.is_alive()):
-            logger.warning(
-                f"AgentOps server port {self.server_port} is stored, but server process is not alive. Returning stored port."
-            )
-        return self.server_port
+
+class BypassableV4Client(V4Client):
+    """
+    V4Client with toggleable post requests.
+    Returns dummy response when `_agentops_service_enabled` is False.
+    """
+
+    def post(self, *args: Any, **kwargs: Any) -> requests.Response:
+        if _agentops_service_enabled:
+            return super().post(*args, **kwargs)
+        else:
+            logger.debug("SwitchableV4Client is switched off, skipping post request.")
+            response = requests.Response()
+            response.status_code = 200
+            response._content = b"{}"
+            return response

agentlightning/logging.py

@@ -1,10 +1,18 @@
 # Copyright (c) Microsoft. All rights reserved.
 
+from __future__ import annotations
+
 import logging
 import os
 import platform
+import sys
+import warnings
+from logging.config import dictConfig
+from typing import Any, Dict, Optional
 
-__all__ = ["configure_logger"]
+from rich.console import Console
+
+__all__ = ["setup", "configure_logger", "setup_module"]
 
 
 def configure_logger(level: int = logging.INFO, name: str = "agentlightning") -> logging.Logger:
@@ -15,6 +23,10 @@ def configure_logger(level: int = logging.INFO, name: str = "agentlightning") ->
     not propagate to the root logger, preventing duplicate log emission when
     applications compose multiple logging configurations.
 
+    !!! danger
+
+        This function is deprecated in favor of [`setup_logging`][agentlightning.setup_logging].
+
     Args:
         level: Logging level applied both to the logger and the installed
             handler. Defaults to `logging.INFO`.
@@ -32,23 +44,327 @@ def configure_logger(level: int = logging.INFO, name: str = "agentlightning") ->
         logger.info("agent-lightning is ready!")
         ```
     """
+    warnings.warn("This function is deprecated in favor of `setup_logging`.", DeprecationWarning, stacklevel=2)
 
+    return setup_module(level=level, name=name, console=True, color=True, propagate=False)
+
+
+DEFAULT_FORMAT = "%(asctime)s [%(levelname)s] (Process-%(process)d %(name)s)   %(message)s"
+DATE_FORMAT = "%H:%M:%S"
+
+
+def _to_level_value(lvl: int | str) -> int:
+    if isinstance(lvl, int):
+        return lvl
+    val = getattr(logging, str(lvl).upper(), None)
+    if val is None:
+        raise ValueError(f"Invalid log level: {lvl}")
+    return val
+
+
+def _ensure_file_handler(
+    logger: logging.Logger,
+    filename: str,
+    *,
+    level: int,
+    formatter: Optional[logging.Formatter],
+) -> None:
+    """Attach a FileHandler to `logger` for `filename` if it doesn't already exist."""
+    abspath = os.path.abspath(filename)
+
+    # Avoid duplicates
+    for h in logger.handlers:
+        if isinstance(h, logging.FileHandler) and getattr(h, "baseFilename", None) == abspath:
+            return
+
+    # Ensure directory exists
+    dirname = os.path.dirname(abspath)
+    if dirname:
+        os.makedirs(dirname, exist_ok=True)
+
+    fh = logging.FileHandler(abspath, encoding="utf-8")
+    fh.setLevel(level)
+    if formatter is not None:
+        fh.setFormatter(formatter)
+    else:
+        fh.setFormatter(logging.Formatter(DEFAULT_FORMAT, DATE_FORMAT))
+
+    logger.addHandler(fh)
+
+
+def setup(
+    level: int | str = "INFO",
+    *,
+    console: bool = True,
+    color: bool | Dict[str, Any] = True,
+    propagate: bool = False,
+    disable_existing_loggers: bool = False,
+    capture_warnings: bool = False,
+    submodule_levels: Optional[dict[str, int | str]] = None,
+    extra_handlers: Optional[list[logging.Handler]] = None,
+    formatter: Optional[logging.Formatter] = None,
+    apply_to: Optional[list[str]] = None,
+    files: Optional[str | dict[str, str]] = None,
+) -> None:
+    """Configures logging for the `agentlightning` logger hierarchy.
+
+    This function provides a one-stop setup utility for configuring the
+    `agentlightning` root logger and optionally its submodules or external
+    loggers. It supports console logging, colored rich output, per-submodule
+    log levels, and optional handler/formatter injection.
+
+    The setup is intentionally isolated: it does not modify the global root
+    logger or loggers belonging to other libraries unless explicitly directed
+    via `apply_to`.
+
+    Args:
+        level:
+            Logging level for the base `agentlightning` logger. Accepts either
+            an integer (e.g., `logging.DEBUG`) or a string level name
+            (e.g., `"INFO"`). Defaults to `"INFO"`.
+        console:
+            Whether to attach a console handler to the logger. Defaults to
+            `True`.
+        color:
+            Enables rich-formatted output using `RichHandler` when `True`
+            or a configuration dict. If `False`, a plain text formatter is
+            used instead. Defaults to `True`.
+        propagate:
+            Whether `agentlightning` logs should propagate to ancestor
+            loggers. Defaults to `False`.
+        disable_existing_loggers:
+            Passed to `logging.config.dictConfig`. If `True`, disables all
+            existing configured loggers before applying this configuration.
+            Defaults to `False`.
+        capture_warnings:
+            If `True`, redirects Python `warnings` emitted via the `warnings`
+            module into the logging system. Defaults to `False`.
+        submodule_levels:
+            Mapping of submodule logger names to logging levels. If a specified
+            submodule level is more verbose than the base level, a warning is emitted.
+        extra_handlers:
+            A list of user-provided handlers to attach to the `agentlightning` logger.
+            Handlers are added idempotently; duplicates are not reattached.
+        formatter:
+            A formatter to apply to any handler under `agentlightning` that does not
+            already have one assigned. Useful for customizing output without overwriting
+            formatters on custom handlers.
+        apply_to:
+            A list of additional logger names to configure identically to
+            `agentlightning` base logger. Their handlers are replaced with copies of the base
+            handlers, and propagation is disabled to avoid duplicate log emission.
+        files:
+            If a string, attach a FileHandler to the base `agentlightning` logger.
+            If a dict, for each `(logger_name, filename)` pair, attach a FileHandler
+            directly to that logger.
+            Each file handler should use the logger's effective level at creation.
+
+    Notes:
+        * On Windows, this function forces UTF-8 mode in the console to prevent
+          issues with rich output or special characters.
+        * Submodule loggers can generate records below the handler's emission
+          threshold. Whether such records appear depends on both the logger's
+          level and the handler's level.
+        * `apply_to` loggers inherit the same handlers but do not propagate
+          upward, yielding isolated, consistent behavior.
+
+    Examples:
+        Basic setup:
+
+        >>> setup()
+
+        Enabling debug mode with no color:
+
+        >>> setup(level="DEBUG", color=False)
+
+        Overriding specific submodule levels:
+
+        >>> setup(submodule_levels={"agentlightning.io": "DEBUG"})
+
+        Attaching an additional file handler:
+
+        >>> fh = logging.FileHandler("app.log")
+        >>> setup(extra_handlers=[fh])
+    """
     # Ensure UTF-8 encoding on Windows consoles
-    # Note: This change does not fully represent support for execution under the windown system.
+    # Note: This change does not fully represent support for execution under the windows system.
     # It only fixes console printing issues caused by special characters.
     # TODO: More comprehensive Windows support may be needed in the future.
     if platform.system() == "Windows":
         os.environ["PYTHONUTF8"] = "1"
 
-    logger = logging.getLogger(name)
-    logger.handlers.clear()  # clear existing handlers
+    base_logger = setup_module(
+        level,
+        name="agentlightning",
+        console=console,
+        color=color,
+        propagate=propagate,
+        disable_existing_loggers=disable_existing_loggers,
+    )
 
-    # log to stdout
-    handler = logging.StreamHandler()
-    handler.setLevel(level)
-    formatter = logging.Formatter("%(asctime)s [%(levelname)s] (Process-%(process)d %(name)s)   %(message)s")
-    handler.setFormatter(formatter)
-    logger.addHandler(handler)
-    logger.setLevel(level)
-    logger.propagate = False  # prevent double logging
-    return logger
+    base_level_value = base_logger.level
+
+    # Apply user-provided formatter (only to handlers without one,
+    # so we don't clobber custom extra_handlers)
+    if formatter is not None:
+        for h in base_logger.handlers:
+            if h.formatter is None:
+                h.setFormatter(formatter)
+
+    # Attach user-provided handler(s) if any, idempotently
+    if extra_handlers:
+        for h in extra_handlers:
+            if h not in base_logger.handlers:
+                base_logger.addHandler(h)
+
+    # Per-submodule levels
+    if submodule_levels:
+        for name, lvl in submodule_levels.items():
+            sub_level = _to_level_value(lvl)
+
+            # Emit a warning if submodule level is lower (more verbose) than the global/base level
+            if sub_level < base_level_value:
+                base_logger.warning(
+                    "Submodule logger '%s' level %s (%s) is more verbose than base "
+                    "logger level %s (%s). Records below the base level may still be "
+                    "filtered out by handlers depending on their own levels.",
+                    name,
+                    lvl,
+                    sub_level,
+                    logging.getLevelName(base_level_value),
+                    base_level_value,
+                )
+
+            # The logger will *create* records down to the logger's level, but a handler
+            # with a higher level will still drop anything below its own threshold.
+            # Effective emission is gated by both: record.level >= logger.level AND handler.level.
+            logging.getLogger(name).setLevel(lvl)
+
+    # Attach file handlers if requested
+    if files is not None:
+        if isinstance(files, str):
+            # Single file for the entire `agentlightning` hierarchy.
+            _ensure_file_handler(
+                logger=base_logger,
+                filename=files,
+                level=base_level_value,
+                formatter=formatter,
+            )
+        else:
+            # Per-logger files
+            for logger_name, filename in files.items():
+                lg = logging.getLogger(logger_name)
+                # Use the logger's *effective* level at creation time
+                effective_level = lg.getEffectiveLevel()
+                _ensure_file_handler(
+                    logger=lg,
+                    filename=filename,
+                    level=effective_level,
+                    formatter=formatter,
+                )
+
+    # Optionally apply the same handler setup to other loggers outside this module
+    if apply_to:
+        for name in apply_to:
+            lg = logging.getLogger(name)
+            # This removes any existing handlers so we don't duplicate output
+            # and ensures these loggers share exactly the same handlers as base_logger.
+            lg.handlers.clear()
+            for h in base_logger.handlers:
+                lg.addHandler(h)
+            lg.setLevel(base_logger.level)
+            # We've attached handlers directly to these loggers; if propagate
+            # stayed True, records would bubble up to ancestor loggers and could be
+            # emitted twice (here and on the parent/root). Setting False isolates them.
+            lg.propagate = False
+
+    # Optionally capture warnings
+    if capture_warnings:
+        logging.captureWarnings(True)
+
+
+def setup_module(
+    level: int | str = "INFO",
+    *,
+    name: str = "agentlightning",
+    console: bool = True,
+    color: bool | Dict[str, Any] = True,
+    propagate: bool = False,
+    disable_existing_loggers: bool = False,
+) -> logging.Logger:
+    """Initializes and returns the base logger for `agentlightning`.
+
+    This function constructs and applies a `dictConfig` configuration for the
+    logger hierarchy rooted at `name`. It supports either rich console
+    formatting (via `RichHandler`) or plain text formatting, based on the
+    `color` argument.
+
+    Unlike [`setup_logging`][agentlightning.setup_logging], this function configures only a single logger namespace
+    and does not attach extra handlers or submodule levels. It is primarily used
+    internally by [`setup_logging`][agentlightning.setup_logging] but is also suitable for direct integration in
+    custom logging workflows.
+    """
+    root_cfg: Dict[str, Any] = {
+        "version": 1,
+        "disable_existing_loggers": disable_existing_loggers,
+        "loggers": {
+            name: {
+                "handlers": [],
+                "level": level,
+                "propagate": propagate,
+            }
+        },
+        "handlers": {},
+        "formatters": {},
+    }
+
+    # Choose formatter / handler definition
+    if color is not False and console:
+        # Console must be true to display colored outputs
+        if isinstance(color, dict):
+            rich_handler_config = color
+        else:
+            rich_handler_config: Dict[str, Any] = {
+                "rich_tracebacks": False,
+                "markup": False,
+                "show_time": True,
+                "show_path": True,
+            }
+
+            if not _has_width():
+                # e.g., in a CI environment.
+                rich_handler_config["console"] = Console(width=200)
+
+        root_cfg["handlers"]["console"] = {
+            "class": "rich.logging.RichHandler",
+            "level": level,
+            **rich_handler_config,
+        }
+        # RichHandler manages its own style; keep formatter None
+    else:
+        fmt_name = "plain"
+        root_cfg["formatters"][fmt_name] = {
+            "format": DEFAULT_FORMAT,
+            "datefmt": DATE_FORMAT,
+        }
+
+        if console:
+            root_cfg["handlers"]["console"] = {
+                "class": "logging.StreamHandler",
+                "level": level,
+                "formatter": fmt_name,
+            }
+
+    # Attach selected handlers to agentlightning
+    handler_names = list(root_cfg["handlers"].keys())
+    root_cfg["loggers"][name]["handlers"] = handler_names
+
+    # Apply dictConfig (this resets the logger handlers)
+    dictConfig(root_cfg)
+
+    return logging.getLogger(name)
+
+
+def _has_width() -> bool:
+    """Automatically determine whether the terminal has a width."""
+    return sys.stdout.isatty()

agentlightning/runner/agent.py

@@ -11,8 +11,22 @@ from __future__ import annotations
 
 import asyncio
 import logging
+import random
+import threading
 import time
-from typing import TYPE_CHECKING, Any, List, Literal, Optional, Sequence, TypeVar, cast
+from contextlib import suppress
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Awaitable,
+    Callable,
+    List,
+    Literal,
+    Optional,
+    Sequence,
+    TypeVar,
+    cast,
+)
 
 from opentelemetry.sdk.trace import ReadableSpan
 
@@ -30,6 +44,7 @@ from agentlightning.types import (
     RolloutRawResult,
     Span,
 )
+from agentlightning.utils.system_snapshot import system_snapshot
 
 if TYPE_CHECKING:
     from agentlightning.execution.events import ExecutionEvent
@@ -52,7 +67,15 @@ class LitAgentRunner(Runner[T_task]):
         worker_id: Identifier for the active worker process, if any.
     """
 
-    def __init__(self, tracer: Tracer, max_rollouts: Optional[int] = None, poll_interval: float = 5.0) -> None:
+    def __init__(
+        self,
+        tracer: Tracer,
+        max_rollouts: Optional[int] = None,
+        poll_interval: float = 5.0,
+        heartbeat_interval: float = 10.0,
+        interval_jitter: float = 0.1,
+        heartbeat_launch_mode: Literal["asyncio", "thread"] = "asyncio",
+    ) -> None:
         """Initialize the agent runner.
 
         Args:
@@ -60,11 +83,21 @@ class LitAgentRunner(Runner[T_task]):
             max_rollouts: Optional cap on iterations processed by
                 [`iter`][agentlightning.LitAgentRunner.iter].
             poll_interval: Seconds to wait between store polls when no work is available.
+            heartbeat_interval: Seconds to wait between sending heartbeats to the store.
+            interval_jitter: Jitter factor for the poll interval. The actual interval will be between
+                poll_interval - interval_jitter and poll_interval + interval_jitter.
+                This is to avoid the overload caused by the synchronization of the runners.
+            heartbeat_launch_mode: Launch mode for the heartbeat loop. Can be "asyncio" or "thread".
+                "asyncio" is the default and recommended mode. Use "thread" if you are experiencing blocking coroutines.
         """
         super().__init__()
         self._tracer = tracer
         self._max_rollouts = max_rollouts
         self._poll_interval = poll_interval
+        self._heartbeat_interval = heartbeat_interval
+        self._interval_jitter = interval_jitter
+        self._heartbeat_launch_mode = heartbeat_launch_mode
+        self._random_state = random.Random()
 
         # Set later
         self._agent: Optional[LitAgent[T_task]] = None
@@ -105,7 +138,7 @@ class LitAgentRunner(Runner[T_task]):
         self._store = store
         self.worker_id = worker_id
 
-        self._tracer.init_worker(worker_id)
+        self._tracer.init_worker(worker_id, store)
 
     def teardown(self, *args: Any, **kwargs: Any) -> None:
         """Teardown the runner and clean up all resources.
@@ -253,8 +286,9 @@ class LitAgentRunner(Runner[T_task]):
         if isinstance(raw_result, float):
             # Preserve the existing spans before another span is emitted
             trace_spans = list(self._tracer.get_last_trace())
-            # This will emit another span to the tracer
-            reward_span = emit_reward(raw_result)
+            # This will NOT emit another span to the tracer
+            reward_span = emit_reward(raw_result, propagate=False)
+            # We add it to the store manually
             await store.add_otel_span(rollout.rollout_id, rollout.attempt.attempt_id, reward_span)
             trace_spans.append(reward_span)
 
@@ -304,6 +338,75 @@ class LitAgentRunner(Runner[T_task]):
 
         return trace_spans
 
+    async def _emit_heartbeat(self, store: LightningStore) -> None:
+        """Send a heartbeat tick to the store."""
+        worker_id = self.get_worker_id()
+
+        try:
+            await store.update_worker(worker_id, system_snapshot())
+        except asyncio.CancelledError:
+            # bypass the exception
+            raise
+        except Exception:
+            logger.exception("%s Unable to update worker heartbeat.", self._log_prefix())
+
+    def _start_heartbeat_loop(self, store: LightningStore) -> Optional[Callable[[], Awaitable[None]]]:
+        """Start a background heartbeat loop and return an async stopper."""
+
+        if self._heartbeat_interval <= 0:
+            return None
+
+        if self.worker_id is None:
+            logger.warning("%s Cannot start heartbeat loop without worker_id.", self._log_prefix())
+            return None
+
+        if self._heartbeat_launch_mode == "asyncio":
+            stop_event = asyncio.Event()
+
+            async def heartbeat_loop() -> None:
+                while not stop_event.is_set():
+                    await self._emit_heartbeat(store)
+                    with suppress(asyncio.TimeoutError):
+                        interval = self._heartbeat_interval + self._random_state.uniform(
+                            -self._interval_jitter, self._interval_jitter
+                        )
+                        interval = max(interval, 0.01)
+                        await asyncio.wait_for(stop_event.wait(), timeout=interval)
+
+            task = asyncio.create_task(heartbeat_loop(), name=f"{self.get_worker_id()}-heartbeat")
+
+            async def stop() -> None:
+                stop_event.set()
+                with suppress(asyncio.CancelledError):
+                    await task
+
+            return stop
+
+        if self._heartbeat_launch_mode == "thread":
+            stop_evt = threading.Event()
+
+            def thread_worker() -> None:
+                loop = asyncio.new_event_loop()
+                asyncio.set_event_loop(loop)
+                while not stop_evt.is_set():
+                    loop.run_until_complete(self._emit_heartbeat(store))
+                    interval = self._heartbeat_interval + self._random_state.uniform(
+                        -self._interval_jitter, self._interval_jitter
+                    )
+                    interval = max(interval, 0.01)
+                    stop_evt.wait(interval)
+
+            thread = threading.Thread(target=thread_worker, name=f"{self.get_worker_id()}-heartbeat", daemon=True)
+            thread.start()
+
+            async def stop() -> None:
+                stop_evt.set()
+                await asyncio.to_thread(thread.join)
+
+            return stop
+
+        raise ValueError(f"Unsupported heartbeat launch mode: {self._heartbeat_launch_mode}")
+
     async def _sleep_until_next_poll(self, event: Optional[ExecutionEvent] = None) -> None:
         """Sleep until the next poll interval, with optional event-based interruption.
 
@@ -314,11 +417,13 @@ class LitAgentRunner(Runner[T_task]):
             event: Optional [`ExecutionEvent`][agentlightning.ExecutionEvent] object that can be used to interrupt the sleep.
                 If set during the sleep period, the method returns immediately.
         """
+        interval = self._poll_interval + self._random_state.uniform(-self._interval_jitter, self._interval_jitter)
+        interval = max(interval, 0.01)
         if event is None:
-            await asyncio.sleep(self._poll_interval)
+            await asyncio.sleep(interval)
             return
         current_time = time.time()
-        next_time = current_time + self._poll_interval
+        next_time = current_time + interval
         while time.time() < next_time:
             await asyncio.sleep(0.1)
             if event.is_set():
@@ -364,7 +469,7 @@ class LitAgentRunner(Runner[T_task]):
 
             start_time = time.time()
             async with self._tracer.trace_context(
-                name=rollout_id, store=store, rollout_id=rollout_id, attempt_id=next_rollout.attempt.attempt_id
+                name=rollout_id, rollout_id=rollout_id, attempt_id=next_rollout.attempt.attempt_id
             ):
                 await self._trigger_hooks(
                     hook_type="on_trace_start", agent=agent, runner=self, tracer=self._tracer, rollout=next_rollout
@@ -450,39 +555,49 @@ class LitAgentRunner(Runner[T_task]):
         logger.info(f"{self._log_prefix()} Started async rollouts (max: {self._max_rollouts or 'unlimited'}).")
         store = self.get_store()
 
-        while not (event is not None and event.is_set()) and (
-            self._max_rollouts is None or num_tasks_processed < self._max_rollouts
-        ):
-            # Retrieve the next rollout
-            next_rollout: Optional[Rollout] = None
-            while not (event is not None and event.is_set()):
-                logger.debug(f"{self._log_prefix()} Try to poll for next rollout.")
-                next_rollout = await store.dequeue_rollout()
+        stop_heartbeat = self._start_heartbeat_loop(store)
+
+        try:
+            while not (event is not None and event.is_set()) and (
+                self._max_rollouts is None or num_tasks_processed < self._max_rollouts
+            ):
+                # Retrieve the next rollout
+                next_rollout: Optional[Rollout] = None
+                while not (event is not None and event.is_set()):
+                    logger.debug(f"{self._log_prefix()} Try to poll for next rollout.")
+                    next_rollout = await store.dequeue_rollout(worker_id=self.get_worker_id())
+                    if next_rollout is None:
+                        logger.debug(
+                            f"{self._log_prefix()} No rollout to poll. Waiting for {self._poll_interval} seconds."
+                        )
+                        await self._sleep_until_next_poll(event)
+                    else:
+                        break
+
                 if next_rollout is None:
-                    logger.debug(f"{self._log_prefix()} No rollout to poll. Waiting for {self._poll_interval} seconds.")
-                    await self._sleep_until_next_poll(event)
-                else:
-                    break
+                    return
 
-            if next_rollout is None:
-                return
+                try:
+                    # Claim the rollout but updating the current worker id
+                    await store.update_attempt(
+                        next_rollout.rollout_id, next_rollout.attempt.attempt_id, worker_id=self.get_worker_id()
+                    )
+                except Exception:
+                    # This exception could happen if the rollout is dequeued and the other end died for some reason
+                    logger.exception(f"{self._log_prefix()} Exception during update_attempt, giving up the rollout.")
+                    continue
 
-            try:
-                # Claim the rollout but updating the current worker id
-                await store.update_attempt(
-                    next_rollout.rollout_id, next_rollout.attempt.attempt_id, worker_id=self.get_worker_id()
-                )
-            except Exception:
-                # This exception could happen if the rollout is dequeued and the other end died for some reason
-                logger.exception(f"{self._log_prefix()} Exception during update_attempt, giving up the rollout.")
-                continue
+                # Execute the step
+                await self._step_impl(next_rollout)
 
-            # Execute the step
-            await self._step_impl(next_rollout)
-
-            num_tasks_processed += 1
-            if num_tasks_processed % 10 == 0 or num_tasks_processed == 1:
-                logger.info(f"{self._log_prefix()} Progress: {num_tasks_processed}/{self._max_rollouts or 'unlimited'}")
+                num_tasks_processed += 1
+                if num_tasks_processed % 10 == 0 or num_tasks_processed == 1:
+                    logger.info(
+                        f"{self._log_prefix()} Progress: {num_tasks_processed}/{self._max_rollouts or 'unlimited'}"
+                    )
+        finally:
+            if stop_heartbeat is not None:
+                await stop_heartbeat()
 
         logger.info(f"{self._log_prefix()} Finished async rollouts. Processed {num_tasks_processed} tasks.")
 
@@ -526,6 +641,12 @@ class LitAgentRunner(Runner[T_task]):
             resources_id = None
 
         attempted_rollout = await self.get_store().start_rollout(input=input, mode=mode, resources_id=resources_id)
+        # Register the attempt as running by the current worker
+        await self.get_store().update_attempt(
+            attempted_rollout.rollout_id,
+            attempted_rollout.attempt.attempt_id,
+            worker_id=self.get_worker_id(),
+        )
         rollout_id = await self._step_impl(attempted_rollout, raise_on_exception=True)
 
         completed_rollout = await store.get_rollout_by_id(rollout_id)

agentlightning/semconv.py

@@ -0,0 +1,158 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Semantic conventions for Agent-lightning spans.
+
+Conventions in this file are added on demand. We generally DO NOT add
+new semantic conventions unless it's absolutely needed for certain algorithms or scenarios.
+"""
+
+from enum import Enum
+
+from pydantic import BaseModel
+
+AGL_ANNOTATION = "agentlightning.annotation"
+"""Agent-lightning's standard span name for annotations.
+
+Annotations are minimal span units for rewards, tags, and metadatas.
+They are used to "annotate" a specific event or a part of rollout.
+"""
+
+AGL_MESSAGE = "agentlightning.message"
+"""Agent-lightning's standard span name for messages and logs."""
+
+AGL_OBJECT = "agentlightning.object"
+"""Agent-lightning's standard span name for customized objects."""
+
+AGL_EXCEPTION = "agentlightning.exception"
+"""Agent-lightning's standard span name for exceptions.
+
+Used by the exception emitter to record exception details.
+"""
+
+AGL_OPERATION = "agentlightning.operation"
+"""Agent-lightning's standard span name for functions.
+Wrap function or code-blocks as operations.
+"""
+
+AGL_VIRTUAL = "agentlightning.virtual"
+"""Agent-lightning's standard span name for virtual operations.
+
+Mostly used in adapter when needing to represent the root or intermediate operations.
+"""
+
+
+class LightningResourceAttributes(Enum):
+    """Resource attribute names used in Agent-lightning spans."""
+
+    ROLLOUT_ID = "agentlightning.rollout_id"
+    """Resource name for rollout ID in Agent-lightning spans."""
+
+    ATTEMPT_ID = "agentlightning.attempt_id"
+    """Resource name for attempt ID in Agent-lightning spans."""
+
+    SPAN_SEQUENCE_ID = "agentlightning.span_sequence_id"
+    """Resource name for span sequence ID in Agent-lightning spans."""
+
+
+class LightningSpanAttributes(Enum):
+    """Attribute names that commonly appear in Agent-lightning spans.
+
+    Exception types can't be found here because they are defined in OpenTelemetry's official semantic conventions.
+    """
+
+    REWARD = "agentlightning.reward"
+    """Attribute prefix for rewards-related data in reward spans.
+
+    It should be used as a prefix. For example, "agentlightning.reward.0.value" can
+    be used to track a specific metric. See [RewardAttributes][agentlightning.semconv.RewardAttributes].
+    """
+
+    LINK = "agentlightning.link"
+    """Attribute name for linking the current span to another span or other objects like requests/responses."""
+
+    TAG = "agentlightning.tag"
+    """Attribute name for tagging spans with customized strings."""
+
+    MESSAGE_BODY = "agentlightning.message.body"
+    """Attribute name for message text in message spans."""
+
+    OBJECT_TYPE = "agentlightning.object.type"
+    """Attribute name for object type (full qualified name) in object spans.
+
+    I think builtin types like str, int, bool, list, dict are self-explanatory and
+    should also be qualified to use here.
+    """
+
+    OBJECT_LITERAL = "agentlightning.object.literal"
+    """Attribute name for object literal value in object spans (for str, int, bool, ...)."""
+
+    OBJECT_JSON = "agentlightning.object.json"
+    """Attribute name for object serialized value (JSON) in object spans."""
+
+    OPERATION_NAME = "agentlightning.operation.name"
+    """Attribute name for operation name in operation spans, normally the function name."""
+
+    OPERATION_INPUT = "agentlightning.operation.input"
+    """Attribute name for operation input in operation spans."""
+
+    OPERATION_OUTPUT = "agentlightning.operation.output"
+    """Attribute name for operation output in operation spans."""
+
+
+class RewardAttributes(Enum):
+    """Multi-dimensional reward attributes will look like:
+
+    ```json
+    {"agentlightning.reward.0.name": "efficiency", "agentlightning.reward.0.value": 0.75}
+    ```
+
+    The first reward in the reward list will automatically be the primary reward.
+    If the reward list has greater than 1, it shall be a multi-dimensional case.
+    """
+
+    REWARD_NAME = "name"
+    """Key for each dimension in multi-dimensional reward spans."""
+
+    REWARD_VALUE = "value"
+    """Value for each dimension in multi-dimensional reward spans."""
+
+
+class RewardPydanticModel(BaseModel):
+    """A stricter implementation of RewardAttributes used in otel helpers."""
+
+    name: str
+    """Name of the reward dimension."""
+
+    value: float
+    """Value of the reward dimension."""
+
+
+class LinkAttributes(Enum):
+    """Standard link types used in Agent-lightning spans.
+
+    The link is more powerful than [OpenTelemetry link](https://opentelemetry.io/docs/specs/otel/trace/api/#link)
+    in that it supports linking to a queryset of spans.
+    It can even link to span object that hasn't been emitted yet.
+    """
+
+    KEY_MATCH = "key_match"
+    """Linking to spans with matching attribute keys.
+
+    `trace_id` and `span_id` are reserved and will be used to link to specific spans directly.
+
+    For example, it can be `gen_ai.response.id` if intended to be link to a chat completion response span.
+    Or it can be `span_id` to link to a specific span by its ID.
+    """
+
+    VALUE_MATCH = "value_match"
+    """Linking to spans with corresponding attribute values on those keys."""
+
+
+class LinkPydanticModel(BaseModel):
+    """A stricter implementation of LinkAttributes used in otel helpers."""
+
+    key_match: str
+    """The attribute key to match on the target spans."""
+
+    value_match: str
+    """The attribute value to match on the target spans."""

agentlightning/server.py

@@ -142,7 +142,13 @@ class ServerDataStore:
         async with self._resources_lock:
             resources = self._resource_versions.get(resources_id)
             if resources:
-                return ResourcesUpdate(resources_id=resources_id, resources=resources)
+                return ResourcesUpdate(
+                    resources_id=resources_id,
+                    resources=resources,
+                    create_time=time.time(),
+                    update_time=time.time(),
+                    version=1,
+                )
             return None
 
     async def get_latest_resources(self) -> Optional[ResourcesUpdate]:
@@ -357,7 +363,9 @@ class AgentLightningServer:
         if not self._store:
             raise RuntimeError("Store not initialized. The server may not be running.")
         resources_id = f"res-{uuid.uuid4()}"
-        update = ResourcesUpdate(resources_id=resources_id, resources=resources)
+        update = ResourcesUpdate(
+            resources_id=resources_id, resources=resources, create_time=time.time(), update_time=time.time(), version=1
+        )
         await self._store.update_resources(update)
         return resources_id
 

agentlightning/store/__init__.py

@@ -1,14 +1,18 @@
 # Copyright (c) Microsoft. All rights reserved.
 
-from .base import LightningStore
+from .base import LightningStore, LightningStoreCapabilities, LightningStoreStatistics
 from .client_server import LightningStoreClient, LightningStoreServer
+from .collection_based import CollectionBasedLightningStore
 from .memory import InMemoryLightningStore
 from .threading import LightningStoreThreaded
 
 __all__ = [
     "LightningStore",
+    "LightningStoreCapabilities",
+    "LightningStoreStatistics",
     "LightningStoreClient",
     "LightningStoreServer",
     "InMemoryLightningStore",
+    "CollectionBasedLightningStore",
     "LightningStoreThreaded",
 ]

agentlightning/store/base.py

@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-from typing import Any, Dict, List, Literal, Optional, Sequence
+from typing import Any, Dict, List, Literal, Optional, Sequence, Tuple, TypedDict
 
 from opentelemetry.sdk.trace import ReadableSpan
 
@@ -14,9 +14,12 @@ from agentlightning.types import (
     ResourcesUpdate,
     Rollout,
     RolloutConfig,
+    RolloutMode,
     RolloutStatus,
     Span,
     TaskInput,
+    Worker,
+    WorkerStatus,
 )
 
 
@@ -52,6 +55,64 @@ UNSET = _UnsetType()
 Unset = _UnsetType  # Alias for convenience
 
 
+class LightningStoreCapabilities(TypedDict, total=False):
+    """Capability of a LightningStore implementation.
+
+    All keys are optional and false by default.
+    """
+
+    thread_safe: bool
+    """Whether the store is thread-safe."""
+    async_safe: bool
+    """Whether the store is async-safe."""
+    zero_copy: bool
+    """Whether the store has only one copy across all threads/processes."""
+    otlp_traces: bool
+    """Whether the store supports OTLP/HTTP traces."""
+
+
+class LightningStoreStatistics(TypedDict, total=False):
+    """Statistics of a LightningStore implementation."""
+
+    name: str
+    """Name of the store implementation."""
+    total_rollouts: int
+    """Total number of rollouts in the store."""
+    total_attempts: int
+    """Total number of attempts in the store."""
+    total_spans: int
+    """Total number of spans in the store."""
+    total_resources: int
+    """Total number of resources in the store."""
+    total_workers: int
+    """Total number of workers in the store."""
+    uptime: float
+    """Uptime of since the store has been started."""
+
+    # Memory-related statistics
+    total_span_bytes: int
+    """Total number of bytes of spans in the store."""
+    eviction_threshold_bytes: int
+    """Eviction threshold for spans in bytes."""
+    safe_threshold_bytes: int
+    """Safe threshold for spans in bytes."""
+    memory_capacity_bytes: int
+    """Memory capacity of the store in bytes."""
+
+
+class _EnqueueRolloutRequestRequired(TypedDict):
+    input: TaskInput
+
+
+class EnqueueRolloutRequest(_EnqueueRolloutRequestRequired, total=False):
+    """Payload describing a rollout to be queued via `enqueue_rollout`."""
+
+    mode: Optional[RolloutMode]
+    resources_id: Optional[str]
+    config: Optional[RolloutConfig]
+    metadata: Optional[Dict[str, Any]]
+
+
 class LightningStore:
     """Contract for the persistent control-plane that coordinates training rollouts.
 
@@ -74,10 +135,42 @@ class LightningStore:
     Unless stated otherwise, missing identifiers should result in a `ValueError`.
     """
 
+    @property
+    def capabilities(self) -> LightningStoreCapabilities:
+        """Return the capabilities of the store."""
+        return LightningStoreCapabilities(
+            thread_safe=False,
+            async_safe=False,
+            zero_copy=False,
+            otlp_traces=False,
+        )
+
+    async def statistics(self) -> LightningStoreStatistics:
+        """Return the statistics of the store."""
+        return {
+            "name": self.__class__.__name__,
+        }
+
+    def otlp_traces_endpoint(self) -> str:
+        """Return the OTLP/HTTP traces endpoint of the store.
+
+        The traces can have rollout ID and attempt ID (and optionally sequence ID)
+        saved in the "resource" of the spans.
+        The store, if it supports OTLP, should be able to receive the traces and save them
+        via [`add_span`][agentlightning.LightningStore.add_span] or
+        [`add_otel_span`][agentlightning.LightningStore.add_otel_span].
+
+        The endpoint should be compatible with [OTLP HTTP protocol](https://opentelemetry.io/docs/specs/otlp/).
+        It's not necessarily compatible with OTLP gRPC protocol.
+
+        The returned endpoint will usually ends with `/v1/traces`.
+        """
+        raise NotImplementedError()
+
     async def start_rollout(
         self,
         input: TaskInput,
-        mode: Literal["train", "val", "test"] | None = None,
+        mode: RolloutMode | None = None,
         resources_id: str | None = None,
         config: RolloutConfig | None = None,
         metadata: Dict[str, Any] | None = None,
@@ -148,7 +241,23 @@ class LightningStore:
         """
         raise NotImplementedError()
 
-    async def dequeue_rollout(self) -> Optional[AttemptedRollout]:
+    async def enqueue_many_rollouts(self, inputs: Sequence[EnqueueRolloutRequest]) -> Sequence[Rollout]:
+        """Persist multiple rollouts in `queuing` state.
+
+        The implementation can delegate to [`enqueue_rollout()`][agentlightning.LightningStore.enqueue_rollout]
+        per request and preserves the input ordering. Subclasses can override to provide
+        more efficient bulk enqueue semantics.
+
+        Args:
+            inputs: Rollout submission payloads mirroring [`enqueue_rollout()`][agentlightning.LightningStore.enqueue_rollout]'s
+                parameters. Each entry requires `input` and can optionally include other fields.
+
+        Returns:
+            Rollouts enqueued in the same order as `inputs`.
+        """
+        raise NotImplementedError()
+
+    async def dequeue_rollout(self, worker_id: Optional[str] = None) -> Optional[AttemptedRollout]:
         """Claim the oldest queued rollout and transition it to `preparing`.
 
         This function do not block.
@@ -161,6 +270,8 @@ class LightningStore:
           the number of attempts already registered for the rollout plus one.
         * Return an [`AttemptedRollout`][agentlightning.AttemptedRollout] snapshot so the
           runner knows both rollout metadata and the attempt identifier.
+        * Optionally refresh the caller's [`Worker`][agentlightning.Worker] telemetry
+          (e.g., `last_dequeue_time`) when `worker_id` is provided.
 
         Returns:
             The next attempt to execute, or `None` when no eligible rollouts are queued.
@@ -170,6 +281,29 @@ class LightningStore:
         """
         raise NotImplementedError()
 
+    async def dequeue_many_rollouts(
+        self,
+        *,
+        limit: int = 1,
+        worker_id: Optional[str] = None,
+    ) -> Sequence[AttemptedRollout]:
+        """Claim up to `limit` queued rollouts without blocking.
+
+        The implementation can repeatedly invokes
+        [`dequeue_rollout()`][agentlightning.LightningStore.dequeue_rollout] until reaching
+        the requested limit or the queue is empty. Subclasses can override it to fetch
+        multiple rollouts atomically.
+
+        Args:
+            limit: Maximum number of rollouts to claim. Non-positive values return an empty list.
+            worker_id: Optional worker identifier passed through to each dequeue call.
+
+        Returns:
+            Attempted rollouts claimed in FIFO order. May contain fewer than `limit` entries
+            when the queue is exhausted.
+        """
+        raise NotImplementedError()
+
     async def start_attempt(self, rollout_id: str) -> AttemptedRollout:
         """Create a manual retry attempt for an existing rollout.
 
@@ -191,7 +325,15 @@ class LightningStore:
         """
         raise NotImplementedError()
 
-    async def add_span(self, span: Span) -> Span:
+    async def add_many_spans(self, spans: Sequence[Span]) -> Sequence[Span]:
+        """Persist a sequence of pre-constructed spans emitted during rollout execution.
+
+        Implementations can simply delegate to [`add_span()`][agentlightning.LightningStore.add_span] for each span.
+        However, if the store supports bulk insertion, it can implement this method to improve performance.
+        """
+        raise NotImplementedError()
+
+    async def add_span(self, span: Span) -> Optional[Span]:
         """Persist a pre-constructed span emitted during rollout execution.
 
         The provided [`Span`][agentlightning.Span] must already contain the `rollout_id`,
@@ -208,6 +350,7 @@ class LightningStore:
 
         Returns:
             The stored span record (implementations may return a copy).
+            Return `None` if the span was not added due to a duplicate.
 
         Raises:
             NotImplementedError: Subclasses must implement span persistence.
@@ -221,7 +364,7 @@ class LightningStore:
         attempt_id: str,
         readable_span: ReadableSpan,
         sequence_id: int | None = None,
-    ) -> Span:
+    ) -> Optional[Span]:
         """Convert and persist an OpenTelemetry span for a particular attempt.
 
         Implementations must transform the `readable_span` into a [`Span`][agentlightning.Span]
@@ -238,7 +381,7 @@ class LightningStore:
                 automatically.
 
         Returns:
-            The stored span record.
+            The stored span record. Return `None` if the span was not added due to a duplicate.
 
         Raises:
             NotImplementedError: Subclasses must implement span persistence.
@@ -247,30 +390,77 @@ class LightningStore:
         raise NotImplementedError()
 
     async def query_rollouts(
-        self, *, status: Optional[Sequence[RolloutStatus]] = None, rollout_ids: Optional[Sequence[str]] = None
-    ) -> List[Rollout]:
+        self,
+        *,
+        status_in: Optional[Sequence[RolloutStatus]] = None,
+        rollout_id_in: Optional[Sequence[str]] = None,
+        rollout_id_contains: Optional[str] = None,
+        filter_logic: Literal["and", "or"] = "and",
+        sort_by: Optional[str] = None,
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
+        # Deprecated fields
+        status: Optional[Sequence[RolloutStatus]] = None,
+        rollout_ids: Optional[Sequence[str]] = None,
+    ) -> Sequence[Rollout]:
         """Retrieve rollouts filtered by status and/or explicit identifiers.
 
+        This interface supports structured filtering, sorting, and pagination so
+        callers can build simple dashboards without copying data out of the
+        store. The legacy parameters `status` and `rollout_ids` remain valid and
+        are treated as aliases for `status_in` and `rollout_id_in`
+        respectively—when both the new and deprecated parameters are supplied
+        the new parameters take precedence.
+
         Args:
-            status: Optional whitelist of [`RolloutStatus`][agentlightning.RolloutStatus] values.
-            rollout_ids: Optional whitelist of rollout identifiers to include.
+            status_in: Optional whitelist of [`RolloutStatus`][agentlightning.RolloutStatus] values.
+            rollout_id_in: Optional whitelist of rollout identifiers to include.
+            rollout_id_contains: Optional substring match for rollout identifiers.
+            filter_logic: Logical operator to combine filters.
+            sort_by: Optional field to sort by. Must reference a numeric or string
+                field on [`Rollout`][agentlightning.Rollout].
+            sort_order: Direction to sort when `sort_by` is provided.
+            limit: Maximum number of rows to return. Use `-1` for "no limit".
+            offset: Number of rows to skip before returning results.
+            status: Deprecated field. Use `status_in` instead.
+            rollout_ids: Deprecated field. Use `rollout_id_in` instead.
 
         Returns:
-            A list of matching rollouts. Ordering is backend-defined but must be deterministic.
+            A sequence of matching rollouts (or [`AttemptedRollout`][agentlightning.AttemptedRollout]
+            when attempts exist). Ordering is deterministic when `sort_by` is set.
+            The return value is not guaranteed to be a list.
 
         Raises:
             NotImplementedError: Subclasses must implement the query.
         """
         raise NotImplementedError()
 
-    async def query_attempts(self, rollout_id: str) -> List[Attempt]:
+    async def query_attempts(
+        self,
+        rollout_id: str,
+        *,
+        sort_by: Optional[str] = "sequence_id",
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
+    ) -> Sequence[Attempt]:
         """Return every attempt ever created for `rollout_id` in ascending sequence order.
 
+        The parameters allow callers to re-order or paginate the attempts so that
+        large retry histories can be streamed lazily.
+
         Args:
             rollout_id: Identifier of the rollout being inspected.
+            sort_by: Field to sort by. Must be a numeric or string field of
+                [`Attempt`][agentlightning.Attempt]. Defaults to `sequence_id` (oldest first).
+            sort_order: Order to sort by.
+            limit: Limit on the number of results. `-1` for unlimited.
+            offset: Offset into the results.
 
         Returns:
-            Attempts sorted by `sequence_id` (oldest first). Returns an empty list when none exist.
+            Sequence of Attempts. Returns an empty sequence when none exist.
+            The return value is not guaranteed to be a list.
 
         Raises:
             NotImplementedError: Subclasses must implement the query.
@@ -307,6 +497,41 @@ class LightningStore:
         """
         raise NotImplementedError()
 
+    async def query_resources(
+        self,
+        *,
+        resources_id: Optional[str] = None,
+        resources_id_contains: Optional[str] = None,
+        # Filter logic is not supported here because I can't see why it's needed.
+        sort_by: Optional[str] = None,
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
+    ) -> Sequence[ResourcesUpdate]:
+        """List every stored resource snapshot in insertion order.
+
+        Supports lightweight filtering, sorting, and pagination for embedding in
+        dashboards.
+
+        Args:
+            resources_id: Optional identifier of the resources to include.
+            resources_id_contains: Optional substring match for resources identifiers.
+            sort_by: Optional field to sort by (must be numeric or string on
+                [`ResourcesUpdate`][agentlightning.ResourcesUpdate]).
+            sort_order: Order to sort by.
+            limit: Limit on the number of results. `-1` for unlimited.
+            offset: Offset into the results.
+
+        Returns:
+            [`ResourcesUpdate`][agentlightning.ResourcesUpdate] objects.
+            By default, resources are sorted in a deterministic but undefined order.
+            The return value is not guaranteed to be a list.
+
+        Raises:
+            NotImplementedError: Subclasses must implement retrieval.
+        """
+        raise NotImplementedError()
+
     async def get_resources_by_id(self, resources_id: str) -> Optional[ResourcesUpdate]:
         """Return a specific named resource snapshot by identifier.
 
@@ -357,6 +582,20 @@ class LightningStore:
         """
         raise NotImplementedError()
 
+    async def get_many_span_sequence_ids(self, rollout_attempt_ids: Sequence[Tuple[str, str]]) -> Sequence[int]:
+        """Bulk allocate the next strictly increasing sequence number used to order spans.
+
+        Implementations may delegate to [`get_next_span_sequence_id()`][agentlightning.LightningStore.get_next_span_sequence_id]
+        for each rollout and attempt.
+
+        Args:
+            rollout_attempt_ids: List of tuples of rollout and attempt identifiers.
+
+        Returns:
+            List of sequence numbers.
+        """
+        raise NotImplementedError()
+
     async def wait_for_rollouts(self, *, rollout_ids: List[str], timeout: Optional[float] = None) -> List[Rollout]:
         """Block until the targeted rollouts reach a terminal status or the timeout expires.
 
@@ -383,19 +622,61 @@ class LightningStore:
         """
         raise NotImplementedError()
 
-    async def query_spans(self, rollout_id: str, attempt_id: str | Literal["latest"] | None = None) -> List[Span]:
+    async def query_spans(
+        self,
+        rollout_id: str,
+        attempt_id: str | Literal["latest"] | None = None,
+        *,
+        # Filtering
+        trace_id: Optional[str] = None,
+        trace_id_contains: Optional[str] = None,
+        span_id: Optional[str] = None,
+        span_id_contains: Optional[str] = None,
+        parent_id: Optional[str] = None,
+        parent_id_contains: Optional[str] = None,
+        name: Optional[str] = None,
+        name_contains: Optional[str] = None,
+        filter_logic: Literal["and", "or"] = "and",
+        # Pagination
+        limit: int = -1,
+        offset: int = 0,
+        # Sorting
+        sort_by: Optional[str] = "sequence_id",
+        sort_order: Literal["asc", "desc"] = "asc",
+    ) -> Sequence[Span]:
         """Return the stored spans for a rollout, optionally scoped to one attempt.
 
-        Spans must be returned in ascending `sequence_id` order. Implementations may raise
-        a `RuntimeError` when spans were evicted or expired.
+        Supports a handful of filters that cover the most common debugging
+        scenarios (matching `trace_id`/`span_id`/`parent_id` or substring
+        matches on the span name). `attempt_id="latest"` acts as a convenience
+        that resolves the most recent attempt before evaluating filters. When
+        `attempt_id=None`, spans across every attempt are eligible. By default
+        results are sorted by `sequence_id` (oldest first). Implementations may
+        raise a `RuntimeError` when spans were evicted or expired.
 
         Args:
             rollout_id: Identifier of the rollout being inspected.
             attempt_id: Attempt identifier to filter by. Pass `"latest"` to retrieve only the
                 most recent attempt, or `None` to return all spans across attempts.
+            trace_id: Optional trace ID to filter by.
+            trace_id_contains: Optional substring match for trace IDs.
+            span_id: Optional span ID to filter by.
+            span_id_contains: Optional substring match for span IDs.
+            parent_id: Optional parent span ID to filter by.
+            parent_id_contains: Optional substring match for parent span IDs.
+            name: Optional span name to filter by.
+            name_contains: Optional substring match for span names.
+            filter_logic: Logical operator to combine the optional filters above.
+                The `rollout_id` argument is always applied with AND semantics.
+            limit: Limit on the number of results. `-1` for unlimited.
+            offset: Offset into the results.
+            sort_by: Field to sort by. Must be a numeric or string field of
+                [`Span`][agentlightning.Span].
+            sort_order: Order to sort by.
 
         Returns:
             An ordered list of spans (possibly empty).
+            The return value is not guaranteed to be a list.
 
         Raises:
             NotImplementedError: Subclasses must implement the query.
@@ -491,12 +772,19 @@ class LightningStore:
 
         When `attempt_id` is `"latest"` the update must target the attempt with the highest
         `sequence_id`; otherwise it must target the specific attempt. Implementations should
-        propagate status changes to the rollout (for example via [`propagate_status()`][agentlightning.store.utils.propagate_status])
+        propagate status changes to the rollout (for example
+        via [`rollout_status_from_attempt()`][agentlightning.store.utils.rollout_status_from_attempt])
         once the latest attempt transitions to a terminal state.
 
         Similar to [`update_rollout()`][agentlightning.LightningStore.update_rollout],
         parameters also default to the sentinel [`UNSET`][agentlightning.store.base.UNSET].
 
+        If `worker_id` is present, the worker status will be updated following the rules:
+
+        1. If attempt status is "succeeded" or "failed", the corresponding worker status will be set to "idle".
+        2. If attempt status is "unresponsive" or "timeout", the corresponding worker status will be set to "unknown".
+        3. Otherwise, the worker status will be set to "busy".
+
         Args:
             rollout_id: Identifier of the rollout whose attempt will be updated.
             attempt_id: Attempt identifier or `"latest"` as a convenience.
@@ -513,3 +801,64 @@ class LightningStore:
             ValueError: Implementations must raise when the rollout or attempt is unknown.
         """
         raise NotImplementedError()
+
+    async def query_workers(
+        self,
+        *,
+        status_in: Optional[Sequence[WorkerStatus]] = None,
+        worker_id_contains: Optional[str] = None,
+        filter_logic: Literal["and", "or"] = "and",
+        sort_by: Optional[str] = None,
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
+    ) -> Sequence[Worker]:
+        """Query all workers in the system.
+
+        Args:
+            status_in: Optional whitelist of [`WorkerStatus`][agentlightning.WorkerStatus] values.
+            worker_id_contains: Optional substring match for worker identifiers.
+            filter_logic: Logical operator to combine the optional filters above.
+            sort_by: Field to sort by. Must be a numeric or string field of [`Worker`][agentlightning.Worker].
+            sort_order: Order to sort by.
+            limit: Limit on the number of results. `-1` for unlimited.
+            offset: Offset into the results.
+
+        Returns:
+            Sequence of Workers. Returns an empty sequence when none exist.
+            The return value is not guaranteed to be a list.
+        """
+        raise NotImplementedError()
+
+    async def get_worker_by_id(self, worker_id: str) -> Optional[Worker]:
+        """Retrieve a single worker by identifier.
+
+        Args:
+            worker_id: Identifier of the worker.
+
+        Returns:
+            The worker record if it exists, otherwise `None`.
+
+        Raises:
+            NotImplementedError: Subclasses must implement lookup semantics.
+        """
+        raise NotImplementedError()
+
+    async def update_worker(
+        self,
+        worker_id: str,
+        heartbeat_stats: Dict[str, Any] | Unset = UNSET,
+    ) -> Worker:
+        """Record a heartbeat for `worker_id` and refresh telemetry.
+
+        Implementations must treat this API as heartbeat-only: it should snapshot
+        the latest stats when provided, stamp `last_heartbeat_time` with the
+        current wall clock, and rely on other store mutations (`dequeue_rollout`,
+        `update_attempt`, etc.) to drive the worker's busy/idle status,
+        assignment, and activity timestamps.
+
+        Args:
+            worker_id: Identifier of the worker to update.
+            heartbeat_stats: Replacement worker heartbeat statistics (non-null when provided).
+        """
+        raise NotImplementedError()

agentlightning/store/collection/__init__.py

@@ -0,0 +1,30 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from .base import (
+    AtomicLabels,
+    AtomicMode,
+    Collection,
+    FilterOptions,
+    KeyValue,
+    LightningCollections,
+    PaginatedResult,
+    Queue,
+    SortOptions,
+)
+from .memory import DequeBasedQueue, DictBasedKeyValue, InMemoryLightningCollections, ListBasedCollection
+
+__all__ = [
+    "AtomicLabels",
+    "AtomicMode",
+    "Collection",
+    "Queue",
+    "KeyValue",
+    "FilterOptions",
+    "SortOptions",
+    "PaginatedResult",
+    "LightningCollections",
+    "ListBasedCollection",
+    "DequeBasedQueue",
+    "DictBasedKeyValue",
+    "InMemoryLightningCollections",
+]

agentlightning/store/collection/base.py

@@ -0,0 +1,414 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    AsyncContextManager,
+    Awaitable,
+    Callable,
+    Dict,
+    Generic,
+    List,
+    Literal,
+    Mapping,
+    MutableMapping,
+    Optional,
+    Sequence,
+    Tuple,
+    Type,
+    TypeVar,
+    cast,
+)
+
+if TYPE_CHECKING:
+    from typing import Self
+
+from agentlightning.types import (
+    Attempt,
+    FilterField,
+    FilterOptions,
+    PaginatedResult,
+    ResourcesUpdate,
+    Rollout,
+    SortOptions,
+    Span,
+    Worker,
+)
+
+T = TypeVar("T")  # Recommended to be a BaseModel
+K = TypeVar("K")
+V = TypeVar("V")
+
+AtomicMode = Literal["r", "w", "rw"]
+"""What is expected within the atomic context. Can be "read", "write", or "read-write"."""
+
+AtomicLabels = Literal["rollouts", "attempts", "spans", "resources", "workers", "rollout_queue", "span_sequence_ids"]
+"""Labels for atomic operations.
+
+These labels are used to identify the collections that are affected by the atomic operation.
+"""
+
+
+class Collection(Generic[T]):
+    """Behaves like a list of items. Supporting addition, updating, and deletion of items."""
+
+    def primary_keys(self) -> Sequence[str]:
+        """Get the primary keys of the collection."""
+        raise NotImplementedError()
+
+    def __repr__(self) -> str:
+        return f"<{self.__class__.__name__}[{self.item_type().__name__}]>"
+
+    def item_type(self) -> Type[T]:
+        """Get the type of the items in the collection."""
+        raise NotImplementedError()
+
+    async def size(self) -> int:
+        """Get the number of items in the collection."""
+        raise NotImplementedError()
+
+    async def query(
+        self,
+        filter: Optional[FilterOptions] = None,
+        sort: Optional[SortOptions] = None,
+        limit: int = -1,
+        offset: int = 0,
+    ) -> PaginatedResult[T]:
+        """Query the collection with the given filters, sort order, and pagination.
+
+        Args:
+            filter:
+                The filters to apply to the collection. See [`FilterOptions`][agentlightning.FilterOptions].
+
+            sort:
+                The options for sorting the collection. See [`SortOptions`][agentlightning.SortOptions].
+                The field must exist in the model. If field might contain null values, in which case the behavior is undefined
+                (i.e., depending on the implementation).
+
+            limit:
+                Max number of items to return. Use -1 for "no limit".
+
+            offset:
+                Number of items to skip from the start of the *matching* items.
+
+        Returns:
+            PaginatedResult with items, limit, offset, and total matched items.
+        """
+        raise NotImplementedError()
+
+    async def get(
+        self,
+        filter: Optional[FilterOptions] = None,
+        sort: Optional[SortOptions] = None,
+    ) -> Optional[T]:
+        """Get the first item that matches the given filters.
+
+        Args:
+            filter: The filters to apply to the collection.
+                See [`FilterOptions`][agentlightning.store.collection.FilterOptions].
+            sort: Sort options. See [`SortOptions`][agentlightning.store.collection.SortOptions].
+
+        Returns:
+            The first item that matches the given filters, or None if no item matches.
+        """
+        raise NotImplementedError()
+
+    async def insert(self, items: Sequence[T]) -> None:
+        """Add the given items to the collection.
+
+        Raises:
+            ValueError: If an item with the same primary key already exists.
+        """
+        raise NotImplementedError()
+
+    async def update(self, items: Sequence[T], update_fields: Sequence[str] | None = None) -> Sequence[T]:
+        """Update the given items in the collection.
+
+        Args:
+            items: The items to update in the collection.
+            update_fields: The fields to update. If not provided, all fields in the type will be updated.
+                Only applicable if the item type is a Pydantic BaseModel.
+
+        Raises:
+            ValueError: If an item with the primary keys does not exist.
+
+        Returns:
+            The items that were updated.
+        """
+        raise NotImplementedError()
+
+    async def upsert(self, items: Sequence[T], update_fields: Sequence[str] | None = None) -> Sequence[T]:
+        """Upsert the given items into the collection.
+
+        If the items with the same primary keys already exist, they will be updated.
+        Otherwise, they will be inserted.
+
+        The operation has three semantics configurable via `update_fields`:
+
+        - `update_or_insert` via `collection.upsert(items, update_fields=["status", "updated_at"])`.
+          If the item with the same primary keys already exists, only the specified fields will be updated.
+          Otherwise, the item will be inserted.
+        - `get_or_insert` via `collection.upsert(items, update_fields=[])`.
+          If the item with the same primary keys already exists, the item will be left unchanged.
+          Otherwise, the item will be inserted.
+        - `replace_ish` via `collection.upsert(items)`.
+          If the item with the same primary keys already exists, all fields from the item will be set.
+          Otherwise, the item will be inserted.
+
+        Returns:
+            The items that were upserted.
+        """
+        raise NotImplementedError()
+
+    async def delete(self, items: Sequence[T]) -> None:
+        """Delete the given items from the collection.
+
+        Args:
+            items: The items to delete from the collection.
+
+        Raises:
+            ValueError: If the items with the primary keys to be deleted do not exist.
+        """
+        raise NotImplementedError()
+
+
+class Queue(Generic[T]):
+    """Behaves like a deque. Supporting appending items to the end and popping items from the front."""
+
+    def __repr__(self) -> str:
+        return f"<{self.__class__.__name__}[{self.item_type().__name__}]>"
+
+    def item_type(self) -> Type[T]:
+        """Get the type of the items in the queue."""
+        raise NotImplementedError()
+
+    async def has(self, item: T) -> bool:
+        """Check if the given item is in the queue."""
+        raise NotImplementedError()
+
+    async def enqueue(self, items: Sequence[T]) -> Sequence[T]:
+        """Append the given items to the end of the queue.
+
+        Args:
+            items: The items to append to the end of the queue.
+
+        Returns:
+            The items that were appended to the end of the queue.
+        """
+        raise NotImplementedError()
+
+    async def dequeue(self, limit: int = 1) -> Sequence[T]:
+        """Pop the given number of items from the front of the queue.
+
+        Args:
+            limit: The number of items to pop from the front of the queue.
+
+        Returns:
+            The items that were popped from the front of the queue.
+            If there are less than `limit` items in the queue, the remaining items will be returned.
+        """
+        raise NotImplementedError()
+
+    async def peek(self, limit: int = 1) -> Sequence[T]:
+        """Peek the given number of items from the front of the queue.
+
+        Args:
+            limit: The number of items to peek from the front of the queue.
+
+        Returns:
+            The items that were peeked from the front of the queue.
+            If there are less than `limit` items in the queue, the remaining items will be returned.
+        """
+        raise NotImplementedError()
+
+    async def size(self) -> int:
+        """Get the number of items in the queue."""
+        raise NotImplementedError()
+
+
+class KeyValue(Generic[K, V]):
+    """Behaves like a dictionary. Supporting addition, updating, and deletion of items."""
+
+    def __repr__(self) -> str:
+        return f"<{self.__class__.__name__}>"
+
+    async def has(self, key: K) -> bool:
+        """Check if the given key is in the dictionary."""
+        raise NotImplementedError()
+
+    async def get(self, key: K, default: V | None = None) -> V | None:
+        """Get the value for the given key, or the default value if the key is not found."""
+        raise NotImplementedError()
+
+    async def set(self, key: K, value: V) -> None:
+        """Set the value for the given key."""
+        raise NotImplementedError()
+
+    async def pop(self, key: K, default: V | None = None) -> V | None:
+        """Pop the value for the given key, or the default value if the key is not found."""
+        raise NotImplementedError()
+
+    async def size(self) -> int:
+        """Get the number of items in the dictionary."""
+        raise NotImplementedError()
+
+
+class LightningCollections:
+    """Collections of rollouts, attempts, spans, resources, and workers.
+
+    [LightningStore][agentlightning.LightningStore] implementations can use this as a storage base
+    to implement the store API.
+    """
+
+    @property
+    def rollouts(self) -> Collection[Rollout]:
+        """Collections of rollouts."""
+        raise NotImplementedError()
+
+    @property
+    def attempts(self) -> Collection[Attempt]:
+        """Collections of attempts."""
+        raise NotImplementedError()
+
+    @property
+    def spans(self) -> Collection[Span]:
+        """Collections of spans."""
+        raise NotImplementedError()
+
+    @property
+    def resources(self) -> Collection[ResourcesUpdate]:
+        """Collections of resources."""
+        raise NotImplementedError()
+
+    @property
+    def workers(self) -> Collection[Worker]:
+        """Collections of workers."""
+        raise NotImplementedError()
+
+    @property
+    def rollout_queue(self) -> Queue[str]:
+        """Queue of rollouts (tasks)."""
+        raise NotImplementedError()
+
+    @property
+    def span_sequence_ids(self) -> KeyValue[str, int]:
+        """Dictionary (counter) of span sequence IDs."""
+        raise NotImplementedError()
+
+    def atomic(
+        self,
+        *,
+        mode: AtomicMode = "rw",
+        snapshot: bool = False,
+        commit: bool = False,
+        labels: Optional[Sequence[AtomicLabels]] = None,
+        **kwargs: Any,
+    ) -> AsyncContextManager[Self]:
+        """Perform a atomic operation on the collections.
+
+        Subclass may use args and kwargs to support multiple levels of atomicity.
+        The arguments can be seen as tags. They only imply the behavior of the operation, not the implementation.
+
+        Args:
+            mode: The mode of atomicity. See [`AtomicMode`][agentlightning.store.collection.AtomicMode].
+            snapshot: Enable read snapshot for repeatable reads. Data consistency is guaranteed. The real behavior is implementation-dependent.
+            commit: Enable commitment for write operations. Unsuccessful operations will be rolled back depending on the implementation.
+                Recommend to use [`execute()`][agentlightning.store.collection.LightningCollections.execute] for this level to enable automatic retries.
+                Remember that the real behavior is implementation-dependent.
+            labels: Labels to add to the atomic operation (commonly used as lock names or collection names).
+            **kwargs: Keyword arguments to pass to the operation.
+        """
+        raise NotImplementedError()
+
+    async def execute(
+        self,
+        callback: Callable[[Self], Awaitable[T]],
+        *,
+        mode: AtomicMode = "rw",
+        snapshot: bool = False,
+        commit: bool = False,
+        labels: Optional[Sequence[AtomicLabels]] = None,
+        **kwargs: Any,
+    ) -> T:
+        """Execute the given callback within an atomic operation. Retry on transient errors is implied.
+
+        See [`atomic()`][agentlightning.store.collection.LightningCollections.atomic] for more details.
+        """
+        async with self.atomic(mode=mode, snapshot=snapshot, commit=commit, labels=labels, **kwargs) as collections:
+            return await callback(collections)
+
+
+FilterMap = Mapping[str, FilterField]
+
+
+def merge_must_filters(target: MutableMapping[str, FilterField], definition: Any) -> None:
+    """Normalize a `_must` filter group into the provided mapping.
+
+    Mainly for validation purposes.
+    """
+    if definition is None:
+        return
+
+    entries: List[Mapping[str, FilterField]] = []
+    if isinstance(definition, Mapping):
+        entries.append(cast(Mapping[str, FilterField], definition))
+    elif isinstance(definition, Sequence) and not isinstance(definition, (str, bytes)):
+        for entry in definition:  # type: ignore
+            if not isinstance(entry, Mapping):
+                raise TypeError("Each `_must` entry must be a mapping of field names to operators")
+            entries.append(cast(Mapping[str, FilterField], entry))
+    else:
+        raise TypeError("`_must` filters must be provided as a mapping or sequence of mappings")
+
+    for entry in entries:
+        for field_name, ops in entry.items():
+            existing = target.get(field_name, {})
+            merged_ops: Dict[str, Any] = dict(existing)
+            for op_name, expected in ops.items():
+                if op_name in merged_ops:
+                    raise ValueError(f"Duplicate operator '{op_name}' for field '{field_name}' in must filters")
+                merged_ops[op_name] = expected
+            target[field_name] = cast(FilterField, merged_ops)
+
+
+def normalize_filter_options(
+    filter_options: Optional[FilterOptions],
+) -> Tuple[Optional[FilterMap], Optional[FilterMap], Literal["and", "or"]]:
+    """Convert FilterOptions to the internal structure and resolve aggregate logic."""
+    if not filter_options:
+        return None, None, "and"
+
+    aggregate = cast(Literal["and", "or"], filter_options.get("_aggregate", "and"))
+    if aggregate not in ("and", "or"):
+        raise ValueError(f"Unsupported filter aggregate '{aggregate}'")
+
+    # Extract normalized filters and must filters from the filter options.
+    normalized: Dict[str, FilterField] = {}
+    must_filters: Dict[str, FilterField] = {}
+    for field_name, ops in filter_options.items():
+        if field_name == "_aggregate":
+            continue
+        if field_name == "_must":
+            merge_must_filters(must_filters, ops)
+            continue
+        normalized[field_name] = cast(FilterField, dict(ops))  # type: ignore
+
+    return (normalized or None, must_filters or None, aggregate)
+
+
+def resolve_sort_options(sort: Optional[SortOptions]) -> Tuple[Optional[str], Literal["asc", "desc"]]:
+    """Extract sort field/order from the caller-provided SortOptions."""
+    if not sort:
+        return None, "asc"
+
+    sort_name = sort.get("name")
+    if not sort_name:
+        raise ValueError("Sort options must include a 'name' field")
+
+    sort_order = sort.get("order", "asc")
+    if sort_order not in ("asc", "desc"):
+        raise ValueError(f"Unsupported sort order '{sort_order}'")
+
+    return sort_name, sort_order

agentlightning/store/collection/memory.py

@@ -0,0 +1,861 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import threading
+import weakref
+from collections import deque
+from contextlib import AsyncExitStack, asynccontextmanager
+from typing import (
+    Any,
+    Deque,
+    Dict,
+    Iterable,
+    List,
+    Literal,
+    Mapping,
+    MutableMapping,
+    Optional,
+    Sequence,
+    Tuple,
+    Type,
+    TypeVar,
+    Union,
+)
+
+from pydantic import BaseModel
+
+from agentlightning.types import (
+    Attempt,
+    FilterField,
+    FilterOptions,
+    PaginatedResult,
+    ResourcesUpdate,
+    Rollout,
+    SortOptions,
+    Span,
+    Worker,
+)
+
+from .base import (
+    AtomicMode,
+    Collection,
+    FilterMap,
+    KeyValue,
+    LightningCollections,
+    Queue,
+    normalize_filter_options,
+    resolve_sort_options,
+)
+
+T = TypeVar("T")  # Recommended to be a BaseModel, not a dict
+K = TypeVar("K")
+V = TypeVar("V")
+
+logger = logging.getLogger(__name__)
+
+# Nested structure type:
+# dict[pk1] -> dict[pk2] -> ... -> item
+ListBasedCollectionItemType = Union[
+    Dict[Any, "ListBasedCollectionItemType[T]"],  # intermediate node
+    Dict[Any, T],  # leaf node dictionary
+]
+
+MutationMode = Literal["insert", "update", "upsert", "delete"]
+
+
+def _item_matches_filters(
+    item: object,
+    filters: Optional[FilterMap],
+    filter_logic: Literal["and", "or"],
+    must_filters: Optional[FilterMap] = None,
+) -> bool:
+    """Check whether an item matches the provided filter definition.
+
+    Filter format:
+
+    ```json
+    {
+        "_aggregate": "or",
+        "field_name": {
+            "exact": <value>,
+            "within": <iterable_of_allowed_values>,
+            "contains": <substring_or_element>,
+        },
+        ...
+    }
+    ```
+
+    Operators within the same field are stored in a unified pool and combined using
+    a universal logical operator.
+    """
+    if must_filters and not _item_matches_filters(item, must_filters, "and"):
+        return False
+
+    if not filters:
+        return True
+
+    all_conditions_match: List[bool] = []
+
+    for field_name, ops in filters.items():
+        item_value = getattr(item, field_name, None)
+
+        for op_name, expected in ops.items():
+            # Ignore no-op filters
+            if expected is None:
+                continue
+
+            if op_name == "exact":
+                all_conditions_match.append(item_value == expected)
+
+            elif op_name == "within":
+                try:
+                    all_conditions_match.append(item_value in expected)  # type: ignore[arg-type]
+                except TypeError:
+                    all_conditions_match.append(False)
+
+            elif op_name == "contains":
+                if item_value is None:
+                    all_conditions_match.append(False)
+                elif isinstance(item_value, str) and isinstance(expected, str):
+                    all_conditions_match.append(expected in item_value)
+                else:
+                    # Fallback: treat as generic iterable containment.
+                    try:
+                        all_conditions_match.append(expected in item_value)  # type: ignore[arg-type]
+                    except TypeError:
+                        all_conditions_match.append(False)
+            else:
+                raise ValueError(f"Unsupported filter operator '{op_name}' for field '{field_name}'")
+
+    return all(all_conditions_match) if filter_logic == "and" else any(all_conditions_match)
+
+
+def _get_sort_value(item: object, sort_by: str) -> Any:
+    """Get a sort key for the given item/field.
+
+    - If the field name ends with '_time', values are treated as comparable timestamps.
+    - For other fields we try to infer a safe default from the Pydantic model annotation.
+    """
+    value = getattr(item, sort_by, None)
+
+    if sort_by.endswith("_time"):
+        # For *_time fields, push missing values to the end.
+        return float("inf") if value is None else value
+
+    if value is None:
+        # Introspect model field type to choose a reasonable default for None.
+        model_fields = getattr(item.__class__, "model_fields", {})
+        if sort_by not in model_fields:
+            raise ValueError(
+                f"Failed to sort items by '{sort_by}': field does not exist " f"on {item.__class__.__name__}"
+            )
+
+        field_type_str = str(model_fields[sort_by].annotation)
+        if "str" in field_type_str or "Literal" in field_type_str:
+            return ""
+        if "int" in field_type_str:
+            return 0
+        if "float" in field_type_str:
+            return 0.0
+        raise ValueError(f"Failed to sort items by '{sort_by}': unsupported field type {field_type_str!r}")
+
+    return value
+
+
+class ListBasedCollection(Collection[T]):
+    """In-memory implementation of Collection using a nested dict for O(1) primary-key lookup.
+
+    The internal structure is:
+
+        {
+            pk1_value: {
+                pk2_value: {
+                    ...
+                        pkN_value: item
+                }
+            }
+        }
+
+    where the nesting depth equals the number of primary keys.
+
+    Sorting behavior:
+
+    1. If no sort_by is provided, the items are returned in the order of insertion.
+    2. If sort_by is provided, the items are sorted by the value of the sort_by field.
+    3. If the sort_by field is a timestamp, the null values are treated as infinity.
+    4. If the sort_by field is not a timestamp, the null values are treated as empty string
+       if the field is str-like, 0 if the field is int-like, 0.0 if the field is float-like.
+    """
+
+    def __init__(self, items: List[T], item_type: Type[T], primary_keys: Sequence[str]):
+        if not primary_keys:
+            raise ValueError("primary_keys must be non-empty")
+
+        self._items: Dict[Any, Any] = {}
+        self._size: int = 0
+        if issubclass(item_type, dict):
+            raise TypeError(f"Expect item to be not a dict, got {item_type.__name__}")
+        self._item_type: Type[T] = item_type
+        self._primary_keys: Tuple[str, ...] = tuple(primary_keys)
+
+        # Pre-populate the collection with the given items.
+        for item in items or []:
+            self._mutate_single(item, mode="insert")
+
+    def primary_keys(self) -> Sequence[str]:
+        """Return the primary key field names for this collection."""
+        return self._primary_keys
+
+    def item_type(self) -> Type[T]:
+        """Return the Pydantic model type of items stored in this collection."""
+        return self._item_type
+
+    async def size(self) -> int:
+        """Return the number of items stored in the collection."""
+        return self._size
+
+    def __repr__(self) -> str:
+        return f"<{self.__class__.__name__}[{self.item_type().__name__}] ({self._size})>"
+
+    # -------------------------------------------------------------------------
+    # Internal helpers
+    # -------------------------------------------------------------------------
+
+    def _ensure_item_type(self, item: T) -> None:
+        """Validate that the item matches the declared item_type."""
+        if not isinstance(item, self._item_type):
+            raise TypeError(f"Expected item of type {self._item_type.__name__}, " f"got {type(item).__name__}")
+
+    def _extract_primary_key_values(self, item: T) -> Tuple[Any, ...]:
+        """Extract the primary key values from an item.
+
+        Raises:
+            ValueError: If any primary key is missing on the item.
+        """
+        values: List[Any] = []
+        for key in self._primary_keys:
+            if not hasattr(item, key):
+                raise ValueError(f"Item {item} does not have primary key field '{key}'")
+            values.append(getattr(item, key))
+        return tuple(values)
+
+    def _render_key_values(self, key_values: Sequence[Any]) -> str:
+        return ", ".join(f"{name}={value!r}" for name, value in zip(self._primary_keys, key_values))
+
+    def _locate_node(
+        self,
+        key_values: Sequence[Any],
+        create_missing: bool,
+    ) -> Tuple[MutableMapping[Any, Any], Any]:
+        """Locate the parent mapping and final key for an item path.
+
+        Args:
+            key_values: The sequence of primary key values.
+            create_missing: Whether to create intermediate dictionaries as needed.
+
+        Returns:
+            (parent_mapping, final_key)
+
+        Raises:
+            KeyError: If the path does not exist and create_missing is False.
+            ValueError: If the internal structure is corrupted (non-dict where dict is expected).
+        """
+        if not key_values:
+            raise ValueError("key_values must be non-empty")
+
+        current: MutableMapping[Any, Any] = self._items
+        for idx, value in enumerate(key_values):
+            is_last = idx == len(key_values) - 1
+            if is_last:
+                # At the final level, current[value] is the item (or will be).
+                return current, value  # type: ignore
+
+            # Intermediate level: current[value] must be a dict.
+            if value not in current:
+                if not create_missing:
+                    raise KeyError(f"Path does not exist for given primary keys: {self._render_key_values(key_values)}")
+                current[value] = {}
+            next_node = current[value]  # type: ignore
+            if not isinstance(next_node, dict):
+                raise ValueError(f"Internal structure corrupted: expected dict, got {type(next_node)!r}")  # type: ignore
+            current = next_node  # type: ignore
+
+        # We should always return inside the loop.
+        raise RuntimeError("Unreachable")
+
+    def _mutate_single(self, item: T, mode: MutationMode, update_fields: Sequence[str] | None = None) -> Optional[T]:
+        """Core mutation logic shared by insert, update, upsert, and delete."""
+        self._ensure_item_type(item)
+        key_values = self._extract_primary_key_values(item)
+
+        if mode in ("insert", "upsert"):
+            parent, final_key = self._locate_node(key_values, create_missing=True)
+            exists = final_key in parent
+
+            if mode == "insert":
+                if exists:
+                    raise ValueError(f"Item already exists with primary key(s): {self._render_key_values(key_values)}")
+                parent[final_key] = item
+                self._size += 1
+            else:  # upsert
+                if not exists:
+                    self._size += 1
+                    parent[final_key] = item
+
+                elif update_fields is None:
+                    # update_or_insert: update all fields
+                    parent[final_key] = item
+
+                else:
+                    if not issubclass(self._item_type, BaseModel):
+                        raise TypeError(
+                            f"When using update_fields, the item type must be a Pydantic BaseModel, got {self._item_type.__name__}"
+                        )
+
+                    # Try to fetch the existing item
+                    existing = parent[final_key]
+                    if not isinstance(existing, self._item_type):
+                        raise ValueError(
+                            f"Internal structure corrupted: expected {self._item_type.__name__}, got {type(existing)!r}"
+                        )
+
+                    if not isinstance(item, self._item_type):
+                        raise TypeError(
+                            f"When using update_fields, the item type must be a Pydantic BaseModel, got {type(item).__name__}"
+                        )
+
+                    parent[final_key] = parent[final_key].model_copy(
+                        update={field: getattr(item, field) for field in update_fields}
+                    )
+
+            return parent[final_key]
+
+        elif mode in ("update", "delete"):
+            # For update/delete we must not create missing paths.
+            try:
+                parent, final_key = self._locate_node(key_values, create_missing=False)
+            except KeyError:
+                raise ValueError(
+                    f"Item does not exist with primary key(s): {self._render_key_values(key_values)}"
+                ) from None
+
+            if final_key not in parent:
+                raise ValueError(f"Item does not exist with primary key(s): {self._render_key_values(key_values)}")
+
+            if mode == "update":
+                if update_fields is None:
+                    # replace the entire item
+                    parent[final_key] = item
+                else:
+                    if not issubclass(self._item_type, BaseModel):
+                        raise TypeError(
+                            f"When using update_fields, the item type must be a Pydantic BaseModel, got {self._item_type.__name__}"
+                        )
+                    if not isinstance(item, self._item_type):
+                        raise TypeError(
+                            f"When using update_fields, the item type must be a Pydantic BaseModel, got {type(item).__name__}"
+                        )
+                    parent[final_key] = parent[final_key].model_copy(
+                        update={field: getattr(item, field) for field in update_fields}
+                    )
+                return parent[final_key]
+            else:  # delete
+                del parent[final_key]
+                self._size -= 1
+        else:
+            raise ValueError(f"Unknown mutation mode: {mode}")
+
+    def _iter_items(
+        self,
+        root: Optional[Mapping[Any, Any]] = None,
+        filters: Optional[FilterMap] = None,
+        must_filters: Optional[FilterMap] = None,
+        filter_logic: Literal["and", "or"] = "and",
+    ) -> Iterable[T]:
+        """Iterate over all items in the nested dictionary structure, optionally applying filters."""
+        if root is None:
+            root = self._items
+        if not root:
+            return
+        stack: List[Mapping[Any, Any]] = [root]
+        while stack:
+            node = stack.pop()
+            for value in node.values():
+                # Leaf nodes contain items; intermediate nodes are dicts.
+                if isinstance(value, self._item_type):
+                    if _item_matches_filters(value, filters, filter_logic, must_filters):
+                        yield value
+                elif isinstance(value, dict):
+                    stack.append(value)  # type: ignore
+                else:
+                    raise ValueError(
+                        f"Internal structure corrupted: expected dict or {self._item_type.__name__}, "
+                        f"got {type(value)!r}"
+                    )
+
+    def _iter_matching_items(
+        self,
+        filters: Optional[FilterMap],
+        must_filters: Optional[FilterMap],
+        filter_logic: Literal["and", "or"],
+    ) -> Iterable[T]:
+        """Efficiently iterate over items matching filters, using primary-key prefix when possible."""
+        # Fast path: when optional filters can't form a prefix, fall back to scanning.
+        if filter_logic != "and" and must_filters is None:
+            return self._iter_items(filters=filters, must_filters=must_filters, filter_logic=filter_logic)
+
+        # Try to derive a primary-key prefix from exact filters.
+        pk_values_prefix: List[Any] = []
+        prefix_sources: List[FilterMap] = []
+        if must_filters:
+            prefix_sources.append(must_filters)
+        if filter_logic == "and" and filters:
+            prefix_sources.append(filters)
+
+        for pk in self._primary_keys:
+            # combined_ops are: [{"exact": value}, {"within": [...]}, ...]
+            combined_ops: List[FilterField] = []
+            for source in prefix_sources:
+                field_ops = source.get(pk)  # type: ignore[union-attr]
+                if field_ops:
+                    combined_ops.append(field_ops)
+            if not combined_ops:
+                break
+            # Only allow a pure {"exact": value} constraint.
+            exact_value: Any | None = None
+            allow_prefix = True
+            for ops in combined_ops:
+                if set(ops.keys()) != {"exact"}:
+                    allow_prefix = False
+                    break
+                candidate = ops.get("exact")
+                if candidate is None:
+                    allow_prefix = False
+                    break
+                if exact_value is not None and candidate != exact_value:
+                    # Contradictory exact filters mean no items can match.
+                    logger.warning(f"Contradictory exact filters for field '{pk}': {exact_value} != {candidate}")
+                    return ()
+                exact_value = candidate
+
+            if not allow_prefix:
+                break
+
+            value = exact_value
+            if value is None:
+                break
+            pk_values_prefix.append(value)
+
+        if not pk_values_prefix:
+            return self._iter_items(filters=filters, must_filters=must_filters, filter_logic=filter_logic)
+
+        try:
+            if len(pk_values_prefix) == len(self._primary_keys):
+                # All primary keys specified -> at most a single item.
+                parent, final_key = self._locate_node(pk_values_prefix, create_missing=False)
+                single_item = parent.get(final_key)
+                if isinstance(single_item, self._item_type) and _item_matches_filters(
+                    single_item,
+                    filters,
+                    filter_logic,
+                    must_filters,
+                ):
+                    return (single_item,)
+                return ()
+            else:
+                # Prefix of primary keys specified -> iterate only the subtree below that prefix.
+                parent, final_key = self._locate_node(pk_values_prefix, create_missing=False)
+                subtree = parent.get(final_key)
+                if isinstance(subtree, dict):
+                    return self._iter_items(
+                        subtree,  # type: ignore
+                        filters=filters,
+                        must_filters=must_filters,
+                        filter_logic=filter_logic,
+                    )
+                return ()
+        except KeyError:
+            # No items exist for this primary-key prefix.
+            return ()
+
+    async def query(
+        self,
+        filter: Optional[FilterOptions] = None,
+        sort: Optional[SortOptions] = None,
+        limit: int = -1,
+        offset: int = 0,
+    ) -> PaginatedResult[T]:
+        """Query the collection with filters, sort order, and pagination.
+
+        Args:
+            filter: Mapping of field name to operator dict along with the optional `_aggregate` logic.
+            sort: Options describing which field to sort by and in which order.
+            limit: Max number of items to return. Use -1 for "no limit".
+            offset: Number of items to skip from the start of the *matching* items.
+        """
+        filters, must_filters, filter_logic = normalize_filter_options(filter)
+        sort_by, sort_order = resolve_sort_options(sort)
+        items_iter: Iterable[T] = self._iter_matching_items(filters, must_filters, filter_logic)
+
+        # No sorting: stream through items and apply pagination on the fly.
+        if not sort_by:
+            matched_items: List[T] = []
+            total_matched = 0
+
+            for item in items_iter:
+                # Count every match for 'total'
+                total_matched += 1
+
+                # Apply offset/limit window
+                if total_matched <= offset:
+                    continue
+                if limit != -1 and len(matched_items) >= limit:
+                    # Still need to finish iteration to get accurate total_matched.
+                    continue
+
+                matched_items.append(item)
+
+            return PaginatedResult(
+                items=matched_items,
+                limit=limit,
+                offset=offset,
+                total=total_matched,
+            )
+
+        # With sorting: we must materialize all matching items to sort them.
+        all_matches: List[T] = list(items_iter)
+
+        total_matched = len(all_matches)
+        reverse = sort_order == "desc"
+        all_matches.sort(key=lambda x: _get_sort_value(x, sort_by), reverse=reverse)
+
+        if limit == -1:
+            paginated_items = all_matches[offset:]
+        else:
+            paginated_items = all_matches[offset : offset + limit]
+
+        return PaginatedResult(
+            items=paginated_items,
+            limit=limit,
+            offset=offset,
+            total=total_matched,
+        )
+
+    async def get(
+        self,
+        filter: Optional[FilterOptions] = None,
+        sort: Optional[SortOptions] = None,
+    ) -> Optional[T]:
+        """Return the first (or best-sorted) item that matches the given filters, or None."""
+        filters, must_filters, filter_logic = normalize_filter_options(filter)
+        sort_by, sort_order = resolve_sort_options(sort)
+        items_iter: Iterable[T] = self._iter_matching_items(filters, must_filters, filter_logic)
+
+        if not sort_by:
+            # Just return the first matching item, if any.
+            for item in items_iter:
+                return item
+            return None
+
+        # Single-pass min/max according to sort_order.
+        best_item: Optional[T] = None
+        best_key: Any = None
+
+        for item in items_iter:
+            key = _get_sort_value(item, sort_by)
+            if best_item is None:
+                best_item = item
+                best_key = key
+                continue
+
+            if sort_order == "asc":
+                if key < best_key:
+                    best_item, best_key = item, key
+            else:
+                if key > best_key:
+                    best_item, best_key = item, key
+
+        return best_item
+
+    async def insert(self, items: Sequence[T]) -> None:
+        """Insert the given items.
+
+        Raises:
+            ValueError: If any item with the same primary keys already exists.
+        """
+        seen_keys: set[Tuple[Any, ...]] = set()
+        prepared: List[T] = []
+        for item in items:
+            self._ensure_item_type(item)
+            key_values = self._extract_primary_key_values(item)
+            if key_values in seen_keys:
+                raise ValueError(
+                    f"Insert payload contains duplicate primary key(s): {self._render_key_values(key_values)}"
+                )
+            seen_keys.add(key_values)
+            prepared.append(item)
+
+        for item in prepared:
+            self._mutate_single(item, mode="insert")
+
+    async def update(self, items: Sequence[T], update_fields: Sequence[str] | None = None) -> Sequence[T]:
+        """Update the given items.
+
+        Raises:
+            ValueError: If any item with the given primary keys does not exist.
+        """
+        updated_items: List[T] = []
+        for item in items:
+            updated = self._mutate_single(item, mode="update", update_fields=update_fields)
+            if updated is None:
+                raise RuntimeError(f"_mutate_single returned None for item {item}. This should never happen.")
+            updated_items.append(updated)
+        return updated_items
+
+    async def upsert(self, items: Sequence[T], update_fields: Sequence[str] | None = None) -> Sequence[T]:
+        """Upsert the given items (insert if missing, otherwise update)."""
+        upserted_items: List[T] = []
+        for item in items:
+            upserted = self._mutate_single(item, mode="upsert", update_fields=update_fields)
+            if upserted is None:
+                raise RuntimeError(f"_mutate_single returned None for item {item}. This should never happen.")
+            upserted_items.append(upserted)
+        return upserted_items
+
+    async def delete(self, items: Sequence[T]) -> None:
+        """Delete the given items.
+
+        Raises:
+            ValueError: If any item with the given primary keys does not exist.
+        """
+        # We use a two-phase approach to avoid partial deletion if one fails:
+        # first compute key_values to validate, then perform deletions.
+        for item in items:
+            # _mutate_single will validate existence and update size.
+            self._mutate_single(item, mode="delete")
+
+
+class DequeBasedQueue(Queue[T]):
+    """Queue implementation backed by collections.deque.
+
+    Provides O(1) amortized enqueue (append) and dequeue (popleft).
+    """
+
+    def __init__(self, item_type: Type[T], items: Optional[Sequence[T]] = None):
+        self._items: Deque[T] = deque()
+        self._item_type: Type[T] = item_type
+        if items:
+            self._items.extend(items)
+
+    def item_type(self) -> Type[T]:
+        return self._item_type
+
+    def __repr__(self) -> str:
+        return f"<{self.__class__.__name__}[{self.item_type().__name__}] ({len(self._items)})>"
+
+    async def has(self, item: T) -> bool:
+        if not isinstance(item, self._item_type):
+            raise TypeError(f"Expected item of type {self._item_type.__name__}, got {type(item).__name__}")
+        return item in self._items
+
+    async def enqueue(self, items: Sequence[T]) -> Sequence[T]:
+        for item in items:
+            if not isinstance(item, self._item_type):
+                raise TypeError(f"Expected item of type {self._item_type.__name__}, got {type(item).__name__}")
+            self._items.append(item)
+        return items
+
+    async def dequeue(self, limit: int = 1) -> Sequence[T]:
+        if limit <= 0:
+            return []
+        out: List[T] = []
+        for _ in range(min(limit, len(self._items))):
+            out.append(self._items.popleft())
+        return out
+
+    async def peek(self, limit: int = 1) -> Sequence[T]:
+        if limit <= 0:
+            return []
+        result: List[T] = []
+        count = min(limit, len(self._items))
+        for idx, item in enumerate(self._items):
+            if idx >= count:
+                break
+            result.append(item)
+        return result
+
+    async def size(self) -> int:
+        return len(self._items)
+
+
+class DictBasedKeyValue(KeyValue[K, V]):
+    """KeyValue implementation backed by a plain dictionary."""
+
+    def __init__(self, data: Optional[Mapping[K, V]] = None):
+        self._values: Dict[K, V] = dict(data) if data else {}
+
+    async def has(self, key: K) -> bool:
+        return key in self._values
+
+    async def get(self, key: K, default: V | None = None) -> V | None:
+        return self._values.get(key, default)
+
+    async def set(self, key: K, value: V) -> None:
+        self._values[key] = value
+
+    async def pop(self, key: K, default: V | None = None) -> V | None:
+        return self._values.pop(key, default)
+
+    async def size(self) -> int:
+        return len(self._values)
+
+
+class InMemoryLightningCollections(LightningCollections):
+    """In-memory implementation of LightningCollections using Python data structures.
+
+    Serves as the storage base for [`InMemoryLightningStore`][agentlightning.InMemoryLightningStore].
+    """
+
+    def __init__(self, lock_type: Literal["thread", "asyncio"]):
+        self._lock = {
+            "rollouts": _LoopAwareAsyncLock() if lock_type == "asyncio" else _ThreadSafeAsyncLock(),
+            "attempts": _LoopAwareAsyncLock() if lock_type == "asyncio" else _ThreadSafeAsyncLock(),
+            "spans": _LoopAwareAsyncLock() if lock_type == "asyncio" else _ThreadSafeAsyncLock(),
+            "resources": _LoopAwareAsyncLock() if lock_type == "asyncio" else _ThreadSafeAsyncLock(),
+            "workers": _LoopAwareAsyncLock() if lock_type == "asyncio" else _ThreadSafeAsyncLock(),
+            "rollout_queue": _LoopAwareAsyncLock() if lock_type == "asyncio" else _ThreadSafeAsyncLock(),
+            "span_sequence_ids": _LoopAwareAsyncLock() if lock_type == "asyncio" else _ThreadSafeAsyncLock(),
+        }
+        self._rollouts = ListBasedCollection(items=[], item_type=Rollout, primary_keys=["rollout_id"])
+        self._attempts = ListBasedCollection(items=[], item_type=Attempt, primary_keys=["rollout_id", "attempt_id"])
+        self._spans = ListBasedCollection(
+            items=[], item_type=Span, primary_keys=["rollout_id", "attempt_id", "span_id"]
+        )
+        self._resources = ListBasedCollection(items=[], item_type=ResourcesUpdate, primary_keys=["resources_id"])
+        self._workers = ListBasedCollection(items=[], item_type=Worker, primary_keys=["worker_id"])
+        self._rollout_queue = DequeBasedQueue(items=[], item_type=str)
+        self._span_sequence_ids = DictBasedKeyValue[str, int](data={})  # rollout_id -> sequence_id
+
+    @property
+    def rollouts(self) -> ListBasedCollection[Rollout]:
+        return self._rollouts
+
+    @property
+    def attempts(self) -> ListBasedCollection[Attempt]:
+        return self._attempts
+
+    @property
+    def spans(self) -> ListBasedCollection[Span]:
+        return self._spans
+
+    @property
+    def resources(self) -> ListBasedCollection[ResourcesUpdate]:
+        return self._resources
+
+    @property
+    def workers(self) -> ListBasedCollection[Worker]:
+        return self._workers
+
+    @property
+    def rollout_queue(self) -> DequeBasedQueue[str]:
+        return self._rollout_queue
+
+    @property
+    def span_sequence_ids(self) -> DictBasedKeyValue[str, int]:
+        return self._span_sequence_ids
+
+    @asynccontextmanager
+    async def atomic(
+        self, *, mode: AtomicMode = "rw", snapshot: bool = False, labels: Optional[Sequence[str]] = None, **kwargs: Any
+    ):
+        """In-memory collections apply a lock outside. It doesn't need to manipulate the collections inside.
+
+        Skip the locking if mode is "r" and snapshot is False.
+
+        This collection implementation does NOT support rollback / commit.
+        """
+        if mode == "r" and not snapshot:
+            yield self
+            return
+        if not labels:
+            # If no labels are provided, use all locks.
+            labels = list(self._lock.keys())
+        managers = [self._lock[label] for label in labels]
+        async with AsyncExitStack() as stack:
+            for manager in managers:
+                await stack.enter_async_context(manager)
+            yield self
+
+    async def evict_spans_for_rollout(self, rollout_id: str) -> None:
+        """Evict all spans for a given rollout ID.
+
+        Uses private API for efficiency.
+        """
+        self._spans._items.pop(rollout_id, [])  # pyright: ignore[reportPrivateUsage]
+
+
+class _LoopAwareAsyncLock:
+    """Async lock that transparently rebinds to the current event loop.
+
+    The lock intentionally remains *thread-unsafe*: callers must only use it from
+    one thread at a time. If multiple threads interact with the store, each
+    thread gets its own event loop specific lock.
+    """
+
+    def __init__(self) -> None:
+        self._locks: weakref.WeakKeyDictionary[asyncio.AbstractEventLoop, asyncio.Lock] = weakref.WeakKeyDictionary()
+
+    # When serializing and deserializing, we don't need to serialize the locks.
+    # Because another process will have its own set of event loops and its own lock.
+    def __getstate__(self) -> dict[str, Any]:
+        return {}
+
+    def __setstate__(self, state: dict[str, Any]) -> None:
+        self._locks = weakref.WeakKeyDictionary()
+
+    def _get_lock_for_current_loop(self) -> asyncio.Lock:
+        loop = asyncio.get_running_loop()
+        lock = self._locks.get(loop)
+        if lock is None:
+            lock = asyncio.Lock()
+            self._locks[loop] = lock
+        return lock
+
+    async def __aenter__(self) -> asyncio.Lock:
+        lock = self._get_lock_for_current_loop()
+        await lock.acquire()
+        return lock
+
+    async def __aexit__(self, exc_type: type[BaseException] | None, exc: BaseException | None, tb: Any) -> None:
+        loop = asyncio.get_running_loop()
+        lock = self._locks.get(loop)
+        if lock is None or not lock.locked():
+            raise RuntimeError("Lock released without being acquired")
+        lock.release()
+
+
+class _ThreadSafeAsyncLock:
+    """A threading.Lock that can be used in both async and sync contexts."""
+
+    def __init__(self):
+        self._lock = threading.Lock()
+
+    def __enter__(self):
+        self._lock.acquire()
+        return self
+
+    def __exit__(self, *args: Any, **kwargs: Any):
+        self._lock.release()
+
+    async def __aenter__(self):
+        # We run the blocking .acquire() in a thread pool so we don't block the event loop
+        loop = asyncio.get_running_loop()
+        await loop.run_in_executor(None, self._lock.acquire)
+        return self
+
+    async def __aexit__(self, *args: Any, **kwargs: Any):
+        # .release() is non-blocking, so we can call it directly
+        self._lock.release()

agentlightning/store/mongo.py

@@ -0,0 +1,165 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import logging
+import time
+import uuid
+from typing import (
+    Any,
+    Callable,
+    Dict,
+    List,
+    Mapping,
+    Optional,
+    Sequence,
+    TypeVar,
+    Union,
+)
+
+from pymongo import AsyncMongoClient
+
+from agentlightning.types import Attempt, AttemptedRollout, Rollout
+
+from .base import LightningStoreCapabilities, is_finished
+from .collection.mongo import MongoClientPool, MongoLightningCollections, MongoOperationPrometheusTracker
+from .collection_based import CollectionBasedLightningStore, healthcheck_before, tracked
+
+T_callable = TypeVar("T_callable", bound=Callable[..., Any])
+
+logger = logging.getLogger(__name__)
+
+
+def _generate_partition_id() -> str:
+    return "pt-" + hashlib.sha1(uuid.uuid4().bytes).hexdigest()[:12]
+
+
+class MongoLightningStore(CollectionBasedLightningStore[MongoLightningCollections]):
+    """
+    MongoDB implementation of LightningStore using MongoDB collections.
+    Data is persistent and can be shared between multiple processes.
+
+    Args:
+        client: The MongoDB client. Could be a string URI or an instance of AsyncMongoClient.
+        database: The MongoDB database. Could be a string name or an instance of AsyncDatabase.
+            You must provide at least one of client or database.
+        partition_id: The partition id. Useful when sharing the database among multiple Agent-lightning trainers.
+    """
+
+    def __init__(
+        self,
+        *,
+        client: AsyncMongoClient[Mapping[str, Any]] | str,
+        database_name: str | None = None,
+        partition_id: str | None = None,
+        prometheus: bool = False,
+    ) -> None:
+        self._enable_prometheus = prometheus
+        self._auto_created_client = False
+        if isinstance(client, str):
+            self._client = AsyncMongoClient[Mapping[str, Any]](client)
+            self._auto_created_client = True
+        else:
+            self._client = client
+        if database_name is None:
+            database_name = "agentlightning"
+            logger.info("No database name provided, using default 'agentlightning'")
+
+        if partition_id is None:
+            partition_id = _generate_partition_id()
+            logger.info("No partition id provided, generated a new one: %s", partition_id)
+
+        self._client_pool = MongoClientPool(self._client)
+
+        super().__init__(
+            collections=MongoLightningCollections(
+                self._client_pool,
+                database_name,
+                partition_id,
+                prometheus_tracker=MongoOperationPrometheusTracker(enabled=self._enable_prometheus),
+            ),
+            prometheus=self._enable_prometheus,
+        )
+
+    @property
+    def capabilities(self) -> LightningStoreCapabilities:
+        """Return the capabilities of the store."""
+        return LightningStoreCapabilities(
+            thread_safe=True,
+            async_safe=True,
+            zero_copy=True,
+            otlp_traces=False,
+        )
+
+    async def close(self) -> None:
+        """Close the store by closing the client pool."""
+        await self._client_pool.close()
+        # If I created the client, I should close it too.
+        if self._auto_created_client:
+            await self._client.close()
+
+    @tracked("wait_for_rollouts")
+    @healthcheck_before
+    async def wait_for_rollouts(self, *, rollout_ids: List[str], timeout: Optional[float] = None) -> List[Rollout]:
+        """Wait for specified rollouts to complete with a timeout.
+
+        Concurrently wait for all rollouts to complete with a timeout.
+        """
+        start_time = time.time()
+        current_time = start_time
+        deadline = start_time + timeout if timeout is not None else None
+
+        finished_rollouts: Dict[str, Rollout] = {}
+        unfinished_rollout_ids = set(rollout_ids)
+
+        while deadline is None or current_time <= deadline:
+            async with self.collections.atomic(
+                mode="r", snapshot=self._read_snapshot, labels=["rollouts"]
+            ) as collections:
+                # Query the rollouts that are not finished in a single query
+                rollouts = await collections.rollouts.query(
+                    filter={"rollout_id": {"within": list(unfinished_rollout_ids)}}
+                )
+            for rollout in rollouts.items:
+                if is_finished(rollout):
+                    finished_rollouts[rollout.rollout_id] = rollout
+                    unfinished_rollout_ids.remove(rollout.rollout_id)
+
+            if not unfinished_rollout_ids:
+                break
+
+            # Poll every 10 seconds by default
+            # Minus 0.1 to make sure the time is still sufficient for another call
+            rest_time = max(0.01, min(deadline - time.time() - 0.1, 10.0)) if deadline is not None else 10.0
+            await asyncio.sleep(rest_time)
+            current_time = time.time()
+
+        # Reorder the rollouts to match the input order
+        return [finished_rollouts[rollout_id] for rollout_id in rollout_ids if rollout_id in finished_rollouts]
+
+    @tracked("_unlocked_many_rollouts_to_attempted_rollouts")
+    async def _unlocked_many_rollouts_to_attempted_rollouts(
+        self, collections: MongoLightningCollections, rollouts: Sequence[Rollout]
+    ) -> List[Union[Rollout, AttemptedRollout]]:
+        """Query the latest attempts for the rollouts, and attach them to the rollout objects."""
+        async with collections.atomic(mode="r", snapshot=self._read_snapshot, labels=["attempts"]) as collections:
+            attempts = await collections.attempts.query(
+                filter={"rollout_id": {"within": [rollout.rollout_id for rollout in rollouts]}},
+                sort={"name": "sequence_id", "order": "desc"},
+            )
+        latest_attempts: Dict[str, Attempt] = {}
+        for attempt in attempts:
+            if attempt.rollout_id not in latest_attempts:
+                latest_attempts[attempt.rollout_id] = attempt
+            # Otherwise we ignore the attempt because there's already a newer attempt
+
+        return [
+            (
+                AttemptedRollout(**rollout.model_dump(), attempt=latest_attempts[rollout.rollout_id])
+                if rollout.rollout_id in latest_attempts
+                else rollout
+            )
+            for rollout in rollouts
+        ]

agentlightning/store/threading.py

@@ -3,7 +3,7 @@
 from __future__ import annotations
 
 import threading
-from typing import Any, Dict, List, Literal, Optional, Sequence
+from typing import Any, Dict, List, Literal, Optional, Sequence, Tuple
 
 from opentelemetry.sdk.trace import ReadableSpan
 
@@ -18,9 +18,11 @@ from agentlightning.types import (
     RolloutStatus,
     Span,
     TaskInput,
+    Worker,
+    WorkerStatus,
 )
 
-from .base import UNSET, LightningStore, Unset
+from .base import UNSET, LightningStore, LightningStoreCapabilities, LightningStoreStatistics, Unset
 
 
 class LightningStoreThreaded(LightningStore):
@@ -35,6 +37,21 @@ class LightningStoreThreaded(LightningStore):
         self.store = store
         self._lock = threading.Lock()
 
+    @property
+    def capabilities(self) -> LightningStoreCapabilities:
+        """Return the capabilities of the store."""
+        capabilities = self.store.capabilities
+        return {
+            **capabilities,
+            "async_safe": True,
+            "thread_safe": True,
+        }
+
+    async def statistics(self) -> LightningStoreStatistics:
+        """Return the statistics of the store."""
+        with self._lock:
+            return await self.store.statistics()
+
     async def start_rollout(
         self,
         input: TaskInput,
@@ -57,9 +74,9 @@ class LightningStoreThreaded(LightningStore):
         with self._lock:
             return await self.store.enqueue_rollout(input, mode, resources_id, config, metadata)
 
-    async def dequeue_rollout(self) -> Optional[AttemptedRollout]:
+    async def dequeue_rollout(self, worker_id: Optional[str] = None) -> Optional[AttemptedRollout]:
         with self._lock:
-            return await self.store.dequeue_rollout()
+            return await self.store.dequeue_rollout(worker_id=worker_id)
 
     async def start_attempt(self, rollout_id: str) -> AttemptedRollout:
         with self._lock:
@@ -68,15 +85,48 @@ class LightningStoreThreaded(LightningStore):
     async def query_rollouts(
         self,
         *,
+        status_in: Optional[Sequence[RolloutStatus]] = None,
+        rollout_id_in: Optional[Sequence[str]] = None,
+        rollout_id_contains: Optional[str] = None,
+        filter_logic: Literal["and", "or"] = "and",
+        sort_by: Optional[str] = None,
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
         status: Optional[Sequence[RolloutStatus]] = None,
         rollout_ids: Optional[Sequence[str]] = None,
-    ) -> List[Rollout]:
+    ) -> Sequence[Rollout]:
         with self._lock:
-            return await self.store.query_rollouts(status=status, rollout_ids=rollout_ids)
+            return await self.store.query_rollouts(
+                status_in=status_in,
+                rollout_id_in=rollout_id_in,
+                rollout_id_contains=rollout_id_contains,
+                filter_logic=filter_logic,
+                sort_by=sort_by,
+                sort_order=sort_order,
+                limit=limit,
+                offset=offset,
+                status=status,
+                rollout_ids=rollout_ids,
+            )
 
-    async def query_attempts(self, rollout_id: str) -> List[Attempt]:
+    async def query_attempts(
+        self,
+        rollout_id: str,
+        *,
+        sort_by: Optional[str] = "sequence_id",
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
+    ) -> Sequence[Attempt]:
         with self._lock:
-            return await self.store.query_attempts(rollout_id)
+            return await self.store.query_attempts(
+                rollout_id,
+                sort_by=sort_by,
+                sort_order=sort_order,
+                limit=limit,
+                offset=offset,
+            )
 
     async def get_rollout_by_id(self, rollout_id: str) -> Optional[Rollout]:
         with self._lock:
@@ -86,6 +136,26 @@ class LightningStoreThreaded(LightningStore):
         with self._lock:
             return await self.store.get_latest_attempt(rollout_id)
 
+    async def query_resources(
+        self,
+        *,
+        resources_id: Optional[str] = None,
+        resources_id_contains: Optional[str] = None,
+        sort_by: Optional[str] = None,
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
+    ) -> Sequence[ResourcesUpdate]:
+        with self._lock:
+            return await self.store.query_resources(
+                resources_id=resources_id,
+                resources_id_contains=resources_id_contains,
+                sort_by=sort_by,
+                sort_order=sort_order,
+                limit=limit,
+                offset=offset,
+            )
+
     async def add_resources(self, resources: NamedResources) -> ResourcesUpdate:
         with self._lock:
             return await self.store.add_resources(resources)
@@ -102,7 +172,11 @@ class LightningStoreThreaded(LightningStore):
         with self._lock:
             return await self.store.get_latest_resources()
 
-    async def add_span(self, span: Span) -> Span:
+    async def add_many_spans(self, spans: Sequence[Span]) -> Sequence[Span]:
+        with self._lock:
+            return await self.store.add_many_spans(spans)
+
+    async def add_span(self, span: Span) -> Optional[Span]:
         with self._lock:
             return await self.store.add_span(span)
 
@@ -112,7 +186,7 @@ class LightningStoreThreaded(LightningStore):
         attempt_id: str,
         readable_span: ReadableSpan,
         sequence_id: int | None = None,
-    ) -> Span:
+    ) -> Optional[Span]:
         with self._lock:
             return await self.store.add_otel_span(rollout_id, attempt_id, readable_span, sequence_id)
 
@@ -124,13 +198,47 @@ class LightningStoreThreaded(LightningStore):
         with self._lock:
             return await self.store.get_next_span_sequence_id(rollout_id, attempt_id)
 
+    async def get_many_span_sequence_ids(self, rollout_attempt_ids: Sequence[Tuple[str, str]]) -> Sequence[int]:
+        with self._lock:
+            return await self.store.get_many_span_sequence_ids(rollout_attempt_ids)
+
     async def query_spans(
         self,
         rollout_id: str,
         attempt_id: str | Literal["latest"] | None = None,
-    ) -> List[Span]:
+        *,
+        trace_id: Optional[str] = None,
+        trace_id_contains: Optional[str] = None,
+        span_id: Optional[str] = None,
+        span_id_contains: Optional[str] = None,
+        parent_id: Optional[str] = None,
+        parent_id_contains: Optional[str] = None,
+        name: Optional[str] = None,
+        name_contains: Optional[str] = None,
+        filter_logic: Literal["and", "or"] = "and",
+        limit: int = -1,
+        offset: int = 0,
+        sort_by: Optional[str] = "sequence_id",
+        sort_order: Literal["asc", "desc"] = "asc",
+    ) -> Sequence[Span]:
         with self._lock:
-            return await self.store.query_spans(rollout_id, attempt_id)
+            return await self.store.query_spans(
+                rollout_id,
+                attempt_id,
+                trace_id=trace_id,
+                trace_id_contains=trace_id_contains,
+                span_id=span_id,
+                span_id_contains=span_id_contains,
+                parent_id=parent_id,
+                parent_id_contains=parent_id_contains,
+                name=name,
+                name_contains=name_contains,
+                filter_logic=filter_logic,
+                limit=limit,
+                offset=offset,
+                sort_by=sort_by,
+                sort_order=sort_order,
+            )
 
     async def update_rollout(
         self,
@@ -171,3 +279,39 @@ class LightningStoreThreaded(LightningStore):
                 last_heartbeat_time=last_heartbeat_time,
                 metadata=metadata,
             )
+
+    async def query_workers(
+        self,
+        *,
+        status_in: Optional[Sequence[WorkerStatus]] = None,
+        worker_id_contains: Optional[str] = None,
+        filter_logic: Literal["and", "or"] = "and",
+        sort_by: Optional[str] = None,
+        sort_order: Literal["asc", "desc"] = "asc",
+        limit: int = -1,
+        offset: int = 0,
+    ) -> Sequence[Worker]:
+        with self._lock:
+            return await self.store.query_workers(
+                status_in=status_in,
+                worker_id_contains=worker_id_contains,
+                sort_by=sort_by,
+                sort_order=sort_order,
+                limit=limit,
+                offset=offset,
+            )
+
+    async def get_worker_by_id(self, worker_id: str) -> Optional[Worker]:
+        with self._lock:
+            return await self.store.get_worker_by_id(worker_id)
+
+    async def update_worker(
+        self,
+        worker_id: str,
+        heartbeat_stats: Dict[str, Any] | Unset = UNSET,
+    ) -> Worker:
+        with self._lock:
+            return await self.store.update_worker(
+                worker_id=worker_id,
+                heartbeat_stats=heartbeat_stats,
+            )

agentlightning/store/utils.py

@@ -1,7 +1,7 @@
 # Copyright (c) Microsoft. All rights reserved.
 
 import time
-from typing import Awaitable, Callable, List, cast
+from typing import Awaitable, Callable, Dict, List, Tuple
 
 from agentlightning.types import Attempt, AttemptedRollout, AttemptStatus, Rollout, RolloutConfig, RolloutStatus
 
@@ -9,66 +9,102 @@ UpdateRolloutStatus = Callable[[str, RolloutStatus], Awaitable[Rollout]]
 UpdateAttemptStatus = Callable[[str, str, AttemptStatus], Awaitable[Attempt]]
 
 
-async def propagate_status(
-    update_rollout_status: UpdateRolloutStatus,  # this should be unlocked
+LATENCY_BUCKETS = [
+    0.000001,
+    0.000002,
+    0.000005,
+    0.00001,
+    0.00002,
+    0.00005,
+    0.0001,
+    0.0002,
+    0.0005,
+    0.001,
+    0.002,
+    0.003,
+    0.005,
+    0.007,
+    0.01,
+    0.015,
+    0.02,
+    0.03,
+    0.05,
+    0.07,
+    0.1,
+    0.2,
+    0.3,
+    0.5,
+    0.7,
+    1.0,
+    2.0,
+    3.0,
+    5.0,
+    7.0,
+    10.0,
+    12.0,
+    15.0,
+    20.0,
+    25.0,
+    30.0,
+    40.0,
+    50.0,
+    60.0,
+    90.0,
+    120.0,
+    180.0,
+    240.0,
+    300.0,
+]
+
+
+async def rollout_status_from_attempt(
     attempt: Attempt,
     config: RolloutConfig,
-) -> Rollout:
+) -> RolloutStatus:
     """
     Propagate the status of an attempt to the rollout.
 
-    The rollout should be made sure in a state to be outdated.
-    Requeue the rollout if it should be retried.
-
-    This operation is completely unlocked. The caller is responsible for locking the store.
+    Returns:
+        The status of the rollout from the perspective of the attempt.
     """
     # Propagate the status directly to the rollout
     if attempt.status == "preparing" or attempt.status == "running" or attempt.status == "succeeded":
-        return await update_rollout_status(
-            attempt.rollout_id,
-            attempt.status,
-        )
+        return attempt.status
 
     if attempt.status == "failed" or attempt.status == "timeout" or attempt.status == "unresponsive":
         # Check if this status should trigger a retry
         if attempt.status in config.retry_condition:
             # If we haven't exceeded max attempts, retry
             if attempt.sequence_id < config.max_attempts:
-                return await update_rollout_status(
-                    attempt.rollout_id,
-                    "requeuing",
-                )
+                return "requeuing"
 
         # If we can't retry or shouldn't retry, mark as failed
-        return await update_rollout_status(
-            attempt.rollout_id,
-            "failed",
-        )
+        return "failed"
 
     raise ValueError(f"Invalid attempt status: {attempt.status}")
 
 
-async def healthcheck(
+async def scan_unhealthy_rollouts(
     rollouts: List[AttemptedRollout],
-    update_rollout_status: UpdateRolloutStatus,
-    update_attempt_status: UpdateAttemptStatus,
-) -> None:
+) -> Dict[Tuple[str, str], AttemptStatus]:
     """
     Perform health check on all running rollouts in the store.
 
     This method should be called periodically to:
 
-    1. Update rollout status to failed to succeeded when the attempt is done
-    2. Check for unresponsive attempts (no heartbeat or spans for a while)
-    3. Check for timed-out rollouts (running too long since start_time)
-    4. Update attempt/rollout status accordingly
+    1. Check for unresponsive attempts (no heartbeat or spans for a while)
+    2. Check for timed-out rollouts (running too long since start_time)
 
     This operation is completely unlocked. The caller is responsible for locking the store.
 
     Args:
-        store: The LightningStore instance to check rollouts from
+        rollouts: The list of running rollouts to check.
+
+    Returns:
+        A dictionary of updates to the rollouts.
     """
     current_time = time.time()
+    updates: Dict[Tuple[str, str], AttemptStatus] = {}
 
     for rollout in rollouts:
         config = rollout.config  # policy for retry and timeout
@@ -76,52 +112,31 @@ async def healthcheck(
         # Get the latest attempt for this rollout
         latest_attempt = rollout.attempt
         if not latest_attempt:
-            continue
-
-        # Check if the attempt has already failed or succeeded
-        if latest_attempt.status == "failed" or latest_attempt.status == "succeeded":
-            await propagate_status(update_rollout_status, latest_attempt, config)
+            # This should not happen
             continue
 
         # Check for timeout condition (based on attempt start_time, instead of rollout start_time)
         if config.timeout_seconds is not None and current_time - latest_attempt.start_time > config.timeout_seconds:
-            await update_attempt_status(
-                latest_attempt.rollout_id,
-                latest_attempt.attempt_id,
-                "timeout",
-            )
+            updates[(latest_attempt.rollout_id, latest_attempt.attempt_id)] = "timeout"
             continue
 
         # Check for unresponsive condition (based on last heartbeat)
-        if latest_attempt.last_heartbeat_time:
-            if latest_attempt.status == "preparing":
-                # If still preparing, mark it as running
-                latest_attempt = await update_attempt_status(
-                    latest_attempt.rollout_id,
-                    latest_attempt.attempt_id,
-                    "running",
-                )
+        # (1) Haven't received heartbeat for a while
+        if (
+            latest_attempt.last_heartbeat_time
+            and config.unresponsive_seconds is not None
+            and current_time - latest_attempt.last_heartbeat_time > config.unresponsive_seconds
+        ):
+            updates[(latest_attempt.rollout_id, latest_attempt.attempt_id)] = "unresponsive"
+            continue
 
-            # Haven't received heartbeat for a while
-            if (
-                config.unresponsive_seconds is not None
-                and current_time - cast(float, latest_attempt.last_heartbeat_time) > config.unresponsive_seconds
-            ):
-                await update_attempt_status(
-                    latest_attempt.rollout_id,
-                    latest_attempt.attempt_id,
-                    "unresponsive",
-                )
-                continue
-
-        # Check if there's no last heartbeat (no spans) at all
+        # (2) Check if there's no last heartbeat (no spans) at all
         if (
             latest_attempt.last_heartbeat_time is None
             and config.unresponsive_seconds is not None
             and current_time - latest_attempt.start_time > config.unresponsive_seconds
         ):
-            await update_attempt_status(
-                latest_attempt.rollout_id,
-                latest_attempt.attempt_id,
-                "unresponsive",
-            )
+            updates[(latest_attempt.rollout_id, latest_attempt.attempt_id)] = "unresponsive"
+            continue
+
+    return updates

agentlightning/tracer/agentops.py

@@ -2,25 +2,24 @@
 
 from __future__ import annotations
 
-import asyncio
 import logging
 import os
-import threading
+import warnings
 from contextlib import asynccontextmanager, contextmanager
-from typing import TYPE_CHECKING, Any, AsyncGenerator, Awaitable, Iterator, List, Optional
+from typing import TYPE_CHECKING, Any, AsyncGenerator, Iterator, List, Optional
 
 import agentops
 import agentops.sdk.core
+import opentelemetry.trace as trace_api
 from agentops.sdk.core import TracingCore
-from agentops.sdk.processors import SpanProcessor
-from opentelemetry.instrumentation.utils import suppress_instrumentation
-from opentelemetry.sdk.trace import ReadableSpan
+from opentelemetry.sdk.trace import TracerProvider as TracerProviderImpl
+from opentelemetry.trace import get_tracer_provider
+from opentelemetry.trace.status import StatusCode
 
 from agentlightning.instrumentation import instrument_all, uninstrument_all
-from agentlightning.instrumentation.agentops import AgentOpsServerManager
 from agentlightning.store.base import LightningStore
 
-from .base import Tracer
+from .otel import LightningSpanProcessor, OtelTracer
 
 if TYPE_CHECKING:
     from agentops.integration.callbacks.langchain import LangchainCallbackHandler
@@ -29,7 +28,7 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
-class AgentOpsTracer(Tracer):
+class AgentOpsTracer(OtelTracer):
     """Traces agent execution using AgentOps.
 
     This tracer provides functionality to capture execution details using the
@@ -56,95 +55,37 @@ class AgentOpsTracer(Tracer):
         self.instrument_managed = instrument_managed
         self.daemon = daemon
 
-        self._agentops_server_manager = AgentOpsServerManager(self.daemon)
-        self._agentops_server_port_val: Optional[int] = None
-
         if not self.agentops_managed:
             logger.warning("agentops_managed=False. You are responsible for AgentOps setup.")
         if not self.instrument_managed:
             logger.warning("instrument_managed=False. You are responsible for all instrumentation.")
 
-    def __getstate__(self):
-        state = self.__dict__.copy()
-        state["_agentops_server_manager"] = None  # Exclude the unpicklable server manager
-        # _agentops_server_port_val (int) is inherently picklable and will be included.
-        logger.debug(f"Getting state for pickling Trainer (PID {os.getpid()}). _agentops_server_manager excluded.")
-        return state
-
-    def __setstate__(self, state: Any):
-        self.__dict__.update(state)
-        # In child process, self._agentops_server_manager will be None.
-        logger.debug(f"Setting state for unpickled Trainer (PID {os.getpid()}). _agentops_server_manager is None.")
-
-    def init(self, *args: Any, **kwargs: Any):
-        if self.agentops_managed and self._agentops_server_manager:
-            self._agentops_server_manager.start()
-            self._agentops_server_port_val = self._agentops_server_manager.get_port()
-            if self._agentops_server_port_val is None:
-                if (
-                    self._agentops_server_manager.server_process is not None
-                    and self._agentops_server_manager.server_process.is_alive()
-                ):
-                    raise RuntimeError("AgentOps server started but port is None. Check server manager logic.")
-                elif (
-                    self._agentops_server_port_val is None and self._agentops_server_manager.server_process is None
-                ):  # Server failed to start
-                    raise RuntimeError("AgentOps server manager indicates server is not running and port is None.")
-
-    def teardown(self):
-        if self.agentops_managed:
-            self._agentops_server_manager.stop()
-            logger.info("AgentOps server stopped.")
-
     def instrument(self, worker_id: int):
         instrument_all()
 
     def uninstrument(self, worker_id: int):
         uninstrument_all()
 
-    def init_worker(self, worker_id: int):
-        super().init_worker(worker_id)
-        logger.info(f"[Worker {worker_id}] Setting up tracer...")  # worker_id included in process name
+    def _initialize_tracer_provider(self, worker_id: int):
+        logger.info(f"[Worker {worker_id}] Setting up AgentOps tracer...")  # worker_id included in process name
 
         if self.instrument_managed:
             self.instrument(worker_id)
             logger.info(f"[Worker {worker_id}] Instrumentation applied.")
 
         if self.agentops_managed:
-            if self._agentops_server_port_val:  # Use the stored, picklable port value
-                base_url = f"http://localhost:{self._agentops_server_port_val}"
-                env_vars_to_set = {
-                    "AGENTOPS_API_KEY": "dummy",
-                    "AGENTOPS_API_ENDPOINT": base_url,
-                    "AGENTOPS_APP_URL": f"{base_url}/notavailable",
-                    "AGENTOPS_EXPORTER_ENDPOINT": f"{base_url}/traces",
-                }
-                for key, value in env_vars_to_set.items():
-                    os.environ[key] = value
-                    logger.info(f"[Worker {worker_id}] Env var set: {key}={value}")
-            else:
-                logger.warning(
-                    f"[Worker {worker_id}] AgentOps managed, but local server port is not available. Client may not connect as expected."
-                )
-
+            os.environ.setdefault("AGENTOPS_API_KEY", "dummy")
             if not agentops.get_client().initialized:
-                agentops.init()  # type: ignore
+                agentops.init(auto_start_session=False)  # type: ignore
                 logger.info(f"[Worker {worker_id}] AgentOps client initialized.")
             else:
                 logger.warning(f"[Worker {worker_id}] AgentOps client was already initialized.")
 
         self._lightning_span_processor = LightningSpanProcessor()
 
-        try:
-            # new versions
-            instance = agentops.sdk.core.tracer
-            # TODO: The span processor cannot be deleted once added.
-            # This might be a problem if the tracer is entered and exited multiple times.
-            instance.provider.add_span_processor(self._lightning_span_processor)  # type: ignore
-        except AttributeError:
-            # old versions
-            instance = TracingCore.get_instance()  # type: ignore
-            instance._provider.add_span_processor(self._lightning_span_processor)  # type: ignore
+        # TODO: The span processor cannot be deleted once added.
+        # This might be a problem if the tracer is entered and exited multiple times.
+        self._get_tracer_provider().add_span_processor(self._lightning_span_processor)  # type: ignore
 
     def teardown_worker(self, worker_id: int) -> None:
         super().teardown_worker(worker_id)
@@ -161,7 +102,7 @@ class AgentOpsTracer(Tracer):
         store: Optional[LightningStore] = None,
         rollout_id: Optional[str] = None,
         attempt_id: Optional[str] = None,
-    ) -> AsyncGenerator[LightningSpanProcessor, None]:
+    ) -> AsyncGenerator[trace_api.Tracer, None]:
         """
         Starts a new tracing context. This should be used as a context manager.
 
@@ -172,12 +113,18 @@ class AgentOpsTracer(Tracer):
             attempt_id: Optional attempt ID to add the spans to.
 
         Yields:
-            The [`LightningSpanProcessor`][agentlightning.tracer.agentops.LightningSpanProcessor] instance to collect spans.
+            The OpenTelemetry tracer instance to collect spans.
         """
-        with self._trace_context_sync(
-            name=name, store=store, rollout_id=rollout_id, attempt_id=attempt_id
-        ) as processor:
-            yield processor
+        if store is not None:
+            warnings.warn(
+                "store is deprecated in favor of init_worker(). It will be removed in the future.",
+                DeprecationWarning,
+                stacklevel=3,
+            )
+        else:
+            store = self._store
+        with self._trace_context_sync(name=name, store=store, rollout_id=rollout_id, attempt_id=attempt_id) as tracer:
+            yield tracer
 
     @contextmanager
     def _trace_context_sync(
@@ -187,31 +134,51 @@ class AgentOpsTracer(Tracer):
         store: Optional[LightningStore] = None,
         rollout_id: Optional[str] = None,
         attempt_id: Optional[str] = None,
-    ) -> Iterator[LightningSpanProcessor]:
+    ) -> Iterator[trace_api.Tracer]:
         """Implementation of `trace_context` for synchronous execution."""
         if not self._lightning_span_processor:
             raise RuntimeError("LightningSpanProcessor is not initialized. Call init_worker() first.")
+        tracer_provider = self._get_tracer_provider()
 
+        kwargs: dict[str, Any] = {}
+        if name is not None:
+            kwargs["trace_name"] = name
+        elif rollout_id is not None:
+            kwargs["trace_name"] = rollout_id
         if store is not None and rollout_id is not None and attempt_id is not None:
+            if store.capabilities.get("otlp_traces", False) is True:
+                logger.debug(f"Tracing to LightningStore rollout_id={rollout_id}, attempt_id={attempt_id}")
+                self._enable_native_otlp_exporter(store, rollout_id, attempt_id)
+            else:
+                self._disable_native_otlp_exporter()
             ctx = self._lightning_span_processor.with_context(store=store, rollout_id=rollout_id, attempt_id=attempt_id)
-            with ctx as processor:
-                yield processor
+            with ctx:
+                # AgentOps end_trace and start_trace must live inside the lightning span processor context.
+                # Otherwise some traces might not be recorded.
+                with self._agentops_trace_context(rollout_id, attempt_id, kwargs):
+                    yield trace_api.get_tracer(__name__, tracer_provider=tracer_provider)
         elif store is None and rollout_id is None and attempt_id is None:
+            # TODO: Add tests to cover both paths
+            self._disable_native_otlp_exporter()
             with self._lightning_span_processor:
-                yield self._lightning_span_processor
+                with self._agentops_trace_context(None, None, kwargs):
+                    yield trace_api.get_tracer(__name__, tracer_provider=tracer_provider)
         else:
             raise ValueError("store, rollout_id, and attempt_id must be either all provided or all None")
 
-    def get_last_trace(self) -> List[ReadableSpan]:
-        """
-        Retrieves the raw list of captured spans from the most recent trace.
-
-        Returns:
-            A list of OpenTelemetry `ReadableSpan` objects.
-        """
-        if not self._lightning_span_processor:
-            raise RuntimeError("LightningSpanProcessor is not initialized. Call init_worker() first.")
-        return self._lightning_span_processor.spans()
+    @contextmanager
+    def _agentops_trace_context(self, rollout_id: Optional[str], attempt_id: Optional[str], kwargs: dict[str, Any]):
+        trace = agentops.start_trace(**kwargs)
+        status = StatusCode.OK  # type: ignore
+        try:
+            yield
+        except Exception as e:
+            # This will catch errors in user code.
+            status = StatusCode.ERROR  # type: ignore
+            logger.error(f"Trace failed for rollout_id={rollout_id}, attempt_id={attempt_id}: {e}")
+            raise  # should reraise the error here so that runner can handle it
+        finally:
+            agentops.end_trace(trace, end_state=status)  # type: ignore
 
     def get_langchain_handler(self, tags: List[str] | None = None) -> LangchainCallbackHandler:
         """
@@ -239,135 +206,26 @@ class AgentOpsTracer(Tracer):
 
     get_langchain_callback_handler = get_langchain_handler  # alias
 
+    def _get_tracer_provider(self) -> TracerProviderImpl:
+        try:
+            # new versions
+            instance = agentops.sdk.core.tracer
+            if instance.provider is None:
+                raise RuntimeError("AgentOps TracerProvider is not initialized.")
 
-class LightningSpanProcessor(SpanProcessor):
-    """Span processor that subclasses OpenTelemetry's `SpanProcessor` and adds support to dump traces
-    to a [`LightningStore`][agentlightning.LightningStore].
-    """
+            if get_tracer_provider() is not instance.provider:
+                logger.error(
+                    "Mismatch between global singleton TracerProvider and AgentOps TracerProvider. "
+                    "AgentOps might not work properly."
+                )
 
-    def __init__(self):
-        self._spans: List[ReadableSpan] = []
+            if not isinstance(instance.provider, TracerProviderImpl):  # type: ignore
+                raise RuntimeError("Unsupported TracerProvider type for AgentOps instrumentation.")
 
-        # Store related context and states
-        self._store: Optional[LightningStore] = None
-        self._rollout_id: Optional[str] = None
-        self._attempt_id: Optional[str] = None
-        self._lock = threading.Lock()
-
-        # private asyncio loop running in a daemon thread
-        self._loop_ready = threading.Event()
-        self._loop: Optional[asyncio.AbstractEventLoop] = None
-        self._loop_thread = threading.Thread(target=self._loop_runner, name="otel-loop", daemon=True)
-        self._loop_thread.start()
-        self._loop_ready.wait()  # loop is ready
-
-    def _loop_runner(self):
-        loop = asyncio.new_event_loop()
-        self._loop = loop
-        asyncio.set_event_loop(loop)
-        self._loop_ready.set()
-        loop.run_forever()
-        loop.close()
-
-    def __enter__(self):
-        self._last_trace = None
-        self._spans = []
-        return self
-
-    def __exit__(self, exc_type: Any, exc_val: Any, exc_tb: Any):
-        self._store = None
-        self._rollout_id = None
-        self._attempt_id = None
-
-    def _await_in_loop(self, coro: Awaitable[Any], timeout: Optional[float] = None) -> Any:
-        # submit to the dedicated loop and wait synchronously
-        if self._loop is None:
-            raise RuntimeError("Loop is not initialized. This should not happen.")
-
-        # If already on the exporter loop thread, schedule and return immediately.
-        # ---------------------------------------------------------------------------
-        # WHY THIS CONDITIONAL EXISTS:
-        # In rare cases, span.end() is triggered from a LangchainCallbackHandler.__del__
-        # (or another finalizer) while the Python garbage collector is running on the
-        # *same thread* that owns our exporter event loop ("otel-loop").
-        #
-        # When that happens, on_end() executes on the exporter loop thread itself.
-        # If we were to call `asyncio.run_coroutine_threadsafe(...).result()` here,
-        # it would deadlock immediately — because the loop cannot both wait on and run
-        # the same coroutine. The Future stays pending forever and the loop stops
-        # processing scheduled callbacks.
-        #
-        # To avoid that self-deadlock, we detect when on_end() runs on the exporter
-        # loop thread. If so, we *schedule* the coroutine on the loop (fire-and-forget)
-        # instead of blocking with .result().
-        #
-        # This situation can occur because Python calls __del__ in whatever thread
-        # releases the last reference, which can easily be our loop thread if the
-        # object is dereferenced during loop._run_once().
-        # ---------------------------------------------------------------------------
-        if threading.current_thread() is self._loop_thread:
-            self._loop.call_soon_threadsafe(asyncio.create_task, coro)  # type: ignore
-            return None
-
-        fut = asyncio.run_coroutine_threadsafe(coro, self._loop)  # type: ignore
-        return fut.result(timeout=timeout)  # raises on error  # type: ignore
-
-    def shutdown(self) -> None:
-        if self._loop:
-            self._loop.call_soon_threadsafe(self._loop.stop)
-            self._loop_thread.join(timeout=5)
-            self._loop = None
-
-    def force_flush(self, timeout_millis: int = 30000) -> bool:
-        return True
-
-    def spans(self) -> List[ReadableSpan]:
-        """
-        Get the list of spans collected by this processor.
-        This is useful for debugging and testing purposes.
-
-        Returns:
-            List of ReadableSpan objects collected during tracing.
-        """
-        return self._spans
-
-    def with_context(self, store: LightningStore, rollout_id: str, attempt_id: str):
-        # simple context manager without nesting into asyncio
-        class _Ctx:
-            def __enter__(_):  # type: ignore
-                with self._lock:
-                    self._store, self._rollout_id, self._attempt_id = store, rollout_id, attempt_id
-                    self._last_trace = None
-                    self._spans = []
-                return self
-
-            def __exit__(_, exc_type, exc, tb):  # type: ignore
-                with self._lock:
-                    self._store = self._rollout_id = self._attempt_id = None
-
-        return _Ctx()
-
-    def on_end(self, span: ReadableSpan) -> None:
-        """
-        Process a span when it ends.
-
-        Args:
-            span: The span that has ended.
-        """
-        # Skip if span is not sampled
-        if not span.context or not span.context.trace_flags.sampled:
-            return
-
-        if self._store and self._rollout_id and self._attempt_id:
-            try:
-                # Submit add_otel_span to the event loop and wait for it to complete
-                with suppress_instrumentation():
-                    self._await_in_loop(
-                        self._store.add_otel_span(self._rollout_id, self._attempt_id, span),
-                        timeout=60.0,
-                    )
-            except Exception:
-                # log; on_end MUST NOT raise
-                logger.exception(f"Error adding span to store: {span.name}")
-
-        self._spans.append(span)
+            self._tracer_provider = instance.provider
+            return self._tracer_provider
+        except AttributeError:
+            # old versions
+            instance = TracingCore.get_instance()  # type: ignore
+            self._tracer_provider = instance._provider  # type: ignore
+            return self._tracer_provider  # type: ignore

agentlightning/tracer/base.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import logging
+from contextlib import contextmanager
 from typing import TYPE_CHECKING, Any, AsyncContextManager, Awaitable, Callable, ContextManager, List, Optional
 
 from opentelemetry.sdk.trace import ReadableSpan
@@ -51,6 +52,18 @@ class Tracer(ParallelWorkerBase):
     ```
     """
 
+    _store: Optional[LightningStore] = None
+
+    def init_worker(self, worker_id: int, store: Optional[LightningStore] = None) -> None:
+        """Initialize the tracer for a worker.
+
+        Args:
+            worker_id: The ID of the worker.
+            store: The store to add the spans to. If it's provided, traces will be added to the store when tracing.
+        """
+        super().init_worker(worker_id)
+        self._store = store
+
     def trace_context(
         self,
         name: Optional[str] = None,
@@ -67,11 +80,9 @@ class Tracer(ParallelWorkerBase):
         within the `with` block are collected and made available via
         [`get_last_trace`][agentlightning.Tracer.get_last_trace].
 
-        If a store is provided, the spans will be added to the store when tracing.
-
         Args:
             name: The name for the root span of this trace context.
-            store: The store to add the spans to.
+            store: The store to add the spans to. Deprecated in favor of passing store to init_worker().
             rollout_id: The rollout ID to add the spans to.
             attempt_id: The attempt ID to add the spans to.
         """
@@ -81,7 +92,6 @@ class Tracer(ParallelWorkerBase):
         self,
         name: Optional[str] = None,
         *,
-        store: Optional[LightningStore] = None,
         rollout_id: Optional[str] = None,
         attempt_id: Optional[str] = None,
     ) -> ContextManager[Any]:
@@ -138,3 +148,30 @@ class Tracer(ParallelWorkerBase):
         """
         logger.warning(f"{self.__class__.__name__} does not provide a LangChain callback handler.")
         return None
+
+    @contextmanager
+    def lifespan(self, store: Optional[LightningStore] = None):
+        """A context manager to manage the lifespan of the tracer.
+
+        This can be used to set up and tear down any necessary resources
+        for the tracer, useful for debugging purposes.
+
+        Args:
+            store: The store to add the spans to. If it's provided, traces will be added to the store when tracing.
+        """
+        has_init = False
+        has_init_worker = False
+        try:
+            self.init()
+            has_init = True
+
+            self.init_worker(0, store)
+            has_init_worker = True
+
+            yield
+
+        finally:
+            if has_init_worker:
+                self.teardown_worker(0)
+            if has_init:
+                self.teardown()

agentlightning/tracer/http.py

@@ -19,6 +19,8 @@ from opentelemetry.trace.span import (
     TraceState,
 )
 
+from agentlightning.store import LightningStore
+
 from .base import Tracer
 
 logger = logging.getLogger(__name__)
@@ -68,14 +70,15 @@ class HttpTracer(Tracer):
         self.subprocess_mode = subprocess_mode
         self.subprocess_timeout = subprocess_timeout
 
-    def init_worker(self, worker_id: int) -> None:
+    def init_worker(self, worker_id: int, store: Optional[LightningStore] = None) -> None:
         """
         Initialize the tracer in a worker process.
 
         Args:
             worker_id: The ID of the worker process.
+            store: The store to add the spans to.
         """
-        super().init_worker(worker_id)
+        super().init_worker(worker_id, store)
         logger.info(f"[Worker {worker_id}] HttpTracer initialized.")
 
     @asynccontextmanager

agentlightning/tracer/otel.py

@@ -2,16 +2,27 @@
 
 from __future__ import annotations
 
+import asyncio
 import logging
+import threading
+import warnings
 from contextlib import asynccontextmanager
-from typing import AsyncGenerator, List, Optional
+from typing import Any, AsyncGenerator, Awaitable, List, Optional
 
 import opentelemetry.trace as trace_api
-from opentelemetry.sdk.trace import ReadableSpan, TracerProvider
+from agentops.sdk.core import BatchSpanProcessor
+from opentelemetry.instrumentation.utils import suppress_instrumentation
+from opentelemetry.sdk.resources import Resource
+from opentelemetry.sdk.trace import ReadableSpan, SpanProcessor
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.sdk.trace import TracerProvider as TracerProviderImpl
+from opentelemetry.sdk.trace.export import SimpleSpanProcessor
 
+from agentlightning.semconv import LightningResourceAttributes
 from agentlightning.store.base import LightningStore
+from agentlightning.utils.otel import get_tracer_provider
+from agentlightning.utils.otlp import LightningStoreOTLPExporter
 
-from .agentops import LightningSpanProcessor  # FIXME: This import should be from otel to agentops
 from .base import Tracer
 
 logger = logging.getLogger(__name__)
@@ -29,25 +40,43 @@ class OtelTracer(Tracer):
         # This provider is only initialized when the worker is initialized.
         self._tracer_provider: Optional[TracerProvider] = None
         self._lightning_span_processor: Optional[LightningSpanProcessor] = None
+        self._simple_span_processor: Optional[SimpleSpanProcessor] = None
+        self._otlp_span_exporter: Optional[LightningStoreOTLPExporter] = None
         self._initialized: bool = False
 
-    def init_worker(self, worker_id: int):
-        super().init_worker(worker_id)
+    def init_worker(self, worker_id: int, store: Optional[LightningStore] = None):
+        super().init_worker(worker_id, store)
+        self._initialize_tracer_provider(worker_id)
+
+    def _initialize_tracer_provider(self, worker_id: int):
         logger.info(f"[Worker {worker_id}] Setting up OpenTelemetry tracer...")
 
         if self._initialized:
-            logger.error("Tracer provider is already initialized. OpenTelemetry may not work as expected.")
+            logger.info(f"[Worker {worker_id}] Tracer provider is already initialized. Skipping initialization.")
+            return
 
-        tracer_provider = TracerProvider()
-        trace_api.set_tracer_provider(tracer_provider)
+        try:
+            get_tracer_provider()
+            logger.error(
+                f"[Worker {worker_id}] Tracer provider is already initialized but not by OtelTracer. OpenTelemetry may not work as expected."
+            )
+        except RuntimeError:
+            logger.debug(f"[Worker {worker_id}] Tracer provider is not initialized by OtelTracer. Initializing it now.")
+
+        self._tracer_provider = TracerProvider()
+        trace_api.set_tracer_provider(self._tracer_provider)
         self._lightning_span_processor = LightningSpanProcessor()
-        tracer_provider.add_span_processor(self._lightning_span_processor)
+        self._tracer_provider.add_span_processor(self._lightning_span_processor)
+        self._otlp_span_exporter = LightningStoreOTLPExporter()
+        self._simple_span_processor = SimpleSpanProcessor(self._otlp_span_exporter)
+        self._tracer_provider.add_span_processor(self._simple_span_processor)
         self._initialized = True
 
+        logger.info(f"[Worker {worker_id}] OpenTelemetry tracer provider initialized.")
+
     def teardown_worker(self, worker_id: int):
         super().teardown_worker(worker_id)
-        logger.info(f"[Worker {worker_id}] Tearing down OpenTelemetry tracer...")
-        self._tracer_provider = None
+        logger.info(f"[Worker {worker_id}] Tearing down OpenTelemetry tracer does NOT remove the tracer provider.")
 
     @asynccontextmanager
     async def trace_context(
@@ -57,7 +86,7 @@ class OtelTracer(Tracer):
         store: Optional[LightningStore] = None,
         rollout_id: Optional[str] = None,
         attempt_id: Optional[str] = None,
-    ) -> AsyncGenerator[LightningSpanProcessor, None]:
+    ) -> AsyncGenerator[trace_api.Tracer, None]:
         """
         Starts a new tracing context. This should be used as a context manager.
 
@@ -68,20 +97,37 @@ class OtelTracer(Tracer):
             attempt_id: Optional attempt ID to add the spans to.
 
         Yields:
-            The LightningSpanProcessor instance to collect spans.
+            The OpenTelemetry tracer instance to collect spans.
         """
         if not self._lightning_span_processor:
             raise RuntimeError("LightningSpanProcessor is not initialized. Call init_worker() first.")
 
-        if store is not None and rollout_id is not None and attempt_id is not None:
-            ctx = self._lightning_span_processor.with_context(store=store, rollout_id=rollout_id, attempt_id=attempt_id)
-            with ctx as processor:
-                yield processor
-        elif store is None and rollout_id is None and attempt_id is None:
-            with self._lightning_span_processor:
-                yield self._lightning_span_processor
+        if store is not None:
+            warnings.warn(
+                "store is deprecated in favor of init_worker(). It will be removed in the future.",
+                DeprecationWarning,
+                stacklevel=3,
+            )
         else:
-            raise ValueError("store, rollout_id, and attempt_id must be either all provided or all None")
+            store = self._store
+
+        if rollout_id is not None and attempt_id is not None:
+            if store is None:
+                raise ValueError("store is required to be initialized when rollout_id and attempt_id are provided")
+            if store.capabilities.get("otlp_traces", False) is True:
+                logger.debug(f"Tracing to LightningStore rollout_id={rollout_id}, attempt_id={attempt_id}")
+                self._enable_native_otlp_exporter(store, rollout_id, attempt_id)
+            else:
+                self._disable_native_otlp_exporter()
+            ctx = self._lightning_span_processor.with_context(store=store, rollout_id=rollout_id, attempt_id=attempt_id)
+            with ctx:
+                yield trace_api.get_tracer(__name__, tracer_provider=self._tracer_provider)
+        elif rollout_id is None and attempt_id is None:
+            self._disable_native_otlp_exporter()
+            with self._lightning_span_processor:
+                yield trace_api.get_tracer(__name__, tracer_provider=self._tracer_provider)
+        else:
+            raise ValueError("rollout_id and attempt_id must be either all provided or all None")
 
     def get_last_trace(self) -> List[ReadableSpan]:
         """
@@ -93,3 +139,246 @@ class OtelTracer(Tracer):
         if not self._lightning_span_processor:
             raise RuntimeError("LightningSpanProcessor is not initialized. Call init_worker() first.")
         return self._lightning_span_processor.spans()
+
+    def _get_tracer_provider(self) -> TracerProviderImpl:
+        if self._tracer_provider is None:
+            raise RuntimeError("TracerProvider is not initialized. Call init_worker() first.")
+        return self._tracer_provider
+
+    def _enable_native_otlp_exporter(self, store: LightningStore, rollout_id: str, attempt_id: str):
+        tracer_provider = self._get_tracer_provider()
+        active_span_processor = tracer_provider._active_span_processor  # pyright: ignore[reportPrivateUsage]
+
+        # Override the resources so that the server knows where the request comes from.
+        tracer_provider._resource = tracer_provider._resource.merge(  # pyright: ignore[reportPrivateUsage]
+            Resource.create(
+                {
+                    LightningResourceAttributes.ROLLOUT_ID.value: rollout_id,
+                    LightningResourceAttributes.ATTEMPT_ID.value: attempt_id,
+                }
+            )
+        )
+        instrumented = False
+        candidates: List[str] = []
+        for processor in active_span_processor._span_processors:  # pyright: ignore[reportPrivateUsage]
+            if isinstance(processor, LightningSpanProcessor):
+                # We don't need the LightningSpanProcessor any more.
+                logger.debug("LightningSpanProcessor already present in TracerProvider, disabling it.")
+                processor.disable_store_submission = True
+            elif isinstance(processor, (SimpleSpanProcessor, BatchSpanProcessor)):
+                # Instead, we rely on the OTLPSpanExporter to send spans to the store.
+                if isinstance(processor.span_exporter, LightningStoreOTLPExporter):
+                    processor.span_exporter.enable_store_otlp(store.otlp_traces_endpoint(), rollout_id, attempt_id)
+                    logger.debug(f"Set LightningStoreOTLPExporter endpoint to {store.otlp_traces_endpoint()}")
+                    instrumented = True
+                else:
+                    candidates.append(
+                        f"{processor.__class__.__name__} with {processor.span_exporter.__class__.__name__}"
+                    )
+            else:
+                candidates.append(f"{processor.__class__.__name__}")
+
+        if not instrumented:
+            raise RuntimeError(
+                "Failed to enable native OTLP exporter: no BatchSpanProcessor or SimpleSpanProcessor with "
+                "LightningStoreOTLPExporter found in TracerProvider. Please try using a non-OTLP store."
+                "Candidates are: " + ", ".join(candidates)
+            )
+
+    def _disable_native_otlp_exporter(self):
+        tracer_provider = self._get_tracer_provider()
+        active_span_processor = tracer_provider._active_span_processor  # pyright: ignore[reportPrivateUsage]
+        tracer_provider._resource = tracer_provider._resource.merge(  # pyright: ignore[reportPrivateUsage]
+            Resource.create(
+                {
+                    LightningResourceAttributes.ROLLOUT_ID.value: "",
+                    LightningResourceAttributes.ATTEMPT_ID.value: "",
+                }
+            )
+        )  # reset resource
+        for processor in active_span_processor._span_processors:  # pyright: ignore[reportPrivateUsage]
+            if isinstance(processor, LightningSpanProcessor):
+                # We will be in need of the LightningSpanProcessor again.
+                logger.debug("Enabling LightningSpanProcessor in TracerProvider.")
+                processor.disable_store_submission = False
+
+
+class LightningSpanProcessor(SpanProcessor):
+    """Span processor that subclasses OpenTelemetry's `SpanProcessor` and adds support to dump traces
+    to a [`LightningStore`][agentlightning.LightningStore].
+
+    It serves two purposes:
+
+    1. Records all the spans in a local buffer.
+    2. Submits the spans to the event loop to be added to the store.
+    """
+
+    def __init__(self, disable_store_submission: bool = False):
+        self._disable_store_submission: bool = disable_store_submission
+        self._spans: List[ReadableSpan] = []
+
+        # Store related context and states
+        self._store: Optional[LightningStore] = None
+        self._rollout_id: Optional[str] = None
+        self._attempt_id: Optional[str] = None
+        self._lock = threading.Lock()
+
+        # private asyncio loop running in a daemon thread
+        self._loop_ready = threading.Event()
+        self._loop: Optional[asyncio.AbstractEventLoop] = None
+        self._loop_thread: Optional[threading.Thread] = None
+
+    def __repr__(self) -> str:
+        return (
+            f"{self.__class__.__name__}("
+            + f"disable_store_submission={self.disable_store_submission}, "
+            + f"store={self.store!r}, "
+            + f"rollout_id={self.rollout_id!r}, "
+            + f"attempt_id={self.attempt_id!r})"
+        )
+
+    @property
+    def store(self) -> Optional[LightningStore]:
+        """The store to submit the spans to."""
+        return self._store
+
+    @property
+    def rollout_id(self) -> Optional[str]:
+        """The rollout ID to submit the spans to."""
+        return self._rollout_id
+
+    @property
+    def attempt_id(self) -> Optional[str]:
+        """The attempt ID to submit the spans to."""
+        return self._attempt_id
+
+    @property
+    def disable_store_submission(self) -> bool:
+        """Whether to disable submitting spans to the store."""
+        return self._disable_store_submission
+
+    @disable_store_submission.setter
+    def disable_store_submission(self, value: bool) -> None:
+        self._disable_store_submission = value
+
+    def _ensure_loop(self) -> None:
+        if self._loop_thread is None or self._loop is None:
+            self._loop_ready.clear()
+            self._loop_thread = threading.Thread(target=self._loop_runner, name="otel-loop", daemon=True)
+            self._loop_thread.start()
+            self._loop_ready.wait()  # loop is ready
+
+    def _loop_runner(self):
+        loop = asyncio.new_event_loop()
+        self._loop = loop
+        asyncio.set_event_loop(loop)
+        self._loop_ready.set()
+        loop.run_forever()
+        loop.close()
+
+    def __enter__(self):
+        self._last_trace = None
+        self._spans = []
+        return self
+
+    def __exit__(self, exc_type: Any, exc_val: Any, exc_tb: Any):
+        self._store = None
+        self._rollout_id = None
+        self._attempt_id = None
+
+    def _await_in_loop(self, coro: Awaitable[Any], timeout: Optional[float] = None) -> Any:
+        # submit to the dedicated loop and wait synchronously
+        self._ensure_loop()
+        if self._loop is None:
+            raise RuntimeError("Loop is not initialized. This should not happen.")
+
+        # If already on the exporter loop thread, schedule and return immediately.
+        # ---------------------------------------------------------------------------
+        # WHY THIS CONDITIONAL EXISTS:
+        # In rare cases, span.end() is triggered from a LangchainCallbackHandler.__del__
+        # (or another finalizer) while the Python garbage collector is running on the
+        # *same thread* that owns our exporter event loop ("otel-loop").
+        #
+        # When that happens, on_end() executes on the exporter loop thread itself.
+        # If we were to call `asyncio.run_coroutine_threadsafe(...).result()` here,
+        # it would deadlock immediately — because the loop cannot both wait on and run
+        # the same coroutine. The Future stays pending forever and the loop stops
+        # processing scheduled callbacks.
+        #
+        # To avoid that self-deadlock, we detect when on_end() runs on the exporter
+        # loop thread. If so, we *schedule* the coroutine on the loop (fire-and-forget)
+        # instead of blocking with .result().
+        #
+        # This situation can occur because Python calls __del__ in whatever thread
+        # releases the last reference, which can easily be our loop thread if the
+        # object is dereferenced during loop._run_once().
+        # ---------------------------------------------------------------------------
+        if threading.current_thread() is self._loop_thread:
+            self._loop.call_soon_threadsafe(asyncio.create_task, coro)  # type: ignore
+            return None
+
+        fut = asyncio.run_coroutine_threadsafe(coro, self._loop)  # type: ignore
+        return fut.result(timeout=timeout)  # raises on error  # type: ignore
+
+    def shutdown(self) -> None:
+        if self._loop:
+            self._loop.call_soon_threadsafe(self._loop.stop)
+            self._loop = None
+        if self._loop_thread:
+            self._loop_thread.join(timeout=5)
+
+    def force_flush(self, timeout_millis: int = 30000) -> bool:
+        return True
+
+    def spans(self) -> List[ReadableSpan]:
+        """
+        Get the list of spans collected by this processor.
+        This is useful for debugging and testing purposes.
+
+        Returns:
+            List of ReadableSpan objects collected during tracing.
+        """
+        return self._spans
+
+    def with_context(self, store: LightningStore, rollout_id: str, attempt_id: str):
+        # simple context manager without nesting into asyncio
+        class _Ctx:
+            def __enter__(_):  # type: ignore
+                # Use _ instead of self to avoid shadowing the instance method.
+                with self._lock:
+                    self._store, self._rollout_id, self._attempt_id = store, rollout_id, attempt_id
+                    self._last_trace = None
+                    self._spans = []
+                return self
+
+            def __exit__(_, exc_type, exc, tb):  # type: ignore
+                with self._lock:
+                    self._store = self._rollout_id = self._attempt_id = None
+
+        return _Ctx()
+
+    def on_end(self, span: ReadableSpan) -> None:
+        """
+        Process a span when it ends.
+
+        Args:
+            span: The span that has ended.
+        """
+        # Skip if span is not sampled
+        if not span.context or not span.context.trace_flags.sampled:
+            return
+
+        if not self._disable_store_submission and self._store and self._rollout_id and self._attempt_id:
+            try:
+                # Submit add_otel_span to the event loop and wait for it to complete
+                with suppress_instrumentation():
+                    self._ensure_loop()
+                    self._await_in_loop(
+                        self._store.add_otel_span(self._rollout_id, self._attempt_id, span),
+                        timeout=60.0,
+                    )
+            except Exception:
+                # log; on_end MUST NOT raise
+                logger.exception(f"Error adding span to store: {span.name}")
+
+        self._spans.append(span)

agentlightning/types/core.py

@@ -10,14 +10,19 @@ from typing import (
     Callable,
     Dict,
     Generic,
+    Iterator,
     List,
     Literal,
+    Mapping,
     Optional,
     Protocol,
+    Sequence,
     SupportsIndex,
+    TypedDict,
     TypeVar,
     Union,
     cast,
+    overload,
 )
 
 from opentelemetry.sdk.trace import ReadableSpan
@@ -49,6 +54,12 @@ __all__ = [
     "Attempt",
     "AttemptedRollout",
     "Hook",
+    "Worker",
+    "WorkerStatus",
+    "PaginatedResult",
+    "FilterOptions",
+    "SortOptions",
+    "FilterField",
 ]
 
 T_co = TypeVar("T_co", covariant=True)
@@ -200,6 +211,32 @@ class AttemptedRollout(Rollout):
         return self
 
 
+WorkerStatus = Literal["idle", "busy", "unknown"]
+
+
+class Worker(BaseModel):
+    """Worker information. This is actually the same as Runner info."""
+
+    worker_id: str
+    """The ID of the worker."""
+    status: WorkerStatus = "unknown"
+    """The status of the worker."""
+    heartbeat_stats: Optional[Dict[str, Any]] = None
+    """Statistics about the worker's heartbeat."""
+    last_heartbeat_time: Optional[float] = None
+    """The last time when the worker has reported the stats."""
+    last_dequeue_time: Optional[float] = None
+    """The last time when the worker has tried to dequeue a rollout."""
+    last_busy_time: Optional[float] = None
+    """The last time when the worker has started an attempt and became busy."""
+    last_idle_time: Optional[float] = None
+    """The last time when the worker has triggered the end of an attempt and became idle."""
+    current_rollout_id: Optional[str] = None
+    """The ID of the current rollout that the worker is processing."""
+    current_attempt_id: Optional[str] = None
+    """The ID of the current attempt that the worker is processing."""
+
+
 TaskInput = Any
 """Task input type. Accepts arbitrary payloads."""
 
@@ -393,3 +430,104 @@ class Hook(ParallelWorkerBase):
         Subclasses can override this method for cleanup or additional
         logging. By default, this is a no-op.
         """
+
+
+class FilterField(TypedDict, total=False):
+    """An operator dict for a single field."""
+
+    exact: Any
+    within: Sequence[Any]
+    contains: str
+
+
+FilterOptions = Mapping[
+    Union[str, Literal["_aggregate", "_must"]],
+    Union[FilterField, Literal["and", "or"], Mapping[str, FilterField]],
+]
+"""A mapping of field name -> operator dict.
+
+Each operator dict can contain:
+
+- "exact": value for exact equality.
+- "within": iterable of allowed values.
+- "contains": substring to search for in string fields.
+
+The filter can also have a special field called "_aggregate" that can be used to specify the logic
+to combine the results of the filters:
+
+- "and": all conditions must match. This is the default value if not specified.
+- "or": at least one condition must match.
+
+All conditions within a field and between different fields are
+stored in a unified pool and combined using `_aggregate`.
+
+The filter can also have a special group called "_must", which is a mapping of filters that must all match,
+no matter whether the aggregate logic is "and" or "or".
+
+Example:
+
+```json
+{
+    "_aggregate": "or",
+    "_must": {
+        "city": {"exact": "New York"},
+        "timezone": {"within": ["America/New_York", "America/Los_Angeles"]},
+    },
+    "status": {"exact": "active"},
+    "id": {"within": [1, 2, 3]},
+    "name": {"contains": "foo"},
+}
+```
+"""
+
+
+class SortOptions(TypedDict):
+    """Options for sorting the collection."""
+
+    name: str
+    """The name of the field to sort by."""
+    order: Literal["asc", "desc"]
+    """The order to sort by."""
+
+
+T_item = TypeVar("T_item")
+
+
+class PaginatedResult(BaseModel, Sequence[T_item]):
+    """Result of a paginated query.
+
+    Behaves like a sequence, but also carries pagination metadata (limit, offset, total).
+    """
+
+    items: Sequence[T_item]
+    """Items in the result."""
+    limit: int
+    """Limit of the result."""
+    offset: int
+    """Offset of the result."""
+    total: int
+    """Total number of items in the collection."""
+
+    def __len__(self) -> int:
+        return len(self.items)
+
+    @overload
+    def __getitem__(self, index: int) -> T_item: ...
+
+    @overload
+    def __getitem__(self, index: slice) -> Sequence[T_item]: ...
+
+    def __getitem__(self, index: Union[int, slice]) -> Union[T_item, Sequence[T_item]]:
+        return self.items[index]
+
+    # Overriding __iter__ enables list(paginated_result) to work as expected,
+    # but changes Pydantic's default dict iteration behavior (which would otherwise
+    # iterate over field names).
+    def __iter__(self) -> Iterator[T_item]:  # type: ignore
+        return iter(self.items)
+
+    def __repr__(self) -> str:
+        first_item_repr = repr(self.items[0]) if self.items else "empty"
+        items_repr = f"[{first_item_repr}, ...]" if len(self.items) > 1 else first_item_repr
+        slice_repr = f"{self.offset}:" if self.limit == -1 else f"{self.offset}:{self.offset + self.limit}"
+        return f"<PaginatedResult ({slice_repr} of {self.total}) {items_repr}>"

agentlightning/types/resources.py

@@ -194,5 +194,11 @@ class ResourcesUpdate(BaseModel):
 
     resources_id: str
     """Identifier used to version the resources."""
+    create_time: float
+    """Timestamp of the creation time of the resources."""
+    update_time: float
+    """Timestamp of the last update time of the resources."""
+    version: int
+    """Version of the resources."""
     resources: NamedResources
     """Mapping of resource names to their definitions."""

agentlightning/types/tracer.py

@@ -16,6 +16,8 @@ from opentelemetry.sdk.trace.id_generator import RandomIdGenerator
 from opentelemetry.trace.status import Status as OtelStatus
 from pydantic import BaseModel, ConfigDict
 
+from agentlightning.semconv import AGL_VIRTUAL
+
 __all__ = [
     "AttributeValue",
     "Attributes",
@@ -379,7 +381,7 @@ class Span(BaseModel):
                 is_remote=False,
                 trace_state={},
             ),
-            name=name or SpanNames.VIRTUAL.value,
+            name=name or AGL_VIRTUAL,
             resource=resource or OtelResource(attributes={}, schema_url=""),
             attributes=attributes,
             status=TraceStatus(status_code="OK"),
@@ -399,7 +401,7 @@ class Span(BaseModel):
 
 
 class SpanNames(str, Enum):
-    """Enumerated span names recognised by Agent-lightning."""
+    """Enumerated span names recognised by Agent-lightning. Deprecated in favor of [semconv][agentlightning.semconv]."""
 
     REWARD = "agentlightning.reward"
     """The name of the reward span."""
@@ -411,10 +413,16 @@ class SpanNames(str, Enum):
     """The name of the exception span."""
     VIRTUAL = "agentlightning.virtual"
     """The name of the virtual span. It represents derived spans without concrete operations."""
+    ROLLOUT_ID = "agentlightning.rollout_id"
+    """The name of the rollout ID."""
+    ATTEMPT_ID = "agentlightning.attempt_id"
+    """The name of the attempt ID."""
+    SPAN_SEQUENCE_ID = "agentlightning.span_sequence_id"
+    """The name of the span sequence ID."""
 
 
 class SpanAttributeNames(str, Enum):
-    """Canonical attribute names written by Agent Lightning emitters."""
+    """Canonical attribute names written by Agent Lightning emitters. Deprecated in favor of [semconv][agentlightning.semconv]."""
 
     MESSAGE = "message"
     """The name of the message attribute."""

agentlightning/utils/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) Microsoft. All rights reserved.

agentlightning/utils/otel.py

@@ -0,0 +1,401 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Utilities shared for OpenTelemetry span (attributes) support."""
+
+import logging
+from typing import Any, Dict, List, Sequence, Union, cast
+from warnings import filterwarnings
+
+import opentelemetry.trace as trace_api
+from agentops.sdk.exporters import OTLPSpanExporter
+from opentelemetry.sdk.trace import ReadableSpan, SpanLimits, SynchronousMultiSpanProcessor, Tracer
+from opentelemetry.sdk.trace import TracerProvider as TracerProviderImpl
+from opentelemetry.sdk.trace.export import BatchSpanProcessor, SimpleSpanProcessor
+from opentelemetry.sdk.util.instrumentation import InstrumentationInfo, InstrumentationScope
+from opentelemetry.trace import get_tracer_provider as otel_get_tracer_provider
+from pydantic import TypeAdapter
+
+from agentlightning.env_var import LightningEnvVar, resolve_bool_env_var
+from agentlightning.semconv import LightningSpanAttributes, LinkAttributes, LinkPydanticModel
+from agentlightning.types import SpanLike
+from agentlightning.utils.otlp import LightningStoreOTLPExporter
+
+logger = logging.getLogger(__name__)
+
+__all__ = [
+    "full_qualified_name",
+    "get_tracer_provider",
+    "get_tracer",
+    "make_tag_attributes",
+    "extract_tags_from_attributes",
+    "make_link_attributes",
+    "query_linked_spans",
+    "extract_links_from_attributes",
+    "filter_attributes",
+    "filter_and_unflatten_attributes",
+    "flatten_attributes",
+    "unflatten_attributes",
+]
+
+
+def full_qualified_name(obj: type) -> str:
+    if str(obj.__module__) == "builtins":
+        return obj.__qualname__
+    return f"{obj.__module__}.{obj.__qualname__}"
+
+
+def get_tracer_provider(inspect: bool = True) -> TracerProviderImpl:
+    """Get the OpenTelemetry tracer provider configured for Agent Lightning.
+
+    Args:
+        inspect: Whether to inspect the tracer provider and log its configuration.
+            When it's on, make sure you also set the logger level to DEBUG to see the logs.
+    """
+    from agentlightning.tracer.otel import LightningSpanProcessor
+
+    if hasattr(trace_api, "_TRACER_PROVIDER") and trace_api._TRACER_PROVIDER is None:  # type: ignore[attr-defined]
+        raise RuntimeError("Tracer is not initialized. Cannot emit a meaningful span.")
+    tracer_provider = otel_get_tracer_provider()
+    if not isinstance(tracer_provider, TracerProviderImpl):
+        logger.error(
+            "Tracer provider is expected to be an instance of opentelemetry.sdk.trace.TracerProvider, found: %s",
+            full_qualified_name(type(tracer_provider)),
+        )
+        return cast(TracerProviderImpl, tracer_provider)
+
+    if not inspect:
+        return tracer_provider
+
+    emitter_debug = resolve_bool_env_var(LightningEnvVar.AGL_EMITTER_DEBUG, fallback=None)
+    logger_effective_level = logger.getEffectiveLevel()
+    if emitter_debug is True and logger_effective_level > logging.DEBUG:
+        logger.warning(
+            "Emitter debug logging is enabled but logging level is not set to DEBUG. Nothing will be logged."
+        )
+
+    if emitter_debug is None:
+        # Set to true by default if the logging level is lower than DEBUG
+        emitter_debug = logging.DEBUG >= logger_effective_level
+
+    if emitter_debug:
+        active_span_processor = tracer_provider._active_span_processor  # pyright: ignore[reportPrivateUsage]
+        processors: List[str] = []
+        active_span_processor_cls = active_span_processor.__class__.__name__
+        for processor in active_span_processor._span_processors:  # pyright: ignore[reportPrivateUsage]
+            if isinstance(processor, LightningSpanProcessor):
+                # The legacy case for tracers without OTLP support.
+                processors.append(f"{active_span_processor_cls} - {processor!r}")
+            elif isinstance(processor, (SimpleSpanProcessor, BatchSpanProcessor)):
+                processor_cls = processor.__class__.__name__
+                if isinstance(processor.span_exporter, LightningStoreOTLPExporter):
+                    # This should be the main path now.
+                    processors.append(f"{active_span_processor_cls} - {processor_cls} - {processor.span_exporter!r}")
+                elif isinstance(processor.span_exporter, OTLPSpanExporter):
+                    # You need to be careful if the code goes into this path.
+                    endpoint = processor.span_exporter._endpoint  # pyright: ignore[reportPrivateUsage]
+                    processors.append(
+                        f"{active_span_processor_cls} - {processor_cls} - "
+                        f"{processor.span_exporter.__class__.__name__}(endpoint={endpoint!r})"
+                    )
+                else:
+                    # Other cases like Console Span Exporter.
+                    processors.append(
+                        f"{active_span_processor_cls} - {processor_cls} - {processor.span_exporter.__class__.__name__}"
+                    )
+            else:
+                processors.append(f"{active_span_processor_cls} - {processor.__class__.__name__}")
+
+        logger.debug(f"Tracer provider: {tracer_provider!r}. Active span processors:")
+        for processor in processors:
+            logger.debug("  * " + processor)
+
+    return tracer_provider
+
+
+def get_tracer(use_active_span_processor: bool = True) -> trace_api.Tracer:
+    """Resolve the OpenTelemetry tracer configured for Agent Lightning.
+
+    Args:
+        use_active_span_processor: Whether to use the active span processor.
+
+    Returns:
+        OpenTelemetry tracer tagged with the `agentlightning` instrumentation name.
+
+    Raises:
+        RuntimeError: If OpenTelemetry was not initialized before calling this helper.
+    """
+    if hasattr(trace_api, "_TRACER_PROVIDER") and trace_api._TRACER_PROVIDER is None:  # type: ignore[attr-defined]
+        raise RuntimeError("Tracer is not initialized. Cannot emit a meaningful span.")
+
+    tracer_provider = get_tracer_provider(inspect=True)  # inspection is on by default
+
+    if use_active_span_processor:
+        return tracer_provider.get_tracer("agentlightning")
+
+    else:
+        filterwarnings(
+            "ignore",
+            message=r"You should use InstrumentationScope. Deprecated since version 1.11.1.",
+            category=DeprecationWarning,
+            module="opentelemetry.sdk.trace",
+        )
+
+        return Tracer(
+            tracer_provider.sampler,
+            tracer_provider.resource,
+            # We use an empty span processor to avoid emitting spans to the tracer
+            SynchronousMultiSpanProcessor(),
+            tracer_provider.id_generator,
+            InstrumentationInfo("agentlightning", "", ""),  # type: ignore
+            SpanLimits(),
+            InstrumentationScope(
+                "agentlightning",
+                "",
+                "",
+                {},
+            ),
+        )
+
+
+def make_tag_attributes(tags: List[str]) -> Dict[str, Any]:
+    """Convert a list of tags into flattened attributes for span tagging.
+
+    There is no syntax enforced for tags, they are just strings. For example:
+
+    ```python
+    ["gen_ai.model:gpt-4", "reward.extrinsic"]
+    ```
+    """
+    return flatten_attributes({LightningSpanAttributes.TAG.value: tags})
+
+
+def extract_tags_from_attributes(attributes: Dict[str, Any]) -> List[str]:
+    """Extract tag attributes from flattened span attributes.
+
+    Args:
+        attributes: A dictionary of flattened span attributes.
+    """
+    maybe_tag_list = filter_and_unflatten_attributes(attributes, LightningSpanAttributes.TAG.value)
+    return TypeAdapter(List[str]).validate_python(maybe_tag_list)
+
+
+def make_link_attributes(links: Dict[str, str]) -> Dict[str, Any]:
+    """Convert a dictionary of links into flattened attributes for span linking.
+
+    Links example:
+
+    ```python
+    {
+        "gen_ai.response.id": "response-123",
+        "span_id": "abcd-efgh-ijkl",
+    }
+    ```
+    """
+    link_list: List[Dict[str, str]] = []
+    for key, value in links.items():
+        if not isinstance(value, str):  # pyright: ignore[reportUnnecessaryIsInstance]
+            raise ValueError(f"Link value must be a string, got {type(value)} for key '{key}'")
+        link_list.append({LinkAttributes.KEY_MATCH.value: key, LinkAttributes.VALUE_MATCH.value: value})
+    return flatten_attributes({LightningSpanAttributes.LINK.value: link_list})
+
+
+def query_linked_spans(spans: Sequence[SpanLike], links: List[LinkPydanticModel]) -> List[SpanLike]:
+    """Query spans that are linked by the given link attributes.
+
+    Args:
+        spans: A sequence of spans to search.
+        links: A list of link attributes to match.
+
+    Returns:
+        A list of spans that match the given link attributes.
+    """
+    matched_spans: List[SpanLike] = []
+
+    for span in spans:
+        span_attributes = span.attributes or {}
+        is_match = True
+        for link in links:
+            # trace_id and span_id must be full match.
+            if link.key_match == "trace_id":
+                if isinstance(span, ReadableSpan):
+                    trace_id = trace_api.format_trace_id(span.context.trace_id) if span.context else None
+                else:
+                    trace_id = span.trace_id
+                if trace_id != link.value_match:
+                    is_match = False
+                    break
+
+            elif link.key_match == "span_id":
+                if isinstance(span, ReadableSpan):
+                    span_id = trace_api.format_span_id(span.context.span_id) if span.context else None
+                else:
+                    span_id = span.span_id
+                if span_id != link.value_match:
+                    is_match = False
+                    break
+
+            else:
+                attribute = span_attributes.get(link.key_match)
+                # attributes must also be a full match currently.
+                if attribute != link.value_match:
+                    is_match = False
+                    break
+
+        if is_match:
+            matched_spans.append(span)
+
+    return matched_spans
+
+
+def extract_links_from_attributes(attributes: Dict[str, Any]) -> List[LinkPydanticModel]:
+    """Extract link attributes from flattened span attributes.
+
+    Args:
+        attributes: A dictionary of flattened span attributes.
+    """
+    maybe_link_list = filter_and_unflatten_attributes(attributes, LightningSpanAttributes.LINK.value)
+    return TypeAdapter(List[LinkPydanticModel]).validate_python(maybe_link_list)
+
+
+def filter_attributes(attributes: Dict[str, Any], prefix: str) -> Dict[str, Any]:
+    """Filter attributes that start with the given prefix.
+
+    The attribute must start with `prefix.` or be exactly `prefix` to be included.
+
+    Args:
+        attributes: A dictionary of span attributes.
+        prefix: The prefix to filter by.
+
+    Returns:
+        A dictionary of attributes that start with the given prefix.
+    """
+    return {k: v for k, v in attributes.items() if k.startswith(prefix + ".") or k == prefix}
+
+
+def filter_and_unflatten_attributes(attributes: Dict[str, Any], prefix: str) -> Union[Dict[str, Any], List[Any]]:
+    """Filter attributes that start with the given prefix and unflatten them.
+    The prefix will be removed during unflattening.
+
+    Args:
+        attributes: A dictionary of span attributes.
+        prefix: The prefix to filter by.
+
+    Returns:
+        A nested dictionary or list of attributes that start with the given prefix.
+    """
+    filtered_attributes = filter_attributes(attributes, prefix)
+    stripped_attributes: Dict[str, Any] = {}
+    for k, v in filtered_attributes.items():
+        if k == prefix:
+            raise ValueError(f"Cannot unflatten attribute with key exactly equal to prefix: {prefix}")
+        else:
+            stripped_key = k[len(prefix) + 1 :]  # +1 to remove the dot
+        stripped_attributes[stripped_key] = v
+    return unflatten_attributes(stripped_attributes)
+
+
+def flatten_attributes(nested_data: Union[Dict[str, Any], List[Any]]) -> Dict[str, Any]:
+    """Flatten a nested dictionary or list into a flat dictionary with dotted keys.
+
+    This function recursively traverses dictionaries and lists, producing a flat
+    key-value mapping where nested paths are represented via dot-separated keys.
+    Lists are indexed numerically.
+
+    Example:
+
+        >>> flatten_attributes({"a": {"b": 1, "c": [2, 3]}})
+        {"a.b": 1, "a.c.0": 2, "a.c.1": 3}
+
+    Args:
+        nested_data: A nested structure composed of dictionaries, lists, or
+            primitive values.
+
+    Returns:
+        A flat dictionary mapping dotted-string paths to primitive values.
+    """
+
+    flat: Dict[str, Any] = {}
+
+    def _walk(value: Any, prefix: str = "") -> None:
+        if isinstance(value, dict):
+            for k, v in cast(Dict[Any, Any], value).items():
+                if not isinstance(k, str):
+                    raise ValueError(
+                        f"Only string keys are supported in dictionaries, got '{k}' of type {type(k)} in {prefix}"
+                    )
+                new_prefix = f"{prefix}.{k}" if prefix else k
+                _walk(v, new_prefix)
+        elif isinstance(value, list):
+            for idx, item in enumerate(cast(List[Any], value)):
+                new_prefix = f"{prefix}.{idx}" if prefix else str(idx)
+                _walk(item, new_prefix)
+        else:
+            flat[prefix] = value
+
+    _walk(nested_data)
+    return flat
+
+
+def unflatten_attributes(flat_data: Dict[str, Any]) -> Union[Dict[str, Any], List[Any]]:
+    """Reconstruct a nested dictionary/list structure from a flat dictionary.
+
+    Keys are dot-separated paths. Segments that are digit strings will only
+    become list indices if *all* keys in that dict form a consecutive
+    0..n-1 range. Otherwise they remain dict keys.
+
+    Example:
+
+        >>> unflatten_attributes({"a.b": 1, "a.c.0": 2, "a.c.1": 3})
+        {"a": {"b": 1, "c": [2, 3]}}
+
+    Args:
+        flat_data: A dictionary whose keys are dot-separated paths and whose
+            values are primitive data elements.
+
+    Returns:
+        A nested dictionary (and lists where appropriate) corresponding to
+        the flattened structure.
+    """
+    # 1) Build a pure dict tree first (no lists yet)
+    root: Dict[str, Any] = {}
+
+    for flat_key, value in flat_data.items():
+        parts = flat_key.split(".")
+        curr: Dict[str, Any] = root
+
+        for part in parts[:-1]:
+            # Ensure intermediate node is a dict
+            if part not in curr or not isinstance(curr[part], dict):
+                curr[part] = {}
+            curr = curr[part]  # type: ignore[assignment]
+
+        curr[parts[-1]] = value
+
+    # 2) Recursively convert dicts-with-consecutive-numeric-keys into lists
+    def convert(node: Union[Dict[str, Any], List[Any]]) -> Union[Dict[str, Any], List[Any]]:
+        if isinstance(node, dict):
+            # First convert children
+            for k, v in list(node.items()):
+                node[k] = convert(v)
+
+            if not node:
+                # empty dict stays dict
+                return node
+
+            # Check if keys are all numeric strings
+            keys = list(node.keys())
+            if all(isinstance(k, str) and k.isdigit() for k in keys):  # pyright: ignore[reportUnnecessaryIsInstance]
+                indices = sorted(int(k) for k in keys)
+                # Must be exactly 0..n-1
+                if indices == list(range(len(indices))):
+                    return [node[str(i)] for i in range(len(indices))]
+
+            return node
+
+        if isinstance(node, list):  # pyright: ignore[reportUnnecessaryIsInstance]
+            return [convert(v) for v in node]
+
+        # Keep as is
+        return node
+
+    return convert(root)

agentlightning/utils/otlp.py

@@ -0,0 +1,474 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import gzip
+import logging
+from typing import Any, Awaitable, Callable, Dict, List, Optional, Sequence, Tuple, Type, TypeVar
+
+from fastapi import Request, Response
+from google.protobuf import json_format
+from google.rpc.status_pb2 import Status
+from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+from opentelemetry.proto.collector.logs.v1.logs_service_pb2 import (
+    ExportLogsServiceRequest,
+    ExportLogsServiceResponse,
+)
+from opentelemetry.proto.collector.metrics.v1.metrics_service_pb2 import (
+    ExportMetricsServiceRequest,
+    ExportMetricsServiceResponse,
+)
+from opentelemetry.proto.collector.trace.v1.trace_service_pb2 import (
+    ExportTraceServiceRequest,
+    ExportTraceServiceResponse,
+)
+from opentelemetry.proto.common.v1.common_pb2 import AnyValue, KeyValue
+from opentelemetry.proto.resource.v1.resource_pb2 import Resource as ProtoResource
+from opentelemetry.proto.trace.v1.trace_pb2 import Span as ProtoSpan
+from opentelemetry.proto.trace.v1.trace_pb2 import Status as ProtoStatus
+from opentelemetry.sdk.resources import Resource
+from opentelemetry.sdk.trace import ReadableSpan
+from opentelemetry.sdk.trace.export import SpanExportResult
+from opentelemetry.util.types import AttributeValue
+
+from agentlightning.semconv import LightningResourceAttributes
+from agentlightning.types.tracer import (
+    Attributes,
+    Event,
+    Link,
+    OtelResource,
+    Span,
+    SpanContext,
+    TraceStatus,
+    convert_timestamp,
+)
+
+PROTOBUF_CT = "application/x-protobuf"
+
+logger = logging.getLogger(__name__)
+
+
+T_request = TypeVar("T_request", ExportLogsServiceRequest, ExportMetricsServiceRequest, ExportTraceServiceRequest)
+T_response = TypeVar("T_response", ExportLogsServiceResponse, ExportMetricsServiceResponse, ExportTraceServiceResponse)
+
+
+async def handle_otlp_export(
+    request: Request,
+    request_message_cls: Type[T_request],
+    response_message_cls: Type[T_response],
+    message_callback: Optional[Callable[[T_request], Awaitable[None]]],
+    signal_name: str,
+) -> Response:
+    """
+    Generic handler for /v1/traces, /v1/metrics, /v1/logs.
+
+    Convert the OTLP Protobuf request to a JSON-like object.
+    """
+    content_type = request.headers.get("Content-Type", "").split(";")[0].strip()
+
+    if content_type != PROTOBUF_CT:
+        # For brevity we only support binary protobuf here.
+        return _bad_request_response(
+            request,
+            f"Unsupported Content-Type '{content_type}', expected '{PROTOBUF_CT}'",
+            content_type=PROTOBUF_CT,
+        )
+
+    raw_body = await request.body()
+    body = _read_body_maybe_gzip(request, raw_body)
+
+    # Empty request is allowed and should still succeed.
+    if not body:
+        req_msg = request_message_cls()
+    else:
+        req_msg = request_message_cls()
+        try:
+            req_msg.ParseFromString(body)
+        except Exception as exc:
+            return _bad_request_response(request, f"Unable to parse OTLP {signal_name} payload: {exc}")
+
+    if message_callback is not None:
+        await message_callback(req_msg)
+
+    # Build success response. Partial success field is left unset.
+    resp_msg = response_message_cls()
+
+    # Encode response in the same Content-Type as request.
+    if content_type == PROTOBUF_CT:
+        resp_bytes = resp_msg.SerializeToString()
+    else:
+        resp_bytes = json_format.MessageToJson(resp_msg).encode("utf-8")
+
+    resp_bytes, headers = _maybe_gzip_response(request, resp_bytes)
+
+    return Response(
+        content=resp_bytes,
+        media_type=content_type,
+        status_code=200,
+        headers=headers,
+    )
+
+
+async def spans_from_proto(
+    request: ExportTraceServiceRequest,
+    sequence_id_bulk_issuer: Callable[[Sequence[Tuple[str, str]]], Awaitable[Sequence[int]]],
+) -> List[Span]:
+    """Parse an OTLP proto payload into List[Span].
+
+    A store is needed here for generating a sequence ID for each span.
+    """
+    output_spans: List[Span] = []
+
+    for resource_spans in request.resource_spans:
+        # Resource-level attributes & IDs
+        resource_attrs = _kv_list_to_dict(resource_spans.resource.attributes)
+        # rollout_id, attempt_id from resource attributes when present.
+        rollout_id_resource = resource_attrs.get(LightningResourceAttributes.ROLLOUT_ID.value)
+        attempt_id_resource = resource_attrs.get(LightningResourceAttributes.ATTEMPT_ID.value)
+        # If sequence id is provided, all the spans will share the same sequence ID.
+        # unless otherwise overridden by span-level attributes.
+        sequence_id_resource = resource_attrs.get(LightningResourceAttributes.SPAN_SEQUENCE_ID.value)
+
+        otel_resource = _resource_from_proto(resource_spans.resource, getattr(resource_spans, "schema_url", ""))
+
+        # Each ScopeSpans contains multiple spans
+        for scope_spans in resource_spans.scope_spans:
+            for proto_span in scope_spans.spans:
+                trace_id_hex = _bytes_to_trace_id_hex(proto_span.trace_id)
+                span_id_hex = _bytes_to_span_id_hex(proto_span.span_id)
+                parent_id_hex = _bytes_to_span_id_hex(proto_span.parent_span_id) if proto_span.parent_span_id else None
+
+                # Status
+                status_code_str = _STATUS_CODE_MAP.get(proto_span.status.code, "UNSET")
+                status = TraceStatus(
+                    status_code=status_code_str,
+                    description=proto_span.status.message or None,
+                )
+
+                # Attributes
+                span_attrs = _kv_list_to_dict(proto_span.attributes)
+
+                # Context
+                context = SpanContext(
+                    trace_id=trace_id_hex,
+                    span_id=span_id_hex,
+                    is_remote=False,
+                    trace_state={},
+                )
+
+                # Try to get if span attributes contain something like rollout_id or attempt_id
+                # Override the resource-level attributes with the span-level attributes if present.
+                rollout_id_span = span_attrs.get(LightningResourceAttributes.ROLLOUT_ID.value)
+                attempt_id_span = span_attrs.get(LightningResourceAttributes.ATTEMPT_ID.value)
+                sequence_id_span = span_attrs.get(LightningResourceAttributes.SPAN_SEQUENCE_ID.value)
+
+                # Normalize to regular strings and ints
+                rollout_id_raw = rollout_id_span if rollout_id_span is not None else rollout_id_resource
+                attempt_id_raw = attempt_id_span if attempt_id_span is not None else attempt_id_resource
+                sequence_id_raw = sequence_id_span if sequence_id_span is not None else sequence_id_resource
+
+                rollout_id, attempt_id = _normalize_rollout_attempt_id(rollout_id_raw, attempt_id_raw)
+                sequence_id = _normalize_sequence_id(sequence_id_raw)
+
+                if rollout_id is None or attempt_id is None:
+                    logger.warning(
+                        "Both rollout_id and attempt_id must be present in resource attributes. "
+                        "Spans will not be able to log to the store because of missing IDs: rollout_id=%s, attempt_id=%s, sequence_id=%s",
+                        rollout_id,
+                        attempt_id,
+                        sequence_id,
+                    )
+                    continue
+
+                # Generate a new sequence ID if not provided
+                if sequence_id is None:
+                    current_sequence_id = -1
+                elif sequence_id < 0:
+                    logger.error(
+                        "Invalid sequence_id value in resource attributes: %r. Must be a positive integer. Regenerating one.",
+                        sequence_id,
+                    )
+                    current_sequence_id = -1
+                else:
+                    current_sequence_id = sequence_id
+
+                # Build Span
+                span = Span(
+                    rollout_id=rollout_id,
+                    attempt_id=attempt_id,
+                    sequence_id=current_sequence_id,
+                    trace_id=trace_id_hex,
+                    span_id=span_id_hex,
+                    parent_id=parent_id_hex,
+                    name=proto_span.name,
+                    status=status,
+                    attributes=span_attrs,
+                    events=_events_from_proto(proto_span),
+                    links=_links_from_proto(proto_span),
+                    start_time=convert_timestamp(proto_span.start_time_unix_nano),
+                    end_time=convert_timestamp(proto_span.end_time_unix_nano),
+                    context=context,
+                    parent=None,  # OTLP only has parent_span_id; we don't have full SpanContext
+                    resource=otel_resource,
+                )
+
+                output_spans.append(span)
+
+    # Finalize the sequence IDs
+    bulk_issue_requests = [(span.rollout_id, span.attempt_id) for span in output_spans if span.sequence_id < 0]
+    bulk_sequence_ids = await sequence_id_bulk_issuer(bulk_issue_requests)
+    for span, sequence_id in zip(
+        [span for span in output_spans if span.sequence_id < 0], bulk_sequence_ids, strict=True
+    ):
+        span.sequence_id = sequence_id
+
+    return output_spans
+
+
+class LightningStoreOTLPExporter(OTLPSpanExporter):
+    """OTLP Exporter that write to a LightningStore-compatible backend.
+
+    The backend requires two special attributes on each span:
+
+    - `agentlightning.rollout_id`: The rollout ID to associate the span with.
+    - `agentlightning.attempt_id`: The attempt ID to associate the span with.
+
+    It can optionally use the following attribute to sequence spans:
+
+    - `agentlightning.span_sequence_id`: A decimal string representing the sequence ID of the span.
+    """
+
+    _default_endpoint: Optional[str] = None
+    _rollout_id: Optional[str] = None
+    _attempt_id: Optional[str] = None
+
+    def __repr__(self) -> str:
+        return (
+            f"{self.__class__.__name__}("
+            + f"endpoint={self.endpoint!r}, "
+            + f"rollout_id={self.rollout_id!r}, "
+            + f"attempt_id={self.attempt_id!r}, "
+            + f"should_bypass={self.should_bypass()!r})"
+        )
+
+    @property
+    def endpoint(self) -> Optional[str]:
+        """The endpoint to submit the spans to."""
+        if hasattr(self, "_endpoint"):
+            return self._endpoint
+        return None
+
+    @property
+    def rollout_id(self) -> Optional[str]:
+        """The rollout ID to submit the spans to."""
+        if hasattr(self, "_rollout_id"):
+            return self._rollout_id
+        return None
+
+    @property
+    def attempt_id(self) -> Optional[str]:
+        """The attempt ID to submit the spans to."""
+        if hasattr(self, "_attempt_id"):
+            return self._attempt_id
+        return None
+
+    def enable_store_otlp(self, endpoint: str, rollout_id: str, attempt_id: str) -> None:
+        """Enable storing OTLP data to a specific LightningStore rollout/attempt."""
+        self._rollout_id = rollout_id
+        self._attempt_id = attempt_id
+
+        self._default_endpoint = self._endpoint
+        self._endpoint = endpoint
+
+    def disable_store_otlp(self) -> None:
+        """Disable storing OTLP data to LightningStore."""
+        self._rollout_id = None
+        self._attempt_id = None
+        if self._default_endpoint is not None:
+            self._endpoint = self._default_endpoint
+
+    def should_bypass(self) -> bool:
+        """Check if the exporter should bypass the default export if rollout_id and attempt_id are not set."""
+        return True
+
+    def export(self, spans: Sequence[ReadableSpan]) -> SpanExportResult:
+        if self._rollout_id is not None and self._attempt_id is not None:
+            # rollout_id and attempt_id are present in resource attributes
+            # It means that the server supports OTLP endpoint.
+            for span in spans:
+                # Override the resources so that the server knows where the request comes from.
+                span._resource = span._resource.merge(  # pyright: ignore[reportPrivateUsage]
+                    Resource.create(
+                        {
+                            LightningResourceAttributes.ROLLOUT_ID.value: self._rollout_id,
+                            LightningResourceAttributes.ATTEMPT_ID.value: self._attempt_id,
+                        }
+                    )
+                )
+            return super().export(spans)
+        elif not self.should_bypass():
+            logger.debug("Rollout ID and Attempt ID not set; using default OTLP exporter behavior.")
+            return super().export(spans)
+        else:
+            logger.debug("Rollout ID and Attempt ID not set; bypassing export.")
+            return SpanExportResult.SUCCESS
+
+
+def _read_body_maybe_gzip(request: Request, raw_body: bytes) -> bytes:
+    """
+    Decompress body if Content-Encoding: gzip; otherwise return as is.
+    """
+    encoding = request.headers.get("Content-Encoding", "").lower()
+    if encoding == "gzip":
+        return gzip.decompress(raw_body)
+    return raw_body
+
+
+def _maybe_gzip_response(request: Request, payload: bytes) -> Tuple[bytes, Dict[str, str]]:
+    """
+    If Accept-Encoding includes gzip, gzip the payload and set Content-Encoding header.
+    """
+    ae = request.headers.get("Accept-Encoding", "")
+    tokens = [token.split(";")[0].strip().lower() for token in ae.split(",") if token.strip()]
+    headers: Dict[str, str] = {}
+    if "gzip" in tokens:
+        payload = gzip.compress(payload)
+        headers["Content-Encoding"] = "gzip"
+    return payload, headers
+
+
+def _bad_request_response(request: Request, message: str, content_type: str = PROTOBUF_CT) -> Response:
+    """
+    Build a 400 response whose body is a protobuf Status message, encoded
+    in the same Content-Type as the request (OTLP/HTTP requirement).
+    """
+    status_msg = Status(message=message)
+
+    if content_type == PROTOBUF_CT:
+        body = status_msg.SerializeToString()
+    else:
+        # Fallback: JSON representation of Status.
+        body = json_format.MessageToJson(status_msg).encode("utf-8")
+
+    body, headers = _maybe_gzip_response(request, body)
+
+    return Response(
+        content=body,
+        status_code=400,
+        media_type=content_type,
+        headers=headers,
+    )
+
+
+def _normalize_rollout_attempt_id(
+    rollout_id: Optional[AttributeValue], attempt_id: Optional[AttributeValue]
+) -> Tuple[Optional[str], Optional[str]]:
+    """Normalize a rollout or attempt ID to a string."""
+    rollout_id_str = str(rollout_id) if rollout_id is not None else None
+    attempt_id_str = str(attempt_id) if attempt_id is not None else None
+    return rollout_id_str, attempt_id_str
+
+
+def _normalize_sequence_id(sequence_id: Optional[AttributeValue]) -> Optional[int]:
+    """Normalize a sequence ID to an integer."""
+    if sequence_id is None:
+        return None
+    try:
+        sequence_id_int = int(str(sequence_id))
+    except (ValueError, TypeError):
+        logger.warning(
+            "Invalid sequence_id value in resource attributes: %r. Must be an integer or string representing an integer. Assuming None.",
+            sequence_id,
+        )
+        sequence_id_int = None
+    return sequence_id_int
+
+
+def _any_value_to_python(value: AnyValue) -> Any:
+    """Convert OTLP AnyValue -> plain Python value."""
+    kind = value.WhichOneof("value")
+    if kind is None:
+        return None
+    if kind == "string_value":
+        return value.string_value
+    if kind == "bool_value":
+        return value.bool_value
+    if kind == "int_value":
+        return int(value.int_value)
+    if kind == "double_value":
+        return float(value.double_value)
+    if kind == "array_value":
+        return [_any_value_to_python(v) for v in value.array_value.values]
+    if kind == "kvlist_value":
+        # Map<string, AnyValue> -> dict
+        return {kv.key: _any_value_to_python(kv.value) for kv in value.kvlist_value.values}
+    if kind == "bytes_value":
+        # Serialize bytes as hex string to stay JSON-friendly
+        return value.bytes_value.hex()
+    return None
+
+
+def _kv_list_to_dict(kvs: Sequence[KeyValue]) -> Attributes:
+    """Convert repeated KeyValue -> Attributes dict."""
+    return {kv.key: _any_value_to_python(kv.value) for kv in kvs}
+
+
+_STATUS_CODE_MAP = {
+    ProtoStatus.STATUS_CODE_UNSET: "UNSET",
+    ProtoStatus.STATUS_CODE_OK: "OK",
+    ProtoStatus.STATUS_CODE_ERROR: "ERROR",
+}
+
+
+def _bytes_to_trace_id_hex(b: bytes) -> str:
+    # OTLP uses 16-byte trace IDs; format as 32-char hex
+    if not b:
+        return "0" * 32
+    return b.hex().rjust(32, "0")
+
+
+def _bytes_to_span_id_hex(b: bytes) -> str:
+    # OTLP uses 8-byte span IDs; format as 16-char hex
+    if not b:
+        return "0" * 16
+    return b.hex().rjust(16, "0")
+
+
+def _events_from_proto(span: ProtoSpan) -> List[Event]:
+    """Event converter from OTLP ProtoSpan to List[Event]."""
+    return [
+        Event(
+            name=e.name,
+            attributes=_kv_list_to_dict(e.attributes),
+            timestamp=convert_timestamp(e.time_unix_nano),
+        )
+        for e in span.events
+    ]
+
+
+def _links_from_proto(span: ProtoSpan) -> List[Link]:
+    """Link converter from OTLP ProtoSpan to List[Link]."""
+    links: List[Link] = []
+    for link in span.links:
+        trace_id_hex = _bytes_to_trace_id_hex(link.trace_id)
+        span_id_hex = _bytes_to_span_id_hex(link.span_id)
+        ctx = SpanContext(
+            trace_id=trace_id_hex,
+            span_id=span_id_hex,
+            is_remote=False,
+            trace_state={},  # OTLP trace_state is currently a string; you can parse if needed
+        )
+        links.append(
+            Link(
+                context=ctx,
+                attributes=_kv_list_to_dict(link.attributes) or None,
+            )
+        )
+    return links
+
+
+def _resource_from_proto(resource: ProtoResource, schema_url: str = "") -> OtelResource:
+    return OtelResource(
+        attributes=_kv_list_to_dict(resource.attributes),
+        schema_url=schema_url or "",
+    )

agentlightning/utils/system_snapshot.py

@@ -0,0 +1,72 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import platform
+import socket
+from contextlib import suppress
+from datetime import datetime
+from typing import Any, Dict, List, cast
+
+import psutil
+from gpustat import GPUStat, GPUStatCollection
+
+
+def system_snapshot(include_gpu: bool = False) -> Dict[str, Any]:
+    # CPU
+    cpu = {
+        "cpu_name": platform.processor(),
+        "cpu_cores": psutil.cpu_count(logical=False),
+        "cpu_threads": psutil.cpu_count(logical=True),
+        "cpu_usage_pct": psutil.cpu_percent(0.05),
+    }
+
+    # Memory
+    vm = psutil.virtual_memory()
+    mem = {
+        "mem_used_gb": round(vm.used / (2**30), 2),
+        "mem_total_gb": round(vm.total / (2**30), 2),
+        "mem_pct": vm.percent,
+    }
+
+    # Disk
+    du = psutil.disk_usage("/")
+    disk = {
+        "disk_used_gb": round(du.used / (2**30), 2),
+        "disk_total_gb": round(du.total / (2**30), 2),
+        "disk_pct": du.percent,
+    }
+
+    # GPU
+    gpus: List[Dict[str, Any]] = []
+    with suppress(Exception):
+        for g in GPUStatCollection.new_query().gpus:  # type: ignore
+            g = cast(GPUStat, g)
+            gpus.append(
+                {
+                    "gpu": g.name,  # type: ignore
+                    "util_pct": g.utilization,
+                    "mem_used_mb": g.memory_used,
+                    "mem_total_mb": g.memory_total,
+                    "temp_c": g.temperature,
+                }
+            )
+
+    # Network
+    net = psutil.net_io_counters()
+    netinfo = {
+        "bytes_sent_mb": round(net.bytes_sent / (2**20), 2),
+        "bytes_recv_mb": round(net.bytes_recv / (2**20), 2),
+    }
+
+    # OS / meta
+    return {
+        "timestamp": datetime.now().isoformat(timespec="seconds"),
+        "host": socket.gethostname(),
+        "os": platform.platform(),
+        **cpu,
+        **mem,
+        **disk,
+        **netinfo,
+        **({"gpus": gpus} if include_gpu else {}),
+    }

agentlightning/verl/daemon.py

@@ -18,14 +18,12 @@ from flask import Flask, Response, abort, request
 from tensordict import TensorDict
 from verl import DataProto
 
-from agentlightning import LLM, AgentLightningServer, NamedResources, RolloutLegacy, configure_logger
+from agentlightning import LLM, AgentLightningServer, NamedResources, RolloutLegacy
 from agentlightning.adapter.triplet import TracerTraceToTriplet, TraceToTripletBase
 from agentlightning.llm_proxy import LLMProxy, ModelConfig
 from agentlightning.store.base import LightningStore
 from agentlightning.types import Rollout, RolloutConfig, Task
 
-configure_logger()
-
 __all__ = [
     "AgentModeDaemon",
     "get_left_padded_ids_and_attention_mask",
@@ -294,7 +292,7 @@ class AgentModeDaemon:
         self._proxy_thread.start()
         print(f"Proxy server running on port {self.proxy_port}")
 
-    def _update_proxy_server_v1(self):
+    async def _update_proxy_server_v1(self):
         model_name = self.train_information.get("model")
         if not model_name:
             raise ValueError("Model name is not set.")
@@ -313,12 +311,7 @@ class AgentModeDaemon:
             ],
         )
 
-        if self.llm_proxy.is_running():
-            # FIXME: Need to switch to a different port right now
-            # because the forked processes carried the old fd
-            self.llm_proxy.restart(_port=_find_available_port())
-        else:
-            self.llm_proxy.start()
+        await self.llm_proxy.restart()
 
     def start(self):
         """Starts the main AgentLightningServer and the proxy server."""
@@ -352,7 +345,7 @@ class AgentModeDaemon:
         if server_addresses != self.backend_llm_server_addresses:
             self.backend_llm_server_addresses = server_addresses
             if self.mode == "v1" and not self.llm_proxy.is_running():
-                self._update_proxy_server_v1()
+                await self._update_proxy_server_v1()
         self.is_train = is_train
 
         # 1. Update resources on the server for clients to use
@@ -564,14 +557,17 @@ class AgentModeDaemon:
         )  # FIXME: Evaluate whether grouping stats by source is actually needed.
 
         for rollout_id, rollout in self._completed_rollouts_v0.items():
+            final_reward_raw: Optional[float] = rollout.final_reward
             final_reward = self._fillna_reward(rollout)
             if not rollout.triplets:
                 print(f"Warning: No triplets found for test rollout {rollout.rollout_id}.")
-                sample_stat_list.append({"reward": final_reward})
+                sample_stat_list.append({"reward": final_reward, "has_reward": final_reward_raw is not None})
                 continue
             response_length_list = [len(triplet.response.get("token_ids", [])) for triplet in rollout.triplets]
+
             if "data_source" in self._task_id_to_original_sample[rollout_id]:
                 # When a test sample includes a 'data_source' field, record per-source statistics for test results.
+                # TODO: This is a flawed design. We should have a better way to handle this.
                 data_source = self._task_id_to_original_sample[rollout_id]["data_source"]
                 sample_stat_list_by_source[data_source].append(
                     {
@@ -579,6 +575,7 @@ class AgentModeDaemon:
                         "mean_response_length": np.mean(response_length_list) if response_length_list else 0,
                         "turn_count": len(rollout.triplets),
                         "reward": final_reward,
+                        "has_reward": final_reward_raw is not None,
                     }
                 )
             sample_stat_list.append(
@@ -587,6 +584,7 @@ class AgentModeDaemon:
                     "mean_response_length": np.mean(response_length_list) if response_length_list else 0,
                     "turn_count": len(rollout.triplets),
                     "reward": final_reward,
+                    "has_reward": final_reward_raw is not None,
                 }
             )
         metric_dict: Dict[str, Any] = {}
@@ -601,6 +599,9 @@ class AgentModeDaemon:
                 {
                     f"val/{data_source}/n_rollouts": len(sample_stats),
                     f"val/{data_source}/n_rollouts_w_trace": len(stats_w_trace_by_source[data_source]),
+                    f"val/{data_source}/n_rollouts_w_reward": len(
+                        [stat for stat in sample_stats if stat["has_reward"]]
+                    ),
                     f"val/{data_source}/reward": np.mean(
                         [stat["reward"] for stat in sample_stats]
                     ),  # each rollout must have a reward (fillna if missing)
@@ -619,6 +620,7 @@ class AgentModeDaemon:
             {
                 "val/n_rollouts": len(sample_stat_list),
                 "val/n_rollouts_w_trace": len(stats_w_trace),
+                "val/n_rollouts_w_reward": len([stat for stat in sample_stat_list if stat["has_reward"]]),
                 "val/reward": np.mean(
                     [stat["reward"] for stat in sample_stat_list]
                 ),  # each rollout must have a reward (fillna if missing)
@@ -643,9 +645,10 @@ class AgentModeDaemon:
         # 1. Reconstruct the `finished_id_to_sample_info` structure from completed rollouts
         finished_id_to_sample_info: Dict[str, Dict[str, Any]] = {}
         finished_id_to_final_reward: Dict[str, float] = {}
+        sample_with_reward_count = 0
         for rollout_id, rollout in self._completed_rollouts_v0.items():
             original_sample = self._task_id_to_original_sample[rollout_id]
-
+            sample_with_reward_count += int(rollout.final_reward is not None)
             final_reward = self._fillna_reward(rollout)
 
             if not rollout.triplets:
@@ -764,6 +767,7 @@ class AgentModeDaemon:
             "training/reward": np.mean(list(finished_id_to_final_reward.values())),
             "training/n_rollouts": len(finished_id_to_final_reward),
             "training/n_rollouts_w_trace": len(finished_id_to_sample_info),
+            "training/n_rollouts_w_reward": sample_with_reward_count,
             "training/n_truncated_triplets": n_trunc_sample_because_of_response,
             "training/n_triplets": n_transition,
         }

agentlightning/verl/entrypoint.py

@@ -2,10 +2,12 @@
 
 # type: ignore
 
+from importlib.metadata import version
 from typing import Any
 
 import hydra
 import ray
+from packaging import version as packaging_version
 from verl.trainer.main_ppo import create_rl_sampler
 from verl.trainer.ppo.reward import load_reward_manager
 
@@ -39,11 +41,17 @@ def run_ppo(
 ) -> None:
     if not ray.is_initialized():
         # this is for local ray cluster
+        try:
+            # verl >= 0.6.0
+            num_cpus = config.ray_kwargs.ray_init.num_cpus
+        except AttributeError:
+            # verl < 0.6.0
+            num_cpus = config.ray_init.num_cpus
         ray.init(
             runtime_env={
                 "env_vars": {"TOKENIZERS_PARALLELISM": "true", "NCCL_DEBUG": "WARN", "VLLM_LOGGING_LEVEL": "WARN"}
             },
-            num_cpus=config.ray_init.num_cpus,
+            num_cpus=num_cpus,
         )
 
     runner = TaskRunner.remote()

agentlightning/verl/trainer.py

@@ -12,6 +12,7 @@ from typing import Dict, Tuple
 
 import numpy as np
 import torch
+import verl
 from codetiming import Timer
 from omegaconf import OmegaConf
 from tqdm import tqdm
@@ -19,7 +20,7 @@ from verl import DataProto
 from verl.protocol import pad_dataproto_to_divisor, unpad_dataproto
 from verl.trainer.ppo.core_algos import agg_loss
 from verl.trainer.ppo.metric_utils import (
-    compute_data_metrics,
+    _compute_response_info,
     compute_throughout_metrics,
     compute_timing_metrics,
 )
@@ -53,6 +54,108 @@ def _timer(name: str, timing_raw: Dict[str, float]):
     timing_raw[name] += timer.last
 
 
+# This function is adapted from verl.
+# We introduce a new parameter `suffix` to distinguish between metrics computed
+# before and after AgentLightning’s post-processing.
+# - "Before" refers to raw reward and advantage values.
+# - "After" refers to values computed following post-processing, which involves:
+#     (1) Dropping prompts that exceed the maximum allowed length.
+#     (2) Adjusting the batch size to be a multiple of the mini PPO size.
+# Different suffixes are used to label these two stages accordingly.
+def compute_data_metrics(batch: DataProto, use_critic: bool = True, suffix: str = "") -> Dict[str, Any]:
+    """
+    Computes various metrics from a batch of data for PPO training.
+
+    This function calculates metrics related to scores, rewards, advantages, returns, values,
+    and sequence lengths from a batch of data. It provides statistical information (mean, max, min)
+    for each metric category.
+
+    Args:
+        batch: A DataProto object containing batch data with token-level scores, rewards, advantages, etc.
+        use_critic: Whether to include critic-specific metrics. Defaults to True.
+
+    Returns:
+        A dictionary of metrics including:
+            - critic/score/mean, max, min: Statistics about sequence scores
+            - critic/rewards/mean, max, min: Statistics about sequence rewards
+            - critic/advantages/mean, max, min: Statistics about advantages
+            - critic/returns/mean, max, min: Statistics about returns
+            - critic/values/mean, max, min: Statistics about critic values (if use_critic=True)
+            - critic/vf_explained_var: Explained variance of the value function (if use_critic=True)
+            - response_length/mean, max, min, clip_ratio: Statistics about response lengths
+            - prompt_length/mean, max, min, clip_ratio: Statistics about prompt lengths
+    """
+    sequence_score = batch.batch["token_level_scores"].sum(-1)
+    sequence_reward = batch.batch["token_level_rewards"].sum(-1)
+
+    advantages = batch.batch["advantages"]
+    returns = batch.batch["returns"]
+
+    max_response_length = batch.batch["responses"].shape[-1]
+
+    prompt_mask = batch.batch["attention_mask"][:, :-max_response_length].bool()
+    response_mask = batch.batch["attention_mask"][:, -max_response_length:].bool()
+
+    max_prompt_length = prompt_mask.size(-1)
+
+    response_info = _compute_response_info(batch)
+    prompt_length = response_info["prompt_length"]
+    response_length = response_info["response_length"]
+
+    valid_adv = torch.masked_select(advantages, response_mask)
+    valid_returns = torch.masked_select(returns, response_mask)
+
+    if use_critic:
+        values = batch.batch["values"]
+        valid_values = torch.masked_select(values, response_mask)
+        return_diff_var = torch.var(valid_returns - valid_values)
+        return_var = torch.var(valid_returns)
+
+    metrics = {
+        # score
+        "critic/score/mean" + suffix: torch.mean(sequence_score).detach().item(),
+        "critic/score/max" + suffix: torch.max(sequence_score).detach().item(),
+        "critic/score/min" + suffix: torch.min(sequence_score).detach().item(),
+        # reward
+        "critic/rewards/mean" + suffix: torch.mean(sequence_reward).detach().item(),
+        "critic/rewards/max" + suffix: torch.max(sequence_reward).detach().item(),
+        "critic/rewards/min" + suffix: torch.min(sequence_reward).detach().item(),
+        # adv
+        "critic/advantages/mean" + suffix: torch.mean(valid_adv).detach().item(),
+        "critic/advantages/max" + suffix: torch.max(valid_adv).detach().item(),
+        "critic/advantages/min" + suffix: torch.min(valid_adv).detach().item(),
+        # returns
+        "critic/returns/mean" + suffix: torch.mean(valid_returns).detach().item(),
+        "critic/returns/max" + suffix: torch.max(valid_returns).detach().item(),
+        "critic/returns/min" + suffix: torch.min(valid_returns).detach().item(),
+        **(
+            {
+                # values
+                "critic/values/mean" + suffix: torch.mean(valid_values).detach().item(),
+                "critic/values/max" + suffix: torch.max(valid_values).detach().item(),
+                "critic/values/min" + suffix: torch.min(valid_values).detach().item(),
+                # vf explained var
+                "critic/vf_explained_var" + suffix: (1.0 - return_diff_var / (return_var + 1e-5)).detach().item(),
+            }
+            if use_critic
+            else {}
+        ),
+        # response length
+        "response_length/mean" + suffix: torch.mean(response_length).detach().item(),
+        "response_length/max" + suffix: torch.max(response_length).detach().item(),
+        "response_length/min" + suffix: torch.min(response_length).detach().item(),
+        "response_length/clip_ratio"
+        + suffix: torch.mean(torch.eq(response_length, max_response_length).float()).detach().item(),
+        # prompt length
+        "prompt_length/mean" + suffix: torch.mean(prompt_length).detach().item(),
+        "prompt_length/max" + suffix: torch.max(prompt_length).detach().item(),
+        "prompt_length/min" + suffix: torch.min(prompt_length).detach().item(),
+        "prompt_length/clip_ratio"
+        + suffix: torch.mean(torch.eq(prompt_length, max_prompt_length).float()).detach().item(),
+    }
+    return metrics
+
+
 class AgentLightningTrainer(RayPPOTrainer):
     """
     Specialized PPO trainer for agent-based reinforcement learning.
@@ -215,6 +318,9 @@ class AgentLightningTrainer(RayPPOTrainer):
                     config=self.config.algorithm,
                 )
 
+            # Calculate the metrics before processing. Refer to the comments of function `compute_data_metrics` for details.
+            metrics.update(compute_data_metrics(batch=batch, use_critic=self.use_critic, suffix="_before_processing"))
+
             # after advantages are assinged, we begin to drop (1) long prompt (2) floor to ppo minisize
             keep_indices = (~batch.batch["is_drop_mask"]).nonzero(as_tuple=True)[0]
             metrics["training/n_triplets_prompt_too_long"] = (
@@ -274,7 +380,7 @@ class AgentLightningTrainer(RayPPOTrainer):
                     )
 
         # compute training metrics
-        metrics.update(compute_data_metrics(batch=batch, use_critic=self.use_critic))
+        metrics.update(compute_data_metrics(batch=batch, use_critic=self.use_critic, suffix="_after_processing"))
         metrics.update(compute_timing_metrics(batch=batch, timing_raw=timing_raw))
         # TODO: implement actual tflpo and theoretical tflpo
         n_gpus = self.resource_pool_manager.get_n_gpus()
@@ -298,14 +404,20 @@ class AgentLightningTrainer(RayPPOTrainer):
         assert self.async_rollout_mode, "If agent mode is enabled, async server must be enabled"
         if self.adapter is not None and not isinstance(self.adapter, TraceToTripletBase):
             raise ValueError("Adapter must be a TraceToTripletBase for currently VERL implementation.")
+        verl_version = verl.__version__
+        if verl_version == "0.5.0":
+            # Note (Zhiyuan): To avoid further patch into vllm async server, using the same sentence to get the naming here.
+            # However, it is possible that verl updates the naming and causes incompatibility.
+            # Reference: https://github.com/volcengine/verl/blob/5b5e09d9cc20625e436d01f69d9cc739ff681c54/verl/workers/rollout/vllm_rollout/vllm_async_server.py#L217
+            model = "/".join(self.config.actor_rollout_ref.model.path.split("/")[-2:])
+        else:
+            # For other versions (e.g., 0.6.0), we use the full path to the model.
+            model = self.config.actor_rollout_ref.model.path
         self.agent_mode_daemon = AgentModeDaemon(
             self.config.agentlightning.port,
             self.config.actor_rollout_ref.rollout.n,
             train_information={
-                # Note (Zhiyuan): To avoid further patch into vllm async server, using the same sentence to get the naming here.
-                # However, it is possible that verl updates the naming and causes incompatibility.
-                # Reference: https://github.com/volcengine/verl/blob/5b5e09d9cc20625e436d01f69d9cc739ff681c54/verl/workers/rollout/vllm_rollout/vllm_async_server.py#L217
-                "model": "/".join(self.config.actor_rollout_ref.model.path.split("/")[-2:]),
+                "model": model,
                 "temperature": self.config.actor_rollout_ref.rollout.temperature,
             },
             tokenizer=self.tokenizer,

dashboard/.gitignore

@@ -0,0 +1,132 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+lerna-debug.log*
+.pnpm-debug.log*
+
+# Diagnostic reports (https://nodejs.org/api/report.html)
+report.[0-9]*.[0-9]*.[0-9]*.[0-9]*.json
+
+# Runtime data
+pids
+*.pid
+*.seed
+*.pid.lock
+
+# Directory for instrumented libs generated by jscoverage/JSCover
+lib-cov
+
+# Coverage directory used by tools like istanbul
+coverage
+*.lcov
+
+# nyc test coverage
+.nyc_output
+
+# Grunt intermediate storage (https://gruntjs.com/creating-plugins#storing-task-files)
+.grunt
+
+# Bower dependency directory (https://bower.io/)
+bower_components
+
+# node-waf configuration
+.lock-wscript
+
+# Compiled binary addons (https://nodejs.org/api/addons.html)
+build/Release
+
+# Dependency directories
+node_modules/
+jspm_packages/
+
+# Snowpack dependency directory (https://snowpack.dev/)
+web_modules/
+
+# TypeScript cache
+*.tsbuildinfo
+
+# Optional npm cache directory
+.npm
+
+# Optional eslint cache
+.eslintcache
+
+# Optional stylelint cache
+.stylelintcache
+
+# Microbundle cache
+.rpt2_cache/
+.rts2_cache_cjs/
+.rts2_cache_es/
+.rts2_cache_umd/
+
+# Optional REPL history
+.node_repl_history
+
+# Output of 'npm pack'
+*.tgz
+
+# Yarn Integrity file
+.yarn-integrity
+
+# dotenv environment variable files
+.env
+.env.development.local
+.env.test.local
+.env.production.local
+.env.local
+
+# parcel-bundler cache (https://parceljs.org/)
+.cache
+.parcel-cache
+
+# Next.js build output
+.next
+out
+
+# Nuxt.js build / generate output
+.nuxt
+dist
+
+# Gatsby files
+.cache/
+# Comment in the public line in if your project uses Gatsby and not Next.js
+# https://nextjs.org/blog/next-9-1#public-directory-support
+# public
+
+# vuepress build output
+.vuepress/dist
+
+# vuepress v2.x temp and cache directory
+.temp
+.cache
+
+# Docusaurus cache and generated files
+.docusaurus
+
+# Serverless directories
+.serverless/
+
+# FuseBox cache
+.fusebox/
+
+# DynamoDB Local files
+.dynamodb/
+
+# TernJS port file
+.tern-port
+
+# Stores VSCode versions used for testing VSCode extensions
+.vscode-test
+
+# yarn v2
+.yarn/cache
+.yarn/unplugged
+.yarn/build-state.yml
+.yarn/install-state.gz
+.pnp.*
+
+.DS_Store

dashboard/.prettierrc.mjs

@@ -0,0 +1,47 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+/** @type {import("@ianvs/prettier-plugin-sort-imports").PrettierConfig} */
+const config = {
+  printWidth: 120,
+  singleQuote: true,
+  tabWidth: 2,
+  useTabs: false,
+  semi: true,
+  quoteProps: 'consistent',
+  jsxSingleQuote: true,
+  trailingComma: 'all',
+  bracketSpacing: true,
+  objectWrap: 'preserve',
+  arrowParens: 'always',
+  proseWrap: 'preserve',
+  endOfLine: 'lf',
+  plugins: ['@ianvs/prettier-plugin-sort-imports'],
+  importOrder: [
+    '.*styles.css$',
+    '',
+    'dayjs',
+    '^react$',
+    '^next$',
+    '^next/.*$',
+    '<BUILTIN_MODULES>',
+    '<THIRD_PARTY_MODULES>',
+    '^@mantine/(.*)$',
+    '^@mantinex/(.*)$',
+    '^@mantine-tests/(.*)$',
+    '^@docs/(.*)$',
+    '^@/.*$',
+    '^../(?!.*.css$).*$',
+    '^./(?!.*.css$).*$',
+    '\\.css$',
+  ],
+  overrides: [
+    {
+      files: '*.mdx',
+      options: {
+        printWidth: 120,
+      },
+    },
+  ],
+};
+
+export default config;

dashboard/.storybook/constants.ts

@@ -0,0 +1,12 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+// Centralized constants that keep Storybook fixtures deterministic so Chromatic
+// snapshots do not drift when the build environment changes.
+export const STORY_DATE_NOW_MS = 1762775145209;
+export const STORY_DATE_NOW_SECONDS = Math.floor(STORY_DATE_NOW_MS / 1000);
+
+// Use a fixed origin so any code that would normally read window.location.*
+// in the app can rely on the same value from Storybook fixtures. Prefer HTTPS
+// so Chromatic (which is served over HTTPS) avoids mixed-content fetch errors.
+export const STORY_BASE_URL = 'https://storybook.agentlightning.invalid';
+export const STORY_LOCATION_HREF = `${STORY_BASE_URL}/storybook`;

dashboard/.storybook/main.ts

@@ -0,0 +1,20 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import type { StorybookConfig } from '@storybook/react-vite';
+
+const config: StorybookConfig = {
+  core: {
+    disableWhatsNewNotifications: true,
+    disableTelemetry: true,
+    enableCrashReports: false,
+  },
+  stories: ['../src/**/*.mdx', '../src/**/*.story.@(js|jsx|ts|tsx)'],
+  staticDirs: ['../static'],
+  addons: ['@storybook/addon-themes', '@storybook/addon-vitest'],
+  framework: {
+    name: '@storybook/react-vite',
+    options: {},
+  },
+};
+
+export default config;

dashboard/.storybook/modes.ts

@@ -0,0 +1,13 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+export const allModes = {
+  MD: {
+    viewport: 'md',
+  },
+  LG: {
+    viewport: 'lg',
+  },
+  XL: {
+    viewport: 'xl',
+  },
+} as const;

dashboard/.storybook/preview.tsx

@@ -0,0 +1,79 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import '@mantine/core/styles.css';
+import 'mantine-datatable/styles.css';
+import '../src/styles/theme.css';
+import '../src/styles/app.css';
+
+import { initialize, mswLoader } from 'msw-storybook-addon';
+import { ColorSchemeScript, MantineProvider } from '@mantine/core';
+import { shadcnCssVariableResolver } from '../src/cssVariableResolver';
+import { theme as mantineTheme } from '../src/theme';
+import { STORY_DATE_NOW_MS } from './constants';
+
+type ColorSchemeValue = 'light' | 'dark';
+
+initialize({
+  onUnhandledRequest: 'bypass',
+  serviceWorker: {
+    url: '/mockServiceWorker.js',
+  },
+});
+
+const fixedDateNow = (() => {
+  const patched = Date.now as typeof Date.now & { __storybookPatched?: boolean };
+  if (patched.__storybookPatched) {
+    return patched;
+  }
+  const replacement = (() => STORY_DATE_NOW_MS) as typeof Date.now & { __storybookPatched?: boolean };
+  replacement.__storybookPatched = true;
+  return replacement;
+})();
+
+Date.now = fixedDateNow;
+
+export const parameters = {
+  layout: 'fullscreen',
+  options: {
+    showPanel: false,
+    // @ts-expect-error – storybook throws build error for (a: any, b: any)
+    storySort: (a, b) => a.title.localeCompare(b.title, undefined, { numeric: true }),
+  },
+  backgrounds: { disable: true },
+  viewport: {
+    options: {
+      md: { name: 'md', styles: { width: '1280px', height: '800px' } },
+      lg: { name: 'lg', styles: { width: '1920px', height: '1080px' } },
+      xl: { name: 'xl', styles: { width: '2560px', height: '1440px' } },
+    },
+  },
+};
+
+export const globalTypes = {
+  theme: {
+    name: 'Theme',
+    description: 'Mantine color scheme',
+    defaultValue: 'light',
+    toolbar: {
+      icon: 'mirror',
+      items: [
+        { value: 'light', title: 'Light' },
+        { value: 'dark', title: 'Dark' },
+      ],
+    },
+  },
+};
+
+export const decorators = [
+  (Story: any, context: any) => {
+    const scheme = (context.parameters.theme ?? context.globals.theme ?? 'light') as ColorSchemeValue;
+    return (
+      <MantineProvider theme={mantineTheme} cssVariablesResolver={shadcnCssVariableResolver} forceColorScheme={scheme}>
+        <ColorSchemeScript />
+        <Story />
+      </MantineProvider>
+    );
+  },
+];
+
+export const loaders = [mswLoader];

dashboard/.storybook/vitest.setup.ts

@@ -0,0 +1,8 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { setProjectAnnotations } from '@storybook/react-vite';
+import * as projectAnnotations from './preview';
+
+// This is an important step to apply the right configuration when testing your stories.
+// More info at: https://storybook.js.org/docs/api/portable-stories/portable-stories-vitest#setprojectannotations
+setProjectAnnotations([projectAnnotations]);

dashboard/.stylelintignore

@@ -0,0 +1,5 @@
+# Generated files
+dist
+
+# Theme files
+theme.css

dashboard/.stylelintrc.json

@@ -0,0 +1,28 @@
+{
+  "extends": ["stylelint-config-standard-scss"],
+  "rules": {
+    "custom-property-pattern": null,
+    "selector-class-pattern": null,
+    "scss/no-duplicate-mixins": null,
+    "declaration-empty-line-before": null,
+    "declaration-block-no-redundant-longhand-properties": null,
+    "alpha-value-notation": null,
+    "custom-property-empty-line-before": null,
+    "property-no-vendor-prefix": null,
+    "color-function-notation": null,
+    "length-zero-no-unit": null,
+    "selector-not-notation": null,
+    "no-descending-specificity": null,
+    "comment-empty-line-before": null,
+    "scss/at-mixin-pattern": null,
+    "scss/at-rule-no-unknown": null,
+    "value-keyword-case": null,
+    "media-feature-range-notation": null,
+    "selector-pseudo-class-no-unknown": [
+      true,
+      {
+        "ignorePseudoClasses": ["global"]
+      }
+    ]
+  }
+}

dashboard/README.md

@@ -0,0 +1,27 @@
+# Agent-lightning Dashboard
+
+This is the dashboard for Agent-lightning. It is a web application that allows you to inspect your Agent-lightning store and debug running experiments.
+
+The dashboard is built with React, Mantine UI, and Storybook.
+
+## npm scripts
+
+## Build and dev scripts
+
+- `dev` – start development server
+- `build` – build production version of the app
+- `preview` – locally preview production build
+
+### Testing scripts
+
+- `eslint` - runs ESLint
+- `stylelint` - runs Stylelint
+- `prettier` - runs Prettier
+- `typecheck` - runs TypeScript typecheck
+- `vitest` – runs vitest tests
+- `chromatic` – runs chromatic tests
+
+### Other scripts
+
+- `storybook` – starts storybook dev server
+- `build-storybook` – build production storybook bundle to `storybook-static`

dashboard/eslint.config.js

@@ -0,0 +1,50 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+// @ts-check
+import stylistic from '@stylistic/eslint-plugin';
+import mantine from 'eslint-config-mantine';
+import { defineConfig } from 'eslint/config';
+import tseslint from 'typescript-eslint';
+
+export default defineConfig([
+  // These are arrays → safe to spread
+  ...tseslint.configs.recommended,
+  stylistic.configs.customize({ semi: true }),
+
+  // mantine is often a single object → include as-is (or spread only if it's actually an array)
+  ...(Array.isArray(mantine) ? mantine : [mantine]),
+
+  // ignores go as their own entry
+  { ignores: ['**/*.{mjs,cjs,js,d.ts,d.mts}'] },
+
+  // file-specific rules
+  {
+    files: ['**/*.story.tsx'],
+    rules: { 'no-console': 'off' },
+  },
+
+  // project/TS settings + your custom rules
+  {
+    languageOptions: {
+      parserOptions: {
+        tsconfigRootDir: process.cwd(),
+        project: ['./tsconfig.json'],
+      },
+    },
+    rules: {
+      // Disabling conflict rules with prettier
+      '@stylistic/brace-style': ['error', '1tbs', { allowSingleLine: false }],
+      '@stylistic/no-trailing-spaces': 'error',
+      '@stylistic/no-multiple-empty-lines': ['error', { max: 2, maxEOF: 1 }],
+      '@stylistic/jsx-quotes': ['error', 'prefer-single'],
+      '@stylistic/multiline-ternary': 'off',
+      '@stylistic/arrow-parens': ['error', 'always'],
+      '@stylistic/jsx-closing-bracket-location': 'off',
+      '@stylistic/operator-linebreak': 'off',
+      '@stylistic/jsx-newline': 'off',
+      '@stylistic/jsx-one-expression-per-line': 'off',
+      '@stylistic/indent': 'off',
+      '@stylistic/indent-binary-ops': 'off',
+    },
+  },
+]);

dashboard/package.json

@@ -0,0 +1,75 @@
+{
+  "name": "agent-lightning-dashboard",
+  "type": "module",
+  "version": "0.3.0",
+  "scripts": {
+    "dev": "vite",
+    "build": "tsc && vite build",
+    "preview": "vite preview",
+    "typecheck": "tsc --noEmit",
+    "eslint": "eslint .",
+    "stylelint": "stylelint '**/*.css'",
+    "prettier": "prettier --check \"**/*.{ts,tsx,mjs,cjs}\"",
+    "vitest": "vitest run --project unit",
+    "vitest-storybook": "vitest run --project storybook",
+    "storybook": "storybook dev -p 6006",
+    "build-storybook": "storybook build",
+    "chromatic": "chromatic"
+  },
+  "dependencies": {
+    "@mantine/core": "8.3.5",
+    "@mantine/hooks": "8.3.5",
+    "@monaco-editor/react": "^4.7.0",
+    "@reduxjs/toolkit": "^2.9.2",
+    "@tabler/icons-react": "^3.35.0",
+    "clsx": "^2.1.1",
+    "dayjs": "^1.11.18",
+    "mantine-datatable": "^8.2.0",
+    "react": "^19.2.0",
+    "react-dom": "^19.2.0",
+    "react-redux": "^9.2.0",
+    "react-router-dom": "^7.9.4"
+  },
+  "devDependencies": {
+    "@eslint/js": "^9.37.0",
+    "@ianvs/prettier-plugin-sort-imports": "^4.7.0",
+    "@storybook/addon-themes": "^9.1.10",
+    "@storybook/addon-vitest": "^9.1.16",
+    "@storybook/react": "^9.1.10",
+    "@storybook/react-vite": "^9.1.10",
+    "@stylistic/eslint-plugin": "^5.5.0",
+    "@testing-library/dom": "^10.4.1",
+    "@testing-library/jest-dom": "^6.9.1",
+    "@testing-library/react": "^16.3.0",
+    "@testing-library/user-event": "^14.6.1",
+    "@types/node": "^24.7.1",
+    "@types/react": "^19.2.2",
+    "@types/react-dom": "^19.2.1",
+    "@vitejs/plugin-react": "^5.0.4",
+    "chromatic": "^13.3.3",
+    "eslint": "^9.37.0",
+    "eslint-config-mantine": "^4.0.3",
+    "eslint-plugin-jsx-a11y": "^6.10.2",
+    "eslint-plugin-react": "^7.37.5",
+    "identity-obj-proxy": "^3.0.0",
+    "jsdom": "^27.0.0",
+    "msw": "^2.11.6",
+    "msw-storybook-addon": "^2.0.6",
+    "postcss": "^8.5.6",
+    "postcss-preset-mantine": "1.18.0",
+    "postcss-simple-vars": "^7.0.1",
+    "prettier": "^3.6.2",
+    "prop-types": "^15.8.1",
+    "storybook": "^9.1.10",
+    "stylelint": "^16.25.0",
+    "stylelint-config-standard-scss": "^16.0.0",
+    "typescript": "^5.9.3",
+    "typescript-eslint": "^8.46.0",
+    "vite": "^7.1.9",
+    "vite-tsconfig-paths": "^5.1.4",
+    "vitest": "^4.0.0",
+    "playwright": "^1.56.1",
+    "@vitest/browser-playwright": "4.0.4",
+    "@vitest/coverage-v8": "4.0.4"
+  }
+}

dashboard/postcss.config.cjs

@@ -0,0 +1,16 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+module.exports = {
+  plugins: {
+    'postcss-preset-mantine': {},
+    'postcss-simple-vars': {
+      variables: {
+        'mantine-breakpoint-xs': '36em',
+        'mantine-breakpoint-sm': '48em',
+        'mantine-breakpoint-md': '62em',
+        'mantine-breakpoint-lg': '75em',
+        'mantine-breakpoint-xl': '88em',
+      },
+    },
+  },
+};

dashboard/public/index.html

@@ -0,0 +1,13 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <link rel="icon" type="image/svg+xml" href="../src/favicon.svg" />
+    <meta name="viewport" content="minimum-scale=1, initial-scale=1, width=device-width, user-scalable=no" />
+    <title>Agent-lightning Dashboard</title>
+  </head>
+  <body>
+    <div id="root"></div>
+    <script type="module" src="/main.tsx"></script>
+  </body>
+</html>

dashboard/public/main.tsx

@@ -0,0 +1,3 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import '../src/main.js';

dashboard/src/App.tsx

@@ -0,0 +1,30 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import '@mantine/core/styles.css';
+import 'mantine-datatable/styles.css';
+import './styles/theme.css';
+import './styles/app.css';
+
+import { MantineProvider } from '@mantine/core';
+import { useColorScheme } from '@mantine/hooks';
+import { shadcnCssVariableResolver } from './cssVariableResolver';
+import { selectThemePreference } from './features/config/selectors';
+import { Router } from './Router';
+import { useAppSelector } from './store/hooks';
+import { shadcnTheme } from './theme';
+
+export default function App() {
+  const themePreference = useAppSelector(selectThemePreference);
+  const systemColorScheme = useColorScheme();
+  const resolvedColorScheme = themePreference === 'system' ? systemColorScheme : themePreference;
+
+  return (
+    <MantineProvider
+      theme={shadcnTheme}
+      cssVariablesResolver={shadcnCssVariableResolver}
+      forceColorScheme={resolvedColorScheme}
+    >
+      <Router />
+    </MantineProvider>
+  );
+}

dashboard/src/Router.tsx

@@ -0,0 +1,46 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { createBrowserRouter, Navigate, RouterProvider } from 'react-router-dom';
+import { AppLayoutWithState } from './layouts/AppLayout';
+import { ResourcesPage } from './pages/Resources.page';
+import { RolloutsPage } from './pages/Rollouts.page';
+import { SettingsPage } from './pages/Settings.page';
+import { TracesPage } from './pages/Traces.page';
+import { WorkersPage } from './pages/Workers.page';
+
+const router = createBrowserRouter([
+  {
+    path: '/',
+    element: <AppLayoutWithState />,
+    children: [
+      {
+        index: true,
+        element: <Navigate to='/rollouts' replace />,
+      },
+      {
+        path: 'rollouts',
+        element: <RolloutsPage />,
+      },
+      {
+        path: 'resources',
+        element: <ResourcesPage />,
+      },
+      {
+        path: 'traces',
+        element: <TracesPage />,
+      },
+      {
+        path: 'runners',
+        element: <WorkersPage />,
+      },
+      {
+        path: 'settings',
+        element: <SettingsPage />,
+      },
+    ],
+  },
+]);
+
+export function Router() {
+  return <RouterProvider router={router} />;
+}

dashboard/src/components/AppAlertBanner.story.tsx

@@ -0,0 +1,91 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import type { Meta, StoryObj } from '@storybook/react';
+import { Provider } from 'react-redux';
+import { initialConfigState } from '@/features/config/slice';
+import { initialRolloutsUiState } from '@/features/rollouts/slice';
+import type { AlertsState, AlertTone } from '@/features/ui/alert';
+import { initialDrawerState } from '@/features/ui/drawer/slice';
+import { createAppStore } from '@/store';
+import { STORY_BASE_URL, STORY_DATE_NOW_MS } from '../../.storybook/constants';
+import { AppAlertBanner } from './AppAlertBanner';
+
+const meta: Meta<typeof AppAlertBanner> = {
+  title: 'Components/AppAlertBanner',
+  component: AppAlertBanner,
+  parameters: {
+    layout: 'fullscreen',
+  },
+};
+
+export default meta;
+
+type Story = StoryObj<typeof AppAlertBanner>;
+
+function renderWithAlert(message: string, tone: AlertTone) {
+  const alertState: AlertsState = {
+    alerts: [
+      {
+        id: 'storybook-alert',
+        message,
+        tone,
+        isVisible: true,
+        createdAt: STORY_DATE_NOW_MS,
+      },
+    ],
+  };
+
+  const store = createAppStore({
+    config: {
+      ...initialConfigState,
+      baseUrl: STORY_BASE_URL,
+    },
+    drawer: initialDrawerState,
+    rollouts: initialRolloutsUiState,
+    alert: alertState,
+  });
+
+  return (
+    <Provider store={store}>
+      <div style={{ padding: 24 }}>
+        <AppAlertBanner />
+      </div>
+    </Provider>
+  );
+}
+
+export const InfoAlert: Story = {
+  render: () => renderWithAlert('Background synchronization completed successfully.', 'info'),
+};
+
+export const WarningAlert: Story = {
+  render: () =>
+    renderWithAlert('Rollout data may be stale. Check your network connection before continuing.', 'warning'),
+};
+
+export const ErrorAlert: Story = {
+  render: () =>
+    renderWithAlert('Unable to reach the Agent-lightning API. Retry or adjust the backend settings.', 'error'),
+};
+
+export const NoAlert: Story = {
+  render: () => {
+    const store = createAppStore({
+      config: {
+        ...initialConfigState,
+        baseUrl: STORY_BASE_URL,
+      },
+      drawer: initialDrawerState,
+      rollouts: initialRolloutsUiState,
+      alert: { alerts: [] },
+    });
+
+    return (
+      <Provider store={store}>
+        <div style={{ padding: 24 }}>
+          <AppAlertBanner />
+        </div>
+      </Provider>
+    );
+  },
+};

dashboard/src/components/AppAlertBanner.tsx

@@ -0,0 +1,84 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { useEffect, useState } from 'react';
+import { IconAlertCircle, IconAlertTriangle, IconInfoCircle } from '@tabler/icons-react';
+import { Notification, Portal, Transition } from '@mantine/core';
+import { hideAlert, selectHighestPriorityAlert, type AppAlert } from '@/features/ui/alert';
+import { useAppDispatch, useAppSelector } from '@/store/hooks';
+
+const ALERT_META = {
+  info: {
+    color: 'blue',
+    icon: IconInfoCircle,
+  },
+  warning: {
+    color: 'yellow',
+    icon: IconAlertTriangle,
+  },
+  error: {
+    color: 'red',
+    icon: IconAlertCircle,
+  },
+} as const;
+
+export function AppAlertBanner() {
+  const dispatch = useAppDispatch();
+  const alert = useAppSelector(selectHighestPriorityAlert);
+  const [transitionAlert, setTransitionAlert] = useState<AppAlert | null>(alert);
+
+  useEffect(() => {
+    if (alert) {
+      setTransitionAlert(alert);
+    }
+  }, [alert]);
+
+  const handleClose = (id?: string) => {
+    if (id) {
+      dispatch(hideAlert({ id }));
+    }
+  };
+
+  const currentAlert = alert ?? transitionAlert;
+  const mounted = Boolean(alert);
+
+  if (!currentAlert) {
+    return null;
+  }
+
+  const meta = ALERT_META[currentAlert.tone];
+  const IconComponent = meta.icon;
+
+  return (
+    <Portal>
+      <Transition
+        mounted={mounted}
+        transition='slide-down'
+        duration={200}
+        timingFunction='ease'
+        onExited={() => setTransitionAlert(null)}
+      >
+        {(styles) => (
+          <Notification
+            icon={<IconComponent size={18} />}
+            color={meta.color}
+            variant='light'
+            withCloseButton
+            onClose={() => handleClose(currentAlert.id)}
+            style={{
+              position: 'fixed',
+              top: 16,
+              right: 16,
+              maxWidth: 450,
+              width: 'calc(100% - 32px)',
+              zIndex: 2000,
+              boxShadow: 'var(--mantine-shadow-md)',
+              ...styles,
+            }}
+          >
+            {currentAlert.message}
+          </Notification>
+        )}
+      </Transition>
+    </Portal>
+  );
+}

dashboard/src/components/AppDrawer.component.tsx

@@ -0,0 +1,583 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { useCallback, useEffect, useMemo, useRef, useState, type ReactNode } from 'react';
+import { Editor } from '@monaco-editor/react';
+import { IconCheck, IconCopy } from '@tabler/icons-react';
+import type { DataTableSortStatus } from 'mantine-datatable';
+import { createSearchParams, Link, useInRouterContext, useLocation } from 'react-router-dom';
+import {
+  ActionIcon,
+  Anchor,
+  Badge,
+  Box,
+  CopyButton,
+  Drawer,
+  Group,
+  Stack,
+  Text,
+  Tooltip,
+  useMantineColorScheme,
+} from '@mantine/core';
+import { useGetSpansQuery } from '@/features/rollouts';
+import { closeDrawer, openDrawer, selectDrawerContent, selectDrawerIsOpen } from '@/features/ui/drawer';
+import { useAppDispatch, useAppSelector } from '@/store/hooks';
+import type { Attempt, AttemptStatus, Rollout, RolloutStatus, Span, Worker } from '@/types';
+import { formatStatusLabel } from '@/utils/format';
+import { TracesTable, type TracesTableRecord } from './TracesTable.component';
+
+const ATTEMPT_STATUS_COLORS: Record<AttemptStatus, string> = {
+  failed: 'red',
+  preparing: 'violet',
+  running: 'blue',
+  succeeded: 'teal',
+  timeout: 'orange',
+  unresponsive: 'orange',
+};
+
+const ROLLOUT_STATUS_COLORS: Record<RolloutStatus, string> = {
+  cancelled: 'gray',
+  failed: 'red',
+  preparing: 'violet',
+  queuing: 'blue',
+  requeuing: 'cyan',
+  running: 'blue',
+  succeeded: 'teal',
+};
+
+const SPAN_STATUS_COLORS: Record<Span['status']['status_code'], string> = {
+  UNSET: 'gray',
+  OK: 'teal',
+  ERROR: 'red',
+};
+
+const WORKER_STATUS_COLORS: Record<Worker['status'], string> = {
+  busy: 'orange',
+  idle: 'teal',
+  unknown: 'gray',
+};
+
+const TRACES_SORT_FIELD_MAP: Record<string, string> = {
+  name: 'name',
+  traceId: 'trace_id',
+  spanId: 'span_id',
+  parentId: 'parent_id',
+  statusCode: 'status_code',
+  startTime: 'start_time',
+  duration: 'duration',
+};
+
+type SortDirection = 'asc' | 'desc';
+
+type LocalSortState = {
+  column: string;
+  direction: SortDirection;
+};
+
+function resolveTracesSortField(column: string): string {
+  return TRACES_SORT_FIELD_MAP[column] ?? 'start_time';
+}
+
+function getStatusBadgeColor(status: RolloutStatus | AttemptStatus, isAttempt: boolean) {
+  if (isAttempt) {
+    return ATTEMPT_STATUS_COLORS[status as AttemptStatus] ?? 'gray';
+  }
+
+  return ROLLOUT_STATUS_COLORS[status as RolloutStatus] ?? 'gray';
+}
+
+function formatJson(value: unknown) {
+  try {
+    return JSON.stringify(value, null, 2);
+  } catch {
+    return String(value);
+  }
+}
+
+export type AppDrawerProps = {
+  opened: boolean;
+  onClose: () => void;
+  title?: ReactNode;
+  body?: ReactNode;
+};
+
+export function AppDrawer({ opened, onClose, title, body }: AppDrawerProps) {
+  return (
+    <Drawer
+      position='right'
+      size='lg'
+      opened={opened}
+      onClose={onClose}
+      overlayProps={{ opacity: 0.5 }}
+      withinPortal
+      styles={{
+        content: {
+          display: 'flex',
+          flexDirection: 'column',
+          maxHeight: '100vh',
+        },
+        body: {
+          flex: 1,
+          display: 'flex',
+          flexDirection: 'column',
+          padding: 'var(--mantine-spacing-md)',
+          minHeight: 0,
+          overflow: 'hidden',
+        },
+      }}
+      title={title}
+    >
+      <Stack gap='md' h='100%' style={{ flex: 1, minHeight: 0 }}>
+        {body}
+      </Stack>
+    </Drawer>
+  );
+}
+
+type TraceDrawerTitleProps = {
+  span: Span;
+};
+
+export function TraceDrawerTitle({ span }: TraceDrawerTitleProps) {
+  const spanStatusCode = span.status?.status_code ?? null;
+  const spanBadgeColor = spanStatusCode ? (SPAN_STATUS_COLORS[spanStatusCode] ?? 'gray') : undefined;
+
+  return (
+    <Stack gap={3}>
+      <Group gap={6}>
+        <Text fw={600}>{span.name ?? span.spanId}</Text>
+        {spanStatusCode ? (
+          <Badge size='sm' variant='light' color={spanBadgeColor}>
+            {spanStatusCode}
+          </Badge>
+        ) : null}
+      </Group>
+      <Group gap={6}>
+        <Text size='sm' c='dimmed'>
+          {span.spanId}
+        </Text>
+        <CopyButton value={span.spanId}>
+          {({ copied, copy }) => (
+            <Tooltip label={copied ? 'Copied' : 'Copy'} withArrow>
+              <ActionIcon
+                aria-label={`Copy span ID ${span.spanId}`}
+                variant='subtle'
+                color={copied ? 'teal' : 'gray'}
+                size='sm'
+                onClick={(event) => {
+                  event.stopPropagation();
+                  copy();
+                }}
+              >
+                {copied ? <IconCheck size={14} /> : <IconCopy size={14} />}
+              </ActionIcon>
+            </Tooltip>
+          )}
+        </CopyButton>
+      </Group>
+      <Group gap='xs'>
+        <Group gap={3}>
+          <Text size='sm' c='dimmed' fw={500}>
+            Rollout
+          </Text>
+          <Text size='sm' c='dimmed'>
+            {span.rolloutId}
+          </Text>
+        </Group>
+        <Group gap={3}>
+          <Text size='sm' c='dimmed' fw={500}>
+            Attempt
+          </Text>
+          <Text size='sm' c='dimmed'>
+            {span.attemptId ?? '—'}
+          </Text>
+        </Group>
+      </Group>
+    </Stack>
+  );
+}
+
+type RolloutAttemptDrawerTitleProps = {
+  rollout: Rollout;
+  attempt: Attempt | null;
+};
+
+export function RolloutAttemptDrawerTitle({ rollout, attempt }: RolloutAttemptDrawerTitleProps) {
+  const rolloutId = rollout.rolloutId;
+  const attemptId = attempt?.attemptId ?? null;
+  const rolloutStatus = rollout.status ?? null;
+  const attemptStatus = attempt?.status ?? null;
+  const rolloutStatusLabel = rolloutStatus ? formatStatusLabel(rolloutStatus) : null;
+  const attemptStatusLabel = attemptStatus ? formatStatusLabel(attemptStatus) : null;
+  const hasStatusMismatch = rolloutStatus !== null && attemptStatus !== null && rolloutStatus !== attemptStatus;
+  const rolloutBadgeColor = rolloutStatus ? getStatusBadgeColor(rolloutStatus, false) : undefined;
+  const attemptBadgeColor = attemptStatus ? getStatusBadgeColor(attemptStatus, true) : undefined;
+  const showRolloutBadgeInHeading = Boolean(rolloutStatusLabel && (!attemptStatus || hasStatusMismatch));
+  const showAttemptBadge = Boolean(attemptStatusLabel && attemptStatus);
+
+  return (
+    <Stack gap={3}>
+      <Group gap={6}>
+        <Text fw={600}>{rolloutId}</Text>
+        <CopyButton value={rolloutId}>
+          {({ copied, copy }) => (
+            <Tooltip label={copied ? 'Copied' : 'Copy'} withArrow>
+              <ActionIcon
+                aria-label={`Copy rollout ID ${rolloutId}`}
+                variant='subtle'
+                color={copied ? 'teal' : 'gray'}
+                size='sm'
+                onClick={(event) => {
+                  event.stopPropagation();
+                  copy();
+                }}
+              >
+                {copied ? <IconCheck size={14} /> : <IconCopy size={14} />}
+              </ActionIcon>
+            </Tooltip>
+          )}
+        </CopyButton>
+        {showRolloutBadgeInHeading && rolloutStatusLabel ? (
+          <Badge size='sm' variant='light' color={rolloutBadgeColor}>
+            {rolloutStatusLabel}
+          </Badge>
+        ) : null}
+      </Group>
+      <Group gap='xs'>
+        {attemptId ? (
+          <Group gap={3}>
+            <Text size='sm' c='dimmed' fw={500}>
+              Attempt
+            </Text>
+            <Text size='sm' c='dimmed'>
+              {attemptId}
+            </Text>
+          </Group>
+        ) : null}
+        {showAttemptBadge && attemptStatusLabel ? (
+          <Badge size='sm' variant='light' color={attemptBadgeColor}>
+            {attemptStatusLabel}
+          </Badge>
+        ) : null}
+        {!showRolloutBadgeInHeading && !attemptStatus && rolloutStatusLabel ? (
+          <Badge size='sm' variant='light' color={rolloutBadgeColor}>
+            {rolloutStatusLabel}
+          </Badge>
+        ) : null}
+      </Group>
+    </Stack>
+  );
+}
+
+type JsonEditorProps = {
+  value: unknown;
+};
+
+export function JsonEditor({ value }: JsonEditorProps) {
+  const { colorScheme } = useMantineColorScheme();
+  const editorTheme = colorScheme === 'dark' ? 'vs-dark' : 'vs-light';
+
+  return (
+    <Box data-testid='json-editor-container' style={{ flex: 1, minHeight: 0 }}>
+      <Editor
+        height='100%'
+        language='json'
+        value={formatJson(value)}
+        theme={editorTheme}
+        options={{
+          readOnly: true,
+          domReadOnly: true,
+          minimap: { enabled: false },
+          automaticLayout: true,
+          scrollBeyondLastLine: false,
+          fontSize: 13,
+        }}
+      />
+    </Box>
+  );
+}
+
+type RolloutTracesDrawerBodyProps = {
+  rollout: Rollout;
+  attempt: Attempt | null;
+  onShowRollout: (record: TracesTableRecord) => void;
+  onShowSpanDetail: (record: TracesTableRecord) => void;
+};
+
+function RolloutTracesDrawerBody({ rollout, attempt, onShowRollout, onShowSpanDetail }: RolloutTracesDrawerBodyProps) {
+  const [page, setPage] = useState(1);
+  const [recordsPerPage, setRecordsPerPage] = useState(100);
+  const [sort, setSort] = useState<LocalSortState>({
+    column: 'startTime',
+    direction: 'desc',
+  });
+
+  useEffect(() => {
+    setPage(1);
+  }, [rollout.rolloutId, attempt?.attemptId]);
+
+  const queryArgs = useMemo(
+    () => ({
+      rolloutId: rollout.rolloutId,
+      attemptId: attempt?.attemptId ?? undefined,
+      limit: recordsPerPage,
+      offset: Math.max(0, (page - 1) * recordsPerPage),
+      sortBy: resolveTracesSortField(sort.column),
+      sortOrder: sort.direction,
+    }),
+    [rollout.rolloutId, attempt?.attemptId, recordsPerPage, page, sort],
+  );
+
+  const { data, isFetching, isError, error, refetch } = useGetSpansQuery(queryArgs);
+  const spans = data?.items ?? [];
+  const totalRecords = data?.total ?? 0;
+  const tracesLinkSearch = useMemo(() => {
+    const params = createSearchParams({
+      rolloutId: rollout.rolloutId,
+      ...(attempt?.attemptId ? { attemptId: attempt.attemptId } : {}),
+    });
+    return params.toString();
+  }, [attempt?.attemptId, rollout.rolloutId]);
+  const tracesLinkHref = tracesLinkSearch ? `/traces?${tracesLinkSearch}` : '/traces';
+  const isWithinRouter = useInRouterContext();
+
+  const handleSortStatusChange = useCallback((status: DataTableSortStatus<TracesTableRecord>) => {
+    setSort({
+      column: status.columnAccessor as string,
+      direction: status.direction,
+    });
+  }, []);
+
+  const handlePageChange = useCallback((nextPage: number) => {
+    setPage(nextPage);
+  }, []);
+
+  const handleRecordsPerPageChange = useCallback((value: number) => {
+    setRecordsPerPage(value);
+    setPage(1);
+  }, []);
+
+  return (
+    <Stack gap='md' style={{ flex: 1, minHeight: 0 }}>
+      <Group justify='space-between' align='center' gap='sm' wrap='nowrap'>
+        <Text size='sm' style={{ flex: 1, minWidth: 0 }}>
+          Showing spans for{' '}
+          <Text component='span' fw={600}>
+            {rollout.rolloutId}
+            {attempt ? ` · Attempt ${attempt.sequenceId} (${attempt.attemptId})` : ' · Latest attempt'}
+          </Text>
+        </Text>
+        {isWithinRouter ? (
+          <Anchor
+            component={Link}
+            to={tracesLinkHref}
+            size='sm'
+            aria-label={`Open traces page for rollout ${rollout.rolloutId}${
+              attempt ? ` attempt ${attempt.sequenceId}` : ''
+            }`}
+          >
+            View full traces
+          </Anchor>
+        ) : (
+          <Anchor
+            href={tracesLinkHref}
+            size='sm'
+            aria-label={`Open traces page for rollout ${rollout.rolloutId}${
+              attempt ? ` attempt ${attempt.sequenceId}` : ''
+            }`}
+          >
+            View full traces
+          </Anchor>
+        )}
+      </Group>
+      <Box data-testid='traces-drawer-table-container' style={{ flex: 1, minHeight: 0, overflow: 'auto' }}>
+        <TracesTable
+          spans={spans}
+          totalRecords={totalRecords}
+          isFetching={isFetching}
+          isError={isError}
+          error={error}
+          searchTerm=''
+          sort={sort}
+          page={page}
+          recordsPerPage={recordsPerPage}
+          onSortStatusChange={handleSortStatusChange}
+          onPageChange={handlePageChange}
+          onRecordsPerPageChange={handleRecordsPerPageChange}
+          onResetFilters={() => {}}
+          onRefetch={refetch}
+          onShowRollout={onShowRollout}
+          onShowSpanDetail={onShowSpanDetail}
+          recordsPerPageOptions={[50, 100, 200, 500]}
+        />
+      </Box>
+    </Stack>
+  );
+}
+
+type WorkerDrawerTitleProps = {
+  worker: Worker;
+};
+
+function WorkerDrawerTitle({ worker }: WorkerDrawerTitleProps) {
+  const badgeColor = WORKER_STATUS_COLORS[worker.status] ?? 'gray';
+  return (
+    <Stack gap={3}>
+      <Group gap={6} align='center'>
+        <Text fw={600}>{worker.workerId}</Text>
+        <CopyButton value={worker.workerId}>
+          {({ copied, copy }) => (
+            <Tooltip label={copied ? 'Copied' : 'Copy'} withArrow>
+              <ActionIcon
+                aria-label={`Copy worker ID ${worker.workerId}`}
+                variant='subtle'
+                color={copied ? 'teal' : 'gray'}
+                size='sm'
+                onClick={(event) => {
+                  event.stopPropagation();
+                  copy();
+                }}
+              >
+                {copied ? <IconCheck size={14} /> : <IconCopy size={14} />}
+              </ActionIcon>
+            </Tooltip>
+          )}
+        </CopyButton>
+        <Badge size='sm' variant='light' color={badgeColor}>
+          {formatStatusLabel(worker.status)}
+        </Badge>
+      </Group>
+      <Group gap='xl'>
+        <Group gap={4}>
+          <Text size='sm' c='dimmed' fw={500}>
+            Rollout
+          </Text>
+          <Text size='sm' c='dimmed'>
+            {worker.currentRolloutId ?? '—'}
+          </Text>
+        </Group>
+        <Group gap={4}>
+          <Text size='sm' c='dimmed' fw={500}>
+            Attempt
+          </Text>
+          <Text size='sm' c='dimmed'>
+            {worker.currentAttemptId ?? '—'}
+          </Text>
+        </Group>
+      </Group>
+    </Stack>
+  );
+}
+
+export function AppDrawerContainer() {
+  const dispatch = useAppDispatch();
+  const isOpen = useAppSelector(selectDrawerIsOpen);
+  const content = useAppSelector(selectDrawerContent);
+  const isRouterAvailable = useInRouterContext();
+
+  const handleClose = useCallback(() => {
+    dispatch(closeDrawer());
+  }, [dispatch]);
+  const handleNavigation = useCallback(() => {
+    if (isOpen) {
+      dispatch(closeDrawer());
+    }
+  }, [dispatch, isOpen]);
+
+  const derivedContent = useMemo(() => {
+    if (!content) {
+      return null;
+    }
+
+    if (content.type === 'worker-detail') {
+      const { worker } = content;
+      const title = <WorkerDrawerTitle worker={worker} />;
+      const body = <JsonEditor value={worker} />;
+      return { title, body };
+    }
+
+    if (content.type === 'trace-detail') {
+      const { span } = content;
+      const title = <TraceDrawerTitle span={span} />;
+      const body = <JsonEditor value={span} />;
+
+      return { title, body };
+    }
+
+    const rollout = content.rollout;
+    const attempt = content.attempt;
+    const title = <RolloutAttemptDrawerTitle rollout={rollout} attempt={attempt} />;
+
+    if (content.type === 'rollout-json') {
+      const jsonValue = content.isNested && content.attempt ? content.attempt : rollout;
+      const body = jsonValue ? <JsonEditor value={jsonValue} /> : null;
+      return { title, body };
+    }
+
+    if (content.type === 'rollout-traces') {
+      const body = (
+        <RolloutTracesDrawerBody
+          rollout={rollout}
+          attempt={attempt}
+          onShowRollout={() => {
+            const attemptForRecord = attempt ?? rollout.attempt ?? null;
+            dispatch(
+              openDrawer({
+                type: 'rollout-json',
+                rollout,
+                attempt: attemptForRecord,
+                isNested: content.isNested,
+              }),
+            );
+          }}
+          onShowSpanDetail={(record) => {
+            const attemptForRecord = attempt ?? rollout.attempt ?? null;
+            dispatch(
+              openDrawer({
+                type: 'trace-detail',
+                span: record,
+                rollout,
+                attempt: attemptForRecord,
+              }),
+            );
+          }}
+        />
+      );
+
+      return { title, body };
+    }
+
+    return null;
+  }, [content, dispatch]);
+
+  if (!content || !derivedContent) {
+    return null;
+  }
+
+  const { title, body } = derivedContent;
+
+  return (
+    <>
+      {isRouterAvailable ? <DrawerLocationWatcher onNavigation={handleNavigation} /> : null}
+      <AppDrawer opened={isOpen} onClose={handleClose} title={title} body={body} />
+    </>
+  );
+}
+
+type DrawerLocationWatcherProps = {
+  onNavigation: () => void;
+};
+
+function DrawerLocationWatcher({ onNavigation }: DrawerLocationWatcherProps) {
+  const location = useLocation();
+  const lastLocationKeyRef = useRef(location.key);
+
+  useEffect(() => {
+    if (lastLocationKeyRef.current === location.key) {
+      return;
+    }
+    lastLocationKeyRef.current = location.key;
+    onNavigation();
+  }, [location.key, onNavigation]);
+
+  return null;
+}