try developer tier

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-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 +7,8 @@ on:
       - Examples - Spider
       - Examples - APO
       - Examples - Unsloth
+      - Examples - Tinker
+      - Examples - Azure
     types: [completed]
 
   workflow_dispatch:
@@ -31,5 +33,7 @@ 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'] },
             ];
             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 +1,19 @@
+# This workflow is used to benchmark the performance of the project.
+# It's kept as a placeholder for now.
+
+name: Benchmark
+permissions:
+  contents: read
+on:
+  workflow_dispatch:
+
+jobs:
+  benchmark:
+    name: Benchmark
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 60
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h

.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-compat.yml

@@ -14,7 +14,7 @@ on:
 run-name: >-
   ${{ github.event_name == 'repository_dispatch'
       && format(
-        'PR #{0} - Label {1} - {2}',
+        'Backward Compatibility - PR #{0} - {1} - {2}',
         github.event.client_payload.pull_number,
         github.event.client_payload.ci_label,
         github.event.client_payload.correlation_id
@@ -56,7 +56,7 @@ jobs:
             --group dev --group experiment --group agents --group torch-gpu-${{ matrix.setup-script }}
       - name: Override VERL (stable)
         run: |
-          uv pip install verl==0.5.0
+          uv pip install verl==0.5.0 vllm==0.10.2
         if: matrix.setup-script == 'stable'
       - name: Freeze dependencies
         run: |

.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/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
@@ -29,6 +29,129 @@ jobs:
       github.event.action == 'ci-all'
     name: GPU Test 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
+          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:
@@ -69,18 +192,10 @@ 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
 
-      - 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: Launch LiteLLM Proxy
         run: |
           ./scripts/litellm_run.sh
@@ -88,10 +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 \
@@ -166,7 +167,7 @@ jobs:
 
       - 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"
 

.gitignore

@@ -213,3 +213,6 @@ agentlightning/dashboard/**/*.css
 agentlightning/dashboard/**/*.js
 agentlightning/dashboard/**/*.html
 agentlightning/dashboard/**/*.svg
+
+# Docker data
+docker/data/

README.md

@@ -4,7 +4,7 @@
 
 # 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)
@@ -34,6 +34,12 @@ 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).
@@ -69,10 +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
 
@@ -92,7 +99,7 @@ If you find Agent Lightning useful in your research or projects, please cite our
 
 ## ⚡ Contributing
 
-This project welcomes contributions and suggestions. Start by reading the [Contributing Guide](docs/community/contributing.md) for 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.
+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

@@ -1,6 +1,6 @@
 # Copyright (c) Microsoft. All rights reserved.
 
-__version__ = "0.2.2"
+__version__ = "0.3.0"
 
 from .adapter import *
 from .algorithm import *
@@ -10,7 +10,9 @@ from .emitter 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

@@ -6,7 +6,7 @@ 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
@@ -670,7 +670,7 @@ 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:
@@ -800,7 +800,7 @@ class LlmProxyTraceToTriplet(TraceToTripletBase):
         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/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/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
 
@@ -18,6 +18,7 @@ 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",
@@ -25,16 +26,66 @@ def main(argv: Iterable[str] | None = None) -> int:
         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()
+    if args.backend == "memory":
+        store = InMemoryLightningStore()
+    elif args.backend == "mongo":
+        from agentlightning.store.mongo import MongoLightningStore
+
+        store = MongoLightningStore(client=args.mongo_uri)
+    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="0.0.0.0",
+        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())

agentlightning/emitter/reward.py

@@ -129,12 +129,13 @@ def reward(fn: FnType) -> FnType:
         return wrapper  # type: ignore
 
 
-def emit_reward(reward: float) -> ReadableSpan:
+def emit_reward(reward: float, auto_export: bool = True) -> ReadableSpan:
     """Emit a reward value as an OpenTelemetry span.
 
     Args:
         reward: Numeric reward to record. Integers and booleans are converted to
             floating point numbers for consistency.
+        auto_export: Whether to export the span automatically.
 
     Returns:
         Readable span capturing the recorded reward.
@@ -150,7 +151,7 @@ def emit_reward(reward: float) -> ReadableSpan:
         raise ValueError(f"Reward must be a number, got: {type(reward)}")
 
     # TODO: This should use the tracer from current context by tracer
-    tracer = get_tracer()
+    tracer = get_tracer(use_active_span_processor=auto_export)
     span = tracer.start_span(SpanNames.REWARD.value, attributes={"reward": reward})
     # Do nothing; it's just a number
     with span:

agentlightning/emitter/utils.py

@@ -2,13 +2,22 @@
 
 """Utilities shared across emitter implementations."""
 
+from typing import cast
+from warnings import filterwarnings
+
 import opentelemetry.trace as trace_api
+from opentelemetry.sdk.trace import SpanLimits, SynchronousMultiSpanProcessor, Tracer
+from opentelemetry.sdk.trace import TracerProvider as TracerProviderImpl
+from opentelemetry.sdk.util.instrumentation import InstrumentationInfo, InstrumentationScope
 from opentelemetry.trace import get_tracer_provider
 
 
-def get_tracer() -> trace_api.Tracer:
+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.
 
@@ -18,5 +27,31 @@ def get_tracer() -> trace_api.Tracer:
     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")
+    tracer_provider = cast(TracerProviderImpl, get_tracer_provider())
+
+    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",
+                "",
+                "",
+                {},
+            ),
+        )

agentlightning/execution/client_server.py

@@ -67,10 +67,11 @@ 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.
 
@@ -94,6 +95,9 @@ 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")
@@ -133,6 +137,7 @@ class ClientServerExecutionStrategy(ExecutionStrategy):
                 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.allowed_exit_codes = tuple(allowed_exit_codes)
 
     async def _execute_algorithm(
         self, algorithm: AlgorithmBundle, store: LightningStore, stop_evt: ExecutionEvent
@@ -338,10 +343,10 @@ 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/instrumentation/agentops.py

@@ -13,7 +13,8 @@ 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 opentelemetry.sdk.trace.export import SpanExportResult
+
+from agentlightning.utils.otlp import LightningStoreOTLPExporter
 
 logger = logging.getLogger(__name__)
 
@@ -32,25 +33,27 @@ def enable_agentops_service(enabled: bool = True) -> None:
     """
     Enable or disable communication with the AgentOps service.
 
-    False (default): AgentOps exporters and clients will run in local mode
-    and will not attempt to communicate with the remote AgentOps service.
-    True: all exporters and clients will operate in normal mode and send data
-    to the AgentOps service as expected.
+    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"Switch set to {enabled} for exporters and clients.")
+    logger.info(f"AgentOps service enabled is set to {enabled}.")
 
 
 def _patch_exporters():
     import agentops.client.api
     import agentops.sdk.core
-    import opentelemetry.exporter.otlp.proto.http.metric_exporter
-    import opentelemetry.exporter.otlp.proto.http.trace_exporter
 
     agentops.sdk.core.AuthenticatedOTLPExporter = BypassableAuthenticatedOTLPExporter  # type: ignore
-    opentelemetry.exporter.otlp.proto.http.metric_exporter.OTLPMetricExporter = BypassableOTLPMetricExporter
-    opentelemetry.exporter.otlp.proto.http.trace_exporter.OTLPSpanExporter = BypassableOTLPSpanExporter
+    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
 
@@ -58,12 +61,11 @@ def _patch_exporters():
 def _unpatch_exporters():
     import agentops.client.api
     import agentops.sdk.core
-    import opentelemetry.exporter.otlp.proto.http.metric_exporter
-    import opentelemetry.exporter.otlp.proto.http.trace_exporter
 
     agentops.sdk.core.AuthenticatedOTLPExporter = AuthenticatedOTLPExporter  # type: ignore
-    opentelemetry.exporter.otlp.proto.http.metric_exporter.OTLPMetricExporter = OTLPMetricExporter
-    opentelemetry.exporter.otlp.proto.http.trace_exporter.OTLPSpanExporter = OTLPSpanExporter
+    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
 
@@ -243,18 +245,15 @@ def uninstrument_agentops():
         pass
 
 
-class BypassableAuthenticatedOTLPExporter(AuthenticatedOTLPExporter):
+class BypassableAuthenticatedOTLPExporter(LightningStoreOTLPExporter, AuthenticatedOTLPExporter):
     """
     AuthenticatedOTLPExporter with switchable service control.
+
     When `_agentops_service_enabled` is False, skip export and return success.
     """
 
-    def export(self, *args: Any, **kwargs: Any) -> SpanExportResult:
-        if _agentops_service_enabled:
-            return super().export(*args, **kwargs)
-        else:
-            logger.debug("SwitchableAuthenticatedOTLPExporter is switched off, skipping export.")
-            return SpanExportResult.SUCCESS
+    def should_bypass(self) -> bool:
+        return not _agentops_service_enabled
 
 
 class BypassableOTLPMetricExporter(OTLPMetricExporter):
@@ -271,18 +270,16 @@ class BypassableOTLPMetricExporter(OTLPMetricExporter):
             return MetricExportResult.SUCCESS
 
 
-class BypassableOTLPSpanExporter(OTLPSpanExporter):
+class BypassableOTLPSpanExporter(LightningStoreOTLPExporter):
     """
     OTLPSpanExporter with switchable service control.
     When `_agentops_service_enabled` is False, skip export and return success.
+
+    This is used instead of BypassableAuthenticatedOTLPExporter on legacy AgentOps versions.
     """
 
-    def export(self, *args: Any, **kwargs: Any) -> SpanExportResult:
-        if _agentops_service_enabled:
-            return super().export(*args, **kwargs)
-        else:
-            logger.debug("SwitchableOTLPSpanExporter is switched off, skipping export.")
-            return SpanExportResult.SUCCESS
+    def should_bypass(self) -> bool:
+        return not _agentops_service_enabled
 
 
 class BypassableV3Client(V3Client):

agentlightning/llm_proxy.py

@@ -4,12 +4,15 @@ from __future__ import annotations
 
 import ast
 import asyncio
+import json
 import logging
 import os
 import re
 import tempfile
 import threading
+import time
 from contextlib import asynccontextmanager
+from datetime import datetime
 from typing import (
     Any,
     AsyncGenerator,
@@ -18,8 +21,11 @@ from typing import (
     Dict,
     Iterable,
     List,
+    Literal,
     Optional,
     Sequence,
+    Tuple,
+    Type,
     TypedDict,
     Union,
     cast,
@@ -29,14 +35,19 @@ import litellm
 import opentelemetry.trace as trace_api
 import yaml
 from fastapi import Request, Response
+from fastapi.responses import StreamingResponse
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.integrations.opentelemetry import OpenTelemetry, OpenTelemetryConfig
 from litellm.proxy.proxy_server import app, save_worker_config  # pyright: ignore[reportUnknownVariableType]
+from litellm.types.utils import CallTypes
+from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+from opentelemetry.sdk.resources import Resource
 from opentelemetry.sdk.trace import ReadableSpan
 from opentelemetry.sdk.trace.export import SpanExporter, SpanExportResult
 from starlette.middleware.base import BaseHTTPMiddleware
+from starlette.types import Scope
 
-from agentlightning.types import LLM, ProxyLLM
+from agentlightning.types import LLM, ProxyLLM, SpanNames
 from agentlightning.utils.server_launcher import (
     LaunchMode,
     PythonServerLauncher,
@@ -183,7 +194,7 @@ class LightningSpanExporter(SpanExporter):
     def __init__(self, _store: Optional[LightningStore] = None):
         self._store: Optional[LightningStore] = _store  # this is only for testing purposes
         self._buffer: List[ReadableSpan] = []
-        self._lock: Optional[threading.RLock] = None
+        self._lock: Optional[threading.Lock] = None
         self._loop_lock_pid: Optional[int] = None
 
         # Single dedicated event loop running in a daemon thread.
@@ -192,6 +203,8 @@ class LightningSpanExporter(SpanExporter):
         self._loop: Optional[asyncio.AbstractEventLoop] = None
         self._loop_thread: Optional[threading.Thread] = None
 
+        self._otlp_exporter = OTLPSpanExporter()
+
     def _ensure_loop(self) -> asyncio.AbstractEventLoop:
         """Lazily initialize the event loop and thread on first use.
 
@@ -205,15 +218,15 @@ class LightningSpanExporter(SpanExporter):
             self._loop_thread.start()
         return self._loop
 
-    def _ensure_lock(self) -> threading.RLock:
+    def _ensure_lock(self) -> threading.Lock:
         """Lazily initialize the lock on first use.
 
         Returns:
-            threading.RLock: The initialized lock.
+            threading.Lock: The initialized lock.
         """
         self._clear_loop_and_lock()
         if self._lock is None:
-            self._lock = threading.RLock()
+            self._lock = threading.Lock()
         return self._lock
 
     def _clear_loop_and_lock(self) -> None:
@@ -275,24 +288,18 @@ class LightningSpanExporter(SpanExporter):
         with self._ensure_lock():
             for span in spans:
                 self._buffer.append(span)
-
-        # Run the async flush on our private loop, synchronously from caller's POV.
-        async def _locked_flush():
-            # Take the lock inside the coroutine to serialize with other flushes.
-            with self._ensure_lock():
-                return await self._maybe_flush()
-
-        try:
-            loop = self._ensure_loop()
-            fut = asyncio.run_coroutine_threadsafe(_locked_flush(), loop)
-            fut.result()  # Bubble up any exceptions from the coroutine.
-        except Exception as e:
-            logger.exception("Export flush failed: %s", e)
-            return SpanExportResult.FAILURE
+            default_endpoint = self._otlp_exporter._endpoint  # pyright: ignore[reportPrivateUsage]
+            try:
+                self._maybe_flush()
+            except Exception as e:
+                logger.exception("Export flush failed: %s", e)
+                return SpanExportResult.FAILURE
+            finally:
+                self._otlp_exporter._endpoint = default_endpoint  # pyright: ignore[reportPrivateUsage]
 
         return SpanExportResult.SUCCESS
 
-    async def _maybe_flush(self):
+    def _maybe_flush(self):
         """Flush ready subtrees from the buffer.
 
         Strategy:
@@ -314,11 +321,20 @@ class LightningSpanExporter(SpanExporter):
             if not subtree_spans:
                 continue
 
+            # Store is initialized lazily here in most cases.
             store = self._store or get_active_llm_proxy().get_store()
             if store is None:
                 logger.warning("Store is not set in LLMProxy. Cannot log spans to store.")
                 continue
 
+            # If the store supports OTLP endpoint, use it.
+            if store.capabilities.get("otlp_traces", False):
+                otlp_traces_endpoint = store.otlp_traces_endpoint()
+                self._otlp_exporter._endpoint = otlp_traces_endpoint  # pyright: ignore[reportPrivateUsage]
+                otlp_enabled = True
+            else:
+                otlp_enabled = False
+
             # Merge all custom headers found in the subtree.
             headers_merged: Dict[str, Any] = {}
 
@@ -352,7 +368,9 @@ class LightningSpanExporter(SpanExporter):
                 headers_merged.update(cast(Dict[str, Any], headers))
 
             if not headers_merged:
-                logger.warning(f"No headers found in {len(subtree_spans)} subtree spans. Cannot log to store.")
+                logger.warning(
+                    f"No headers found in {len(subtree_spans)} subtree spans of root {root_span_id}. Cannot log to store."
+                )
                 continue
 
             # Validate and normalize required header fields.
@@ -372,10 +390,34 @@ class LightningSpanExporter(SpanExporter):
             sequence_id_decimal = int(sequence_id)
 
             # Persist each span in the subtree with the resolved identifiers.
-            for span in subtree_spans:
-                await store.add_otel_span(
-                    rollout_id=rollout_id, attempt_id=attempt_id, sequence_id=sequence_id_decimal, readable_span=span
-                )
+            if otlp_enabled:
+                # If store has OTLP support, directly use OTLP exporter and export in batch
+                for span in subtree_spans:
+                    span._resource = span._resource.merge(  # pyright: ignore[reportPrivateUsage]
+                        Resource.create(
+                            {
+                                SpanNames.ROLLOUT_ID: rollout_id,
+                                SpanNames.ATTEMPT_ID: attempt_id,
+                                SpanNames.SPAN_SEQUENCE_ID: sequence_id_decimal,
+                            }
+                        )
+                    )
+                export_result = self._otlp_exporter.export(subtree_spans)
+                if export_result != SpanExportResult.SUCCESS:
+                    raise RuntimeError(f"Failed to export spans via OTLP exporter. Result: {export_result}")
+
+            else:
+                # The old way: store does not support OTLP endpoint
+                for span in subtree_spans:
+                    loop = self._ensure_loop()
+                    add_otel_span_task = store.add_otel_span(
+                        rollout_id=rollout_id,
+                        attempt_id=attempt_id,
+                        sequence_id=sequence_id_decimal,
+                        readable_span=span,
+                    )
+                    fut = asyncio.run_coroutine_threadsafe(add_otel_span_task, loop)
+                    fut.result()  # Bubble up any exceptions from the coroutine.
 
     def _get_root_span_ids(self) -> Iterable[int]:
         """Yield span_ids for root spans currently in the buffer.
@@ -454,6 +496,23 @@ class LightningOpenTelemetry(OpenTelemetry):
 
         super().__init__(config=config)  # pyright: ignore[reportUnknownMemberType]
 
+    async def async_pre_call_deployment_hook(
+        self, kwargs: Dict[str, Any], call_type: Optional[CallTypes] = None
+    ) -> Optional[Dict[str, Any]]:
+        """The root span is sometimes missing (e.g., when Anthropic endpoint is used).
+        It is created in an auth module in LiteLLM. If it's missing, we create it here.
+        """
+        if "metadata" not in kwargs or "litellm_parent_otel_span" not in kwargs["metadata"]:
+            parent_otel_span = self.create_litellm_proxy_request_started_span(  # type: ignore
+                start_time=datetime.now(),
+                headers=kwargs.get("headers", {}),
+            )
+            updated_metadata = {**kwargs.get("metadata", {}), "litellm_parent_otel_span": parent_otel_span}
+
+            return {**kwargs, "metadata": updated_metadata}
+        else:
+            return kwargs
+
 
 class RolloutAttemptMiddleware(BaseHTTPMiddleware):
     """
@@ -499,6 +558,432 @@ class RolloutAttemptMiddleware(BaseHTTPMiddleware):
         return response
 
 
+class MessageInspectionMiddleware(BaseHTTPMiddleware):
+    """Middleware to inspect the request and response bodies.
+
+    It's for debugging purposes. Add it via "message_inspection" middleware alias.
+    """
+
+    async def dispatch(self, request: Request, call_next: Callable[[Request], Awaitable[Response]]) -> Response:
+        ti = time.time()
+        logger.info(f"Received request with scope: {request.scope}")
+        logger.info(f"Received request with body: {await request.body()}")
+        response = await call_next(request)
+        elapsed = time.time() - ti
+        logger.info(f"Response to request took {elapsed} seconds")
+        logger.info(f"Received response with status code: {response.status_code}")
+        logger.info(f"Received response with body: {response.body}")
+        return response
+
+
+class StreamConversionMiddleware(BaseHTTPMiddleware):
+    """Middleware to convert streaming responses to non-streaming responses.
+
+    Useful for backend that only supports non-streaming responses.
+
+    LiteLLM's OpenTelemetry is also buggy with streaming responses.
+    The conversion will hopefully bypass the bug.
+    """
+
+    async def dispatch(self, request: Request, call_next: Callable[[Request], Awaitable[Response]]) -> Response:
+        # Only process POST requests to completion endpoints
+        if request.method != "POST":
+            return await call_next(request)
+
+        # Check if it's a chat completions or messages endpoint
+        endpoint_format: Literal["openai", "anthropic", "unknown"] = "unknown"
+        if request.url.path.endswith("/chat/completions") or "/chat/completions?" in request.url.path:
+            endpoint_format = "openai"
+        elif request.url.path.endswith("/messages") or "/messages?" in request.url.path:
+            endpoint_format = "anthropic"
+        else:
+            endpoint_format = "unknown"
+
+        if endpoint_format == "unknown":
+            # Directly bypass the middleware
+            return await call_next(request)
+
+        # Read the request body
+        try:
+            json_body = await request.json()
+        except json.JSONDecodeError:
+            logger.warning(f"Request body is not valid JSON: {request.body}")
+            return await call_next(request)
+
+        # Check if streaming is requested
+        is_streaming = json_body.get("stream", False)
+
+        # Simple case: no streaming requested, just return the response
+        if not is_streaming:
+            return await call_next(request)
+
+        # Now the stream case
+        return await self._handle_stream_case(request, json_body, endpoint_format, call_next)
+
+    async def _handle_stream_case(
+        self,
+        request: Request,
+        json_body: Dict[str, Any],
+        endpoint_format: Literal["openai", "anthropic"],
+        call_next: Callable[[Request], Awaitable[Response]],
+    ) -> Response:
+        # 1) Modify the request body to force stream=False
+        modified_json = dict(json_body)
+        modified_json["stream"] = False
+        modified_body = json.dumps(modified_json).encode("utf-8")
+
+        # 2) Build a new scope + receive that yields our modified body
+        scope: Scope = dict(request.scope)
+        # rewrite headers for accept/content-length
+        new_headers: List[Tuple[bytes, bytes]] = []
+        saw_accept = False
+        for k, v in scope["headers"]:
+            kl = k.lower()
+            if kl == b"accept":
+                saw_accept = True
+                new_headers.append((k, b"application/json"))
+            elif kl == b"content-length":
+                # replace with new length
+                continue
+            else:
+                new_headers.append((k, v))
+        if not saw_accept:
+            new_headers.append((b"accept", b"application/json"))
+        new_headers.append((b"content-length", str(len(modified_body)).encode("ascii")))
+        scope["headers"] = new_headers
+
+        # Directly modify the request body
+        # Creating a new request won't work because request is cached in the base class
+        request._body = modified_body  # type: ignore
+
+        response = await call_next(request)
+
+        buffered: Optional[bytes] = None
+        # 4) If OK, buffer the response body (it should be JSON because we forced stream=False)
+        if 200 <= response.status_code < 300:
+            try:
+                if hasattr(response, "body_iterator"):
+                    # Buffer body safely
+                    body_chunks: List[bytes] = []
+                    async for chunk in response.body_iterator:  # type: ignore
+                        body_chunks.append(chunk)  # type: ignore
+                    buffered = b"".join(body_chunks)
+                else:
+                    buffered = response.body  # type: ignore
+
+                data = json.loads(buffered or b"{}")
+
+                if endpoint_format == "anthropic":
+                    return StreamingResponse(
+                        self.anthropic_stream_generator(data),
+                        media_type="text/event-stream",
+                        headers={"Cache-Control": "no-cache", "Connection": "keep-alive", "X-Accel-Buffering": "no"},
+                    )
+                else:
+                    # openai format
+                    return StreamingResponse(
+                        self.openai_stream_generator(data),
+                        media_type="text/event-stream",
+                        headers={"Cache-Control": "no-cache", "Connection": "keep-alive", "X-Accel-Buffering": "no"},
+                    )
+            except Exception as e:
+                # If anything goes wrong, fall back to non-streaming JSON
+                logger.exception(f"Error converting to stream; returning non-stream response: {e}")
+                # Rebuild the consumed response
+                return Response(
+                    content=buffered if buffered is not None else b"",
+                    status_code=response.status_code,
+                    headers=dict(response.headers),
+                    media_type=response.media_type,
+                    background=response.background,
+                )
+        else:
+            return response
+
+    async def anthropic_stream_generator(self, original_response: Dict[str, Any]):
+        """Generate Anthropic SSE-formatted chunks from complete content blocks
+
+        This is a dirty hack for Anthropic-style streaming from non-streaming response.
+        The sse format is subject to change based on Anthropic's implementation.
+        If so, try to use `MessageInspectionMiddleware` to inspect the update and fix accordingly.
+        """
+        # Anthropic format - handle multiple content blocks (text + tool_use)
+        content_blocks: List[Dict[str, Any]] = original_response.get("content", [])
+        message_id = original_response.get("id", f"msg_{int(time.time() * 1000)}")
+        model = original_response.get("model", "claude")
+
+        # Send message_start event
+        message_start: Dict[str, Any] = {
+            "type": "message_start",
+            "message": {
+                "id": message_id,
+                "type": "message",
+                "role": "assistant",
+                "content": [],
+                "model": model,
+                "stop_reason": None,
+                "stop_sequence": None,
+                "usage": original_response.get("usage", {"input_tokens": 0, "output_tokens": 0}),
+            },
+        }
+        yield f"event: message_start\ndata: {json.dumps(message_start)}\n\n"
+
+        # Send ping to keep connection alive
+        ping = {"type": "ping"}
+        yield f"event: ping\ndata: {json.dumps(ping)}\n\n"
+
+        # Process each content block
+        for block_index, block in enumerate(content_blocks):
+            block_type = block.get("type", "text")
+
+            if block_type == "text":
+                # Handle text block
+                content = block.get("text", "")
+
+                # Send content_block_start event
+                content_block_start = {
+                    "type": "content_block_start",
+                    "index": block_index,
+                    "content_block": {"type": "text", "text": ""},
+                }
+                yield f"event: content_block_start\ndata: {json.dumps(content_block_start)}\n\n"
+
+                # Stream text content in chunks
+                if content:
+                    words = content.split()
+                    chunk_size = 5
+
+                    for i in range(0, len(words), chunk_size):
+                        chunk_words = words[i : i + chunk_size]
+                        text_chunk = " ".join(chunk_words)
+
+                        # Add space after chunk unless it's the last one
+                        if i + chunk_size < len(words):
+                            text_chunk += " "
+
+                        content_block_delta = {
+                            "type": "content_block_delta",
+                            "index": block_index,
+                            "delta": {"type": "text_delta", "text": text_chunk},
+                        }
+                        yield f"event: content_block_delta\ndata: {json.dumps(content_block_delta)}\n\n"
+                        await asyncio.sleep(0.02)
+
+                # Send content_block_stop event
+                content_block_stop = {"type": "content_block_stop", "index": block_index}
+                yield f"event: content_block_stop\ndata: {json.dumps(content_block_stop)}\n\n"
+
+            elif block_type == "tool_use":
+                # Handle tool_use block
+                tool_name = block.get("name", "")
+                tool_input = block.get("input", {})
+                tool_id = block.get("id", f"toolu_{int(time.time() * 1000)}")
+
+                # Send content_block_start event for tool use
+                content_block_start: Dict[str, Any] = {
+                    "type": "content_block_start",
+                    "index": block_index,
+                    "content_block": {"type": "tool_use", "id": tool_id, "name": tool_name, "input": {}},
+                }
+                yield f"event: content_block_start\ndata: {json.dumps(content_block_start)}\n\n"
+
+                # Stream tool input as JSON string chunks
+                input_json = json.dumps(tool_input)
+                chunk_size = 20  # characters per chunk for JSON
+
+                for i in range(0, len(input_json), chunk_size):
+                    json_chunk = input_json[i : i + chunk_size]
+
+                    content_block_delta = {
+                        "type": "content_block_delta",
+                        "index": block_index,
+                        "delta": {"type": "input_json_delta", "partial_json": json_chunk},
+                    }
+                    yield f"event: content_block_delta\ndata: {json.dumps(content_block_delta)}\n\n"
+                    await asyncio.sleep(0.01)
+
+                # Send content_block_stop event
+                content_block_stop = {"type": "content_block_stop", "index": block_index}
+                yield f"event: content_block_stop\ndata: {json.dumps(content_block_stop)}\n\n"
+
+        # Send message_delta event with stop reason
+        message_delta = {
+            "type": "message_delta",
+            "delta": {"stop_reason": original_response.get("stop_reason", "end_turn"), "stop_sequence": None},
+            "usage": {"output_tokens": original_response.get("usage", {}).get("output_tokens", 0)},
+        }
+        yield f"event: message_delta\ndata: {json.dumps(message_delta)}\n\n"
+
+        # Send message_stop event
+        message_stop = {"type": "message_stop"}
+        yield f"event: message_stop\ndata: {json.dumps(message_stop)}\n\n"
+
+    async def openai_stream_generator(self, response_json: Dict[str, Any]) -> AsyncGenerator[str, Any]:
+        """
+        Convert a *complete* OpenAI chat.completions choice into a stream of
+        OpenAI-compatible SSE chunks.
+
+        This emits:
+
+          - an initial delta with the role ("assistant"),
+          - a sequence of deltas for message.content (split into small chunks),
+          - deltas for any tool_calls (including id/name and chunked arguments),
+          - a terminal chunk with finish_reason,
+          - and finally the literal '[DONE]'.
+
+        Notes:
+
+        - We only handle a *single* choice (index 0 typically).
+        - We purposefully don't attempt to stream logprobs.
+        - Chunking strategy is simple and conservative to avoid splitting
+          multi-byte characters: we slice on spaces where possible, then fall
+          back to fixed-size substrings.
+        """
+        choice = cast(Dict[str, Any], (response_json.get("choices") or [{}])[0])
+        model = response_json.get("model", "unknown")
+        created: int = int(time.time())
+        index: int = choice.get("index", 0)
+
+        message: Dict[str, Any] = choice.get("message", {}) or {}
+        role: str = message.get("role", "assistant")
+        content: str = message.get("content") or ""
+        tool_calls: List[Any] = message.get("tool_calls") or []
+        finish_reason: Optional[str] = choice.get(
+            "finish_reason"
+        )  # e.g., "stop", "length", "tool_calls", "content_filter"
+
+        def sse_chunk(obj: Dict[str, Any]) -> str:
+            return f"data: {json.dumps(obj, ensure_ascii=False)}\n\n"
+
+        # 1) initial chunk with the role
+        yield sse_chunk(
+            {
+                "id": f"chatcmpl-{created}",
+                "object": "chat.completion.chunk",
+                "created": created,
+                "model": model,
+                "choices": [{"index": index, "delta": {"role": role}, "finish_reason": None}],
+            }
+        )
+
+        # 2) stream textual content as small deltas
+        async def stream_content(text: str):
+            if not text:
+                return
+            # prefer splitting on spaces in ~20–40 char pieces
+            approx = 28
+            start = 0
+            n = len(text)
+            while start < n:
+                end = min(start + approx, n)
+                if end < n:
+                    # try to break on a space going forward
+                    space = text.rfind(" ", start, end)
+                    if space > start:
+                        end = space + 1
+                delta_text = text[start:end]
+                start = end
+                if not delta_text:
+                    break
+                yield sse_chunk(
+                    {
+                        "id": f"chatcmpl-{created}",
+                        "object": "chat.completion.chunk",
+                        "created": created,
+                        "model": model,
+                        "choices": [{"index": index, "delta": {"content": delta_text}, "finish_reason": None}],
+                    }
+                )
+                # tiny pause helps some UIs animate smoothly; keep very small
+                await asyncio.sleep(0.0)
+
+        async for piece in stream_content(content):  # type: ignore[misc]
+            yield piece  # pass through the produced chunks
+
+        # 3) stream tool_calls if present (id/name first, then arguments piecemeal)
+        for tc_index, tc in enumerate(tool_calls):
+            tc_type = tc.get("type", "function")
+            tc_id = tc.get("id") or f"call_{created}_{tc_index}"
+            fn: Dict[str, Any] = (tc.get("function") or {}) if tc_type == "function" else {}
+            fn_name: str = fn.get("name", "")
+            fn_args: str = fn.get("arguments", "") or ""
+
+            # (a) delta that announces the tool call id/type/name
+            yield sse_chunk(
+                {
+                    "id": f"chatcmpl-{created}",
+                    "object": "chat.completion.chunk",
+                    "created": created,
+                    "model": model,
+                    "choices": [
+                        {
+                            "index": index,
+                            "delta": {
+                                "tool_calls": [
+                                    {"index": tc_index, "id": tc_id, "type": tc_type, "function": {"name": fn_name}}
+                                ]
+                            },
+                            "finish_reason": None,
+                        }
+                    ],
+                }
+            )
+
+            # (b) stream arguments in small substrings
+            arg_chunk_size = 40
+            for pos in range(0, len(fn_args), arg_chunk_size):
+                partial = fn_args[pos : pos + arg_chunk_size]
+                yield sse_chunk(
+                    {
+                        "id": f"chatcmpl-{created}",
+                        "object": "chat.completion.chunk",
+                        "created": created,
+                        "model": model,
+                        "choices": [
+                            {
+                                "index": index,
+                                "delta": {"tool_calls": [{"index": tc_index, "function": {"arguments": partial}}]},
+                                "finish_reason": None,
+                            }
+                        ],
+                    }
+                )
+                await asyncio.sleep(0.0)
+
+        # 4) terminal chunk with finish_reason (default to "stop" if missing)
+        yield sse_chunk(
+            {
+                "id": f"chatcmpl-{created}",
+                "object": "chat.completion.chunk",
+                "created": created,
+                "model": model,
+                "choices": [
+                    {
+                        "index": index,
+                        "delta": {},
+                        "finish_reason": finish_reason or ("tool_calls" if tool_calls else "stop"),
+                    }
+                ],
+            }
+        )
+
+        # 5) literal DONE sentinel
+        yield "data: [DONE]\n\n"
+
+
+_MIDDLEWARE_REGISTRY: Dict[str, Type[BaseHTTPMiddleware]] = {
+    "rollout_attempt": RolloutAttemptMiddleware,
+    "stream_conversion": StreamConversionMiddleware,
+    "message_inspection": MessageInspectionMiddleware,
+}
+
+
+_CALLBACK_REGISTRY = {
+    "return_token_ids": AddReturnTokenIds,
+    "opentelemetry": LightningOpenTelemetry,
+}
+
+
 class LLMProxy:
     """Host a LiteLLM OpenAI-compatible proxy bound to a LightningStore.
 
@@ -522,7 +1007,10 @@ class LLMProxy:
 
     !!! warning
 
-        The LLM Proxy does support streaming, but the tracing is still problematic when streaming is enabled.
+        By default (or when "stream_conversion" middleware is enabled), the LLM Proxy will convert OpenAI and Anthropic requests with `stream=True`
+        to a non-streaming request before going through the LiteLLM proxy. This is because the OpenTelemetry tracer provided by
+        LiteLLM is buggy with streaming responses. You can disable this by removing the "stream_conversion" middleware.
+        In that case, you might lose some tracing information like token IDs.
 
     !!! danger
 
@@ -544,6 +1032,13 @@ class LLMProxy:
             `launch_mode="asyncio"` launches the server in the current thread as an asyncio task.
             It is NOT recommended because it often causes hanging requests. Only use it if you know what you are doing.
         launcher_args: Arguments for the server launcher. If this is provided, host, port, and launch_mode will be ignored. Cannot be used together with port, host, and launch_mode.
+        middlewares: List of FastAPI middleware classes or strings to register. You can specify the class aliases or classes that have been imported.
+            If not provided, the default middlewares (RolloutAttemptMiddleware and StreamConversionMiddleware) will be used.
+            Available middleware aliases are: "rollout_attempt", "stream_conversion", "message_inspection".
+            Middlewares are the **first layer** of request processing. They are applied to all requests before the LiteLLM proxy.
+        callbacks: List of LiteLLM callback classes or strings to register. You can specify the class aliases or classes that have been imported.
+            If not provided, the default callbacks (AddReturnTokenIds and LightningOpenTelemetry) will be used.
+            Available callback aliases are: "return_token_ids", "opentelemetry".
     """
 
     def __init__(
@@ -557,7 +1052,8 @@ class LLMProxy:
         num_workers: int = 1,
         launch_mode: LaunchMode = "mp",
         launcher_args: PythonServerLauncherArgs | None = None,
-        _add_return_token_ids: bool = True,
+        middlewares: List[Union[Type[BaseHTTPMiddleware], str]] | None = None,
+        callbacks: List[Union[Type[CustomLogger], str]] | None = None,
     ):
         self.store = store
 
@@ -589,7 +1085,33 @@ class LLMProxy:
 
         self._config_file = None
 
-        self._add_return_token_ids = _add_return_token_ids
+        self.middlewares: List[Type[BaseHTTPMiddleware]] = []
+        if middlewares is None:
+            middlewares = ["rollout_attempt", "stream_conversion"]
+        for middleware in middlewares:
+            if isinstance(middleware, str):
+                if middleware not in _MIDDLEWARE_REGISTRY:
+                    raise ValueError(
+                        f"Invalid middleware alias: {middleware}. Available aliases are: {list(_MIDDLEWARE_REGISTRY.keys())}"
+                    )
+                middleware = _MIDDLEWARE_REGISTRY[middleware]
+                self.middlewares.append(middleware)
+            else:
+                self.middlewares.append(middleware)
+
+        self.callbacks: List[Type[CustomLogger]] = []
+        if callbacks is None:
+            callbacks = ["return_token_ids", "opentelemetry"]
+        for callback in callbacks:
+            if isinstance(callback, str):
+                if callback not in _CALLBACK_REGISTRY:
+                    raise ValueError(
+                        f"Invalid callback alias: {callback}. Available aliases are: {list(_CALLBACK_REGISTRY.keys())}"
+                    )
+                callback = _CALLBACK_REGISTRY[callback]
+                self.callbacks.append(callback)
+            else:
+                self.callbacks.append(callback)
 
     def get_store(self) -> Optional[LightningStore]:
         """Get the store used by the proxy.
@@ -641,21 +1163,18 @@ class LLMProxy:
         set_active_llm_proxy(self)
 
         # Install middleware if it's not already installed.
-        installed: bool = False
+        installation_status: Dict[Any, bool] = {}
         for mw in app.user_middleware:
-            if mw.cls is RolloutAttemptMiddleware:
-                # Check whether the middleware is installed.
-                # It could be installed by other LLM Proxy instances, but it doesn't matter.
-                logger.info("Found existing RolloutAttemptMiddleware installed. Will not install a new one.")
-                installed = True
-                break
+            installation_status[mw.cls] = True
 
-        if not installed:
-            # Fallback to adding a new middleware
-            logger.info("Adding a new middleware to the FastAPI app.")
-            app.add_middleware(RolloutAttemptMiddleware)
+        for mw in self.middlewares:
+            if mw not in installation_status:
+                logger.info(f"Adding middleware {mw} to the FastAPI app.")
+                app.add_middleware(mw)
+            else:
+                logger.info(f"Middleware {mw} is already installed. Will not install a new one.")
 
-        if not initialize_llm_callbacks(self._add_return_token_ids):
+        if not initialize_llm_callbacks(self.callbacks):
             # If it's not the first time to initialize the callbacks, also
             # reset LiteLLM's logging worker so its asyncio.Queue binds to the new loop.
             _reset_litellm_logging_worker()
@@ -720,16 +1239,16 @@ class LLMProxy:
         if self.store is None:
             raise ValueError("Store is not set. Please set the store before starting the LLMProxy.")
 
-        store_capabilities = self.store.capabilities()
-        if self.server_launcher.args.launch_mode == "mp" and not store_capabilities["zero_copy"]:
+        store_capabilities = self.store.capabilities
+        if self.server_launcher.args.launch_mode == "mp" and not store_capabilities.get("zero_copy", False):
             raise RuntimeError(
                 "The store does not support zero-copy. Please use another store, or use asyncio or thread mode to launch the server."
             )
-        elif self.server_launcher.args.launch_mode == "thread" and not store_capabilities["thread_safe"]:
+        elif self.server_launcher.args.launch_mode == "thread" and not store_capabilities.get("thread_safe", False):
             raise RuntimeError(
                 "The store is not thread-safe. Please use another store, or use asyncio mode to launch the server."
             )
-        elif self.server_launcher.args.launch_mode == "asyncio" and not store_capabilities["async_safe"]:
+        elif self.server_launcher.args.launch_mode == "asyncio" and not store_capabilities.get("async_safe", False):
             raise RuntimeError("The store is not async-safe. Please use another store.")
 
         logger.info(
@@ -849,7 +1368,7 @@ def set_active_llm_proxy(proxy: LLMProxy) -> None:
     _global_llm_proxy = proxy
 
 
-def initialize_llm_callbacks(_add_return_token_ids: bool = True) -> bool:
+def initialize_llm_callbacks(callback_classes: List[Type[CustomLogger]]) -> bool:
     """Restore `litellm.callbacks` to a state that is just initialized by agent-lightning.
 
     When litellm is restarted multiple times in the same process, more and more callbacks
@@ -857,8 +1376,7 @@ def initialize_llm_callbacks(_add_return_token_ids: bool = True) -> bool:
     This function remembers the initial state of `litellm.callbacks` and always restore to that state.
 
     Args:
-        _add_return_token_ids: Whether to add the return token ids callback. Internal use only.
-        Ideally the callback should automatically be enabled when the backend supports it.
+        callback_classes: List of callback classes to register.
 
     Returns:
         Whether the callbacks are initialized for the first time.
@@ -866,31 +1384,31 @@ def initialize_llm_callbacks(_add_return_token_ids: bool = True) -> bool:
     global _callbacks_before_litellm_start
 
     if _callbacks_before_litellm_start is None:
-        litellm.callbacks.extend(  # type: ignore
-            [
-                AddReturnTokenIds(),
-                LightningOpenTelemetry(),
-            ]
-            if _add_return_token_ids
-            else [
-                LightningOpenTelemetry(),
-            ]
-        )
+        litellm.callbacks.extend([cls() for cls in callback_classes])  # type: ignore
         _callbacks_before_litellm_start = [*litellm.callbacks]  # type: ignore
         return True
 
+    else:
+        # Put whatever is missing in the new callback classes to the existing callbacks.
+        for cls in callback_classes:
+            if not any(isinstance(cb, cls) for cb in _callbacks_before_litellm_start):
+                logger.info(f"Adding missing callback {cls} to the existing callbacks.")
+                _callbacks_before_litellm_start.append(cls())
+
     _reset_litellm_logging_callback_manager()
 
-    # Check if tracer provider is malformed due to global tracer clear in tests.
-    if not _check_tracer_provider():
-        logger.warning(
-            "Global tracer provider might have been cleared outside. Re-initializing OpenTelemetry callback."
-        )
-        _callbacks_before_litellm_start = [
-            cb for cb in _callbacks_before_litellm_start if not isinstance(cb, LightningOpenTelemetry)
-        ] + [LightningOpenTelemetry()]
-    else:
-        logger.debug("Global tracer provider is valid. Reusing existing OpenTelemetry callback.")
+    if LightningOpenTelemetry in callback_classes:
+        # Check if tracer provider is malformed due to global tracer clear in tests.
+        if not _check_tracer_provider():
+            logger.warning(
+                "Global tracer provider might have been cleared outside. Re-initializing OpenTelemetry callback."
+            )
+            _callbacks_before_litellm_start = [
+                cb for cb in _callbacks_before_litellm_start if not isinstance(cb, LightningOpenTelemetry)
+            ] + [LightningOpenTelemetry()]
+        else:
+            logger.debug("Global tracer provider is valid. Reusing existing OpenTelemetry callback.")
+    # Otherwise, we just skip the check for opentelemetry and use the existing callback.
 
     litellm.callbacks.clear()  # type: ignore
     litellm.callbacks.extend(_callbacks_before_litellm_start)  # type: ignore

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, auto_export=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/store/__init__.py

@@ -2,6 +2,7 @@
 
 from .base import LightningStore, LightningStoreCapabilities
 from .client_server import LightningStoreClient, LightningStoreServer
+from .collection_based import CollectionBasedLightningStore
 from .memory import InMemoryLightningStore
 from .threading import LightningStoreThreaded
 
@@ -11,5 +12,6 @@ __all__ = [
     "LightningStoreClient",
     "LightningStoreServer",
     "InMemoryLightningStore",
+    "CollectionBasedLightningStore",
     "LightningStoreThreaded",
 ]

agentlightning/store/base.py

@@ -17,6 +17,8 @@ from agentlightning.types import (
     RolloutStatus,
     Span,
     TaskInput,
+    Worker,
+    WorkerStatus,
 )
 
 
@@ -52,8 +54,11 @@ UNSET = _UnsetType()
 Unset = _UnsetType  # Alias for convenience
 
 
-class LightningStoreCapabilities(TypedDict):
-    """Capability of a LightningStore implementation."""
+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."""
@@ -61,6 +66,8 @@ class LightningStoreCapabilities(TypedDict):
     """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 LightningStore:
@@ -85,14 +92,32 @@ 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,
         )
 
+    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,
@@ -167,7 +192,7 @@ class LightningStore:
         """
         raise NotImplementedError()
 
-    async def dequeue_rollout(self) -> Optional[AttemptedRollout]:
+    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.
@@ -180,6 +205,8 @@ 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.
@@ -266,30 +293,77 @@ 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.
@@ -326,11 +400,35 @@ class LightningStore:
         """
         raise NotImplementedError()
 
-    async def query_resources(self) -> List[ResourcesUpdate]:
+    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:
-            A chronological list of [`ResourcesUpdate`][agentlightning.ResourcesUpdate] objects.
+            [`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.
@@ -413,19 +511,61 @@ 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.
@@ -527,6 +667,12 @@ class LightningStore:
         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.
@@ -543,3 +689,64 @@ 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 +1,18 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from .base import Collection, FilterOptions, KeyValue, LightningCollections, PaginatedResult, Queue, SortOptions
+from .memory import DequeBasedQueue, DictBasedKeyValue, InMemoryLightningCollections, ListBasedCollection
+
+__all__ = [
+    "Collection",
+    "Queue",
+    "KeyValue",
+    "FilterOptions",
+    "SortOptions",
+    "PaginatedResult",
+    "LightningCollections",
+    "ListBasedCollection",
+    "DequeBasedQueue",
+    "DictBasedKeyValue",
+    "InMemoryLightningCollections",
+]

agentlightning/store/collection/base.py

@@ -0,0 +1,356 @@
+# 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")
+
+
+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]) -> None:
+        """Update the given items in the collection.
+
+        Raises:
+            ValueError: If an item with the primary keys does not exist.
+        """
+        raise NotImplementedError()
+
+    async def upsert(self, items: Sequence[T]) -> None:
+        """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.
+        """
+        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, *args: Any, **kwargs: Any) -> AsyncContextManager[Self]:
+        """Perform a atomic operation on the collections.
+
+        Subclass may use args and kwargs to support multiple levels of atomicity.
+
+        Args:
+            *args: Arguments to pass to the operation.
+            **kwargs: Keyword arguments to pass to the operation.
+        """
+        raise NotImplementedError()
+
+    async def execute(self, callback: Callable[[Self], Awaitable[T]]) -> T:
+        """Execute the given callback within an atomic operation."""
+        async with self.atomic() 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 +1,744 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import weakref
+from collections import deque
+from contextlib import asynccontextmanager
+from typing import (
+    Any,
+    Deque,
+    Dict,
+    Iterable,
+    List,
+    Literal,
+    Mapping,
+    MutableMapping,
+    Optional,
+    Sequence,
+    Tuple,
+    Type,
+    TypeVar,
+    Union,
+)
+
+from agentlightning.types import (
+    Attempt,
+    FilterField,
+    FilterOptions,
+    PaginatedResult,
+    ResourcesUpdate,
+    Rollout,
+    SortOptions,
+    Span,
+    Worker,
+)
+
+from .base import (
+    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) -> None:
+        """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 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":
+                parent[final_key] = item
+            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.
+        """
+        for item in items:
+            self._mutate_single(item, mode="insert")
+
+    async def update(self, items: Sequence[T]) -> None:
+        """Update the given items.
+
+        Raises:
+            ValueError: If any item with the given primary keys does not exist.
+        """
+        for item in items:
+            self._mutate_single(item, mode="update")
+
+    async def upsert(self, items: Sequence[T]) -> None:
+        """Upsert the given items (insert if missing, otherwise update)."""
+        for item in items:
+            self._mutate_single(item, mode="upsert")
+
+    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):
+        self._lock = _LoopAwareAsyncLock()
+        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, *args: Any, **kwargs: Any):
+        """In-memory collections apply a lock outside. It doesn't need to manipulate the collections inside."""
+        async with self._lock:
+            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()

agentlightning/store/mongo.py

@@ -0,0 +1,82 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import hashlib
+import logging
+import uuid
+from typing import (
+    Any,
+    Callable,
+    Mapping,
+    TypeVar,
+)
+
+from pymongo import AsyncMongoClient
+
+from .base import LightningStoreCapabilities
+from .collection.mongo import MongoClientPool, MongoLightningCollections
+from .collection_based import CollectionBasedLightningStore
+
+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,
+    ) -> None:
+        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))
+
+    @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()

agentlightning/store/threading.py

@@ -18,6 +18,8 @@ from agentlightning.types import (
     RolloutStatus,
     Span,
     TaskInput,
+    Worker,
+    WorkerStatus,
 )
 
 from .base import UNSET, LightningStore, LightningStoreCapabilities, Unset
@@ -35,9 +37,10 @@ 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()
+        capabilities = self.store.capabilities
         return {
             **capabilities,
             "async_safe": True,
@@ -66,9 +69,9 @@ 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:
@@ -77,15 +80,48 @@ 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:
@@ -95,6 +131,26 @@ 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)
@@ -137,9 +193,39 @@ class LightningStoreThreaded(LightningStore):
         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,
@@ -180,3 +266,39 @@ 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/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.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
@@ -67,9 +66,8 @@ class AgentOpsTracer(Tracer):
     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)
@@ -85,16 +83,9 @@ class AgentOpsTracer(Tracer):
 
         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)
@@ -111,7 +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.
 
@@ -122,12 +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(
@@ -137,47 +134,52 @@ 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:
+                # 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:
+                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")
+
+    @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:
-            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
-            else:
-                raise ValueError("store, rollout_id, and attempt_id must be either all provided or all None")
+            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}, error={e}")
+            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_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()
-
     def get_langchain_handler(self, tags: List[str] | None = None) -> LangchainCallbackHandler:
         """
         Get the Langchain callback handler for integrating with Langchain.
@@ -204,135 +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 +2,26 @@
 
 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.store.base import LightningStore
+from agentlightning.types.tracer import SpanNames
+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__)
@@ -29,21 +39,31 @@ 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.")
 
-        tracer_provider = TracerProvider()
-        trace_api.set_tracer_provider(tracer_provider)
+        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...")
@@ -57,7 +77,7 @@ 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 +88,37 @@ 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 +130,222 @@ 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(
+                {
+                    SpanNames.ROLLOUT_ID: rollout_id,
+                    SpanNames.ATTEMPT_ID: 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(
+                {
+                    SpanNames.ROLLOUT_ID: "",
+                    SpanNames.ATTEMPT_ID: "",
+                }
+            )
+        )  # 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
+
+    @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/tracer.py

@@ -411,6 +411,12 @@ 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):

agentlightning/utils/otlp.py

@@ -0,0 +1,428 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+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.store.base import LightningStore
+from agentlightning.types.tracer import (
+    Attributes,
+    Event,
+    Link,
+    OtelResource,
+    Span,
+    SpanContext,
+    SpanNames,
+    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, store: LightningStore) -> 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(SpanNames.ROLLOUT_ID)
+        attempt_id_resource = resource_attrs.get(SpanNames.ATTEMPT_ID)
+        # 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(SpanNames.SPAN_SEQUENCE_ID)
+
+        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(SpanNames.ROLLOUT_ID)
+                attempt_id_span = span_attrs.get(SpanNames.ATTEMPT_ID)
+                sequence_id_span = span_attrs.get(SpanNames.SPAN_SEQUENCE_ID)
+
+                # 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 = await store.get_next_span_sequence_id(
+                        rollout_id=rollout_id, attempt_id=attempt_id
+                    )
+                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)
+
+    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 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(
+                        {
+                            SpanNames.ROLLOUT_ID: self._rollout_id,
+                            SpanNames.ATTEMPT_ID: 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/server_launcher.py

@@ -15,7 +15,7 @@ import traceback
 from contextlib import asynccontextmanager, suppress
 from dataclasses import dataclass
 from multiprocessing.process import BaseProcess
-from typing import Any, AsyncContextManager, AsyncIterator, Dict, Literal, Optional
+from typing import Any, AsyncContextManager, AsyncIterator, Dict, Literal, Optional, cast
 
 import aiohttp
 import requests
@@ -53,6 +53,8 @@ class PythonServerLauncherArgs:
     """
     log_level: int = logging.INFO
     """The log level to use."""
+    access_log: bool = False
+    """Whether to turn on access logs."""
     startup_timeout: float = 60.0
     """The timeout to wait for the server to start up."""
     kill_unhealthy_server: bool = True
@@ -63,6 +65,8 @@ class PythonServerLauncherArgs:
     """The timeout to wait for the thread to join."""
     process_join_timeout: float = 10.0
     """The timeout to wait for the process to join."""
+    timeout_keep_alive: int = 30
+    """The timeout to keep the connection alive."""
 
 
 @dataclass
@@ -156,7 +160,9 @@ async def run_uvicorn_asyncio(
 
         if not uvicorn_server.started:
             # Normally, the program will not reach this point, as the server will throw the exception itself earlier.
-            raise RuntimeError(f"Server did not start up within {timeout:.2f} seconds.") from server_start_exception
+            raise RuntimeError(
+                f"Server did not start up within {time.time() - start_time:.2f} seconds."
+            ) from server_start_exception
 
         logger.info(f"Server started up in {time.time() - start_time:.2f} seconds.")
 
@@ -608,6 +614,13 @@ class PythonServerLauncher:
         self._host: Optional[str] = self.args.host
         self._port: Optional[int] = self.args.port
         self._access_host: Optional[str] = self.args.access_host
+        self.initialize()
+
+    def initialize(self):
+        # ensure the host/port/access_host are set
+        self._ensure_host()
+        self._ensure_port()
+        self._ensure_access_host()
 
         # uvicorn (in-proc asyncio)
         self._uvicorn_server: Optional[uvicorn.Server] = None
@@ -626,6 +639,26 @@ class PythonServerLauncher:
         # is_running flag
         self._is_running: bool = False
 
+    def __getstate__(self):
+        """Control pickling to prevent server state from being sent to subprocesses."""
+        return {
+            "app": self.app,
+            "args": self.args,
+            "serve_context": self.serve_context,
+            "_host": self._host,
+            "_port": self._port,
+            "_access_host": self._access_host,
+        }
+
+    def __setstate__(self, state: Dict[str, Any]):
+        self.app = state["app"]
+        self.args = cast(PythonServerLauncherArgs, state["args"])
+        self.serve_context = state["serve_context"]
+        self._host = state["_host"]
+        self._port = state["_port"]
+        self._access_host = state["_access_host"]
+        self.initialize()
+
     @property
     def endpoint(self) -> str:
         """Return the externally advertised host:port pair regardless of accessibility."""
@@ -744,17 +777,18 @@ class PythonServerLauncher:
         return self._port
 
     def _ensure_access_host(self) -> str:
-        if self.args.access_host is None:
-            if self._ensure_host() in ("0.0.0.0", "::"):
-                # Probe host normalization for 0.0.0.0
-                logger.warning("No access host provided, using default outbound IPv4 address for this machine.")
-                self._access_host = _get_default_ipv4_address()
+        if self._access_host is None:
+            if self.args.access_host is None:
+                if self._ensure_host() in ("0.0.0.0", "::"):
+                    # Probe host normalization for 0.0.0.0
+                    logger.warning("No access host provided, using default outbound IPv4 address for this machine.")
+                    self._access_host = _get_default_ipv4_address()
+                else:
+                    logger.warning("No access host provided, using the host provided.")
+                    self._access_host = self._ensure_host()
             else:
-                logger.warning("No access host provided, using the host provided.")
-                self._access_host = self._ensure_host()
-        else:
-            self._access_host = self.args.access_host
-        return self._access_host
+                self._access_host = self.args.access_host
+        return self._access_host  # type: ignore
 
     def _create_uvicorn_server(self) -> uvicorn.Server:
         config = uvicorn.Config(
@@ -762,7 +796,9 @@ class PythonServerLauncher:
             host=self._ensure_host(),
             port=self._ensure_port(),
             log_level=self.args.log_level,
+            access_log=self.args.access_log,
             loop="asyncio",
+            timeout_keep_alive=self.args.timeout_keep_alive,
         )
         return uvicorn.Server(config)
 
@@ -834,17 +870,19 @@ class PythonServerLauncher:
             evt: ChildEvent = await asyncio.to_thread(self._thread_event_queue.get, True, timeout)
         except queue.Empty:
             if not self._thread.is_alive():
-                logger.error("Threaded server failed to start and is not alive. No error event was received.")
-                return
-            logger.error("Threaded server failed to start and sends no event. This should not happen.")
+                raise RuntimeError("Threaded server failed to start and is not alive. No error event was received.")
+            logger.error(
+                "Threaded server failed to start and sends no event. This should not happen. Shutting down server."
+            )
             await self._stop_uvicorn_thread()
-            return
+            raise RuntimeError("Threaded server failed to start and sends no event. This should not happen.")
 
         if evt.kind == "error":
             logger.error("Threaded server failed to start (%s): %s\n%s", evt.exc_type, evt.message, evt.traceback)
             await asyncio.to_thread(self._thread.join, self.args.thread_join_timeout)
             if self._thread.is_alive():
-                raise RuntimeError(evt.message or "Threaded server failed to start and refused to shut down.")
+                logger.error("Threaded server failed to start and refused to shut down.")
+            raise RuntimeError(evt.message)
         else:
             logger.info("Threaded server started successfully.")
             self._is_running = True
@@ -893,7 +931,7 @@ class PythonServerLauncher:
                 "workers": int(self.args.n_workers),
                 "worker_class": "uvicorn_worker.UvicornWorker",
                 "loglevel": logging.getLevelName(self.args.log_level).lower(),
-                "accesslog": None,
+                "accesslog": "-" if self.args.access_log else None,
                 "errorlog": "-",
                 "preload_app": True,
                 "graceful_timeout": int(
@@ -939,11 +977,12 @@ class PythonServerLauncher:
             evt: ChildEvent = await asyncio.to_thread(self._mp_event_queue.get, True, timeout)
         except queue.Empty:
             if not self._proc.is_alive():
-                logger.error("Server process failed to start and is not alive. No error event was received.")
-                return
-            logger.error("Server process failed to start and sends no event. This should not happen.")
+                raise RuntimeError("Server process failed to start and is not alive. No error event was received.")
+            logger.error(
+                "Server process failed to start and sends no event. This should not happen. Shutting down server."
+            )
             await self._stop_serving_process()
-            return
+            raise RuntimeError("Server process failed to start and sends no event. This should not happen.")
 
         if evt.kind == "error":
             logger.error(
@@ -955,7 +994,8 @@ class PythonServerLauncher:
             )
             await asyncio.to_thread(self._proc.join, self.args.process_join_timeout)
             if self._proc.is_alive():
-                raise RuntimeError(evt.message or "Server process failed to start and refused to shut down.")
+                logger.error("Server process failed to start and refused to shut down.")
+            raise RuntimeError(evt.message)
         else:
             logger.info("Subprocess server started successfully.")
             self._is_running = True

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",
@@ -559,14 +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})
                 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(
                     {
@@ -574,6 +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(
@@ -582,6 +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] = {}
@@ -596,6 +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)
@@ -614,6 +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)
@@ -638,9 +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:
@@ -759,6 +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/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
@@ -403,14 +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/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "agent-lightning-dashboard",
-  "version": "0.2.2",
+  "version": "0.3.0",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "agent-lightning-dashboard",
-      "version": "0.2.2",
+      "version": "0.3.0",
       "dependencies": {
         "@mantine/core": "8.3.5",
         "@mantine/hooks": "8.3.5",

dashboard/package.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-lightning-dashboard",
   "type": "module",
-  "version": "0.2.2",
+  "version": "0.3.0",
   "scripts": {
     "dev": "vite",
     "build": "tsc && vite build",

dashboard/src/Router.tsx

@@ -6,6 +6,7 @@ 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([
   {
@@ -28,6 +29,10 @@ const router = createBrowserRouter([
         path: 'traces',
         element: <TracesPage />,
       },
+      {
+        path: 'runners',
+        element: <WorkersPage />,
+      },
       {
         path: 'settings',
         element: <SettingsPage />,

dashboard/src/components/AppDrawer.component.tsx

@@ -21,7 +21,7 @@ import {
 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 } from '@/types';
+import type { Attempt, AttemptStatus, Rollout, RolloutStatus, Span, Worker } from '@/types';
 import { formatStatusLabel } from '@/utils/format';
 import { TracesTable, type TracesTableRecord } from './TracesTable.component';
 
@@ -50,6 +50,12 @@ const SPAN_STATUS_COLORS: Record<Span['status']['status_code'], string> = {
   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',
@@ -408,6 +414,60 @@ function RolloutTracesDrawerBody({ rollout, attempt, onShowRollout, onShowSpanDe
   );
 }
 
+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);
@@ -428,6 +488,13 @@ export function AppDrawerContainer() {
       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} />;

dashboard/src/components/TracesTable.component.tsx

@@ -22,6 +22,7 @@ const DEFAULT_RECORDS_PER_PAGE_OPTIONS = [50, 100, 200, 500];
 
 const COLUMN_VISIBILITY: Record<string, ColumnVisibilityConfig> = {
   name: { minWidth: 12.5, priority: 0 },
+  sequenceId: { fixedWidth: 6, priority: 1 },
   spanId: { fixedWidth: 14, priority: 1 },
   traceId: { fixedWidth: 24, priority: 3 },
   parentId: { fixedWidth: 12, priority: 2 },
@@ -86,6 +87,12 @@ function createTracesColumns({
         </Text>
       ),
     },
+    {
+      accessor: 'sequenceId',
+      title: 'Seq.',
+      sortable: true,
+      render: ({ sequenceId }) => <Text size='sm'>{sequenceId}</Text>,
+    },
     {
       accessor: 'traceId',
       title: 'Trace ID',

dashboard/src/components/WorkersTable.component.tsx

@@ -0,0 +1,362 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { useCallback, useEffect, useMemo } from 'react';
+import { IconCheck, IconCopy, IconInfoCircle, IconRefresh } from '@tabler/icons-react';
+import { DataTable, type DataTableColumn, type DataTableSortStatus } from 'mantine-datatable';
+import { ActionIcon, Badge, Box, Button, CopyButton, Group, Stack, Text, Tooltip } from '@mantine/core';
+import { useElementSize, useViewportSize } from '@mantine/hooks';
+import { getLayoutAwareWidth } from '@/layouts/helper';
+import type { Worker } from '@/types';
+import { getErrorDescriptor } from '@/utils/error';
+import { formatDateTime, formatRelativeTime, formatStatusLabel } from '@/utils/format';
+import { createResponsiveColumns, type ColumnVisibilityConfig } from '@/utils/table';
+
+const DEFAULT_RECORDS_PER_PAGE_OPTIONS = [50, 100, 200, 500];
+
+const COLUMN_VISIBILITY: Record<string, ColumnVisibilityConfig> = {
+  workerId: { fixedWidth: 12, priority: 0 },
+  status: { fixedWidth: 6, priority: 1 },
+  currentRolloutId: { fixedWidth: 14, priority: 3 },
+  currentAttemptId: { fixedWidth: 14, priority: 3 },
+  lastHeartbeatTime: { fixedWidth: 10, priority: 2 },
+  lastBusyTime: { fixedWidth: 10, priority: 3 },
+  lastIdleTime: { fixedWidth: 10, priority: 3 },
+  lastDequeueTime: { fixedWidth: 10, priority: 1 },
+  actions: { fixedWidth: 5, priority: 0 },
+};
+
+export type WorkersTableRecord = Worker & {
+  timestamps: Record<
+    'lastHeartbeatTime' | 'lastBusyTime' | 'lastIdleTime' | 'lastDequeueTime',
+    { absolute: string; relative: string }
+  >;
+};
+
+const buildTimestampMeta = (value: Worker['lastHeartbeatTime']) => ({
+  absolute: formatDateTime(value),
+  relative: formatRelativeTime(value),
+});
+
+function buildWorkerRecord(worker: Worker): WorkersTableRecord {
+  return {
+    ...worker,
+    timestamps: {
+      lastHeartbeatTime: buildTimestampMeta(worker.lastHeartbeatTime),
+      lastBusyTime: buildTimestampMeta(worker.lastBusyTime),
+      lastIdleTime: buildTimestampMeta(worker.lastIdleTime),
+      lastDequeueTime: buildTimestampMeta(worker.lastDequeueTime),
+    },
+  };
+}
+
+type WorkersColumnsOptions = {
+  onShowDetails: (worker: Worker) => void;
+};
+
+const STATUS_COLORS: Record<Worker['status'], string> = {
+  busy: 'orange',
+  idle: 'teal',
+  unknown: 'gray',
+};
+
+function createWorkersColumns({ onShowDetails }: WorkersColumnsOptions): DataTableColumn<WorkersTableRecord>[] {
+  return [
+    {
+      accessor: 'workerId',
+      title: 'Runner ID',
+      sortable: true,
+      render: ({ workerId }) => (
+        <Group gap={2} wrap='nowrap'>
+          <Text fw={500} size='sm'>
+            {workerId}
+          </Text>
+          <CopyButton value={workerId}>
+            {({ copied, copy }) => (
+              <Tooltip label={copied ? 'Copied' : 'Copy'} withArrow>
+                <ActionIcon
+                  aria-label={`Copy worker ID ${workerId}`}
+                  variant='subtle'
+                  color={copied ? 'teal' : 'gray'}
+                  size='sm'
+                  onClick={(event) => {
+                    event.stopPropagation();
+                    copy();
+                  }}
+                >
+                  {copied ? <IconCheck size={14} /> : <IconCopy size={14} />}
+                </ActionIcon>
+              </Tooltip>
+            )}
+          </CopyButton>
+        </Group>
+      ),
+    },
+    {
+      accessor: 'status',
+      title: 'Status',
+      sortable: true,
+      render: ({ status }) => {
+        const color = STATUS_COLORS[status] ?? 'gray';
+        return (
+          <Badge size='sm' variant='light' color={color} radius='sm'>
+            {formatStatusLabel(status)}
+          </Badge>
+        );
+      },
+    },
+    {
+      accessor: 'currentRolloutId',
+      title: 'Current Rollout',
+      sortable: true,
+      render: ({ currentRolloutId }) => <Text size='sm'>{currentRolloutId ?? '—'}</Text>,
+    },
+    {
+      accessor: 'currentAttemptId',
+      title: 'Current Attempt',
+      sortable: true,
+      render: ({ currentAttemptId }) => <Text size='sm'>{currentAttemptId ?? '—'}</Text>,
+    },
+    {
+      accessor: 'lastHeartbeatTime',
+      title: 'Heartbeat',
+      sortable: true,
+      render: ({ timestamps }) => (
+        <Stack gap={0} justify='center'>
+          <Text size='sm'>{timestamps.lastHeartbeatTime.relative}</Text>
+          {timestamps.lastHeartbeatTime.absolute !== '—' && (
+            <Text size='xs' c='dimmed'>
+              {timestamps.lastHeartbeatTime.absolute}
+            </Text>
+          )}
+        </Stack>
+      ),
+    },
+    {
+      accessor: 'lastBusyTime',
+      title: 'Last Busy',
+      sortable: true,
+      render: ({ timestamps }) => (
+        <Stack gap={0} justify='center'>
+          <Text size='sm'>{timestamps.lastBusyTime.relative}</Text>
+          {timestamps.lastBusyTime.absolute !== '—' && (
+            <Text size='xs' c='dimmed'>
+              {timestamps.lastBusyTime.absolute}
+            </Text>
+          )}
+        </Stack>
+      ),
+    },
+    {
+      accessor: 'lastIdleTime',
+      title: 'Last Idle',
+      sortable: true,
+      render: ({ timestamps }) => (
+        <Stack gap={0} justify='center'>
+          <Text size='sm'>{timestamps.lastIdleTime.relative}</Text>
+          {timestamps.lastIdleTime.absolute !== '—' && (
+            <Text size='xs' c='dimmed'>
+              {timestamps.lastIdleTime.absolute}
+            </Text>
+          )}
+        </Stack>
+      ),
+    },
+    {
+      accessor: 'lastDequeueTime',
+      title: 'Last Dequeue',
+      sortable: true,
+      render: ({ timestamps }) => (
+        <Stack gap={0} justify='center'>
+          <Text size='sm'>{timestamps.lastDequeueTime.relative}</Text>
+          {timestamps.lastDequeueTime.absolute !== '—' && (
+            <Text size='xs' c='dimmed'>
+              {timestamps.lastDequeueTime.absolute}
+            </Text>
+          )}
+        </Stack>
+      ),
+    },
+    {
+      accessor: 'actions',
+      title: 'Actions',
+      textAlign: 'left',
+      render: (record) => (
+        <Tooltip label='Show runner detail' withArrow disabled={!onShowDetails}>
+          <ActionIcon
+            aria-label='Show runner detail'
+            variant='subtle'
+            color='gray'
+            onClick={(event) => {
+              event.stopPropagation();
+              onShowDetails(record);
+            }}
+          >
+            <IconInfoCircle size={16} />
+          </ActionIcon>
+        </Tooltip>
+      ),
+    },
+  ];
+}
+
+export type WorkersTableProps = {
+  workers: Worker[] | undefined;
+  totalRecords: number;
+  isFetching: boolean;
+  isError: boolean;
+  error: unknown;
+  searchTerm: string;
+  sort: { column: string; direction: 'asc' | 'desc' };
+  page: number;
+  recordsPerPage: number;
+  onSortStatusChange: (status: DataTableSortStatus<WorkersTableRecord>) => void;
+  onPageChange: (page: number) => void;
+  onRecordsPerPageChange: (value: number) => void;
+  onResetFilters: () => void;
+  onRefetch: () => void;
+  onShowDetails: (worker: Worker) => void;
+  recordsPerPageOptions?: number[];
+};
+
+export function WorkersTable({
+  workers,
+  totalRecords,
+  isFetching,
+  isError,
+  error,
+  searchTerm,
+  sort,
+  page,
+  recordsPerPage,
+  onSortStatusChange,
+  onPageChange,
+  onRecordsPerPageChange,
+  onResetFilters,
+  onRefetch,
+  onShowDetails,
+  recordsPerPageOptions = DEFAULT_RECORDS_PER_PAGE_OPTIONS,
+}: WorkersTableProps) {
+  const { ref: tableContainerRef, width: containerWidth } = useElementSize();
+  const { width: viewportWidth } = useViewportSize();
+
+  const layoutAwareContainerWidth = useMemo(
+    () => getLayoutAwareWidth(containerWidth, viewportWidth),
+    [containerWidth, viewportWidth],
+  );
+
+  const workerRecords = useMemo<WorkersTableRecord[]>(() => {
+    if (!workers) {
+      return [];
+    }
+    return workers.map((worker) => buildWorkerRecord(worker));
+  }, [workers]);
+
+  const columns = useMemo(() => createWorkersColumns({ onShowDetails }), [onShowDetails]);
+  const responsiveColumns = useMemo(
+    () => createResponsiveColumns(columns, layoutAwareContainerWidth, COLUMN_VISIBILITY),
+    [columns, layoutAwareContainerWidth],
+  );
+
+  const totalPages = useMemo(
+    () => Math.max(1, Math.ceil(Math.max(0, totalRecords) / Math.max(1, recordsPerPage))),
+    [recordsPerPage, totalRecords],
+  );
+
+  useEffect(() => {
+    if (page > totalPages) {
+      onPageChange(totalPages);
+    }
+  }, [onPageChange, page, totalPages]);
+
+  const hasActiveFilters = searchTerm.trim().length > 0;
+
+  const sortStatus: DataTableSortStatus<WorkersTableRecord> = {
+    columnAccessor: sort.column,
+    direction: sort.direction,
+  };
+
+  const handleSortStatusChange = useCallback(
+    (status: DataTableSortStatus<WorkersTableRecord>) => {
+      onSortStatusChange(status);
+    },
+    [onSortStatusChange],
+  );
+
+  const errorDescriptor = isError ? getErrorDescriptor(error) : null;
+  const errorMessage = isError
+    ? `Workers are temporarily unavailable${errorDescriptor ? ` (${errorDescriptor})` : ''}.`
+    : 'Workers are temporarily unavailable.';
+
+  const emptyState = (
+    <Stack gap='sm' align='center' py='lg'>
+      {isError ? (
+        <>
+          <Text fw={600} size='sm'>
+            {errorMessage}
+          </Text>
+          <Text size='sm' c='dimmed' ta='center'>
+            Use the retry button to try again, or adjust the search to broaden the results.
+          </Text>
+          <Group gap='xs'>
+            <Button size='xs' variant='light' color='gray' leftSection={<IconRefresh size={14} />} onClick={onRefetch}>
+              Retry
+            </Button>
+            {hasActiveFilters ? (
+              <Button size='xs' variant='subtle' onClick={onResetFilters}>
+                Clear filters
+              </Button>
+            ) : null}
+          </Group>
+        </>
+      ) : (
+        <>
+          <Text fw={600} size='sm'>
+            No workers found
+          </Text>
+          <Text size='sm' c='dimmed' ta='center'>
+            {hasActiveFilters
+              ? 'Try adjusting the search to see more results.'
+              : 'Try refreshing to fetch the latest worker status.'}
+          </Text>
+          <Group gap='xs'>
+            <Button size='xs' variant='light' leftSection={<IconRefresh size={14} />} onClick={onRefetch}>
+              Refresh
+            </Button>
+            {hasActiveFilters ? (
+              <Button size='xs' variant='subtle' onClick={onResetFilters}>
+                Clear filters
+              </Button>
+            ) : null}
+          </Group>
+        </>
+      )}
+    </Stack>
+  );
+
+  return (
+    <Box ref={tableContainerRef}>
+      <DataTable<WorkersTableRecord>
+        classNames={{ root: 'workers-table' }}
+        withTableBorder
+        withColumnBorders
+        highlightOnHover
+        verticalAlign='center'
+        minHeight={workerRecords.length === 0 ? 400 : undefined}
+        idAccessor='workerId'
+        records={workerRecords}
+        columns={responsiveColumns}
+        totalRecords={totalRecords}
+        recordsPerPage={recordsPerPage}
+        page={page}
+        onPageChange={onPageChange}
+        onRecordsPerPageChange={onRecordsPerPageChange}
+        recordsPerPageOptions={recordsPerPageOptions}
+        sortStatus={sortStatus}
+        onSortStatusChange={handleSortStatusChange}
+        fetching={isFetching}
+        loaderSize='sm'
+        emptyState={workerRecords.length === 0 ? emptyState : undefined}
+      />
+    </Box>
+  );
+}

dashboard/src/components/WorkersTable.story.tsx

@@ -0,0 +1,154 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { useMemo, useState } from 'react';
+import type { Meta, StoryObj } from '@storybook/react';
+import { IconSearch } from '@tabler/icons-react';
+import { Box, Stack, TextInput, Title } from '@mantine/core';
+import type { Worker } from '@/types';
+import { WorkersTable } from './WorkersTable.component';
+
+const meta: Meta<typeof WorkersTable> = {
+  title: 'Components/WorkersTable',
+  component: WorkersTable,
+  parameters: {
+    layout: 'fullscreen',
+  },
+};
+
+export default meta;
+
+type Story = StoryObj<typeof WorkersTable>;
+
+const now = Math.floor(Date.now() / 1000);
+
+const sampleWorkers: Worker[] = [
+  {
+    workerId: 'worker-east',
+    status: 'busy',
+    heartbeatStats: { queueDepth: 2, gpuUtilization: 0.82 },
+    lastHeartbeatTime: now - 20,
+    lastDequeueTime: now - 60,
+    lastBusyTime: now - 120,
+    lastIdleTime: now - 600,
+    currentRolloutId: 'ro-story-001',
+    currentAttemptId: 'at-story-010',
+  },
+  {
+    workerId: 'worker-west',
+    status: 'busy',
+    heartbeatStats: { queueDepth: 1 },
+    lastHeartbeatTime: now - 45,
+    lastDequeueTime: now - 300,
+    lastBusyTime: now - 200,
+    lastIdleTime: now - 4800,
+    currentRolloutId: 'ro-story-003',
+    currentAttemptId: 'at-story-033',
+  },
+  {
+    workerId: 'worker-north',
+    status: 'idle',
+    heartbeatStats: { queueDepth: 0 },
+    lastHeartbeatTime: now - 90,
+    lastDequeueTime: now - 3600,
+    lastBusyTime: now - 5400,
+    lastIdleTime: now - 5400,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+  {
+    workerId: 'worker-south',
+    status: 'idle',
+    heartbeatStats: null,
+    lastHeartbeatTime: now - 900,
+    lastDequeueTime: now - 7200,
+    lastBusyTime: now - 8600,
+    lastIdleTime: now - 8600,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+  {
+    workerId: 'worker-standby',
+    status: 'unknown',
+    heartbeatStats: { queueDepth: 0 },
+    lastHeartbeatTime: now - 15,
+    lastDequeueTime: now - 4000,
+    lastBusyTime: null,
+    lastIdleTime: null,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+];
+
+type WorkersTableStoryWrapperProps = {
+  maxWidth: number;
+  initialSort?: { column: string; direction: 'asc' | 'desc' };
+};
+
+function WorkersTableStoryWrapper({ maxWidth, initialSort }: WorkersTableStoryWrapperProps) {
+  const [searchTerm, setSearchTerm] = useState('');
+  const [page, setPage] = useState(1);
+  const [recordsPerPage, setRecordsPerPage] = useState(5);
+  const [sort, setSort] = useState<{ column: string; direction: 'asc' | 'desc' }>(
+    () => initialSort ?? { column: 'lastHeartbeatTime', direction: 'desc' },
+  );
+
+  const filteredWorkers = useMemo(() => {
+    const normalized = searchTerm.trim().toLowerCase();
+    if (!normalized) {
+      return sampleWorkers;
+    }
+    return sampleWorkers.filter((worker) => worker.workerId.toLowerCase().includes(normalized));
+  }, [searchTerm]);
+
+  return (
+    <Stack gap='md' p='lg'>
+      <Title order={2}>Workers ({maxWidth}px max width)</Title>
+      <TextInput
+        placeholder='Search'
+        leftSection={<IconSearch size={16} />}
+        value={searchTerm}
+        onChange={(event) => setSearchTerm(event.currentTarget.value)}
+        w='100%'
+        style={{ maxWidth: 360 }}
+      />
+      <Box style={{ maxWidth }}>
+        <WorkersTable
+          workers={filteredWorkers}
+          totalRecords={filteredWorkers.length}
+          isFetching={false}
+          isError={false}
+          error={null}
+          searchTerm={searchTerm}
+          sort={sort}
+          page={page}
+          recordsPerPage={recordsPerPage}
+          onSortStatusChange={(status) => {
+            return setSort({ column: status.columnAccessor as string, direction: status.direction });
+          }}
+          onPageChange={setPage}
+          onRecordsPerPageChange={setRecordsPerPage}
+          onResetFilters={() => {
+            setSearchTerm('');
+            setPage(1);
+          }}
+          onRefetch={() => {}}
+          onShowDetails={() => {}}
+        />
+      </Box>
+    </Stack>
+  );
+}
+
+export const Wide: Story = {
+  render: () => <WorkersTableStoryWrapper maxWidth={1600} />,
+};
+
+export const Narrow: Story = {
+  render: () => <WorkersTableStoryWrapper maxWidth={780} />,
+};
+
+export const SortedByCurrentRollout: Story = {
+  render: () => (
+    <WorkersTableStoryWrapper maxWidth={1200} initialSort={{ column: 'currentRolloutId', direction: 'asc' }} />
+  ),
+};

dashboard/src/features/rollouts/api.ts

@@ -13,6 +13,8 @@ import type {
   RolloutStatus,
   Span,
   Timestamp,
+  Worker,
+  WorkerStatus,
 } from '../../types';
 
 const rawBaseQuery = fetchBaseQuery({
@@ -122,6 +124,21 @@ const normalizeResources = (value: unknown): Resources => {
   };
 };
 
+const normalizeWorker = (value: unknown): Worker => {
+  const camelized = camelCaseKeys(value) as Worker;
+  return {
+    workerId: camelized.workerId,
+    status: camelized.status,
+    heartbeatStats: camelized.heartbeatStats ?? null,
+    lastHeartbeatTime: camelized.lastHeartbeatTime ?? null,
+    lastDequeueTime: camelized.lastDequeueTime ?? null,
+    lastBusyTime: camelized.lastBusyTime ?? null,
+    lastIdleTime: camelized.lastIdleTime ?? null,
+    currentRolloutId: camelized.currentRolloutId ?? null,
+    currentAttemptId: camelized.currentAttemptId ?? null,
+  };
+};
+
 const normalizePaginatedResponse = <T>(value: unknown, normalizer: (item: unknown) => T): PaginatedResponse<T> => {
   if (!value || typeof value !== 'object') {
     throw new Error('Expected paginated response payload');
@@ -183,6 +200,15 @@ export type GetResourcesQueryArgs = {
   resourcesIdContains?: string | null;
 };
 
+export type GetWorkersQueryArgs = {
+  limit: number;
+  offset: number;
+  sortBy?: string | null;
+  sortOrder?: 'asc' | 'desc';
+  workerIdContains?: string | null;
+  statusIn?: WorkerStatus[];
+};
+
 export type GetRolloutAttemptsQueryArgs = {
   rolloutId: string;
   limit?: number;
@@ -208,7 +234,7 @@ export type GetSpansQueryArgs = {
 export const rolloutsApi = createApi({
   reducerPath: 'rolloutsApi',
   baseQuery: dynamicBaseQuery,
-  tagTypes: ['Rollout', 'Span', 'Resources'],
+  tagTypes: ['Rollout', 'Span', 'Resources', 'Worker'],
   endpoints: (builder) => ({
     getResources: builder.query<PaginatedResponse<Resources>, GetResourcesQueryArgs>({
       query: ({ limit, offset, sortBy, sortOrder, resourcesIdContains }) => {
@@ -238,6 +264,37 @@ export const rolloutsApi = createApi({
             ]
           : [{ type: 'Resources' as const, id: 'LIST' }],
     }),
+    getWorkers: builder.query<PaginatedResponse<Worker>, GetWorkersQueryArgs>({
+      query: ({ limit, offset, sortBy, sortOrder, workerIdContains, statusIn }) => {
+        const searchParams = new URLSearchParams();
+        searchParams.set('limit', String(typeof limit === 'number' ? limit : -1));
+        searchParams.set('offset', String(typeof offset === 'number' ? offset : 0));
+        if (sortBy) {
+          searchParams.set('sort_by', sortBy);
+        }
+        if (sortOrder) {
+          searchParams.set('sort_order', sortOrder);
+        }
+        if (workerIdContains && workerIdContains.trim().length > 0) {
+          searchParams.set('worker_id_contains', workerIdContains.trim());
+        }
+        if (statusIn && statusIn.length > 0) {
+          statusIn.forEach((status) => searchParams.append('status_in', status));
+        }
+
+        const queryString = searchParams.toString();
+        const url = queryString.length > 0 ? `v1/agl/workers?${queryString}` : 'v1/agl/workers';
+        return { url, method: 'GET' };
+      },
+      transformResponse: (response: unknown) => normalizePaginatedResponse(response, normalizeWorker),
+      providesTags: (result) =>
+        result
+          ? [
+              { type: 'Worker' as const, id: 'LIST' },
+              ...result.items.map((worker) => ({ type: 'Worker' as const, id: worker.workerId })),
+            ]
+          : [{ type: 'Worker' as const, id: 'LIST' }],
+    }),
     getRollouts: builder.query<PaginatedResponse<Rollout>, GetRolloutsQueryArgs>({
       query: ({ limit, offset, sortBy, sortOrder, statusIn, rolloutIdContains, modeIn }) => {
         const searchParams = new URLSearchParams();
@@ -343,4 +400,10 @@ export const rolloutsApi = createApi({
   }),
 });
 
-export const { useGetResourcesQuery, useGetRolloutsQuery, useGetRolloutAttemptsQuery, useGetSpansQuery } = rolloutsApi;
+export const {
+  useGetResourcesQuery,
+  useGetWorkersQuery,
+  useGetRolloutsQuery,
+  useGetRolloutAttemptsQuery,
+  useGetSpansQuery,
+} = rolloutsApi;

dashboard/src/features/ui/drawer/slice.ts

@@ -1,9 +1,9 @@
 // Copyright (c) Microsoft. All rights reserved.
 
 import { createSlice, type PayloadAction } from '@reduxjs/toolkit';
-import type { Attempt, Rollout, Span } from '@/types';
+import type { Attempt, Rollout, Span, Worker } from '@/types';
 
-export type DrawerType = 'rollout-json' | 'rollout-traces' | 'trace-detail';
+export type DrawerType = 'rollout-json' | 'rollout-traces' | 'trace-detail' | 'worker-detail';
 
 export type DrawerContent =
   | {
@@ -17,6 +17,10 @@ export type DrawerContent =
       span: Span;
       rollout: Rollout | null;
       attempt: Attempt | null;
+    }
+  | {
+      type: 'worker-detail';
+      worker: Worker;
     };
 
 export type DrawerState = {

dashboard/src/features/workers/index.ts

@@ -0,0 +1,5 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+export * from './slice';
+export * from './selectors';
+export { useGetWorkersQuery } from '../rollouts';

dashboard/src/features/workers/selectors.ts

@@ -0,0 +1,45 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { createSelector } from '@reduxjs/toolkit';
+import type { GetWorkersQueryArgs } from '@/features/rollouts';
+import type { RootState } from '@/store';
+import type { WorkersSortState } from './slice';
+
+const WORKERS_SORT_FIELD_MAP: Record<string, string> = {
+  workerId: 'worker_id',
+  status: 'status',
+  currentRolloutId: 'current_rollout_id',
+  currentAttemptId: 'current_attempt_id',
+  lastHeartbeatTime: 'last_heartbeat_time',
+  lastDequeueTime: 'last_dequeue_time',
+  lastBusyTime: 'last_busy_time',
+  lastIdleTime: 'last_idle_time',
+};
+
+const resolveWorkersSortField = (sort: WorkersSortState): string =>
+  WORKERS_SORT_FIELD_MAP[sort.column] ?? 'last_heartbeat_time';
+
+export const selectWorkersUiState = (state: RootState) => state.workers;
+
+export const selectWorkersSearchTerm = (state: RootState) => selectWorkersUiState(state).searchTerm;
+export const selectWorkersPage = (state: RootState) => selectWorkersUiState(state).page;
+export const selectWorkersRecordsPerPage = (state: RootState) => selectWorkersUiState(state).recordsPerPage;
+export const selectWorkersSort = (state: RootState) => selectWorkersUiState(state).sort;
+
+export const selectWorkersQueryArgs = createSelector(
+  [selectWorkersSearchTerm, selectWorkersPage, selectWorkersRecordsPerPage, selectWorkersSort],
+  (searchTerm, page, recordsPerPage, sort): GetWorkersQueryArgs => {
+    const normalizedSearch = searchTerm.trim();
+    const limit = Math.max(1, recordsPerPage);
+    const offset = Math.max(0, (page - 1) * limit);
+    const sortBy = resolveWorkersSortField(sort);
+
+    return {
+      limit,
+      offset,
+      sortBy,
+      sortOrder: sort.direction,
+      workerIdContains: normalizedSearch.length > 0 ? normalizedSearch : undefined,
+    };
+  },
+);

dashboard/src/features/workers/slice.ts

@@ -0,0 +1,59 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { createSlice, type PayloadAction } from '@reduxjs/toolkit';
+
+export type SortDirection = 'asc' | 'desc';
+
+export type WorkersSortState = {
+  column: string;
+  direction: SortDirection;
+};
+
+export type WorkersUiState = {
+  searchTerm: string;
+  page: number;
+  recordsPerPage: number;
+  sort: WorkersSortState;
+};
+
+export const initialWorkersUiState: WorkersUiState = {
+  searchTerm: '',
+  page: 1,
+  recordsPerPage: 50,
+  sort: {
+    column: 'lastHeartbeatTime',
+    direction: 'desc',
+  },
+};
+
+const workersSlice = createSlice({
+  name: 'workers',
+  initialState: initialWorkersUiState,
+  reducers: {
+    setWorkersSearchTerm(state, action: PayloadAction<string>) {
+      state.searchTerm = action.payload;
+      state.page = 1;
+    },
+    setWorkersPage(state, action: PayloadAction<number>) {
+      state.page = action.payload;
+    },
+    setWorkersRecordsPerPage(state, action: PayloadAction<number>) {
+      state.recordsPerPage = action.payload;
+      state.page = 1;
+    },
+    setWorkersSort(state, action: PayloadAction<WorkersSortState>) {
+      state.sort = action.payload;
+    },
+    resetWorkersFilters(state) {
+      state.searchTerm = initialWorkersUiState.searchTerm;
+      state.page = initialWorkersUiState.page;
+      state.recordsPerPage = initialWorkersUiState.recordsPerPage;
+      state.sort = initialWorkersUiState.sort;
+    },
+  },
+});
+
+export const { setWorkersSearchTerm, setWorkersPage, setWorkersRecordsPerPage, setWorkersSort, resetWorkersFilters } =
+  workersSlice.actions;
+
+export const workersReducer = workersSlice.reducer;

dashboard/src/features/workers/workers.test.ts

@@ -0,0 +1,102 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { createServerBackedStore } from '@test-utils';
+import { describe, expect, it } from 'vitest';
+import { rolloutsApi } from '@/features/rollouts';
+import type { Worker } from '@/types';
+import { selectWorkersQueryArgs } from './selectors';
+import {
+  resetWorkersFilters,
+  setWorkersPage,
+  setWorkersRecordsPerPage,
+  setWorkersSearchTerm,
+  setWorkersSort,
+} from './slice';
+
+const extractWorkerIds = (workers: Worker[]): string[] => workers.map((worker) => worker.workerId);
+
+describe('workers feature integration', () => {
+  it('builds default query arguments from the UI state', () => {
+    const store = createServerBackedStore();
+    const queryArgs = selectWorkersQueryArgs(store.getState());
+
+    expect(queryArgs).toMatchObject({
+      limit: 50,
+      offset: 0,
+      sortBy: 'last_heartbeat_time',
+      sortOrder: 'desc',
+      workerIdContains: undefined,
+    });
+  });
+
+  it('fetches workers from the Python LightningStore server', async () => {
+    const store = createServerBackedStore();
+    const queryArgs = selectWorkersQueryArgs(store.getState());
+
+    const subscription = store.dispatch(rolloutsApi.endpoints.getWorkers.initiate(queryArgs));
+    const data = await subscription.unwrap();
+    subscription.unsubscribe();
+
+    expect(data.total).toBeGreaterThanOrEqual(4);
+    expect(data.items).toHaveLength(Math.min(queryArgs.limit, data.total));
+
+    const workerIds = extractWorkerIds(data.items);
+    expect(workerIds).toEqual(expect.arrayContaining(['worker-east', 'worker-west']));
+
+    const heartbeatTimes = data.items.map((worker) => worker.lastHeartbeatTime ?? 0);
+    const sortedHeartbeatTimes = [...heartbeatTimes].sort((a, b) => b - a);
+    expect(heartbeatTimes).toEqual(sortedHeartbeatTimes);
+  });
+
+  it('paginates worker results based on UI state', async () => {
+    const store = createServerBackedStore();
+    store.dispatch(setWorkersRecordsPerPage(2));
+    store.dispatch(setWorkersPage(2));
+
+    const queryArgs = selectWorkersQueryArgs(store.getState());
+    expect(queryArgs).toMatchObject({ limit: 2, offset: 2 });
+
+    const subscription = store.dispatch(rolloutsApi.endpoints.getWorkers.initiate(queryArgs));
+    const data = await subscription.unwrap();
+    subscription.unsubscribe();
+
+    expect(data.items).toHaveLength(2);
+    expect(data.total).toBeGreaterThanOrEqual(4);
+  });
+
+  it('applies search and sorting preferences', async () => {
+    const store = createServerBackedStore();
+    store.dispatch(resetWorkersFilters());
+    store.dispatch(setWorkersSearchTerm('worker-west'));
+    store.dispatch(setWorkersSort({ column: 'workerId', direction: 'asc' }));
+
+    const queryArgs = selectWorkersQueryArgs(store.getState());
+    expect(queryArgs).toMatchObject({
+      limit: 50,
+      offset: 0,
+      sortBy: 'worker_id',
+      sortOrder: 'asc',
+      workerIdContains: 'worker-west',
+    });
+
+    const subscription = store.dispatch(rolloutsApi.endpoints.getWorkers.initiate(queryArgs));
+    const data = await subscription.unwrap();
+    subscription.unsubscribe();
+
+    expect(data.items).toHaveLength(1);
+    expect(data.items[0].workerId).toBe('worker-west');
+  });
+
+  it('maps current rollout/attempt sorting to backend fields', () => {
+    const store = createServerBackedStore();
+    store.dispatch(setWorkersSort({ column: 'currentRolloutId', direction: 'desc' }));
+    let queryArgs = selectWorkersQueryArgs(store.getState());
+    expect(queryArgs.sortBy).toBe('current_rollout_id');
+    expect(queryArgs.sortOrder).toBe('desc');
+
+    store.dispatch(setWorkersSort({ column: 'currentAttemptId', direction: 'asc' }));
+    queryArgs = selectWorkersQueryArgs(store.getState());
+    expect(queryArgs.sortBy).toBe('current_attempt_id');
+    expect(queryArgs.sortOrder).toBe('asc');
+  });
+});

dashboard/src/layouts/AppLayout.story.tsx

@@ -39,6 +39,10 @@ const ROUTES = [
         path: 'traces',
         element: <Placeholder title='Traces' description='Browse telemetry spans across attempts.' />,
       },
+      {
+        path: 'runners',
+        element: <Placeholder title='Runners' description='Monitor runner activity and status.' />,
+      },
       {
         path: 'settings',
         element: (

dashboard/src/layouts/AppLayout.tsx

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft. All rights reserved.
 
 import { useEffect, useMemo, useState, type ReactNode } from 'react';
-import { IconCpu, IconRouteSquare, IconSettings, IconTimeline } from '@tabler/icons-react';
+import { IconCpu, IconRouteSquare, IconRun, IconSettings, IconTimeline } from '@tabler/icons-react';
 import { Outlet, NavLink as RouterNavLink, useLocation, useNavigate } from 'react-router-dom';
 import { AppShell, Badge, Group, Image, NavLink as MantineNavLink, Stack, Text, UnstyledButton } from '@mantine/core';
 import { AppAlertBanner } from '@/components/AppAlertBanner';
@@ -22,6 +22,7 @@ const NAV_ITEMS: NavItem[] = [
   { label: 'Rollouts', to: '/rollouts', icon: <IconRouteSquare size={16} /> },
   { label: 'Resources', to: '/resources', icon: <IconCpu size={16} /> },
   { label: 'Traces', to: '/traces', icon: <IconTimeline size={16} /> },
+  { label: 'Runners', to: '/runners', icon: <IconRun size={16} /> },
   { label: 'Settings', to: '/settings', icon: <IconSettings size={16} /> },
 ];
 

dashboard/src/pages/Workers.page.story.tsx

@@ -0,0 +1,270 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import type { Meta, StoryObj } from '@storybook/react';
+import { waitFor, within } from '@testing-library/dom';
+import userEvent from '@testing-library/user-event';
+import { Provider } from 'react-redux';
+import { createMemoryRouter, RouterProvider } from 'react-router-dom';
+import { AppAlertBanner } from '@/components/AppAlertBanner';
+import { AppDrawerContainer } from '@/components/AppDrawer.component';
+import { initialConfigState } from '@/features/config/slice';
+import { initialWorkersUiState } from '@/features/workers/slice';
+import { AppLayout } from '@/layouts/AppLayout';
+import { createAppStore } from '@/store';
+import type { Worker } from '@/types';
+import { createWorkersHandlers } from '@/utils/mock';
+import { STORY_BASE_URL, STORY_DATE_NOW_SECONDS } from '../../.storybook/constants';
+import { allModes } from '../../.storybook/modes';
+import { WorkersPage } from './Workers.page';
+
+const meta: Meta<typeof WorkersPage> = {
+  title: 'Pages/WorkersPage',
+  component: WorkersPage,
+  parameters: {
+    layout: 'fullscreen',
+    chromatic: {
+      modes: allModes,
+    },
+  },
+};
+
+export default meta;
+
+type Story = StoryObj<typeof WorkersPage>;
+
+const now = STORY_DATE_NOW_SECONDS;
+
+const sampleWorkers: Worker[] = [
+  {
+    workerId: 'worker-east',
+    status: 'busy',
+    heartbeatStats: { queueDepth: 2, gpuUtilization: 0.82 },
+    lastHeartbeatTime: now - 20,
+    lastDequeueTime: now - 120,
+    lastBusyTime: now - 60,
+    lastIdleTime: now - 600,
+    currentRolloutId: 'ro-story-001',
+    currentAttemptId: 'at-story-010',
+  },
+  {
+    workerId: 'worker-west',
+    status: 'busy',
+    heartbeatStats: { queueDepth: 1 },
+    lastHeartbeatTime: now - 45,
+    lastDequeueTime: now - 300,
+    lastBusyTime: now - 120,
+    lastIdleTime: now - 4800,
+    currentRolloutId: 'ro-story-003',
+    currentAttemptId: 'at-story-033',
+  },
+  {
+    workerId: 'worker-north',
+    status: 'idle',
+    heartbeatStats: { queueDepth: 0 },
+    lastHeartbeatTime: now - 120,
+    lastDequeueTime: now - 3600,
+    lastBusyTime: now - 5400,
+    lastIdleTime: now - 180,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+  {
+    workerId: 'worker-south',
+    status: 'idle',
+    heartbeatStats: null,
+    lastHeartbeatTime: now - 900,
+    lastDequeueTime: now - 7200,
+    lastBusyTime: now - 8600,
+    lastIdleTime: now - 8600,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+  {
+    workerId: 'worker-central',
+    status: 'busy',
+    heartbeatStats: { queueDepth: 3, cpuUtilization: 0.55 },
+    lastHeartbeatTime: now - 8,
+    lastDequeueTime: now - 45,
+    lastBusyTime: now - 10,
+    lastIdleTime: now - 900,
+    currentRolloutId: 'ro-story-005',
+    currentAttemptId: 'at-story-013',
+  },
+  {
+    workerId: 'worker-standby',
+    status: 'idle',
+    heartbeatStats: { queueDepth: 0, threads: 32 },
+    lastHeartbeatTime: now - 300,
+    lastDequeueTime: now - 10800,
+    lastBusyTime: now - 14400,
+    lastIdleTime: now - 200,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+  {
+    workerId: 'worker-observer',
+    status: 'unknown',
+    heartbeatStats: { queueDepth: 0 },
+    lastHeartbeatTime: now - 30,
+    lastDequeueTime: now - 6400,
+    lastBusyTime: null,
+    lastIdleTime: null,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+];
+
+const defaultHandlers = createWorkersHandlers(sampleWorkers);
+
+function createStoryStore(configOverrides?: Partial<typeof initialConfigState>) {
+  return createAppStore({
+    config: {
+      ...initialConfigState,
+      baseUrl: STORY_BASE_URL,
+      autoRefreshMs: 0,
+      ...configOverrides,
+    },
+    workers: initialWorkersUiState,
+  });
+}
+
+function renderWithStore(configOverrides?: Partial<typeof initialConfigState>) {
+  const store = createStoryStore(configOverrides);
+
+  return (
+    <Provider store={store}>
+      <>
+        <WorkersPage />
+        <AppAlertBanner />
+        <AppDrawerContainer />
+      </>
+    </Provider>
+  );
+}
+
+function renderWithinAppLayout(configOverrides?: Partial<typeof initialConfigState>) {
+  const store = createStoryStore(configOverrides);
+  const router = createMemoryRouter(
+    [
+      {
+        path: '/',
+        element: (
+          <AppLayout
+            config={{
+              baseUrl: store.getState().config.baseUrl,
+              autoRefreshMs: store.getState().config.autoRefreshMs,
+            }}
+          />
+        ),
+        children: [
+          {
+            path: '/runners',
+            element: <WorkersPage />,
+          },
+        ],
+      },
+    ],
+    { initialEntries: ['/runners'] },
+  );
+
+  return (
+    <Provider store={store}>
+      <>
+        <RouterProvider router={router} />
+        <AppDrawerContainer />
+      </>
+    </Provider>
+  );
+}
+
+const manyWorkers = Array.from({ length: 80 }, (_, index) => {
+  const suffix = (index + 1).toString().padStart(3, '0');
+  const busy = index % 2 === 0;
+  return {
+    workerId: `worker-batch-${suffix}`,
+    status: busy ? 'busy' : 'idle',
+    heartbeatStats: busy ? { queueDepth: (index % 5) + 1 } : { queueDepth: 0 },
+    lastHeartbeatTime: now - (index * 5 + 15),
+    lastDequeueTime: now - (index * 20 + 60),
+    lastBusyTime: busy ? now - (index * 10 + 30) : null,
+    lastIdleTime: busy ? null : now - (index * 10 + 45),
+    currentRolloutId: busy ? `ro-many-${suffix}` : null,
+    currentAttemptId: busy ? `at-many-${suffix}` : null,
+  } satisfies Worker;
+});
+
+export const Default: Story = {
+  render: () => renderWithinAppLayout(),
+  parameters: {
+    msw: {
+      handlers: defaultHandlers,
+    },
+  },
+};
+
+export const Search: Story = {
+  render: () => renderWithStore(),
+  parameters: {
+    msw: {
+      handlers: defaultHandlers,
+    },
+  },
+  play: async ({ canvasElement }) => {
+    const canvas = within(canvasElement);
+    await canvas.findByText('worker-east');
+
+    const searchInput = canvas.getByPlaceholderText('Search by Runner ID');
+    await userEvent.type(searchInput, 'worker-west');
+
+    await waitFor(() => {
+      if (canvas.queryByText('worker-east')) {
+        throw new Error('Expected filtered table to hide worker-east');
+      }
+      if (!canvas.queryByText('worker-west')) {
+        throw new Error('Expected worker-west to remain visible');
+      }
+    });
+  },
+};
+
+export const DrawerOpen: Story = {
+  render: () => renderWithStore(),
+  parameters: {
+    msw: {
+      handlers: defaultHandlers,
+    },
+  },
+  play: async ({ canvasElement }) => {
+    const canvas = within(canvasElement);
+    await canvas.findByText('worker-east');
+
+    const detailsButtons = await canvas.findAllByRole('button', { name: /detail/i });
+    await userEvent.click(detailsButtons[0]);
+
+    const body = within(document.body);
+    await waitFor(() => {
+      if (!body.queryByTestId('json-editor-container')) {
+        throw new Error('Expected worker detail drawer with JSON view');
+      }
+    });
+  },
+};
+
+export const ManyWorkers: Story = {
+  render: () => renderWithStore(),
+  parameters: {
+    msw: {
+      handlers: createWorkersHandlers(manyWorkers),
+    },
+  },
+};
+
+export const DarkTheme: Story = {
+  render: () => renderWithStore({ theme: 'dark' }),
+  parameters: {
+    theme: 'dark',
+    msw: {
+      handlers: defaultHandlers,
+    },
+  },
+};

dashboard/src/pages/Workers.page.tsx

@@ -0,0 +1,159 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+import { useCallback, useEffect } from 'react';
+import { IconSearch } from '@tabler/icons-react';
+import type { DataTableSortStatus } from 'mantine-datatable';
+import { Skeleton, Stack, TextInput, Title } from '@mantine/core';
+import { WorkersTable, type WorkersTableRecord } from '@/components/WorkersTable.component';
+import { selectAutoRefreshMs } from '@/features/config';
+import { hideAlert, showAlert } from '@/features/ui/alert';
+import { openDrawer } from '@/features/ui/drawer';
+import {
+  resetWorkersFilters,
+  selectWorkersPage,
+  selectWorkersQueryArgs,
+  selectWorkersRecordsPerPage,
+  selectWorkersSearchTerm,
+  selectWorkersSort,
+  setWorkersPage,
+  setWorkersRecordsPerPage,
+  setWorkersSearchTerm,
+  setWorkersSort,
+  useGetWorkersQuery,
+} from '@/features/workers';
+import { useAppDispatch, useAppSelector } from '@/store/hooks';
+import type { PaginatedResponse, Worker } from '@/types';
+import { getErrorDescriptor } from '@/utils/error';
+
+export function WorkersPage() {
+  const dispatch = useAppDispatch();
+  const autoRefreshMs = useAppSelector(selectAutoRefreshMs);
+  const searchTerm = useAppSelector(selectWorkersSearchTerm);
+  const page = useAppSelector(selectWorkersPage);
+  const recordsPerPage = useAppSelector(selectWorkersRecordsPerPage);
+  const sort = useAppSelector(selectWorkersSort);
+  const queryArgs = useAppSelector(selectWorkersQueryArgs);
+
+  const workersQueryResult = useGetWorkersQuery(queryArgs, {
+    pollingInterval: autoRefreshMs > 0 ? autoRefreshMs : undefined,
+  });
+
+  const workersData = workersQueryResult.data as PaginatedResponse<Worker> | undefined;
+  const { isLoading, isFetching, isError, error, refetch } = workersQueryResult;
+
+  const handleSearchTermChange = useCallback(
+    (value: string) => {
+      dispatch(setWorkersSearchTerm(value));
+    },
+    [dispatch],
+  );
+
+  const handleSortStatusChange = useCallback(
+    (status: DataTableSortStatus<WorkersTableRecord>) => {
+      dispatch(
+        setWorkersSort({
+          column: status.columnAccessor,
+          direction: status.direction,
+        }),
+      );
+    },
+    [dispatch],
+  );
+
+  const handlePageChange = useCallback(
+    (nextPage: number) => {
+      dispatch(setWorkersPage(nextPage));
+    },
+    [dispatch],
+  );
+
+  const handleRecordsPerPageChange = useCallback(
+    (value: number) => {
+      dispatch(setWorkersRecordsPerPage(value));
+    },
+    [dispatch],
+  );
+
+  const handleResetFilters = useCallback(() => {
+    dispatch(resetWorkersFilters());
+  }, [dispatch]);
+
+  const handleShowWorkerDetails = useCallback(
+    (worker: Worker) => {
+      dispatch(
+        openDrawer({
+          type: 'worker-detail',
+          worker,
+        }),
+      );
+    },
+    [dispatch],
+  );
+
+  const hasWorkers = Array.isArray(workersData?.items) && workersData.items.length > 0;
+  const showSkeleton = isLoading && !hasWorkers;
+
+  useEffect(() => {
+    if (isError) {
+      const descriptor = getErrorDescriptor(error);
+      const suffix = descriptor ? ` (${descriptor})` : '';
+      dispatch(
+        showAlert({
+          id: 'workers-fetch',
+          message: `Unable to refresh workers${suffix}. The table may be out of date until the connection recovers.`,
+          tone: 'error',
+        }),
+      );
+      return;
+    }
+
+    if (!isLoading && !isFetching) {
+      dispatch(hideAlert({ id: 'workers-fetch' }));
+    }
+  }, [dispatch, error, isError, isFetching, isLoading]);
+
+  useEffect(
+    () => () => {
+      dispatch(hideAlert({ id: 'workers-fetch' }));
+    },
+    [dispatch],
+  );
+
+  return (
+    <Stack gap='md'>
+      <Title order={1}>Runners</Title>
+
+      <TextInput
+        placeholder='Search by Runner ID'
+        value={searchTerm}
+        onChange={(event) => handleSearchTermChange(event.currentTarget.value)}
+        leftSection={<IconSearch size={16} />}
+        data-testid='workers-search-input'
+        w='100%'
+        style={{ maxWidth: 360 }}
+      />
+
+      {showSkeleton ? (
+        <Skeleton height={360} radius='md' />
+      ) : (
+        <WorkersTable
+          workers={workersData?.items}
+          totalRecords={workersData?.total ?? 0}
+          isFetching={isFetching}
+          isError={isError}
+          error={error}
+          searchTerm={searchTerm}
+          sort={sort}
+          page={page}
+          recordsPerPage={recordsPerPage}
+          onSortStatusChange={handleSortStatusChange}
+          onPageChange={handlePageChange}
+          onRecordsPerPageChange={handleRecordsPerPageChange}
+          onResetFilters={handleResetFilters}
+          onRefetch={refetch}
+          onShowDetails={handleShowWorkerDetails}
+        />
+      )}
+    </Stack>
+  );
+}

dashboard/src/store/index.ts

@@ -7,6 +7,7 @@ import { rolloutsApi, rolloutsReducer } from '../features/rollouts';
 import { tracesReducer } from '../features/traces';
 import { alertReducer } from '../features/ui/alert';
 import { drawerReducer } from '../features/ui/drawer';
+import { workersReducer } from '../features/workers';
 
 const rootReducer = combineReducers({
   config: configReducer,
@@ -14,6 +15,7 @@ const rootReducer = combineReducers({
   alert: alertReducer,
   rollouts: rolloutsReducer,
   resources: resourcesReducer,
+  workers: workersReducer,
   traces: tracesReducer,
   [rolloutsApi.reducerPath]: rolloutsApi.reducer,
 });

dashboard/src/types.ts

@@ -28,6 +28,24 @@ export type Attempt = {
   metadata: Record<string, any> | null;
 };
 
+export type WorkerStatus = 'idle' | 'busy' | 'unknown';
+
+/**
+ * Synced with agentlightning.types.core.Worker
+ * with camel case and snake case conversions
+ */
+export type Worker = {
+  workerId: string;
+  status: WorkerStatus;
+  heartbeatStats: Record<string, any> | null;
+  lastHeartbeatTime: Timestamp | null;
+  lastDequeueTime: Timestamp | null;
+  lastBusyTime: Timestamp | null;
+  lastIdleTime: Timestamp | null;
+  currentRolloutId: string | null;
+  currentAttemptId: string | null;
+};
+
 /**
  * Synced with agentlightning.types.core.Rollout
  * with camel case and snake case conversions

dashboard/src/utils/mock.test.ts

@@ -8,27 +8,32 @@
  */
 
 import { describe, expect, it } from 'vitest';
-import type { Attempt, Resources, Rollout, Span } from '@/types';
+import type { Attempt, Resources, Rollout, Span, Worker } from '@/types';
 import {
   buildAttemptsResponse,
   buildResourcesResponse,
   buildRolloutsResponse,
   buildSpansResponse,
+  buildWorkersResponse,
   createMockHandlers,
   createResourcesHandlers,
   createRolloutsHandlers,
   createSpansHandlers,
+  createWorkersHandlers,
   filterResourcesForParams,
   filterRolloutsForParams,
   filterSpansForParams,
+  filterWorkersForParams,
   getResourcesSortValue,
   getRolloutSortValue,
   getSpanSortValue,
+  getWorkerSortValue,
   parseNumberParam,
   sortAttemptsForParams,
   sortResourcesForParams,
   sortRolloutsForParams,
   sortSpansForParams,
+  sortWorkersForParams,
 } from './mock';
 
 const now = Math.floor(Date.now() / 1000);
@@ -219,6 +224,53 @@ const sampleResources: Resources[] = [
   },
 ];
 
+const sampleWorkers: Worker[] = [
+  {
+    workerId: 'worker-alpha',
+    status: 'busy',
+    heartbeatStats: { queueDepth: 2 },
+    lastHeartbeatTime: now - 30,
+    lastDequeueTime: now - 300,
+    lastBusyTime: now - 60,
+    lastIdleTime: now - 600,
+    currentRolloutId: 'ro-001',
+    currentAttemptId: 'at-001',
+  },
+  {
+    workerId: 'worker-beta',
+    status: 'idle',
+    heartbeatStats: { queueDepth: 0 },
+    lastHeartbeatTime: now - 120,
+    lastDequeueTime: now - 1200,
+    lastBusyTime: now - 3600,
+    lastIdleTime: now - 180,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+  {
+    workerId: 'worker-gamma',
+    status: 'busy',
+    heartbeatStats: null,
+    lastHeartbeatTime: now - 10,
+    lastDequeueTime: now - 60,
+    lastBusyTime: now - 20,
+    lastIdleTime: now - 4000,
+    currentRolloutId: 'ro-003',
+    currentAttemptId: 'at-003',
+  },
+  {
+    workerId: 'worker-delta',
+    status: 'unknown',
+    heartbeatStats: { queueDepth: 0 },
+    lastHeartbeatTime: now - 5,
+    lastDequeueTime: now - 80,
+    lastBusyTime: null,
+    lastIdleTime: null,
+    currentRolloutId: null,
+    currentAttemptId: null,
+  },
+];
+
 describe('parseNumberParam', () => {
   it('returns default value when param is missing', () => {
     const params = new URLSearchParams();
@@ -725,6 +777,115 @@ describe('createResourcesHandlers', () => {
   });
 });
 
+describe('filterWorkersForParams', () => {
+  it('returns all workers without filters', () => {
+    const params = new URLSearchParams();
+    const result = filterWorkersForParams(sampleWorkers, params);
+    expect(result).toHaveLength(4);
+  });
+
+  it('filters by status and worker ID substring using AND logic', () => {
+    const params = new URLSearchParams('status_in=busy&worker_id_contains=gamma');
+    const result = filterWorkersForParams(sampleWorkers, params);
+    expect(result).toHaveLength(1);
+    expect(result[0].workerId).toBe('worker-gamma');
+  });
+
+  it('supports filter_logic=or', () => {
+    const params = new URLSearchParams('status_in=idle&worker_id_contains=gamma&filter_logic=or');
+    const result = filterWorkersForParams(sampleWorkers, params);
+    expect(result).toHaveLength(2);
+  });
+
+  it('filters by unknown status', () => {
+    const params = new URLSearchParams('status_in=unknown');
+    const result = filterWorkersForParams(sampleWorkers, params);
+    expect(result).toHaveLength(1);
+    expect(result[0].workerId).toBe('worker-delta');
+  });
+});
+
+describe('getWorkerSortValue', () => {
+  const worker = sampleWorkers[0];
+
+  it('returns worker_id', () => {
+    expect(getWorkerSortValue(worker, 'worker_id')).toBe('worker-alpha');
+  });
+
+  it('returns status', () => {
+    expect(getWorkerSortValue(worker, 'status')).toBe('busy');
+  });
+
+  it('returns timestamp fields', () => {
+    expect(getWorkerSortValue(worker, 'last_busy_time')).toBe(worker.lastBusyTime);
+    expect(getWorkerSortValue(worker, 'last_idle_time')).toBe(worker.lastIdleTime);
+    expect(getWorkerSortValue(worker, 'last_dequeue_time')).toBe(worker.lastDequeueTime);
+  });
+
+  it('returns rollout and attempt identifiers', () => {
+    expect(getWorkerSortValue(worker, 'current_rollout_id')).toBe(worker.currentRolloutId);
+    expect(getWorkerSortValue(worker, 'current_attempt_id')).toBe(worker.currentAttemptId);
+  });
+
+  it('falls back to last_heartbeat_time', () => {
+    expect(getWorkerSortValue(worker, 'unknown')).toBe(worker.lastHeartbeatTime);
+  });
+});
+
+describe('sortWorkersForParams', () => {
+  it('sorts by last heartbeat ascending by default', () => {
+    const result = sortWorkersForParams(sampleWorkers, null, 'asc');
+    expect(result.map((worker) => worker.workerId)).toEqual([
+      'worker-beta',
+      'worker-alpha',
+      'worker-gamma',
+      'worker-delta',
+    ]);
+  });
+
+  it('sorts descending by worker_id when requested', () => {
+    const result = sortWorkersForParams(sampleWorkers, 'worker_id', 'desc');
+    expect(result.map((worker) => worker.workerId)).toEqual([
+      'worker-gamma',
+      'worker-delta',
+      'worker-beta',
+      'worker-alpha',
+    ]);
+  });
+
+  it('sorts by current_rollout_id', () => {
+    const result = sortWorkersForParams(sampleWorkers, 'current_rollout_id', 'asc');
+    expect(result.map((worker) => worker.currentRolloutId)).toEqual([null, null, 'ro-001', 'ro-003']);
+  });
+});
+
+describe('buildWorkersResponse', () => {
+  it('applies filters before pagination', () => {
+    const request = new Request('http://localhost/v1/agl/workers?worker_id_contains=beta&limit=5');
+    const response = buildWorkersResponse(sampleWorkers, request);
+    expect(response.items).toHaveLength(1);
+    const items = response.items as Array<Record<string, unknown>>;
+    expect(items[0].worker_id).toBe('worker-beta');
+  });
+
+  it('applies sort and pagination parameters', () => {
+    const request = new Request('http://localhost/v1/agl/workers?sort_by=worker_id&limit=2&offset=1');
+    const response = buildWorkersResponse(sampleWorkers, request);
+    expect(response.items).toHaveLength(2);
+    const items = response.items as Array<Record<string, unknown>>;
+    expect(items[0].worker_id).toBe('worker-beta');
+    expect(response.total).toBe(4);
+  });
+});
+
+describe('createWorkersHandlers', () => {
+  it('creates handler for workers endpoint', () => {
+    const handlers = createWorkersHandlers(sampleWorkers);
+    expect(handlers).toHaveLength(1);
+    expect(handlers[0].info.header).toContain('GET');
+  });
+});
+
 describe('createRolloutsHandlers', () => {
   it('creates handlers that return correct rollout data', async () => {
     const attemptsByRollout = { 'ro-001': sampleAttempts };

dashboard/src/utils/mock.ts

@@ -14,7 +14,7 @@
  */
 
 import { delay, http, HttpResponse } from 'msw';
-import type { Attempt, Resources, Rollout, Span } from '@/types';
+import type { Attempt, Resources, Rollout, Span, Worker } from '@/types';
 import { snakeCaseKeys } from './format';
 
 /**
@@ -434,6 +434,117 @@ export function buildResourcesResponse(resources: Resources[], request: Request)
   });
 }
 
+/**
+ * Filter workers based on query parameters.
+ * Supports: status_in, worker_id_contains
+ */
+export function filterWorkersForParams(workers: Worker[], params: URLSearchParams): Worker[] {
+  const statusFilters = params.getAll('status_in');
+  const workerIdContains = params.get('worker_id_contains');
+  const filterLogic = params.get('filter_logic') === 'or' ? 'or' : 'and';
+
+  return workers.filter((worker) => {
+    const checks: boolean[] = [];
+    if (statusFilters.length > 0) {
+      checks.push(statusFilters.includes(worker.status));
+    }
+    if (workerIdContains) {
+      checks.push(worker.workerId.toLowerCase().includes(workerIdContains.toLowerCase()));
+    }
+    if (checks.length === 0) {
+      return true;
+    }
+    return filterLogic === 'or' ? checks.some(Boolean) : checks.every(Boolean);
+  });
+}
+
+/**
+ * Resolve a worker sort value for the given column.
+ */
+export function getWorkerSortValue(worker: Worker, sortBy: string): string | number | null {
+  switch (sortBy) {
+    case 'worker_id':
+      return worker.workerId;
+    case 'status':
+      return worker.status;
+    case 'current_rollout_id':
+      return worker.currentRolloutId ?? '';
+    case 'current_attempt_id':
+      return worker.currentAttemptId ?? '';
+    case 'last_busy_time':
+      return worker.lastBusyTime ?? null;
+    case 'last_idle_time':
+      return worker.lastIdleTime ?? null;
+    case 'last_dequeue_time':
+      return worker.lastDequeueTime ?? null;
+    case 'last_heartbeat_time':
+    default:
+      return worker.lastHeartbeatTime ?? null;
+  }
+}
+
+/**
+ * Sort workers based on query parameters.
+ * Default sort_by is 'last_heartbeat_time'.
+ */
+export function sortWorkersForParams(workers: Worker[], sortBy: string | null, sortOrder: 'asc' | 'desc'): Worker[] {
+  const resolvedSortBy = sortBy ?? 'last_heartbeat_time';
+  const sorted = [...workers].sort((a, b) => {
+    const aValue = getWorkerSortValue(a, resolvedSortBy);
+    const bValue = getWorkerSortValue(b, resolvedSortBy);
+    if (aValue === bValue) {
+      return 0;
+    }
+    if (aValue == null) {
+      return -1;
+    }
+    if (bValue == null) {
+      return 1;
+    }
+    if (typeof aValue === 'number' && typeof bValue === 'number') {
+      return aValue - bValue;
+    }
+    return String(aValue).localeCompare(String(bValue));
+  });
+
+  if (sortOrder === 'desc') {
+    sorted.reverse();
+  }
+
+  return sorted;
+}
+
+/**
+ * Build a paginated workers response matching the Python server's format.
+ */
+export function buildWorkersResponse(workers: Worker[], request: Request): Record<string, unknown> {
+  const url = new URL(request.url);
+  const params = url.searchParams;
+  const filtered = filterWorkersForParams(workers, params);
+  const sortBy = params.get('sort_by');
+  const sortOrder = params.get('sort_order') === 'desc' ? 'desc' : 'asc';
+  const sorted = sortWorkersForParams(filtered, sortBy, sortOrder);
+  const limitParam = parseNumberParam(params, 'limit', sorted.length);
+  const offsetParam = parseNumberParam(params, 'offset', 0);
+  const effectiveLimit = limitParam < 0 ? sorted.length : limitParam;
+  const offset = offsetParam < 0 ? 0 : offsetParam;
+  const paginated = effectiveLimit >= 0 ? sorted.slice(offset, offset + effectiveLimit) : [...sorted];
+
+  return snakeCaseKeys({
+    items: paginated,
+    limit: effectiveLimit,
+    offset,
+    total: filtered.length,
+  });
+}
+
+/**
+ * Create MSW handlers for workers endpoints.
+ */
+export function createWorkersHandlers(workers: Worker[]) {
+  return [http.get('*/v1/agl/workers', ({ request }) => HttpResponse.json(buildWorkersResponse(workers, request)))];
+}
+
 /**
  * Create MSW handlers for resources endpoints.
  *

dashboard/test-utils/python-server.py

@@ -33,10 +33,11 @@ from agentlightning.types import (
     RolloutConfig,
     Span,
     TraceStatus,
+    Worker,
 )
 
 
-def inject_mock_data(store: InMemoryLightningStore, now: float | None = None) -> None:
+async def inject_mock_data(store: InMemoryLightningStore, now: float | None = None) -> None:
     """
     Inject mock data directly into the InMemoryLightningStore.
 
@@ -216,20 +217,12 @@ def inject_mock_data(store: InMemoryLightningStore, now: float | None = None) ->
     )
 
     # Inject rollouts directly into store
-    store._rollouts["ro-story-001"] = rollout1
-    store._rollouts["ro-story-002"] = rollout2
-    store._rollouts["ro-story-003"] = rollout3
-    store._rollouts["ro-story-004"] = rollout4
-    store._rollouts["ro-story-005"] = rollout5
-    store._rollouts["ro-story-006"] = rollout6
+    await store.collections.rollouts.insert([rollout1, rollout2, rollout3, rollout4, rollout5, rollout6])
 
     # Inject attempts directly into store
-    store._attempts["ro-story-001"] = [attempt1]
-    store._attempts["ro-story-002"] = [attempt2_1, attempt2_2]
-    store._attempts["ro-story-003"] = [attempt3_1, attempt3_2, attempt3_3]
-    store._attempts["ro-story-004"] = []  # No attempt for preparing rollout
-    store._attempts["ro-story-005"] = [attempt5]
-    store._attempts["ro-story-006"] = [attempt6]
+    await store.collections.attempts.insert(
+        [attempt1, attempt2_1, attempt2_2, attempt3_1, attempt3_2, attempt3_3, attempt5, attempt6]
+    )
 
     # Create and inject spans with diverse data
     # Spans for ro-story-001 (Running) - Multiple nested spans with ongoing execution
@@ -544,11 +537,7 @@ def inject_mock_data(store: InMemoryLightningStore, now: float | None = None) ->
         ),
     ]
 
-    store._spans["ro-story-001"] = spans_ro1
-    store._spans["ro-story-002"] = spans_ro2_a1 + spans_ro2_a2
-    store._spans["ro-story-003"] = spans_ro3_a3
-    store._spans["ro-story-005"] = spans_ro5
-    store._spans["ro-story-006"] = spans_ro6
+    await store.collections.spans.insert(spans_ro1 + spans_ro2_a1 + spans_ro2_a2 + spans_ro3_a3 + spans_ro5 + spans_ro6)
 
     # Create and inject resources with diverse types
     resource1 = ResourcesUpdate(
@@ -627,13 +616,70 @@ def inject_mock_data(store: InMemoryLightningStore, now: float | None = None) ->
         },
     )
 
-    store._resources["rs-story-001"] = resource1
-    store._resources["rs-story-002"] = resource2
-    store._resources["rs-story-003"] = resource3
-    store._resources["rs-story-004"] = resource4
-    store._resources["rs-story-005"] = resource5
+    await store.collections.resources.insert([resource1, resource2, resource3, resource4, resource5])
     store._latest_resources_id = "rs-story-005"
 
+    # Register workers with diverse states and activity windows.
+    workers = [
+        Worker(
+            worker_id="worker-east",
+            status="busy",
+            heartbeat_stats={"queue_depth": 2, "gpu_utilization": 0.82},
+            last_heartbeat_time=now - 20,
+            last_dequeue_time=now - 60,
+            last_busy_time=now - 120,
+            last_idle_time=now - 600,
+            current_rollout_id="ro-story-001",
+            current_attempt_id="at-story-010",
+        ),
+        Worker(
+            worker_id="worker-north",
+            status="idle",
+            heartbeat_stats={"queue_depth": 0, "gpu_utilization": 0.15},
+            last_heartbeat_time=now - 90,
+            last_dequeue_time=now - 3600,
+            last_busy_time=now - 5400,
+            last_idle_time=now - 5400,
+            current_rollout_id=None,
+            current_attempt_id=None,
+        ),
+        Worker(
+            worker_id="worker-west",
+            status="busy",
+            heartbeat_stats={"queue_depth": 1, "gpu_utilization": 0.41},
+            last_heartbeat_time=now - 45,
+            last_dequeue_time=now - 300,
+            last_busy_time=now - 200,
+            last_idle_time=now - 4800,
+            current_rollout_id="ro-story-003",
+            current_attempt_id="at-story-033",
+        ),
+        Worker(
+            worker_id="worker-south",
+            status="idle",
+            heartbeat_stats={"queue_depth": 0},
+            last_heartbeat_time=now - 900,
+            last_dequeue_time=now - 7200,
+            last_busy_time=now - 8600,
+            last_idle_time=now - 8600,
+            current_rollout_id=None,
+            current_attempt_id=None,
+        ),
+        Worker(
+            worker_id="worker-observer",
+            status="unknown",
+            heartbeat_stats={"queue_depth": 0},
+            last_heartbeat_time=now - 15,
+            last_dequeue_time=now - 4000,
+            last_busy_time=None,
+            last_idle_time=None,
+            current_rollout_id=None,
+            current_attempt_id=None,
+        ),
+    ]
+
+    await store.collections.workers.insert(workers)
+
 
 async def main():
     parser = argparse.ArgumentParser(description="Run a Python server for the LightningStore")
@@ -641,7 +687,7 @@ async def main():
     args = parser.parse_args()
 
     store = InMemoryLightningStore()
-    inject_mock_data(store, now=args.now)
+    await inject_mock_data(store, now=args.now)
 
     # Start server
     server = LightningStoreServer(store, "127.0.0.1", 8765, "*")

docker/Dockerfile.dev

@@ -0,0 +1,39 @@
+# Use a Python image with uv pre-installed
+FROM ghcr.io/astral-sh/uv:python3.12-bookworm
+
+# Setup a non-root user
+RUN groupadd --system --gid 999 nonroot \
+ && useradd --system --gid 999 --uid 999 --create-home nonroot
+
+# Install the project into `/app`
+WORKDIR /app
+
+# Enable bytecode compilation
+ENV UV_COMPILE_BYTECODE=1
+
+# Copy from the cache instead of linking since it's a mounted volume
+ENV UV_LINK_MODE=copy
+
+# Ensure installed tools can be executed out of the box
+ENV UV_TOOL_BIN_DIR=/usr/local/bin
+
+# Install the project's dependencies using the lockfile and settings
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,source=uv.lock,target=uv.lock \
+    --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
+    uv sync --locked --no-install-project --group dev --extra mongo --group core-stable
+
+# Then, add the rest of the project source code and install it
+# Installing separately from its dependencies allows optimal layer caching
+COPY . /app
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv sync --locked --group dev --extra mongo --group core-stable
+
+# Place executables in the environment at the front of the path
+ENV PATH="/app/.venv/bin:$PATH"
+
+# Reset the entrypoint, don't invoke `uv`
+ENTRYPOINT []
+
+# Use the non-root user to run our application
+USER nonroot

docker/compose.mongo.yml

@@ -0,0 +1,23 @@
+# Docker-compose file to launch a MongoDB server for development.
+# It's used to test the MongoDB store implementation.
+
+services:
+
+  mongo:
+    image: mongo:latest
+    ulimits:
+      nofile:
+        soft: 65535
+        hard: 65535
+    ports:
+      - "27017:27017"
+    command: ["mongod", "--bind_ip_all", "--replSet", "rs0"]
+    volumes:
+      - ../scripts/mongodb_init_rs_host.js:/docker-entrypoint-initdb.d/init-rs.js:ro
+      - ./data/mongo-host:/data/db
+    healthcheck:
+      test: ["CMD", "mongosh", "--eval", "db.adminCommand('ping')"]
+      interval: 10s
+      timeout: 5s
+      retries: 5
+      start_period: 30s

docker/compose.prometheus-memory-store.yml

@@ -0,0 +1,33 @@
+services:
+
+  app:
+    extends:
+      file: compose.store.yml
+      service: app
+
+    command: agl store --host 0.0.0.0 --port 4747 --prometheus --backend memory
+
+  node-exporter:
+    image: prom/node-exporter:latest
+    # In CI you might not have full /proc, but this is OK for container-level stats
+    pid: "host"
+    network_mode: "service:app" # share network with app for simplicity
+    command:
+      - '--path.rootfs=/host'
+    volumes:
+      - '/:/host:ro,rslave'
+
+  prometheus:
+    image: prom/prometheus:latest
+    command:
+      - '--config.file=/etc/prometheus/prometheus.yml'
+      - '--storage.tsdb.path=/prometheus'
+      - '--storage.tsdb.retention.time=1h'
+    volumes:
+      - ./prometheus.memory-store.yml:/etc/prometheus/prometheus.yml:ro
+      - ./data/prometheus:/prometheus
+    depends_on:
+      - app
+      - node-exporter
+    ports:
+      - "9090:9090"

docker/compose.prometheus-mongo-store.yml

@@ -0,0 +1,62 @@
+services:
+
+  mongo:
+    extends:
+      file: compose.mongo.yml
+      service: mongo
+
+    hostname: mongo
+    volumes:
+      - ./data/mongo-container:/data/db
+      - ../scripts/mongodb_init_rs_profiling.js:/docker-entrypoint-initdb.d/init-rs.js:ro
+
+    # This forces "mongo" to resolve to localhost ONLY inside this container.
+    # This allows rs.initiate to succeed using the hostname "mongo".
+    extra_hosts:
+      - "mongo:127.0.0.1"
+
+  app:
+    extends:
+      file: compose.store.yml
+      service: app
+
+    depends_on:
+      - mongo
+
+    command: agl store --host 0.0.0.0 --port 4747 --prometheus --backend mongo --mongo-uri mongodb://mongo:27017/?replicaSet=rs0 --n-workers 4
+
+  mongodb-exporter:
+    image: percona/mongodb_exporter:0.47.1
+    command:
+      - '--mongodb.uri=mongodb://mongo:27017/'
+      - '--collect-all'
+      - '--mongodb.collstats-colls=agentlightning.rollouts,agentlightning.attempts,agentlightning.spans,agentlightning.resources,agentlightning.workers,agentlightning.rollout_queue,agentlightning.span_sequence_ids'
+    depends_on:
+      - mongo
+    ports:
+      - "9216:9216"
+
+  node-exporter:
+    image: prom/node-exporter:latest
+    # In CI you might not have full /proc, but this is OK for container-level stats
+    pid: "host"
+    command:
+      - '--path.rootfs=/host'
+    volumes:
+      - '/:/host:ro,rslave'
+
+  prometheus:
+    image: prom/prometheus:latest
+    command:
+      - '--config.file=/etc/prometheus/prometheus.yml'
+      - '--storage.tsdb.path=/prometheus'
+      - '--storage.tsdb.retention.time=1h'
+    volumes:
+      - ./prometheus.mongo-store.yml:/etc/prometheus/prometheus.yml:ro
+      - ./data/prometheus:/prometheus
+    depends_on:
+      - app
+      - mongodb-exporter
+      - node-exporter
+    ports:
+      - "9090:9090"

docker/compose.store.yml

@@ -0,0 +1,26 @@
+services:
+
+  app:
+    build:
+      context: ../
+      dockerfile: docker/Dockerfile.dev
+
+    ports:
+      - "4747:4747"
+
+    command: agl store --host 0.0.0.0 --port 4747
+
+    develop:
+      watch:
+        # Sync the working directory with the `/app` directory in the container
+        - action: sync
+          path: ..
+          target: /app
+          # Exclude the project virtual environment — it could be for a
+          # different platform in the container
+          ignore:
+            - .venv/
+
+        # Rebuild the image if dependencies change by checking uv.lock
+        - action: rebuild
+          path: ../uv.lock

docker/prometheus.memory-store.yml

@@ -0,0 +1,17 @@
+global:
+  scrape_interval: 2s
+  evaluation_interval: 2s
+
+scrape_configs:
+  - job_name: app
+    static_configs:
+      - targets: ["app:4747"]
+    metrics_path: /v1/prometheus
+
+  - job_name: mongodb
+    static_configs:
+      - targets: ["mongodb-exporter:9216"]
+
+  - job_name: node
+    static_configs:
+      - targets: ["node-exporter:9100"]

docker/prometheus.mongo-store.yml

@@ -0,0 +1,13 @@
+global:
+  scrape_interval: 2s
+  evaluation_interval: 2s
+
+scrape_configs:
+  - job_name: app
+    static_configs:
+      - targets: ["app:4747"]
+    metrics_path: /v1/prometheus
+
+  - job_name: node
+    static_configs:
+      - targets: ["node-exporter:9100"]

docker/setup.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+
+set -euo pipefail
+
+# Create data directories
+mkdir -p data/prometheus data/mongo-container data/mongo-host
+
+# Change permissions
+chmod 777 data/prometheus data/mongo-container data/mongo-host

docs/changelog.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## Agent-lightning v0.2.2 (11/12/2025)
+
+Agent-lightning v0.2.2 is a stabilization release for v0.2.1. It introduces several bug fixes.
+
+* Fix compatibility issues with VERL 0.6.0.
+* Fix model name for pre-downloaded models in VERL.
+* Fix preparing status transition on rollout when creating attempts.
+* Fix OpenAI Agents SDK compatibility issues.
+
+**Full Changelog**: https://github.com/microsoft/agent-lightning/compare/v0.2.1...v0.2.2
+
+---
+
 ## Agent-lightning v0.2.1 (10/30/2025)
 
 Agent-lightning v0.2.1 is a stabilization release for v0.2.0. It introduces several bug fixes and new features, plus a number of unlisted CI improvements.

docs/community/contributing.md

@@ -1,18 +1,98 @@
 # Contributing Guide
 
-Agent Lightning thrives on community improvements, whether you are polishing docs, fixing bugs, or building new features. This guide shows the shortest path from cloning the repository to shipping a polished pull request.
+Agent Lightning gets better every time someone files a clear bug, polishes docs, improves tests, or lands a new feature. This guide collects the expectations, checklists, and tips that help you go from “I have an idea” to “my pull request just merged.”
 
-## Step 1. Prepare Your Environment
+## Before You Start
 
-### Prerequisites
+Agent-lightning is built by a small Microsoft Research team with limited reviewer hours and GPU budget. For any sizeable change (new algorithm, example, or API surface) please first discuss scope with us in [Discord](https://discord.gg/RYk7CdvDR7). Early alignment keeps your effort from being blocked late in the process.
 
-- **Python** 3.10 or newer (we test on 3.10–3.13).
-- **uv** for dependency and virtual environment management. Install it from the [official uv docs](https://docs.astral.sh/uv/getting-started/installation/).
+## Where You Can Help
+
+Pick a lane, or combine several. Just keep the discussion-first principle in mind for anything non-trivial.
+
+### Documentation Improvements
+
+Documentation improvements are the easiest way to get started. You can find more about how to write good documentations and organize documentations in the following sections. Here are some general contribution points we can think of:
+
+- Tighten language, fix typos, clarify confusing sections, or add missing links. Fresh eyes catch docs gaps best.
+- Organize content using the directories listed below so readers can actually find it.
+- Avoid duplicate prose, unrelated “how-to” guides, or translations (we cannot maintain them today).
+
+!!! note "Changes that are usually rejected"
+
+    - Copy/pasting existing docs with shallow edits.
+    - Adding a `how-to` guide that is not tied to a new example.
+    - Adding doc translations to other languages (no capacity to review/maintain yet).
+
+### Bug Fixes
+
+Bug fixes are the fastest way to get familiar with the codebase. To get started, you can:
+
+- Browse the ["good first issue"](https://github.com/microsoft/agent-lightning/labels/good%20first%20issue) and ["bug"](https://github.com/microsoft/agent-lightning/labels/bug) labels; drop a comment before you start so we can mark it as taken.
+- For fresh bugs, open an issue with reproduction steps, logs, and expected behavior before submitting a fix.
+- Keep each pull request focused, ideally avoiding breaking API changes. Larger refactors should be discussed via RFC or maintainer sync.
+
+### New Examples
+
+Examples must be curated so that we can maintain them. We generally merge only those that meet at least one (ideally several) of these criteria:
+
+- Demonstrates an agent framework or workflow that is materially different from what already exists. ([LangChain](https://www.langchain.com/) vs. [LlamaIndex](https://www.llamaindex.ai/) is not different enough; [LangChain](https://www.langchain.com/) vs. [n8n](https://n8n.io/) or [Vercel AI SDK](https://ai-sdk.dev/) is, because they either have different orchestration paradigms or differ in programming languages.)
+- Shows measurable performance gains on a **real-world** problem with a **real-world** dataset, such as tuning a search agent with Google Search API or improving a coding agent’s (e.g., Claude Code) SWE-Bench score.
+- Integrates a new algorithm, training backend, or serving stack (see “New Algorithms” below).
+- Validates scenarios that are rarely tested, such as multi-modality agents or long-lived memory/workflow agents.
+
+Bonus points for examples that:
+
+- Ship CI or self-test coverage so we know they still work as the core evolves.  **Otherwise, we would have to mark the example as unmaintained because we won't be able to test the examples manually before each release.**
+- Include a [`docs/how-to/`]({{ src("docs/how-to/") }}) guide (or a detailed README if no how-to exists) without duplicating content in multiple places.
+- Favor simple, dependency-light code over heavy abstractions.
+
+!!! warning "Please discuss first"
+
+    Examples tend to be the most time-consuming contributions for both you and reviewers. Sync with us on Discord or through an issue before diving into a new one.
+
+### Fresh Implementations of Core Modules
+
+If you are looking to extend [`Runner`][agentlightning.Runner], [`Tracer`][agentlightning.Tracer], [`Adapter`][agentlightning.Adapter], [`LightningStore`][agentlightning.LightningStore], or another core interface, here are the steps:
+
+1. File an issue or proposal first.
+2. Explain which interface you are extending, why existing implementations are insufficient, and how you intend to test compatibility with the rest of the stack (unit tests, documentation updates, example refreshes, etc.).
+3. Any API changes must be reviewed up front. DO NOT begin coding large changes before the discussion lands!
+
+### New Algorithms
+
+If you are integrating a new training/serving backend, check whether it already lives in the [Algorithm Zoo](../algorithm-zoo/index.md) or is covered in the [Examples Catalog](../how-to/examples-catalog.md). We especially welcome:
+
+- Currently unsupported or under-tested algorithms such as Supervised Fine-tuning (SFT), Direct Policy Optimization (DPO), or Monte Carlo Tree Search (MCTS).
+- Tuning [Resource][agentlightning.Resource]s that are not supported yet, such as workflows or memory.
+- Expansions of supported stacks, e.g., adding multi-modality to APO or multi-agent prompt tuning.
+- Reinforcement-learning integrations beyond our current stack of [VERL](https://github.com/volcengine/verl), [vLLM](https://vllm.ai/), [Azure OpenAI](https://azure.microsoft.com/en-us/products/ai-foundry/models/openai), and [Tinker](https://tinker-docs.thinkingmachines.ai/). Contributions using [SGLang](https://github.com/sgl-project/sglang), [TRL](https://github.com/huggingface/trl), [SkyRL](https://github.com/NovaSky-AI/SkyRL), [RLinf](https://github.com/RLinf/RLinf), [litgpt](https://github.com/Lightning-AI/litgpt), or similar are welcome.
+
+Most brand-new algorithms ultimately land as “new examples,” so read that section too. Post an issue or design doc to scope the work, reuse existing utilities, and avoid duplicating efforts. Mature, battle-tested examples graduate into the [Algorithm Zoo](../algorithm-zoo/index.md).
+
+### Ecosystem Projects
+
+Have a project that builds on Agent-lightning but does not belong in the main repo? Fork it or depend on it externally, then let us know. We can showcase notable projects in [Community Projects](../index.md) and the main [README]({{ src("README.md") }}).
+
+### Other Contribution Ideas
+
+- **Tests.** Add or improve cases in [`tests/`]({{ src("tests") }}) (unit, integration, or end-to-end).
+- **Benchmarks.** Expand [`tests/benchmark`]({{ src("tests/benchmark") }}) to stress large-scale training or rollouts.
+- **Issue triage.** Reproduce bugs, confirm whether they reproduce on `main`, or suggest short-term mitigations so maintainers can prioritize.
+
+## Contribution Workflow
+
+The steps below keep changes reviewable and CI-friendly. Follow them in order; rerun the relevant pieces if you revisit a branch later.
+
+### 1. Prepare Your Environment
+
+Minimum tooling:
+
+- **Python** 3.10+ (3.12 recommended).
+- **uv** for dependency and virtual-environment management. Install it using the [official uv docs](https://docs.astral.sh/uv/getting-started/installation/).
 - **Git** configured with your GitHub credentials.
 
-### Clone the Repository
-
-Fork the repo, then clone your fork and register the upstream remote so you can stay current:
+Clone your fork and point `upstream` at the official repo:
 
 ```bash
 git clone git@github.com:<your-username>/agent-lightning.git
@@ -20,15 +100,13 @@ cd agent-lightning
 git remote add upstream https://github.com/microsoft/agent-lightning.git
 ```
 
-### Install Dependencies
-
-Install the standard development toolchain:
+Install the default development stack:
 
 ```bash
 uv sync --group dev
 ```
 
-Want GPU extras, example dependencies, or other optional features? Pin everything in one pass:
+Need GPU extras or specific optional dependencies? Lock them in with one command:
 
 ```bash
 uv sync --frozen \
@@ -41,28 +119,22 @@ uv sync --frozen \
     --no-default-groups
 ```
 
-After `uv sync`, run commands with `uv run ...` (or `uv run --no-sync` once the environment is locked), or activate the virtual environment in `.venv/`.
+After `uv sync`, run commands via `uv run ...` (add `--no-sync` once the environment is locked) or activate `.venv/`.
 
----
+### 2. Install and Run Pre-commit
 
-## Step 2. Install and Run Pre-commit
-
-We enforce formatting and linting with [pre-commit](https://pre-commit.com/). Install the hooks once, then run them before every push:
+Formatting and linting are enforced through [pre-commit](https://pre-commit.com/). Install once, then run before each push:
 
 ```bash
 uv run pre-commit install
-
-# The following will auto-run if you have set up the pre-commit hooks to run automatically on commit.
 uv run pre-commit run --all-files --show-diff-on-failure --color=always
 ```
 
-Running them locally saves a CI round-trip and keeps diffs tidy.
+Once installed, the hooks run automatically on every `git commit`. Running the pre-commit hooks locally keeps CI green and diffs manageable.
 
----
+### 3. Branch From a Fresh `main`
 
-## Step 3. Branching Workflow
-
-Start from a fresh `main`, then branch for your change:
+Start all work from the latest upstream state:
 
 ```bash
 git fetch upstream
@@ -70,20 +142,32 @@ git checkout main
 git merge upstream/main
 ```
 
-Create a topic branch with one of these prefixes:
+Branch naming convention:
 
-- `feature/<short-description>` for new features
-- `fix/<short-description>` for bug fixes
-- `docs/<short-description>` for documentation-only work
-- `chore/<short-description>` for tooling or maintenance
+- `feature/<short-description>` for new features.
+- `fix/<short-description>` for bug fixes.
+- `docs/<short-description>` for documentation-only updates.
+- `chore/<short-description>` for tooling or maintenance.
 
-Stick to lowercase words separated by hyphens, e.g. `feature/async-runner-hooks`.
+Use lowercase with hyphens, e.g., `feature/async-runner-hooks`.
 
----
+!!! note "Where should docs or examples live?"
 
-## Step 4. Test Your Changes
+    Many new contributors get confused about what to put in the `docs/how-to/` directory and what to put in the `examples/` directory (particularly README files). Here is a quick reference you can refer to:
 
-Most updates should ship with automated checks. Preface commands with `uv run` so they use the project environment.
+    | Location | Description |
+    | --- | --- |
+    | `docs/algorithm-zoo/` | Documentation for **built-in algorithms** shipped with Agent-lightning. |
+    | `docs/how-to/` | Step-by-step **how-to guides**, usually tied to an example in `examples/`. |
+    | `docs/tutorials/` | Conceptual walkthroughs for components or workflows. See [debugging](../tutorials/debug.md) or [parallelization](../tutorials/parallelize.md) for examples. |
+    | `docs/deep-dive/` | Advanced explanations and in-depth concepts. |
+    | `examples/<name>/README.md` | Example-specific README. If any related how-to if that exists, link to it avoid duplicating the same instructions twice; write only brief instructions on how to install and run the example. Otherwise, you can make the README more detailed and self-explanatory. |
+
+    Remember to register new docs in [`mkdocs.yml`]({{ src("mkdocs.yml") }}), add examples to [examples/README]({{ src("examples/README.md") }}), and update the [Examples Catalog](../how-to/examples-catalog.md).
+
+### 4. Test and Validate
+
+Most contributions require automated checks. Prefix commands with `uv run` so they use the project environment.
 
 **Full test suite**
 
@@ -97,53 +181,43 @@ uv run pytest -v
 uv run pytest tests/path/to/test_file.py -k test_name
 ```
 
-**Optional/gated tests**
+**Optional/gated tests:** GPU-specific suites or API-dependent tests run automatically when the required hardware or environment variables (such as `OPENAI_API_KEY`) are present.
 
-GPU-specific suites or API-dependent tests run automatically when the required hardware or environment variables (such as `OPENAI_API_KEY`) are present.
-
-**Static analysis**
+**Static analysis:**
 
 ```bash
 uv run pyright
 ```
 
-Touching code under `examples/`? Each directory includes a README with example-specific smoke tests—run those too.
+If you have touched code under `examples/`, you should run the example-specific smoke tests. Each directory includes a README with example-specific smoke tests—run those too.
 
----
+!!! note "Build documentation when needed"
 
-## Step 5. Build Documentation (When Applicable)
+    Keep API references under [docs/reference]({{ src("docs/reference/") }}) up to date. Doc-only changes should still build cleanly:
 
-Doc changes should build cleanly before you push:
+    ```bash
+    uv run mkdocs serve --strict   # live reload
+    uv run mkdocs build --strict   # CI-equivalent
+    ```
 
-```bash
-uv run mkdocs serve --strict  # live reload while editing
-uv run mkdocs build --strict  # CI-equivalent validation
-```
+    `--strict` elevates warnings to errors so you catch issues before CI.
 
-`--strict` matches CI and promotes warnings to errors so you catch them early.
-
----
-
-## Step 6. Final Local Checks
+Before opening a PR, double-check the basics:
 
+- Run `uv lock` if you changed dependencies.
 - Run `uv run pre-commit run --all-files` (hooks installed via `pre-commit install` run automatically on `git commit`, but rerun them if you amended history).
-- Execute the relevant test commands from Step 4.
-- Validate any affected examples by following the instructions in `examples/<name>/README`.
+- Execute the relevant commands from the test list above.
+- Validate each affected example via its README instructions.
 
----
+### 5. Open a Pull Request
 
-## Step 7. Open a Pull Request
-
-1. Push your branch to your fork:
+1. Push your branch:
    ```bash
    git push origin <branch-name>
    ```
 2. Open a PR against `microsoft/agent-lightning:main`.
-3. Complete the PR template with:
-    - A concise summary of the change.
-    - The tests or commands you ran (copy from Step 4/6).
-    - Linked issues (use `Fixes #123` to auto-close).
-4. Attach screenshots or terminal output when it clarifies behavior.
-5. Address review feedback promptly. Use focused commits, and consider `git commit --fixup` for follow-up adjustments.
+3. Fill out the template with a concise summary, the commands/tests you ran, and linked issues (use `Fixes #123` syntax to auto-close).
+4. Include screenshots or logs if they clarify behavior.
+5. Address review feedback promptly. Follow-up tweaks work best as focused commits; `git commit --fixup` is handy for reviewer-suggested edits.
 
-Thanks for contributing—every improvement grows the Agent Lightning community!
+Thanks for contributing! every improvement strengthens the Agent Lightning community!

docs/deep-dive/store.md

@@ -72,6 +72,15 @@ Each attempt begins in **preparing**, created either when a rollout is dequeued
 
 This simple model allows the system to distinguish between normal termination, abnormal stalling, and recoverable interruption without additional state flags.
 
+## Worker Telemetry
+
+Workers track runner-level activity timestamps (`last_heartbeat_time`, `last_dequeue_time`, `last_busy_time`, `last_idle_time`) plus their current rollout assignment. Those fields are now derived automatically:
+
+- [`dequeue_rollout(worker_id=...)`][agentlightning.LightningStore.dequeue_rollout] records which worker polled the queue and refreshes `last_dequeue_time`.
+- [`update_attempt(..., worker_id=...)`][agentlightning.LightningStore.update_attempt] drives the worker status machine. Assigning an attempt marks the worker **busy** and stamps `last_busy_time`; finishing with `status in {"succeeded","failed"}` switches to **idle**, while watchdog transitions such as `timeout`/`unresponsive` make the worker **unknown** and clear `current_rollout_id` / `current_attempt_id`.
+- [`update_worker(...)`][agentlightning.LightningStore.update_worker] is reserved for heartbeats. It snapshots optional `heartbeat_stats` and always updates `last_heartbeat_time`.
+
+Because every transition flows through these APIs, worker status is derived automatically from rollout execution and heartbeats. Note, however, that calling `update_worker` with a new `worker_id` will create a new worker record with status "unknown" if one does not exist. Thus, while manual status changes are not allowed, new worker records can be created externally via heartbeats.
 
 ## Rollout Transition Map
 
@@ -143,6 +152,13 @@ Programmatically this is encapsulated by [`Span.from_opentelemetry(readable_span
 
     [`add_span`][agentlightning.LightningStore.add_span] or [`add_otel_span`][agentlightning.LightningStore.add_otel_span] both appends a span *and* acts as a heartbeat that can revive `unresponsive` → `running`.
 
+## OTLP Compatibility
+
+Some of the LightningStore implementations support exporting traces via the [OTLP/HTTP specification](https://opentelemetry.io/docs/specs/otlp/). For example, [`LightningStoreServer`][agentlightning.LightningStoreServer] exposes `/v1/traces` endpoint, it implements the binary Protobuf variant defined by the spec, including the required `Content-Type: application/x-protobuf`, optional `Content-Encoding: gzip`, and status responses encoded as `google.rpc.Status`. Agent-lightning helps parsing `ExportTraceServiceRequest` messages, validate identifiers, normalize resource metadata, and allocate sequence
+numbers so store implementations only need to persist [`Span`][agentlightning.Span] objects in order.
+
+Because the interface speaks standard OTLP, any OpenTelemetry-compatible SDK or collector can emit spans directly to a LightningStore OTLP endpoint without custom shims. The server responds according to the OTLP contract (status code, encoding, and error payloads), which keeps Agent-lightning interoperable with existing observability tooling. This compatibility serves as a strong complement to the OpenTelemetry conversion discussed above.
+
 ## Store Implementations
 
 Currently, the only out-of-the-box implementation is [`InMemoryLightningStore`][agentlightning.InMemoryLightningStore]:
@@ -153,6 +169,8 @@ Currently, the only out-of-the-box implementation is [`InMemoryLightningStore`][
 
 For production you will likely want persistence. We’re actively building a SQLite-backed store that keeps the same API surface while adding durability, crash recovery, and better historical span queries. If you need something sooner, implement your own store by subclassing [`LightningStore`][agentlightning.LightningStore] and providing concrete storage for the small set of abstract methods (`enqueue_rollout`, `dequeue_rollout`, `update_attempt`, `add_span`, etc.). This document plus the tests in `tests/store/` illustrate the expected behavior.
 
+Different store implementations may have different capabilities. For example, [`InMemoryLightningStore`][agentlightning.InMemoryLightningStore] does not support exporting traces via OTLP. Try to distinguish the capabilities of a store implementation by checking the [`capabilities`][agentlightning.LightningStore.capabilities] property.
+
 ## Thread Safety
 
 **[`LightningStoreThreaded`][agentlightning.LightningStoreThreaded]** is a subclass of [`LightningStore`][agentlightning.LightningStore] that wraps another underlying store to make a store instance safe for multi-threaded callers. It wraps every state-mutating call in a mutex. Specifically:
@@ -187,7 +205,7 @@ await server.start()      # starts uvicorn in a daemon thread and waits for /hea
 # Client (same or different process)
 client = agl.LightningStoreClient("http://localhost:4747")
 
-print(await client.query_rollouts(status=["queuing"]))
+print(await client.query_rollouts(status_in=["queuing"]))
 
 await client.close()
 await server.stop()

docs/how-to/examples-catalog.md

@@ -0,0 +1,81 @@
+# Examples Catalog
+
+!!! tip "Want to Contribute?"
+
+    We welcome contributions to the examples catalog! Please refer to the [Contributing](../community/contributing.md) guide for more details.
+
+<div class="grid cards" markdown>
+
+-   :material-robot:{ .lg .middle } __APO room selector__
+
+    ---
+
+    Prompt-optimize a room-booking agent with the built-in APO algorithm, then contrast it with the write-your-own algorithm and debugging workflows in the tutorials. Pairs well with the [Train the First Agent how-to]({{ src("docs/how-to/train-first-agent.md") }}) and the [Write the First Algorithm guide]({{ src("docs/how-to/write-first-algorithm.md") }}).
+
+    [:octicons-repo-24: Browse source]({{ src("examples/apo") }})
+
+-   :material-cloud-sync:{ .lg .middle } __Azure OpenAI SFT__
+
+    ---
+
+    Run a supervised fine-tuning loop against Azure OpenAI: roll out the capital-lookup agent, turn traces into JSONL, launch fine-tunes, and redeploy the resulting checkpoints through Azure CLI.
+
+    [:octicons-repo-24: Browse source]({{ src("examples/azure") }})
+
+-   :material-calculator:{ .lg .middle } __Calc-X VERL math__
+
+    ---
+
+    VERL-based reinforcement learning setup for a math-reasoning agent that uses AutoGen plus an MCP calculator tool to solve Calc-X problems end to end.
+
+    [:octicons-repo-24: Browse source]({{ src("examples/calc_x") }})
+
+-   :material-view-grid:{ .lg .middle } __Minimal building blocks__
+
+    ---
+
+    Bite-sized scripts that isolate Agent-lightning primitives (e.g., LightningStore usage, LLM proxying, minimal vLLM host) so you can study each part before composing larger workflows.
+
+    [:octicons-repo-24: Browse source]({{ src("examples/minimal") }})
+
+-   :material-book-open-page-variant:{ .lg .middle } __RAG (MuSiQue)__
+
+    ---
+
+    Retrieval-Augmented Generation pipeline that preps a Wikipedia retriever via MCP and trains a MuSiQue QA agent with GRPO. Documented for historical reference (verified on Agent-lightning v0.1.x).
+
+    [:octicons-repo-24: Browse source]({{ src("examples/rag") }})
+
+-   :material-magnify:{ .lg .middle } __Search-R1 RL__
+
+    ---
+
+    Reproduction of the Search-R1 workflow that prepares its own retrieval backend, runs the rollout script, and coordinates GRPO-style training without extra orchestration layers (last validated on v0.1.x).
+
+    [:octicons-repo-24: Browse source]({{ src("examples/search_r1") }})
+
+-   :material-database:{ .lg .middle } __Spider SQL agent__
+
+    ---
+
+    LangGraph-powered text-to-SQL workflow for the Spider benchmark, combining LangChain tooling with Agent-lightning rollouts; follow along with the [how-to for training SQL agents]({{ src("docs/how-to/train-sql-agent.md") }}).
+
+    [:octicons-repo-24: Browse source]({{ src("examples/spider") }})
+
+-   :material-thought-bubble:{ .lg .middle } __Tinker integration__
+
+    ---
+
+    Adapter package ([`agl_tinker`]({{ src("examples/tinker/agl_tinker") }})) with Tinker plus sample CrewAI/OpenAI agents that feed Agent-lightning traces into Tinker’s reinforcement-learning backend for both toy and 20-Questions-style workflows.
+
+    [:octicons-repo-24: Browse source]({{ src("examples/tinker") }})
+
+-   :material-fast-forward:{ .lg .middle } __Unsloth SFT__
+
+    ---
+
+    Supervised fine-tuning loop that ranks math-agent rollouts, fine-tunes with Unsloth’s 4-bit LoRA stack, and mirrors the [Fine-tune with Unsloth recipe]({{ src("docs/how-to/unsloth-sft.md") }}).
+
+    [:octicons-repo-24: Browse source]({{ src("examples/unsloth") }})
+
+</div>

docs/how-to/train-sql-agent.md

@@ -321,6 +321,44 @@ For the LLaMA profile, export an `HF_TOKEN` before running so VERL can download
     env RAY_DEBUG=legacy HYDRA_FULL_ERROR=1 VLLM_USE_V1=1 ray start --head --dashboard-host=0.0.0.0
     ```
 
+!!! note "Launching Training with NPUs"
+
+    The example also supports running with **Huawei Ascend NPUs**. This feature is contributed by [Teams from Huawei](https://github.com/microsoft/agent-lightning/pull/272). To use it, resort to the function `config_train_npu` in the script.
+
+    **Hardware Supported:** Atlas 200T A2 Box16, Atlas 900 A2 PODc, Atlas 800T A3. At least **a single 40GB NPU** is required to run the **Qwen2.5-Coder-1.5B-Instruct** model.
+
+    **Environment Setup:** Python 3.11.13, CANN 8.2.RC1, torch 2.7.1+cpu, torch_npu 2.7.1.dev20250724. For basic environment preparation, please refer to this [document](https://gitcode.com/Ascend/pytorch).
+
+    Before installing dependencies, configure the following pip mirrors:
+
+    ```bash
+    pip config set global.index-url http://repo.huaweicloud.com/repository/pypi/simple
+    pip config set global.extra-index-url "https://download.pytorch.org/whl/cpu/ https://mirrors.huaweicloud.com/ascend/repos/pypi"
+    ```
+
+    Then install vLLM, vLLM-Ascend and VERL:
+
+    ```bash
+    pip install vllm==0.10.0 --trusted-host repo.huaweicloud.com
+    pip install vllm-Ascend==0.10.0rc1 --trusted-host repo.huaweicloud.com
+    pip install verl==0.5.0
+    ```
+
+    To ensure the VERL framework runs correctly on NPU, add the following lines to `verl/utils/vllm_utils.py`:
+
+    ```python
+    from vllm_ascend.patch import platform
+    from vllm_ascend.patch import worker
+    ```
+
+    See the following reference for more details: [https://github.com/vllm-project/vllm-ascend/issues/1776](https://github.com/vllm-project/vllm-ascend/issues/1776).
+
+    After the above dependencies have been installed, from [`examples/spider`]({{ src("examples/spider") }}) run the following script command:
+
+    ```bash
+    python train_sql_agent.py npu
+    ```
+
 ### Debugging the Agent without VERL
 
 [`sql_agent.py`]({{ src("examples/spider/sql_agent.py") }}) also provides a `debug_sql_agent()` helper to run the LangGraph workflow directly against a local or hosted OpenAI-compatible endpoint before using VERL.

docs/how-to/write-first-algorithm.md

@@ -99,7 +99,7 @@ async def find_best_prompt(store, prompts_to_test, task_input):
         await store.wait_for_rollouts([rollout.rollout_id])
 
         # 4. Query the completed rollout and its spans
-        completed_rollout = (await store.query_rollouts([rollout.rollout_id]))[0]
+        completed_rollout = await store.get_rollout_by_id(rollout.rollout_id)
         print(f"[Algo] Received Result: {completed_rollout.model_dump_json(indent=None)}")
 
         spans = await store.query_spans(rollout.rollout_id)

docs/reference/internal.md

@@ -28,11 +28,17 @@
 
 ::: agentlightning.llm_proxy.AddReturnTokenIds
 
+::: agentlightning.llm_proxy.StreamConversionMiddleware
+
+::: agentlightning.llm_proxy.MessageInspectionMiddleware
+
+::: agentlightning.llm_proxy.RolloutAttemptMiddleware
+
 ::: agentlightning.store.base.UNSET
 
 ::: agentlightning.store.utils.propagate_status
 
-::: agentlightning.tracer.agentops.LightningSpanProcessor
+::: agentlightning.tracer.otel.LightningSpanProcessor
 
 ::: agentlightning.utils.server_launcher.PythonServerLauncher
 
@@ -40,6 +46,10 @@
 
 ::: agentlightning.utils.server_launcher.LaunchMode
 
+::: agentlightning.utils.otlp.handle_otlp_export
+
+::: agentlightning.utils.otlp.spans_from_proto
+
 ## Deprecated APIs
 
 ::: agentlightning.server.AgentLightningServer

docs/reference/store.md

@@ -2,10 +2,14 @@
 
 ::: agentlightning.LightningStore
 
+::: agentlightning.LightningStoreCapabilities
+
 ## Store Implementations
 
 ::: agentlightning.InMemoryLightningStore
 
+::: agentlightning.CollectionBasedLightningStore
+
 ## Client-Server and Thread-safe Wrappers
 
 ::: agentlightning.LightningStoreServer
@@ -13,3 +17,21 @@
 ::: agentlightning.LightningStoreClient
 
 ::: agentlightning.LightningStoreThreaded
+
+## Collections and Collection Implementations
+
+::: agentlightning.store.collection.Collection
+
+::: agentlightning.store.collection.Queue
+
+::: agentlightning.store.collection.KeyValue
+
+::: agentlightning.store.collection.LightningCollections
+
+::: agentlightning.store.collection.ListBasedCollection
+
+::: agentlightning.store.collection.DequeBasedQueue
+
+::: agentlightning.store.collection.DictBasedKeyValue
+
+::: agentlightning.store.collection.InMemoryLightningCollections

docs/reference/trainer.md

@@ -23,3 +23,11 @@
 ## CLI Builder
 
 ::: agentlightning.lightning_cli
+
+## Logging
+
+::: agentlightning.configure_logger
+
+::: agentlightning.setup_module_logging
+
+::: agentlightning.setup_logging

docs/reference/types.md

@@ -26,8 +26,20 @@
 
 ::: agentlightning.AttemptedRollout
 
+::: agentlightning.Worker
+
+::: agentlightning.WorkerStatus
+
 ::: agentlightning.Hook
 
+::: agentlightning.PaginatedResult
+
+::: agentlightning.FilterOptions
+
+::: agentlightning.SortOptions
+
+::: agentlightning.FilterField
+
 ## Resources
 
 ::: agentlightning.Resource

docs/stylesheets/extra.css

@@ -66,7 +66,7 @@
   padding: 1em;
   font-weight: 500;
   position: relative;
-  z-index: 1000;
+  z-index: 3;
   border: 2px solid #FFD15D;
   border-radius: 0.25em;
   margin-bottom: 2em;
@@ -84,3 +84,24 @@ canvas[data-chart] {
   width: 100%;
   display: block;
 }
+
+/* Grid behavior */
+.md-typeset .grid {
+  grid-template-columns: repeat(auto-fit, minmax(24rem, 1fr));
+}
+
+/* Make cards fill equal height and push footer link to bottom */
+.md-typeset .grid.cards > ul > li {
+  display: flex;
+  flex-direction: column;
+  gap: 0;
+}
+
+.md-typeset .grid.cards > ul > li > hr {
+  margin: 0.5em 0;
+}
+
+.md-typeset .grid.cards > ul > li > :last-child {
+  margin-top: auto; /* pushes the last element (Browse source) to bottom */
+  padding-top: 0.5em;
+}

examples/README.md

@@ -4,13 +4,14 @@ This catalog highlights the examples shipped with Agent-lightning.
 
 | Example | Description | CI Maintenance |
 |---------|-------------|----------------|
-| [apo](./apo) | Automatic Prompt Optimization tutorials covering built-in, custom, and debugging workflows. | [![apo workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-apo.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-apo.yml) |
-| [azure](./azure) | Supervised fine-tuning with Azure OpenAI. | **Unmaintained** — last verified with Agent-lightning v0.2.1 |
-| [calc_x](./calc_x) | VERL-powered math reasoning agent training that uses AutoGen with an MCP calculator tool. | [![calc_x workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-calc-x.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-calc-x.yml) |
+| [apo](./apo) | Automatic Prompt Optimization tutorials covering built-in, custom, and debugging workflows. | [![apo workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-apo.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/examples-apo.yml) |
+| [azure](./azure) | Supervised fine-tuning with Azure OpenAI. | [![azure workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-azure.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/examples-azure.yml) |
+| [calc_x](./calc_x) | VERL-powered math reasoning agent training that uses AutoGen with an MCP calculator tool. | [![calc_x workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-calc-x.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/examples-calc-x.yml) |
+| [minimal](./minimal) | Bite-sized programs that demonstrate how individual Agent-lightning building blocks behave in isolation. | [![minimal 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) |
 | [rag](./rag) | Retrieval-Augmented Generation pipeline targeting the MuSiQue dataset with Wikipedia retrieval. | **Unmaintained** — last verified with Agent-lightning v0.1.1 |
 | [search_r1](./search_r1) | Framework-free Search-R1 reinforcement learning training workflow with a retrieval backend. | **Unmaintained** — last verified with Agent-lightning v0.1.2 |
-| [spider](./spider) | Text-to-SQL reinforcement learning training on the Spider dataset using LangGraph. | [![spider workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-spider.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-spider.yml) |
-| [unsloth](./unsloth) | Supervised fine-tuning example powered by Unsloth with 4-bit quantization and LoRA. | [![unsloth workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-unsloth.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/badge-unsloth.yml) |
-| [tinker](./tinker) | Reinforcement learning with Tinker as the backend training service. | **Unmaintained** — last verified with Agent-lightning v0.2.1 |
+| [spider](./spider) | Text-to-SQL reinforcement learning training on the Spider dataset using LangGraph. | [![spider workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-spider.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/examples-spider.yml) |
+| [tinker](./tinker) | Reinforcement learning with Tinker as the backend training service. | [![tinker workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-tinker.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/examples-tinker.yml) |
+| [unsloth](./unsloth) | Supervised fine-tuning example powered by Unsloth with 4-bit quantization and LoRA. | [![unsloth workflow status](https://github.com/microsoft/agent-lightning/actions/workflows/badge-unsloth.yml/badge.svg)](https://github.com/microsoft/agent-lightning/actions/workflows/examples-unsloth.yml) |
 
-*NOTE: CI status avoid taking any workflow running with latest dependencies into account. That's why we reference the corresponding `badge-*` workflows instead. Each example's own README also displays its `examples-*` workflow status whenever the project is maintained by CI.*
+*NOTE: CI status avoids taking any workflow running with latest dependencies into account. That's why we reference the corresponding `badge-*` workflows instead. Each example's own README also displays its `examples-*` workflow status whenever the project is maintained by CI.*

examples/apo/apo_custom_algorithm.py

@@ -29,7 +29,7 @@ python apo_custom_algorithm_trainer.py
 
 import argparse
 import asyncio
-from typing import List, Optional
+from typing import Optional, Sequence
 
 from openai import AsyncOpenAI
 from rich.console import Console
@@ -117,7 +117,7 @@ async def apo_rollout(task: str, prompt_template: agl.PromptTemplate) -> float:
     return await llm_judge(task, text)
 
 
-async def log_llm_span(spans: List[agl.Span]) -> None:
+async def log_llm_span(spans: Sequence[agl.Span]) -> None:
     """Logs the LLM related spans that records prompts and responses."""
     for span in spans:
         if "chat.completion" in span.name:
@@ -181,5 +181,5 @@ async def main():
 
 
 if __name__ == "__main__":
-    agl.configure_logger()
+    agl.setup_logging()
     asyncio.run(main())

examples/apo/apo_custom_algorithm_trainer.py

@@ -19,7 +19,7 @@ python apo_custom_algorithm.py runner
 from apo_custom_algorithm import apo_algorithm, apo_rollout
 from rich.console import Console
 
-from agentlightning import Trainer, configure_logger
+from agentlightning import Trainer, setup_logging
 from agentlightning.algorithm import algo
 from agentlightning.store import LightningStore
 
@@ -39,6 +39,6 @@ async def apo_algorithm_usable_in_trainer(*, store: LightningStore):
 
 
 if __name__ == "__main__":
-    configure_logger()
+    setup_logging()
     trainer = Trainer(n_workers=1, algorithm=apo_algorithm_usable_in_trainer)
     trainer.fit(apo_rollout)

examples/apo/apo_debug.py

@@ -8,7 +8,7 @@ from typing import cast
 
 from apo_custom_algorithm import apo_rollout
 
-from agentlightning import Trainer, configure_logger
+from agentlightning import Trainer, setup_logging
 from agentlightning.litagent import LitAgent
 from agentlightning.runner import LitAgentRunner
 from agentlightning.store import InMemoryLightningStore
@@ -105,7 +105,7 @@ def debug_with_trainer():
 
 
 if __name__ == "__main__":
-    configure_logger()
+    setup_logging()
 
     parser = argparse.ArgumentParser(description="Debug APO with runner or trainer approach.")
     parser.add_argument(

examples/apo/legacy_apo_client.py

@@ -12,7 +12,7 @@ from typing import Any
 import dotenv
 from openai import OpenAI
 
-from agentlightning import configure_logger
+from agentlightning import setup_logging
 from agentlightning.litagent import LitAgent
 from agentlightning.trainer import Trainer
 
@@ -40,7 +40,7 @@ class SimpleAgent(LitAgent[Any]):
 
 
 if __name__ == "__main__":
-    configure_logger()
+    setup_logging()
     dotenv.load_dotenv()
     agent = SimpleAgent()
     # Use 2 workers to simulate multiple clients

examples/apo/room_selector_apo.py

@@ -8,7 +8,7 @@ from typing import Tuple, cast
 from openai import AsyncOpenAI
 from room_selector import RoomSelectionTask, load_room_tasks, prompt_template_baseline, room_selector
 
-from agentlightning import Trainer, configure_logger
+from agentlightning import Trainer, setup_logging
 from agentlightning.adapter import TraceToMessages
 from agentlightning.algorithm.apo import APO
 from agentlightning.types import Dataset
@@ -33,7 +33,7 @@ def setup_apo_logger(file_path: str = "apo.log") -> None:
 
 
 def main() -> None:
-    configure_logger()
+    setup_logging()
     setup_apo_logger()
 
     openai_client = AsyncOpenAI()