merge litellm_internal_staging into bedrock text tool-call fix

.circleci/config.yml

@@ -2400,6 +2400,11 @@ jobs:
     environment:
       DATABASE_URL: "postgresql://e2euser:e2epassword@localhost:5432/litellm_e2e"
       CI: "true"
+      # Boot the proxy with an external logout URL so proxyLogoutUrl.spec.ts can
+      # assert the redirect. Set at job level so both the proxy boot step and the
+      # Playwright step (whose skip guard reads this) see the same value. Safe for
+      # the rest of the suite: nothing else performs a logout.
+      PROXY_LOGOUT_URL: "https://www.example.com"
     steps:
       - checkout
       - setup_google_dns
@@ -2476,7 +2481,8 @@ jobs:
             MOCK_LLM_URL: "http://127.0.0.1:8090/v1"
             DISABLE_SCHEMA_UPDATE: "true"
             SERVER_ROOT_PATH: ""
-            PROXY_LOGOUT_URL: ""
+            # PROXY_LOGOUT_URL is inherited from the job-level environment so the
+            # proxy and proxyLogoutUrl.spec.ts agree on the logout target.
           # LITELLM_LICENSE is forwarded from the project env so premium-gated
           # UI flows can be exercised. license.spec.ts asserts the resulting
           # JWT carries premium_user=true; if it ever stops being passed, that
@@ -2541,7 +2547,6 @@ jobs:
           paths:
             - litellm-docker-database.tar.zst
 
-
   test_bad_database_url:
     machine:
       image: ubuntu-2204:2024.04.1

.github/pull_request_template.md

@@ -10,9 +10,9 @@
 
 **Please complete all items before asking a LiteLLM maintainer to review your PR**
 
-- [ ] I have Added testing in the [`tests/test_litellm/`](https://github.com/BerriAI/litellm/tree/main/tests/test_litellm) directory, **Adding at least 1 test is a hard requirement** - [see details](https://docs.litellm.ai/docs/extras/contributing_code)
+- [ ] I have added meaningful tests
 - [ ] My PR passes all unit tests on [`make test-unit`](https://docs.litellm.ai/docs/extras/contributing_code)
-- [ ] My PR's scope is as isolated as possible, it only solves 1 specific problem
+- [ ] My PR's scope is as isolated as possible; it only solves 1 specific problem
 - [ ] I have requested a Greptile review by commenting `@greptileai` and received a **Confidence Score of at least 4/5** before requesting a maintainer review
 
 ## Delays in PR merge?

.github/workflows/codeql.yml

@@ -53,3 +53,31 @@ jobs:
         uses: github/codeql-action/analyze@ebcb5b36ded6beda4ceefea6a8bc4cc885255bb3 # v3
         with:
           category: "/language:${{ matrix.language }}"
+          output: sarif-results
+          upload: failure-only
+
+      # py/weak-sensitive-data-hashing (CWE-328) fires on the OCI signing call at
+      # litellm/llms/oci/common_utils.py, which hashes the HTTP request body to
+      # produce the x-content-sha256 header required by the OCI HTTP signing spec —
+      # a content-integrity hash, not a password or secret hash. SHA-256 is mandated
+      # by Oracle for this header; see
+      # https://docs.oracle.com/en-us/iaas/Content/API/Concepts/signingrequests.htm
+      # The `usedforsecurity=False` flag on the hashlib.sha256 call already declares
+      # non-security intent, but CodeQL's taint flow still re-fires when callers
+      # further up the stack are modified. The suppression is scoped to this one
+      # file/rule pair via SARIF post-filtering so every other callsite of
+      # py/weak-sensitive-data-hashing in the repository continues to be analyzed.
+      - name: Filter SARIF (OCI sha256)
+        if: matrix.language == 'python'
+        uses: advanced-security/filter-sarif@2da736ff05ef065cb2894ac6892e47b5eac2c3c0 # v1.1
+        with:
+          patterns: |
+            -litellm/llms/oci/common_utils.py:py/weak-sensitive-data-hashing
+          input: sarif-results/python.sarif
+          output: sarif-results/python.sarif
+
+      - name: Upload SARIF
+        uses: github/codeql-action/upload-sarif@ebcb5b36ded6beda4ceefea6a8bc4cc885255bb3 # v3
+        with:
+          sarif_file: sarif-results
+          category: "/language:${{ matrix.language }}"

.github/workflows/create_daily_oss_agent_shin_branch.yml

@@ -0,0 +1,47 @@
+name: Create Daily oss-agent-shin Branch
+
+on:
+  schedule:
+    - cron: "0 0 * * *" # Runs every day at midnight UTC
+  workflow_dispatch: # Allow manual trigger
+
+jobs:
+  create-oss-agent-shin-branch:
+    if: github.repository == 'BerriAI/litellm'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          fetch-depth: 0
+          persist-credentials: false
+
+      - name: Create daily oss-agent-shin branch
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          # Configure Git user
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+
+          # Generate branch name with MM_DD_YYYY format
+          BRANCH_NAME="litellm_oss_agent_shin_$(date +'%m_%d_%Y')"
+          echo "Creating branch: $BRANCH_NAME"
+
+          # Fetch all branches
+          git fetch --all
+
+          # Check if the branch already exists
+          if git show-ref --verify --quiet refs/remotes/origin/$BRANCH_NAME; then
+            echo "Branch $BRANCH_NAME already exists. Skipping creation."
+          else
+            echo "Creating new branch: $BRANCH_NAME"
+            # Create the new branch from main
+            git checkout -b $BRANCH_NAME origin/main
+            # Push the new branch
+            git push origin $BRANCH_NAME
+            echo "Successfully created and pushed branch: $BRANCH_NAME"
+          fi

.github/workflows/test-unit-proxy-endpoints.yml

@@ -7,6 +7,7 @@ on:
       - litellm_internal_staging
       - litellm_oss_branch
       - "litellm_**"
+  workflow_dispatch:
 
 permissions:
   contents: read
@@ -32,6 +33,7 @@ jobs:
         tests/test_litellm/proxy/image_endpoints
         tests/test_litellm/proxy/vector_store_endpoints
         tests/test_litellm/proxy/agent_endpoints
+        tests/test_litellm/proxy/a2a
         tests/test_litellm/proxy/discovery_endpoints
         tests/test_litellm/proxy/health_endpoints
         tests/test_litellm/proxy/public_endpoints
@@ -42,3 +44,16 @@ jobs:
       workers: 2
       reruns: 2
       artifact-name: proxy-endpoints
+
+  # Behavior-pinning tests for litellm/proxy/proxy_server.py. Owns its
+  # own job (not a path on the proxy-endpoints job above) so its budget
+  # is independent and its coverage artifact is uploaded separately.
+  # See: https://www.notion.so/36c43b8acdab81ee845fd5365128a2fc
+  proxy-server:
+    uses: ./.github/workflows/_test-unit-base.yml
+    with:
+      test-path: tests/test_litellm/proxy/proxy_server
+      workers: 4
+      reruns: 2
+      timeout-minutes: 60
+      artifact-name: proxy-server

.github/workflows/test_server_root_path.yml

@@ -101,6 +101,31 @@ jobs:
           docker logs litellm-test
           exit 1
 
+      - name: Setup Node for Playwright
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
+        with:
+          node-version: "20"
+
+      - name: Install UI deps and Chromium
+        working-directory: ui/litellm-dashboard
+        run: |
+          npm ci
+          npx playwright install --with-deps chromium
+
+      - name: Run SERVER_ROOT_PATH redirect e2e
+        working-directory: ui/litellm-dashboard
+        env:
+          SERVER_ROOT_PATH: ${{ matrix.root_path }}
+        run: npx playwright test --config=e2e_tests/serverRootPath.config.ts
+
+      - name: Upload Playwright artifacts on failure
+        if: failure()
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: playwright-trace-${{ strategy.job-index }}
+          path: ui/litellm-dashboard/test-results/
+          retention-days: 7
+
       - name: Cleanup
         if: always()
         run: |

.gitignore

@@ -28,6 +28,8 @@ litellm/tests/config_*.yaml
 litellm/tests/langfuse.log
 langfuse.log
 .langfuse.log
+.pin_list.txt
+.cov_new.xml
 litellm/tests/test_custom_logger.py
 litellm/tests/langfuse.log
 litellm/tests/dynamo*.log
@@ -120,4 +122,5 @@ crash.log
 crash.*.log
 # .terraform.lock.hcl is intentionally NOT ignored — it pins provider versions
 # and should be committed.
-.vscode
+.vscode
+.pin_list.txt

AGENTS.md

@@ -1,293 +1 @@
-# INSTRUCTIONS FOR LITELLM
-
-This document provides comprehensive instructions for AI agents working in the LiteLLM repository.
-
-## OVERVIEW
-
-LiteLLM is a unified interface for 100+ LLMs that:
-- Translates inputs to provider-specific completion, embedding, and image generation endpoints
-- Provides consistent OpenAI-format output across all providers
-- Includes retry/fallback logic across multiple deployments (Router)
-- Offers a proxy server (LLM Gateway) with budgets, rate limits, and authentication
-- Supports advanced features like function calling, streaming, caching, and observability
-
-## REPOSITORY STRUCTURE
-
-### Core Components
-- `litellm/` - Main library code
-  - `llms/` - Provider-specific implementations (OpenAI, Anthropic, Azure, etc.)
-  - `proxy/` - Proxy server implementation (LLM Gateway)
-  - `router_utils/` - Load balancing and fallback logic
-  - `types/` - Type definitions and schemas
-  - `integrations/` - Third-party integrations (observability, caching, etc.)
-
-### Key Directories
-- `tests/` - Comprehensive test suites
-- `ui/litellm-dashboard/` - Admin dashboard UI
-- `enterprise/` - Enterprise-specific features
-
-Documentation lives in the separate [BerriAI/litellm-docs](https://github.com/BerriAI/litellm-docs) repository and is served at [docs.litellm.ai](https://docs.litellm.ai).
-
-## DEVELOPMENT GUIDELINES
-
-### MAKING CODE CHANGES
-
-1. **Provider Implementations**: When adding/modifying LLM providers:
-   - Follow existing patterns in `litellm/llms/{provider}/`
-   - Implement proper transformation classes that inherit from `BaseConfig`
-   - Support both sync and async operations
-   - Handle streaming responses appropriately
-   - Include proper error handling with provider-specific exceptions
-
-2. **Type Safety**: 
-   - Use proper type hints throughout
-   - Update type definitions in `litellm/types/`
-   - Ensure compatibility with both Pydantic v1 and v2
-
-3. **Testing**:
-   - Add tests in appropriate `tests/` subdirectories
-   - Include both unit tests and integration tests
-   - Test provider-specific functionality thoroughly
-   - Consider adding load tests for performance-critical changes
-
-### MAKING CODE CHANGES FOR THE UI (IGNORE FOR BACKEND)
-
-1. **Always use `antd` for new UI components — Tremor is DEPRECATED**
-   - We are migrating off of `@tremor/react`. Do not introduce new `Badge`, `Text`, `Card`, `Grid`, `Title`, or other imports from `@tremor/react` in any new or modified file.
-   - Use `antd` equivalents: `Tag` for labels, plain `<span>`/`<div>` with Tailwind classes (or `Typography.Text`) for text, `Card` from `antd`, etc. Note that `antd` has no `"yellow"` Tag color — use `"gold"` for amber/yellow.
-   - The only exception is the Tremor Table component and its required Tremor Table sub components.
-
-2. **Use Common Components as much as possible**:
-   - These are usually defined in the `common_components` directory
-   - Use these components as much as possible and avoid building new components unless needed
-
-3. **Testing**:
-   - The codebase uses **Vitest** and **React Testing Library**
-   - **Query Priority Order**: Use query methods in this order: `getByRole`, `getByLabelText`, `getByPlaceholderText`, `getByText`, `getByTestId`
-   - **Always use `screen`** instead of destructuring from `render()` (e.g., use `screen.getByText()` not `getByText`)
-   - **Wrap user interactions in `act()`**: Always wrap `fireEvent` calls with `act()` to ensure React state updates are properly handled
-   - **Use `query` methods for absence checks**: Use `queryBy*` methods (not `getBy*`) when expecting an element to NOT be present
-   - **Test names must start with "should"**: All test names should follow the pattern `it("should ...")`
-   - **Mock external dependencies**: Check `setupTests.ts` for global mocks and mock child components/networking calls as needed
-   - **Structure tests properly**:
-     - First test should verify the component renders successfully
-     - Subsequent tests should focus on functionality and user interactions
-     - Use `waitFor` for async operations that aren't already awaited
-   - **Avoid using `querySelector`**: Prefer React Testing Library queries over direct DOM manipulation
-
-### IMPORTANT PATTERNS
-
-1. **Function/Tool Calling**:
-   - LiteLLM standardizes tool calling across providers
-   - OpenAI format is the standard, with transformations for other providers
-   - See `litellm/llms/anthropic/chat/transformation.py` for complex tool handling
-
-2. **Streaming**:
-   - All providers should support streaming where possible
-   - Use consistent chunk formatting across providers
-   - Handle both sync and async streaming
-
-3. **Error Handling**:
-   - Use provider-specific exception classes
-   - Maintain consistent error formats across providers
-   - Include proper retry logic and fallback mechanisms
-
-4. **Configuration**:
-   - Support both environment variables and programmatic configuration
-   - Use `BaseConfig` classes for provider configurations
-   - Allow dynamic parameter passing
-
-## PROXY SERVER (LLM GATEWAY)
-
-The proxy server is a critical component that provides:
-- Authentication and authorization
-- Rate limiting and budget management
-- Load balancing across multiple models/deployments
-- Observability and logging
-- Admin dashboard UI
-- Enterprise features
-
-Key files:
-- `litellm/proxy/proxy_server.py` - Main server implementation
-- `litellm/proxy/auth/` - Authentication logic
-- `litellm/proxy/management_endpoints/` - Admin API endpoints
-
-**Database (proxy)**: Use Prisma model methods (`prisma_client.db.<model>.upsert`, `.find_many`, `.find_unique`, etc.), not raw SQL (`execute_raw`/`query_raw`). See COMMON PITFALLS for details.
-
-## MCP (MODEL CONTEXT PROTOCOL) SUPPORT
-
-LiteLLM supports MCP for agent workflows:
-- MCP server integration for tool calling
-- Transformation between OpenAI and MCP tool formats
-- Support for external MCP servers (Zapier, Jira, Linear, etc.)
-- See `litellm/experimental_mcp_client/` and `litellm/proxy/_experimental/mcp_server/`
-
-## RUNNING SCRIPTS
-
-Use `uv run python script.py` to run Python scripts in the project environment (for non-test files).
-
-## GITHUB TEMPLATES
-
-When opening issues or pull requests, follow these templates:
-
-### Bug Reports (`.github/ISSUE_TEMPLATE/bug_report.yml`)
-- Describe what happened vs. expected behavior
-- Include relevant log output
-- Specify LiteLLM version
-- Indicate if you're part of an ML Ops team (helps with prioritization)
-
-### Feature Requests (`.github/ISSUE_TEMPLATE/feature_request.yml`)
-- Clearly describe the feature
-- Explain motivation and use case with concrete examples
-
-### Pull Requests (`.github/pull_request_template.md`)
-- Add at least 1 test in `tests/litellm/`
-- Ensure `make test-unit` passes
-
-
-## TESTING CONSIDERATIONS
-
-1. **Provider Tests**: Test against real provider APIs when possible
-2. **Proxy Tests**: Include authentication, rate limiting, and routing tests
-3. **Performance Tests**: Load testing for high-throughput scenarios
-4. **Integration Tests**: End-to-end workflows including tool calling
-
-## DOCUMENTATION
-
-- Keep documentation in sync with code changes
-- Update provider documentation when adding new providers
-- Include code examples for new features
-- Update changelog and release notes
-
-## SECURITY CONSIDERATIONS
-
-- Handle API keys securely
-- Validate all inputs, especially for proxy endpoints
-- Consider rate limiting and abuse prevention
-- Follow security best practices for authentication
-
-## ENTERPRISE FEATURES
-
-- Some features are enterprise-only
-- Check `enterprise/` directory for enterprise-specific code
-- Maintain compatibility between open-source and enterprise versions
-
-## COMMON PITFALLS TO AVOID
-
-1. **Breaking Changes**: LiteLLM has many users - avoid breaking existing APIs
-2. **Provider Specifics**: Each provider has unique quirks - handle them properly
-3. **Rate Limits**: Respect provider rate limits in tests
-4. **Memory Usage**: Be mindful of memory usage in streaming scenarios
-5. **Dependencies**: Keep dependencies minimal and well-justified
-6. **UI/Backend Contract Mismatch**: When adding a new entity type to the UI, always check whether the backend endpoint accepts a single value or an array. Match the UI control accordingly (single-select vs. multi-select) to avoid silently dropping user selections
-7. **Missing Tests for New Entity Types**: When adding a new entity type (e.g., in `EntityUsage`, `UsageViewSelect`), always add corresponding tests in the existing test files and update any icon/component mocks
-8. **Raw SQL in proxy DB code**: Do not use `execute_raw` or `query_raw` for proxy database access. Use Prisma model methods (e.g. `prisma_client.db.litellm_tooltable.upsert()`, `.find_many()`, `.find_unique()`) so behavior stays consistent with the schema, the client stays mockable in tests, and you avoid the pitfalls of hand-written SQL (parameter ordering, type casting, schema drift)
-
-8. **Do not hardcode model-specific flags**: Put model-specific capability flags in `model_prices_and_context_window.json` and read them via `get_model_info` (or existing helpers like `supports_reasoning`). This prevents users from needing to upgrade LiteLLM each time a new model supports a feature.
-
-   **Example of BAD** (hardcoded model checks):
-
-   ```python
-   @staticmethod
-   def _is_effort_supported_model(model: str) -> bool:
-       """Check if the model supports the output_config.effort parameter..."""
-       model_lower = model.lower()
-       if AnthropicConfig._is_claude_4_6_model(model):
-           return True
-       return any(
-           v in model_lower for v in ("opus-4-5", "opus_4_5", "opus-4.5", "opus_4.5")
-       )
-   ```
-
-   **Example of GOOD** (config-driven or helper that reads from config):
-
-   ```python
-   if (
-       "claude-3-7-sonnet" in model
-       or AnthropicConfig._is_claude_4_6_model(model)
-       or supports_reasoning(
-           model=model,
-           custom_llm_provider=self.custom_llm_provider,
-       )
-   ):
-       ...
-   ```
-
-   Using helpers like `supports_reasoning` (which read from `model_prices_and_context_window.json` / `get_model_info`) allows future model updates to "just work" without code changes.
-
-9. **Never close HTTP/SDK clients on cache eviction**: Do not add `close()`, `aclose()`, or `create_task(close_fn())` inside `LLMClientCache._remove_key()` or any cache eviction path. Evicted clients may still be held by in-flight requests; closing them causes `RuntimeError: Cannot send a request, as the client has been closed.` in production after the cache TTL (1 hour) expires. Connection cleanup is handled at shutdown by `close_litellm_async_clients()`. See PR #22247 for the full incident history.
-
-## HELPFUL RESOURCES
-
-- Main documentation: https://docs.litellm.ai/ (source: [BerriAI/litellm-docs](https://github.com/BerriAI/litellm-docs))
-- Provider-specific docs: https://docs.litellm.ai/docs/providers/
-- Admin UI for testing proxy features
-
-## WHEN IN DOUBT
-
-- Follow existing patterns in the codebase
-- Check similar provider implementations
-- Ensure comprehensive test coverage
-- Update documentation appropriately
-- Consider backward compatibility impact
-
-## Cursor Cloud specific instructions
-
-### Environment
-
-- uv is installed in `~/.local/bin`; the update script ensures it is on `PATH`.
-- Python 3.12, Node 22 are pre-installed.
-- The project virtual environment lives under `.venv/`.
-
-### Running the proxy server
-
-Create a minimal config file and start the proxy:
-
-```yaml
-# config.yaml
-model_list:
-  - model_name: fake-openai-endpoint
-    litellm_params:
-      model: openai/fake-model
-      api_key: fake-key
-      api_base: https://fake-api.example.com
-
-general_settings:
-  master_key: sk-1234
-
-litellm_settings:
-  drop_params: True
-  telemetry: False
-```
-
-```bash
-uv run litellm --config config.yaml --port 4000
-```
-
-The proxy takes ~15-20 seconds to fully start (it runs Prisma migrations on boot). Wait for `/health` to return before sending requests. Without a PostgreSQL `DATABASE_URL`, the proxy connects to a default Neon dev database embedded in the `litellm-proxy-extras` package.
-
-### Running tests
-
-See `CLAUDE.md` and the `Makefile` for standard commands. Key notes:
-
-- `uv sync --group proxy-dev --extra proxy` installs the Prisma and proxy-side test dependencies used by the standard local workflow.
-- The `--timeout` pytest flag is NOT available; don't pass it.
-- Unit tests: `uv run pytest tests/test_litellm/ -x -vv -n 4`
-- **Before committing, always run `uv run black .` to format your code.** Black formatting is enforced in CI.
-- If `uv sync` fails because the lockfile is outdated, run `uv lock` and retry.
-
-### Lint
-
-```bash
-cd litellm && uv run ruff check .
-```
-
-Ruff is the primary fast linter. For the full lint suite (including mypy, black, circular imports), run `make lint` per `CLAUDE.md`.
-
-### UI Dashboard development
-
-- The UI is at `ui/litellm-dashboard/`. Run `npm run dev` from that directory for the Next.js dev server on port 3000.
-- The proxy at port 4000 serves a **pre-built** static UI from `litellm/proxy/_experimental/out/`. After making UI code changes, you must run `npm run build` in the dashboard directory and copy the output: `cp -r ui/litellm-dashboard/out/* litellm/proxy/_experimental/out/` for the proxy to serve the updated UI.
-- SVGs used as provider logos (loaded via `<img>` tags) must NOT use `fill="currentColor"` — replace with an explicit color like `#000000` or use the `-color` variant from lobehub icons, since CSS color inheritance does not work inside `<img>` elements.
-- Provider logos live in `ui/litellm-dashboard/public/assets/logos/` (source) and `litellm/proxy/_experimental/out/assets/logos/` (pre-built). Both locations must have the file for it to work in dev and proxy-served modes.
-- UI Vitest tests: `cd ui/litellm-dashboard && npx vitest run`
+Read @CLAUDE.md for coding guidelines

CLAUDE.md

@@ -1,181 +1,70 @@
-# CLAUDE.md
+Do not write comments unless they are absolutely necessary to explain some very complex business logic. Please clean up if there are comments that are not absolutely necessary. Do not remove comments that are unrelated to the addition of the code of this PR
 
-This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+Explanation: code comments are, in a way, a violation of DRY code. You must update logic in two locations to change the code and "hard to change" is literally the definition of tech debt. We should instead aim to write code that is intuitive to the reader, while being both easy to maintain and high performance
 
-## Documentation
+Don't assume that the existing code is correct or the right way of doing things / good coding patterns. In fact, there are a lot of bad coding practices, overly complex code, code smells, etc. If something doesn't look right, speak up. Feel free to break existing patterns or question weird existing code to make new code high quality, as in:
+- correct
+- secure
+- performant
+- readable
+- easy to maintain/change
+- modern
+In that order of importance
 
-Documentation lives in a separate repository: [BerriAI/litellm-docs](https://github.com/BerriAI/litellm-docs). It is served at [docs.litellm.ai](https://docs.litellm.ai). Do not create or edit documentation files in this repository — open doc PRs against `BerriAI/litellm-docs` instead.
+When adding new features, add meaningful tests. Don't add tests that don't check anything substantial and is there just to make the code coverage pass. Yes, code coverage is important, but I'd rather have no signal whether the code is working than tests that don't fail when code is broken. The goal is to have tests that would fail before the feature was added/if the code was mutated in a way that breaks the feature and succeed only when the feature is fully working. I should run mutation testing and see > 90% kill rate
 
-## Development Commands
+Same thing for bug fixes. The tests should make it so that this specific bug can never happen again without failing tests (i.e., regression)
 
-### Installation
-- `make install-dev` - Install core development dependencies
-- `make install-proxy-dev` - Install proxy development dependencies with full feature set
-- `make install-test-deps` - Install the full local test environment and generate the Prisma client
+When creating PRs, don't set base to `main`. `litellm_internal_staging` serves that purpose
 
-### Testing
-- `make test` - Run all tests
-- `make test-unit` - Run unit tests (tests/test_litellm) with 4 parallel workers
-- `make test-integration` - Run integration tests (excludes unit tests)
-- `pytest tests/` - Direct pytest execution
+Always use @.github/pull_request_template.md as a guide for your PR body
 
-### Code Quality
-- `make lint` - Run all linting (Ruff, MyPy, Black, circular imports, import safety)
-- `make format` - Apply Black code formatting
-- `make lint-ruff` - Run Ruff linting only
-- `make lint-mypy` - Run MyPy type checking only
-- **Before committing, always run `uv run black .` to format your code.** Black formatting is enforced in CI.
+Never use `pytest` commands or the like as "Screenshots / Proof of Fix". We prefer curl'ing a live proxy instance running on localhost:4000 (I like to run it with `python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log`) and showing both the command run and the output. Also, it should hit real LLM provider APIs, not mocks, and cost real $$$ because that is the most realistic test. The proof of fix should be exactly what the end user / customer would see / do. The run logs in PR #27703 is a prime example of how to do it (not a huge fan of using a python test script that future me and the team will have no visibility into; I prefer just curl commands or a short list of bash commands (e.g., using `for`)). If it's a UI thing, just tell me which URLs to go to (e.g., http://localhost:4000/ui/?page=logs), where to click, what fields to fill out, etc. along with the other commands to run in an ordered list, and I'll do it myself and post the screenshots after you make the PR
 
-### Single Test Files
-- `uv run pytest tests/path/to/test_file.py -v` - Run specific test file
-- `uv run pytest tests/path/to/test_file.py::test_function -v` - Run specific test
+If you ever make public-facing PR descriptions, comments, issues, commit messages, etc., always follow these guidelines to sound less AI-y:
+- don't use emojis
+- don't use "—". Instead, reach for ";", ".", etc.
+- don't use the pattern "It's not X, it's Y", "You're not X, you're Y", etc.
+- don't use bulleted or numbered lists unless it would be nonsensical not to. Instead, prefer prose
 
-### Running Scripts
-- `uv run python script.py` - Run Python scripts (use for non-test files)
+Don't hesitate to use values in .env to get needed API keys and other secrets, as long as you never add them to conversation history, commit them, or include them in GitHub issues / PRs
 
-### GitHub Issue & PR Templates
-When contributing to the project, use the appropriate templates:
+Run tests, format your code, and lint your code before each commit
 
-**Bug Reports** (`.github/ISSUE_TEMPLATE/bug_report.yml`):
-- Describe what happened vs. what you expected
-- Include relevant log output
-- Specify your LiteLLM version
+Ask to commit and push your work when you're done (or if you're confident that your code is good and works, just do it)
 
-**Feature Requests** (`.github/ISSUE_TEMPLATE/feature_request.yml`):
-- Describe the feature clearly
-- Explain the motivation and use case
+When you must use real LLM models to, for example, write e2e tests, write a QA runbook, etc., make sure to use the latest models (doesn't have to be smartest, can also be a modern small, fast one. No strong preference for smart vs fast here, just use something modern) as of the year and month of the current date. Do a web search as necessary to figure that out
 
-**Pull Requests** (`.github/pull_request_template.md`):
-- Add at least 1 test in `tests/litellm/`
-- Ensure `make test-unit` passes
+If you're an internal contributor, when creating a new PR, the typical flow is to branch off litellm_internal_staging and create a branch prefixed with litellm_. Do not create a branch prefixed with claude/ and generally do not have / in your branch names
 
-## Architecture Overview
+Do not add `Co-Authored-By: Claude` or any Claude attribution to commit messages. Never use a `claude/` prefix or put a `/` in a branch name. Do not add "Generated with Claude Code" (or any similar attribution) to PR descriptions. Do not create a new PR/branch off the existing PR to fix/add something that is related and could've just been committed directly to the existing PR's branch
 
-LiteLLM is a unified interface for 100+ LLM providers with two main components:
+When working on a PR, keep the PR description in sync with new commits being made
 
-### Core Library (`litellm/`)
-- **Main entry point**: `litellm/main.py` - Contains core completion() function
-- **Provider implementations**: `litellm/llms/` - Each provider has its own subdirectory
-- **Router system**: `litellm/router.py` + `litellm/router_utils/` - Load balancing and fallback logic
-- **Type definitions**: `litellm/types/` - Pydantic models and type hints
-- **Integrations**: `litellm/integrations/` - Third-party observability, caching, logging
-- **Caching**: `litellm/caching/` - Multiple cache backends (Redis, in-memory, S3, etc.)
+Monkeypatching attributes of a class to do testing is an anti-pattern. Prefer dependency-injecting things into classes. That way, at unit test time, you can pass a mocked dependency in
 
-### Proxy Server (`litellm/proxy/`)
-- **Main server**: `proxy_server.py` - FastAPI application
-- **Authentication**: `auth/` - API key management, JWT, OAuth2
-- **Database**: `db/` - Prisma ORM with PostgreSQL/SQLite support
-- **Management endpoints**: `management_endpoints/` - Admin APIs for keys, teams, models
-- **Pass-through endpoints**: `pass_through_endpoints/` - Provider-specific API forwarding
-- **Guardrails**: `guardrails/` - Safety and content filtering hooks
-- **UI Dashboard**: Served from `_experimental/out/` (Next.js build)
+Do not put names of customers or customer company names in code, PRs, and issues. The codebase is public
 
-## Key Patterns
+CI supply-chain safety: Never pipe a remote script into a shell (`curl ... | bash`, `wget ... | sh`); download the artifact to a file, verify its SHA-256 checksum, then install. Pin every external tool to a specific version with a full URL (not `latest` or `stable`). Verify checksums for all downloaded binaries, using the provider's official `.sha256` / `.sha256sum` sidecar when available. These rules apply to every download in CI
 
-### Provider Implementation
-- Providers inherit from base classes in `litellm/llms/base.py`
-- Each provider has transformation functions for input/output formatting
-- Support both sync and async operations
-- Handle streaming responses and function calling
+## Think Before Coding
 
-### Error Handling
-- Provider-specific exceptions mapped to OpenAI-compatible errors
-- Fallback logic handled by Router system
-- Comprehensive logging through `litellm/_logging.py`
+**Don't assume. Don't hide confusion. Surface tradeoffs.**
 
-### Configuration
-- YAML config files for proxy server (see `proxy/example_config_yaml/`)
-- Environment variables for API keys and settings
-- Database schema managed via Prisma (`proxy/schema.prisma`)
+Before implementing:
+- State your assumptions explicitly. If uncertain, ask.
+- If multiple interpretations exist, present them. Don't pick silently.
+- If a simpler approach exists, say so. Push back when warranted.
+- If something is unclear, stop. Name what's confusing. Ask.
 
-## Development Notes
+## Simplicity First
 
-### Code Style
-- Uses Black formatter, Ruff linter, MyPy type checker
-- Pydantic v2 for data validation
-- Async/await patterns throughout
-- Type hints required for all public APIs
-- **Avoid imports within methods** — place all imports at the top of the file (module-level). Inline imports inside functions/methods make dependencies harder to trace and hurt readability. The only exception is avoiding circular imports where absolutely necessary.
-- **Use dict spread for immutable copies** — prefer `{**original, "key": new_value}` over `dict(obj)` + mutation. The spread produces the final dict in one step and makes intent clear.
-- **Guard at resolution time** — when resolving an optional value through a fallback chain (`a or b or ""`), raise immediately if the resolved result being empty is an error. Don't pass empty strings or sentinel values downstream for the callee to deal with.
-- **Extract complex comprehensions to named helpers** — a set/dict comprehension that calls into the DB or manager (e.g. "which of these server IDs are OAuth2?") belongs in a named helper function, not inline in the caller.
-- **FastAPI parameter declarations** — mark required query/form params with `= Query(...)` / `= Form(...)` explicitly when other params in the same handler are optional. Mixing `str` (required) with `Optional[str] = None` in the same signature causes silent 422s when the required param is missing.
+**Minimum code that solves the problem. Nothing speculative.**
 
-### Testing Strategy
-- Unit tests in `tests/test_litellm/`
-- Integration tests for each provider in `tests/llm_translation/`
-- Proxy tests in `tests/proxy_unit_tests/`
-- Load tests in `tests/load_tests/`
-- **Always add tests when adding new entity types or features** — if the existing test file covers other entity types, add corresponding tests for the new one
-- **Keep monkeypatch stubs in sync with real signatures** — when a function gains a new optional parameter, update every `fake_*` / `stub_*` in tests that patch it to also accept that kwarg (even as `**kwargs`). Stale stubs fail with `unexpected keyword argument` and mask real bugs.
-- **Test all branches of name→ID resolution** — when adding server/resource lookup that resolves names to UUIDs, test: (1) name resolves and UUID is allowed, (2) name resolves but UUID is not allowed, (3) name does not resolve at all. The silent-fallback path is where access-control bugs hide.
+- No features beyond what was asked.
+- No abstractions for single-use code.
+- No "flexibility" or "configurability" that wasn't requested.
+- No error handling for impossible scenarios.
+- If you write 200 lines and it could be 50, rewrite it.
 
-### UI / Backend Consistency
-- When wiring a new UI entity type to an existing backend endpoint, verify the backend API contract (single value vs. array, required vs. optional params) and ensure the UI controls match — e.g., use a single-select dropdown when the backend accepts a single value, not a multi-select
-
-### UI Component Library
-- **Always use `antd` for new UI components** — we are migrating off of `@tremor/react`. Do not introduce new `Badge`, `Text`, `Card`, `Grid`, `Title`, or other imports from `@tremor/react` in any new or modified file. Use `antd` equivalents: `Tag` for labels, `Typography.Text` / `Typography.Title` / `Typography.Paragraph` for textual content (avoid plain text-only `<span>`, `<p>`, `<h*>` when Typography fits), and `Card` from `antd`. Note that `antd` has no `"yellow"` Tag color — use `"gold"` for amber/yellow.
-
-### MCP OAuth / OpenAPI Transport Mapping
-- **`available_on_public_internet: false` with `delegate_auth_to_upstream: true` (oauth2, interactive — not `client_credentials`)** — LiteLLM still allows the anonymous upstream PKCE path (no proxy API key for `/authorize` and matching MCP routes). The internal-only flag mainly affects other surfaces (e.g. IP-based discovery). Rely on the upstream IdP and network policy; the dashboard shows a warning when both are set, and the proxy logs a warning when the server is loaded from config or the database.
-- `TRANSPORT.OPENAPI` is a UI-only concept. The backend only accepts `"http"`, `"sse"`, or `"stdio"`. Always map it to `"http"` before any API call (including pre-OAuth temp-session calls).
-- FastAPI validation errors return `detail` as an array of `{loc, msg, type}` objects. Error extractors must handle: array (map `.msg`), string, nested `{error: string}`, and fallback.
-- When an MCP server already has `authorization_url` stored, skip OAuth discovery (`_discovery_metadata`) — the server URL for OpenAPI MCPs is the spec file, not the API base, and fetching it causes timeouts.
-- `client_id` should be optional in the `/authorize` endpoint — if the server has a stored `client_id` in credentials, use that. Never require callers to re-supply it.
-
-### MCP Credential Storage
-- OAuth credentials and BYOK credentials share the `litellm_mcpusercredentials` table, distinguished by a `"type"` field in the JSON payload (`"oauth2"` vs plain string).
-- When deleting OAuth credentials, check type before deleting to avoid accidentally deleting a BYOK credential for the same `(user_id, server_id)` pair.
-- Always pass the raw `expires_at` timestamp to the client — never set it to `None` for expired credentials. Let the frontend compute the "Expired" display state from the timestamp.
-- Use `RecordNotFoundError` (not bare `except Exception`) when catching "already deleted" in credential delete endpoints.
-
-### Browser Storage Safety (UI)
-- Never write LiteLLM access tokens or API keys to `localStorage` — use `sessionStorage` only. `localStorage` survives browser close and is readable by any injected script (XSS).
-- Shared utility functions (e.g. `extractErrorMessage`) belong in `src/utils/` — never define them inline in hooks or duplicate them across files.
-
-### Database Migrations
-- Prisma handles schema migrations
-- Migration files auto-generated with `prisma migrate dev`
-- Always test migrations against both PostgreSQL and SQLite
-
-### Proxy database access
-- **Do not write raw SQL** for proxy DB operations. Use Prisma model methods instead of `execute_raw` / `query_raw`.
-- Use the generated client: `prisma_client.db.<model>` (e.g. `litellm_tooltable`, `litellm_usertable`) with `.upsert()`, `.find_many()`, `.find_unique()`, `.update()`, `.update_many()` as appropriate. This avoids schema/client drift, keeps code testable with simple mocks, and matches patterns used in spend logs and other proxy code.
-- **No N+1 queries.** Never query the DB inside a loop. Batch-fetch with `{"in": ids}` and distribute in-memory.
-- **Batch writes.** Use `create_many`/`update_many`/`delete_many` instead of individual calls (these return counts only; `update_many`/`delete_many` no-op silently on missing rows). When multiple separate writes target the same table (e.g. in `batch_()`), order by primary key to avoid deadlocks.
-- **Push work to the DB.** Filter, sort, group, and aggregate in SQL, not Python. Verify Prisma generates the expected SQL — e.g. prefer `group_by` over `find_many(distinct=...)` which does client-side processing.
-- **Bound large result sets.** Prisma materializes full results in memory. For results over ~10 MB, paginate with `take`/`skip` or `cursor`/`take`, always with an explicit `order`. Prefer cursor-based pagination (`skip` is O(n)). Don't paginate naturally small result sets.
-- **Limit fetched columns on wide tables.** Use `select` to fetch only needed fields — returns a partial object, so downstream code must not access unselected fields.
-- **Check index coverage.** For new or modified queries, check `schema.prisma` for a supporting index. Prefer extending an existing index (e.g. `@@index([a])` → `@@index([a, b])`) over adding a new one, unless it's a `@@unique`. Only add indexes for large/frequent queries.
-- **Keep schema files in sync.** Apply schema changes to all `schema.prisma` copies (`schema.prisma`, `litellm/proxy/`, `litellm-proxy-extras/`) with a migration under `litellm-proxy-extras/litellm_proxy_extras/migrations/`.
-
-### Setup Wizard (`litellm/setup_wizard.py`)
-- The wizard is implemented as a single `SetupWizard` class with `@staticmethod` methods — keep it that way. No module-level functions except `run_setup_wizard()` (the public entrypoint) and pure helpers (color, ANSI).
-- Use `litellm.utils.check_valid_key(model, api_key)` for credential validation — never roll a custom completion call.
-- Do not hardcode provider env-key names or model lists that already exist in the codebase. Add a `test_model` field to each provider entry to drive `check_valid_key`; set it to `None` for providers that can't be validated with a single API key (Azure, Bedrock, Ollama).
-
-### Enterprise Features
-- Enterprise-specific code in `enterprise/` directory
-- Optional features enabled via environment variables
-- Separate licensing and authentication for enterprise features
-
-### CI Supply-Chain Safety
-- **Never pipe a remote script into a shell** (`curl ... | bash`, `wget ... | sh`). Download the artifact to a file, verify its SHA-256 checksum, then install.
-- **Pin every external tool to a specific version** with a full URL (not `latest` or `stable`). Unversioned downloads silently change under you.
-- **Verify checksums for all downloaded binaries.** Use the provider's official `.sha256` / `.sha256sum` sidecar file when available; otherwise compute and hardcode the digest.
-- **Prefer reusable CircleCI commands** (`commands:` section) so a tool is installed and verified in exactly one place, then referenced everywhere with `- install_<tool>` or `- wait_for_service`.
-- **Don't add tools just because they were there before.** Audit whether an external dependency is still needed. If it can be replaced with a shell one-liner or a tool already in the image, remove it.
-- These rules apply to every download in CI: binaries, install scripts, language version managers, package repos. No exceptions.
-
-### HTTP Client Cache Safety
-- **Never close HTTP/SDK clients on cache eviction.** `LLMClientCache._remove_key()` must not call `close()`/`aclose()` on evicted clients — they may still be used by in-flight requests. Doing so causes `RuntimeError: Cannot send a request, as the client has been closed.` after the 1-hour TTL expires. Cleanup happens at shutdown via `close_litellm_async_clients()`.
-
-### Troubleshooting: DB schema out of sync after proxy restart
-`litellm-proxy-extras` runs `prisma migrate deploy` on startup using **its own** bundled migration files, which may lag behind schema changes in the current worktree. Symptoms: `Unknown column`, `Invalid prisma invocation`, or missing data on new fields.
-
-**Diagnose:** Run `\d "TableName"` in psql and compare against `schema.prisma` — missing columns confirm the issue.
-
-**Fix options:**
-1. **Create a Prisma migration** (permanent) — run `prisma migrate dev --name <description>` in the worktree. The generated file will be picked up by `prisma migrate deploy` on next startup.
-2. **Apply manually for local dev** — `psql -d litellm -c "ALTER TABLE ... ADD COLUMN IF NOT EXISTS ..."` after each proxy start. Fine for dev, not for production.
-3. **Update litellm-proxy-extras** — if the package is installed from PyPI, its migration directory must include the new file. Either update the package or run the migration manually until the next release ships it.
+Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.

GEMINI.md

@@ -1,108 +1 @@
-# GEMINI.md
-
-This file provides guidance to Gemini when working with code in this repository.
-
-## Development Commands
-
-### Installation
-- `make install-dev` - Install core development dependencies
-- `make install-proxy-dev` - Install proxy development dependencies with full feature set
-- `make install-test-deps` - Install all test dependencies
-
-### Testing
-- `make test` - Run all tests
-- `make test-unit` - Run unit tests (tests/test_litellm) with 4 parallel workers
-- `make test-integration` - Run integration tests (excludes unit tests)
-- `pytest tests/` - Direct pytest execution
-
-### Code Quality
-- `make lint` - Run all linting (Ruff, MyPy, Black, circular imports, import safety)
-- `make format` - Apply Black code formatting
-- `make lint-ruff` - Run Ruff linting only
-- `make lint-mypy` - Run MyPy type checking only
-
-### Single Test Files
-- `uv run pytest tests/path/to/test_file.py -v` - Run specific test file
-- `uv run pytest tests/path/to/test_file.py::test_function -v` - Run specific test
-
-### Running Scripts
-- `uv run python script.py` - Run Python scripts (use for non-test files)
-
-### GitHub Issue & PR Templates
-When contributing to the project, use the appropriate templates:
-
-**Bug Reports** (`.github/ISSUE_TEMPLATE/bug_report.yml`):
-- Describe what happened vs. what you expected
-- Include relevant log output
-- Specify your LiteLLM version
-
-**Feature Requests** (`.github/ISSUE_TEMPLATE/feature_request.yml`):
-- Describe the feature clearly
-- Explain the motivation and use case
-
-**Pull Requests** (`.github/pull_request_template.md`):
-- Add at least 1 test in `tests/litellm/`
-- Ensure `make test-unit` passes
-
-## Architecture Overview
-
-LiteLLM is a unified interface for 100+ LLM providers with two main components:
-
-### Core Library (`litellm/`)
-- **Main entry point**: `litellm/main.py` - Contains core completion() function
-- **Provider implementations**: `litellm/llms/` - Each provider has its own subdirectory
-- **Router system**: `litellm/router.py` + `litellm/router_utils/` - Load balancing and fallback logic
-- **Type definitions**: `litellm/types/` - Pydantic models and type hints
-- **Integrations**: `litellm/integrations/` - Third-party observability, caching, logging
-- **Caching**: `litellm/caching/` - Multiple cache backends (Redis, in-memory, S3, etc.)
-
-### Proxy Server (`litellm/proxy/`)
-- **Main server**: `proxy_server.py` - FastAPI application
-- **Authentication**: `auth/` - API key management, JWT, OAuth2
-- **Database**: `db/` - Prisma ORM with PostgreSQL/SQLite support
-- **Management endpoints**: `management_endpoints/` - Admin APIs for keys, teams, models
-- **Pass-through endpoints**: `pass_through_endpoints/` - Provider-specific API forwarding
-- **Guardrails**: `guardrails/` - Safety and content filtering hooks
-- **UI Dashboard**: Served from `_experimental/out/` (Next.js build)
-
-## Key Patterns
-
-### Provider Implementation
-- Providers inherit from base classes in `litellm/llms/base.py`
-- Each provider has transformation functions for input/output formatting
-- Support both sync and async operations
-- Handle streaming responses and function calling
-
-### Error Handling
-- Provider-specific exceptions mapped to OpenAI-compatible errors
-- Fallback logic handled by Router system
-- Comprehensive logging through `litellm/_logging.py`
-
-### Configuration
-- YAML config files for proxy server (see `proxy/example_config_yaml/`)
-- Environment variables for API keys and settings
-- Database schema managed via Prisma (`proxy/schema.prisma`)
-
-## Development Notes
-
-### Code Style
-- Uses Black formatter, Ruff linter, MyPy type checker
-- Pydantic v2 for data validation
-- Async/await patterns throughout
-- Type hints required for all public APIs
-
-### Testing Strategy
-- Unit tests in `tests/test_litellm/`
-- Integration tests for each provider in `tests/llm_translation/`
-- Proxy tests in `tests/proxy_unit_tests/`
-- Load tests in `tests/load_tests/`
-
-### Database Migrations
-- Prisma handles schema migrations
-- Migration files auto-generated with `prisma migrate dev`
-- Always test migrations against both PostgreSQL and SQLite
-
-### Enterprise Features
-- Enterprise-specific code in `enterprise/` directory
-- Optional features enabled via environment variables
-- Separate licensing and authentication for enterprise features
+Read @CLAUDE.md for coding guidelines

backend/Dockerfile

@@ -12,17 +12,27 @@ USER root
 
 COPY --from=uvbin /uv /uvx /usr/local/bin/
 
-RUN apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile
+# nodejs/npm so `prisma generate` uses Wolfi's Node via PRISMA_USE_GLOBAL_NODE
+# instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
+# (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
+RUN for i in 1 2 3; do \
+      apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile nodejs npm && break; \
+      [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
+      sleep 5; \
+    done
 
 # UV_COMPILE_BYTECODE=1   precompiles .pyc at install time → faster cold start.
 # UV_LINK_MODE=copy       avoids hardlink warnings when uv installs from a
 #                         BuildKit cache mount (different filesystem).
 # UV_PYTHON_DOWNLOADS=0   force uv to use the apk-installed CPython instead of
 #                         silently pulling a managed interpreter.
+# PRISMA_USE_GLOBAL_NODE  explicit (matches default) so an env override can't
+#                         silently re-enable nodeenv's Node download.
 ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
     UV_LINK_MODE=copy \
     UV_COMPILE_BYTECODE=1 \
     UV_PYTHON_DOWNLOADS=0 \
+    PRISMA_USE_GLOBAL_NODE=true \
     PATH="/app/.venv/bin:${PATH}"
 
 # Stage 1 — install dependencies only.
@@ -58,7 +68,11 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
 
 USER root
 
-RUN apk add --no-cache bash openssl tzdata python3 libsndfile libatomic
+RUN for i in 1 2 3; do \
+      apk add --no-cache bash openssl tzdata python3 libsndfile libatomic && break; \
+      [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
+      sleep 5; \
+    done
 
 # wolfi-base ships an unprivileged `nonroot` account (UID/GID 65532) with
 # /home/nonroot. We run the backend as that user

cookbook/gollem_go_agent_framework/go.mod

@@ -1,5 +1,5 @@
 module github.com/BerriAI/litellm/cookbook/gollem_go_agent_framework
 
-go 1.25.1
+go 1.26.3
 
 require github.com/fugue-labs/gollem v0.1.0

deploy/charts/litellm-helm/Chart.yaml

@@ -24,7 +24,7 @@ version: 1.1.0
 # incremented each time you make changes to the application. Versions are not expected to
 # follow Semantic Versioning. They should reflect the version the application is using.
 # It is recommended to use it with quotes.
-appVersion: v1.80.12
+appVersion: v1.85.1
 
 annotations:
   org.opencontainers.image.source: "https://github.com/BerriAI/litellm"

deploy/charts/litellm-helm/templates/deployment.yaml

@@ -30,7 +30,7 @@ spec:
         checksum/config: {{ include (print $.Template.BasePath "/configmap-litellm.yaml") . | sha256sum }}
         {{- end }}
         {{- with .Values.podAnnotations }}
-        {{- toYaml . | nindent 8 }}
+        {{- tpl (toYaml .) $ | nindent 8 }}
         {{- end }}
       labels:
         {{- include "litellm.labels" . | nindent 8 }}
@@ -53,7 +53,7 @@ spec:
         - name: {{ include "litellm.name" . }}
           securityContext:
             {{- toYaml .Values.securityContext | nindent 12 }}
-          image: "{{ .Values.image.repository }}:{{ .Values.image.tag | default (printf "main-%s" .Chart.AppVersion) }}"
+          image: "{{ .Values.image.repository }}:{{ .Values.image.tag | default .Chart.AppVersion }}"
           imagePullPolicy: {{ .Values.image.pullPolicy }}
           env:
             - name: HOST

deploy/charts/litellm-helm/templates/migrations-job.yaml

@@ -41,7 +41,7 @@ spec:
       {{- end }} 
       containers:
         - name: prisma-migrations
-          image: "{{ .Values.image.repository }}:{{ .Values.image.tag | default (printf "main-%s" .Chart.AppVersion) }}"
+          image: "{{ .Values.image.repository }}:{{ .Values.image.tag | default .Chart.AppVersion }}"
           imagePullPolicy: {{ .Values.image.pullPolicy }}
           securityContext:
             {{- toYaml .Values.securityContext | nindent 12 }}

deploy/charts/litellm-helm/tests/deployment_tests.yaml

@@ -377,3 +377,28 @@ tests:
           content:
             name: sidecar-tpl
             image: "ghcr.io/berriai/litellm-database:test"
+  - it: should support tpl in podAnnotations
+    template: deployment.yaml
+    set:
+      image:
+        repository: ghcr.io/berriai/litellm-database
+        tag: test
+      # Mirrors the real-world scenario this feature unblocks:
+      # user disables the built-in ConfigMap (and its built-in checksum/config
+      # annotation) and re-implements checksum/config themselves via tpl.
+      proxyConfigMap:
+        create: false
+      podAnnotations:
+        checksum/config: "{{ .Values.image.tag }}"
+        example.com/some-key: "{{ .Values.image.repository }}"
+        example.com/literal: "plain-string-value"
+    asserts:
+      - equal:
+          path: spec.template.metadata.annotations["checksum/config"]
+          value: "test"
+      - equal:
+          path: spec.template.metadata.annotations["example.com/some-key"]
+          value: "ghcr.io/berriai/litellm-database"
+      - equal:
+          path: spec.template.metadata.annotations["example.com/literal"]
+          value: "plain-string-value"

deploy/charts/litellm-helm/values.yaml

@@ -10,7 +10,7 @@ image:
   repository: ghcr.io/berriai/litellm-database
   pullPolicy: Always
   # Overrides the image tag whose default is the chart appVersion.
-  # tag: "main-latest"
+  # tag: "latest"
   tag: ""
 
 imagePullSecrets: []

docker/Dockerfile.non_root

@@ -55,22 +55,10 @@ COPY . .
 # Set non-root flag for build time consistency
 ENV LITELLM_NON_ROOT=true
 
-# Stage the pre-built Admin UI from the checked-in Next.js static export.
-# _experimental/out/ is regenerated as part of the release runbook.
-# Restructure extensionless routes (foo.html -> foo/index.html) to match the layout
-# proxy_server.py expects, and drop a readiness marker.
 RUN mkdir -p /var/lib/litellm/ui /var/lib/litellm/assets && \
     cp -r /app/litellm/proxy/_experimental/out/. /var/lib/litellm/ui/ && \
     cp /app/litellm/proxy/logo.jpg /var/lib/litellm/assets/logo.jpg && \
-    ( cd /var/lib/litellm/ui && \
-      for html_file in *.html; do \
-        if [ "$html_file" != "index.html" ] && [ -f "$html_file" ]; then \
-          folder_name="${html_file%.html}" && \
-          mkdir -p "$folder_name" && \
-          mv "$html_file" "$folder_name/index.html"; \
-        fi; \
-      done && \
-      touch .litellm_ui_ready )
+    touch /var/lib/litellm/ui/.litellm_ui_ready
 
 RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
     if [ "$PROXY_EXTRAS_SOURCE" = "published" ]; then \

enterprise/litellm_enterprise/enterprise_callbacks/send_emails/resend_email.py

@@ -19,12 +19,26 @@ RESEND_API_ENDPOINT = "https://api.resend.com/emails"
 
 
 class ResendEmailLogger(BaseEmailLogger):
+    """
+    Send emails using Resend's API.
+
+    Required env vars:
+    - RESEND_API_KEY
+
+    Optional env vars:
+    - RESEND_FROM_EMAIL: Override the default sender address. Must be on a
+      domain verified in your Resend account. When unset, falls back to the
+      `from_email` argument passed by the caller (which defaults to
+      `notifications@alerts.litellm.ai` and only works on LiteLLM Cloud).
+    """
+
     def __init__(self, internal_usage_cache=None, **kwargs):
         super().__init__(internal_usage_cache=internal_usage_cache, **kwargs)
         self.async_httpx_client = get_async_httpx_client(
             llm_provider=httpxSpecialProvider.LoggingCallback
         )
         self.resend_api_key = os.getenv("RESEND_API_KEY")
+        self.resend_from_email = os.getenv("RESEND_FROM_EMAIL")
 
     async def send_email(
         self,
@@ -33,13 +47,14 @@ class ResendEmailLogger(BaseEmailLogger):
         subject: str,
         html_body: str,
     ):
+        sender_email = self.resend_from_email or from_email
         verbose_logger.debug(
-            f"Sending email from {from_email} to {to_email} with subject {subject}"
+            f"Sending email from {sender_email} to {to_email} with subject {subject}"
         )
         response = await self.async_httpx_client.post(
             url=RESEND_API_ENDPOINT,
             json={
-                "from": from_email,
+                "from": sender_email,
                 "to": to_email,
                 "subject": subject,
                 "html": html_body,

enterprise/litellm_enterprise/proxy/hooks/managed_files.py

@@ -658,7 +658,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                     if isinstance(content, str):
                         continue
                     for c in content:
-                        if c["type"] == "file":
+                        if c.get("type") == "file":
                             file_object = cast(ChatCompletionFileObject, c)
                             file_object_file_field = file_object["file"]
                             file_id = file_object_file_field.get("file_id")

enterprise/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "litellm-enterprise"
-version = "0.1.41"
+version = "0.1.42"
 description = "Package for LiteLLM Enterprise features"
 readme = "README.md"
 requires-python = ">=3.9"
@@ -26,7 +26,7 @@ required-version = ">=0.10.9"
 module-root = ""
 
 [tool.commitizen]
-version = "0.1.41"
+version = "0.1.42"
 version_files = [
     "pyproject.toml:^version",
     "../pyproject.toml:litellm-enterprise==",

gateway/Dockerfile

@@ -12,17 +12,27 @@ USER root
 
 COPY --from=uvbin /uv /uvx /usr/local/bin/
 
-RUN apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile
+# nodejs/npm so `prisma generate` uses Wolfi's Node via PRISMA_USE_GLOBAL_NODE
+# instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
+# (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
+RUN for i in 1 2 3; do \
+      apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile nodejs npm && break; \
+      [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
+      sleep 5; \
+    done
 
 # UV_COMPILE_BYTECODE=1   precompiles .pyc at install time → faster cold start.
 # UV_LINK_MODE=copy       avoids hardlink warnings when uv installs from a
 #                         BuildKit cache mount (different filesystem).
 # UV_PYTHON_DOWNLOADS=0   force uv to use the apk-installed CPython instead of
 #                         silently pulling a managed interpreter.
+# PRISMA_USE_GLOBAL_NODE  explicit (matches default) so an env override can't
+#                         silently re-enable nodeenv's Node download.
 ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
     UV_LINK_MODE=copy \
     UV_COMPILE_BYTECODE=1 \
     UV_PYTHON_DOWNLOADS=0 \
+    PRISMA_USE_GLOBAL_NODE=true \
     PATH="/app/.venv/bin:${PATH}"
 
 # Stage 1 — install dependencies only.
@@ -58,7 +68,11 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
 
 USER root
 
-RUN apk add --no-cache bash openssl tzdata python3 libsndfile libatomic
+RUN for i in 1 2 3; do \
+      apk add --no-cache bash openssl tzdata python3 libsndfile libatomic && break; \
+      [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
+      sleep 5; \
+    done
 
 # wolfi-base ships an unprivileged `nonroot` account (UID/GID 65532) with
 # /home/nonroot. We run the proxy as that user.

helm/litellm/templates/_helpers.tpl

@@ -56,16 +56,34 @@ app.kubernetes.io/component: ui
 {{- end -}}
 
 {{/*
-Shared ServiceAccount name used by all three component Deployments. When
-`serviceAccount.create` is true and `serviceAccount.name` is empty, default
-to the chart fullname. When `create` is false, fall back to the provided
-name or the namespace's `default` SA.
+Per-component ServiceAccount name helpers.
+
+Each component (gateway, backend, ui) has its own SA config under
+.Values.serviceAccounts.<component>. When `create` is true and `name` is
+empty the chart defaults to "<release>-litellm-<component>". When `create`
+is false the chart uses the provided name, or the namespace `default` SA.
 */}}
-{{- define "litellm.serviceAccountName" -}}
-{{- if .Values.serviceAccount.create -}}
-{{ default (include "litellm.fullname" .) .Values.serviceAccount.name }}
+{{- define "litellm.gateway.serviceAccountName" -}}
+{{- if .Values.serviceAccounts.gateway.create -}}
+{{ default (include "litellm.gateway.fullname" .) .Values.serviceAccounts.gateway.name }}
 {{- else -}}
-{{ default "default" .Values.serviceAccount.name }}
+{{ default "default" .Values.serviceAccounts.gateway.name }}
+{{- end -}}
+{{- end -}}
+
+{{- define "litellm.backend.serviceAccountName" -}}
+{{- if .Values.serviceAccounts.backend.create -}}
+{{ default (include "litellm.backend.fullname" .) .Values.serviceAccounts.backend.name }}
+{{- else -}}
+{{ default "default" .Values.serviceAccounts.backend.name }}
+{{- end -}}
+{{- end -}}
+
+{{- define "litellm.ui.serviceAccountName" -}}
+{{- if .Values.serviceAccounts.ui.create -}}
+{{ default (include "litellm.ui.fullname" .) .Values.serviceAccounts.ui.name }}
+{{- else -}}
+{{ default "default" .Values.serviceAccounts.ui.name }}
 {{- end -}}
 {{- end -}}
 

helm/litellm/templates/backend/deployment.yaml

@@ -19,7 +19,8 @@ spec:
       labels:
         {{- include "litellm.backend.selectorLabels" . | nindent 8 }}
     spec:
-      serviceAccountName: {{ include "litellm.serviceAccountName" . }}
+      serviceAccountName: {{ include "litellm.backend.serviceAccountName" . }}
+      automountServiceAccountToken: {{ .Values.serviceAccounts.backend.automount }}
       {{- with .Values.imagePullSecrets }}
       imagePullSecrets:
         {{- toYaml . | nindent 8 }}

helm/litellm/templates/gateway/deployment.yaml

@@ -22,7 +22,8 @@ spec:
       labels:
         {{- include "litellm.gateway.selectorLabels" . | nindent 8 }}
     spec:
-      serviceAccountName: {{ include "litellm.serviceAccountName" . }}
+      serviceAccountName: {{ include "litellm.gateway.serviceAccountName" . }}
+      automountServiceAccountToken: {{ .Values.serviceAccounts.gateway.automount }}
       {{- with .Values.imagePullSecrets }}
       imagePullSecrets:
         {{- toYaml . | nindent 8 }}

helm/litellm/templates/migrations-job.yaml

@@ -28,7 +28,7 @@ spec:
         app.kubernetes.io/component: migrations
     spec:
       restartPolicy: Never
-      serviceAccountName: {{ include "litellm.serviceAccountName" . }}
+      serviceAccountName: {{ include "litellm.backend.serviceAccountName" . }}
       {{- with .Values.imagePullSecrets }}
       imagePullSecrets:
         {{- toYaml . | nindent 8 }}

helm/litellm/templates/serviceaccount.yaml

@@ -1,13 +1,51 @@
-{{- if .Values.serviceAccount.create -}}
+{{- $prev := false -}}
+{{- if .Values.serviceAccounts.gateway.create -}}
+{{- $prev = true }}
 apiVersion: v1
 kind: ServiceAccount
 metadata:
-  name: {{ include "litellm.serviceAccountName" . }}
+  name: {{ include "litellm.gateway.serviceAccountName" . }}
   labels:
     {{- include "litellm.commonLabels" . | nindent 4 }}
-  {{- with .Values.serviceAccount.annotations }}
+    app.kubernetes.io/component: gateway
+  {{- with .Values.serviceAccounts.gateway.annotations }}
   annotations:
     {{- toYaml . | nindent 4 }}
   {{- end }}
-automountServiceAccountToken: {{ .Values.serviceAccount.automount }}
+automountServiceAccountToken: {{ .Values.serviceAccounts.gateway.automount }}
+{{- end }}
+{{- if .Values.serviceAccounts.backend.create }}
+{{- if $prev }}
+---
+{{- end }}
+{{- $prev = true }}
+apiVersion: v1
+kind: ServiceAccount
+metadata:
+  name: {{ include "litellm.backend.serviceAccountName" . }}
+  labels:
+    {{- include "litellm.commonLabels" . | nindent 4 }}
+    app.kubernetes.io/component: backend
+  {{- with .Values.serviceAccounts.backend.annotations }}
+  annotations:
+    {{- toYaml . | nindent 4 }}
+  {{- end }}
+automountServiceAccountToken: {{ .Values.serviceAccounts.backend.automount }}
+{{- end }}
+{{- if .Values.serviceAccounts.ui.create }}
+{{- if $prev }}
+---
+{{- end }}
+apiVersion: v1
+kind: ServiceAccount
+metadata:
+  name: {{ include "litellm.ui.serviceAccountName" . }}
+  labels:
+    {{- include "litellm.commonLabels" . | nindent 4 }}
+    app.kubernetes.io/component: ui
+  {{- with .Values.serviceAccounts.ui.annotations }}
+  annotations:
+    {{- toYaml . | nindent 4 }}
+  {{- end }}
+automountServiceAccountToken: {{ .Values.serviceAccounts.ui.automount }}
 {{- end }}

helm/litellm/templates/ui/deployment.yaml

@@ -19,7 +19,8 @@ spec:
       labels:
         {{- include "litellm.ui.selectorLabels" . | nindent 8 }}
     spec:
-      serviceAccountName: {{ include "litellm.serviceAccountName" . }}
+      serviceAccountName: {{ include "litellm.ui.serviceAccountName" . }}
+      automountServiceAccountToken: {{ .Values.serviceAccounts.ui.automount }}
       {{- with .Values.imagePullSecrets }}
       imagePullSecrets:
         {{- toYaml . | nindent 8 }}

helm/litellm/values.yaml

@@ -14,16 +14,33 @@ ingress:
   host: ""        # optional; if set, becomes the rule's host
   tls: []
 
-# Shared ServiceAccount used by all three component Deployments. Set
-# `create: true` to have the chart provision it (e.g. when wiring an EKS
-# Pod Identity association by SA name). Set `name` to use an existing SA
-# (chart-created or out-of-band). When both are empty / false, pods run
-# with the namespace's `default` SA.
-serviceAccount:
-  create: false
-  automount: true
-  annotations: {}
-  name: ""
+# Per-component ServiceAccounts for gateway, backend, and ui.
+#
+# Each section mirrors the old shared serviceAccount shape. Set `create:
+# true` to have the chart provision the SA (useful for EKS Pod Identity /
+# GKE Workload Identity annotations). Set `name` to bind an existing SA.
+# When both are unset the component pod runs with the namespace `default` SA.
+#
+# The UI SA deliberately defaults to `automount: false` — the static nginx
+# container does not need the K8s API and should not carry a projected
+# ServiceAccount token that a compromised container could use to call the
+# cloud-provider metadata service or the K8s API.
+serviceAccounts:
+  gateway:
+    create: false
+    automount: true
+    annotations: {}
+    name: ""
+  backend:
+    create: false
+    automount: true
+    annotations: {}
+    name: ""
+  ui:
+    create: false
+    automount: false
+    annotations: {}
+    name: ""
 
 # Pre-install / pre-upgrade Helm hook that runs `prisma migrate deploy`
 # against the writer database, creating the LiteLLM schema (tables that

litellm/__init__.py

@@ -225,6 +225,11 @@ use_chat_completions_url_for_anthropic_messages: bool = bool(
 route_all_chat_openai_to_responses: bool = (
     os.getenv("LITELLM_ROUTE_ALL_CHAT_OPENAI_TO_RESPONSES", "false").lower() == "true"
 )  # When True, routes all OpenAI /chat/completions requests through the Responses API bridge
+# When True, Gemini/Vertex Live setup is deferred until client `session.update`.
+# Default False preserves historical behavior (auto-send setup on connect).
+gemini_live_defer_setup: bool = (
+    os.getenv("LITELLM_GEMINI_LIVE_DEFER_SETUP", "false").lower() == "true"
+)
 use_legacy_interactions_schema: bool = (
     os.getenv("LITELLM_USE_LEGACY_INTERACTIONS_SCHEMA", "false").lower() == "true"
 )  # When True, sends Api-Revision: 2026-05-07 to Google so responses use the legacy `outputs`
@@ -1877,6 +1882,9 @@ if TYPE_CHECKING:
     from .llms.azure.completion.transformation import (
         AzureOpenAITextConfig as AzureOpenAITextConfig,
     )
+    from .llms.azure.audio_transcription.transformation import (
+        AzureSpeechAudioTranscriptionConfig as AzureSpeechAudioTranscriptionConfig,
+    )
     from .llms.hosted_vllm.chat.transformation import (
         HostedVLLMChatConfig as HostedVLLMChatConfig,
     )

litellm/_lazy_imports_registry.py

@@ -273,6 +273,7 @@ LLM_CONFIG_NAMES = (
     "AzureOpenAIConfig",
     "AzureOpenAIGPT5Config",
     "AzureOpenAITextConfig",
+    "AzureSpeechAudioTranscriptionConfig",
     "HostedVLLMChatConfig",
     "HostedVLLMEmbeddingConfig",
     # Alias for backwards compatibility
@@ -1054,6 +1055,10 @@ _LLM_CONFIGS_IMPORT_MAP = {
         ".llms.azure.completion.transformation",
         "AzureOpenAITextConfig",
     ),
+    "AzureSpeechAudioTranscriptionConfig": (
+        ".llms.azure.audio_transcription.transformation",
+        "AzureSpeechAudioTranscriptionConfig",
+    ),
     "HostedVLLMChatConfig": (
         ".llms.hosted_vllm.chat.transformation",
         "HostedVLLMChatConfig",

litellm/_service_logger.py

@@ -24,6 +24,22 @@ else:
     UserAPIKeyAuth = Any
 
 
+def _get_otel_v2_class() -> Optional[type]:
+    """Return the ``OpenTelemetryV2`` class, or ``None`` if the OTel SDK is absent.
+
+    Imported lazily: ``litellm.integrations.otel.logger`` imports the OpenTelemetry
+    SDK at module scope, so importing it eagerly would break installs without the
+    SDK. The V2 logger only exists when ``LITELLM_OTEL_V2`` is enabled (which
+    requires the SDK), so a failed import simply means "no V2 logger in play".
+    """
+    try:
+        from litellm.integrations.otel.logger import OpenTelemetryV2
+
+        return OpenTelemetryV2
+    except Exception:
+        return None
+
+
 class ServiceLogging(CustomLogger):
     """
     Separate class used for monitoring health of litellm-adjacent services (redis/postgres).
@@ -38,6 +54,37 @@ class ServiceLogging(CustomLogger):
         if "prometheus_system" in litellm.service_callback:
             self.prometheusServicesLogger = PrometheusServicesLogger()
 
+    def _resolve_otel_service_logger(self, callback: Any) -> Optional[Any]:
+        """Resolve the OTel logger (legacy or V2) to emit a service span on.
+
+        Returns the logger instance whose ``async_service_*_hook`` should fire for
+        this ``callback``, or ``None`` when ``callback`` is not an OTel callback.
+
+        The V2 ``OpenTelemetryV2`` logger is a plain ``CustomLogger`` and is NOT a
+        subclass of the legacy ``OpenTelemetry``, so the legacy ``isinstance``
+        check alone misses it — which is why redis/postgres service spans never
+        showed up under ``LITELLM_OTEL_V2``. Match both the legacy and V2 types,
+        whether the callback is the logger instance itself or the ``"otel"`` string
+        (which routes to the proxy's registered ``open_telemetry_logger``).
+        """
+        otel_v2_cls = _get_otel_v2_class()
+
+        def _is_otel_logger(obj: Any) -> bool:
+            if isinstance(obj, OpenTelemetry):
+                return True
+            return otel_v2_cls is not None and isinstance(obj, otel_v2_cls)
+
+        if _is_otel_logger(callback):
+            return callback
+        if callback == "otel":
+            from litellm.proxy.proxy_server import open_telemetry_logger
+
+            if open_telemetry_logger is not None and _is_otel_logger(
+                open_telemetry_logger
+            ):
+                return open_telemetry_logger
+        return None
+
     def service_success_hook(
         self,
         service: ServiceTypes,
@@ -129,6 +176,13 @@ class ServiceLogging(CustomLogger):
             event_metadata=event_metadata,
         )
 
+        # OTel loggers already fired this event. ``service_callback`` can hold more
+        # than one reference that resolves to the *same* logger — the ``"otel"``
+        # string AND the registered instance both map to ``open_telemetry_logger``
+        # (the V2 logger self-registers its instance even when the string is
+        # present, unlike V1). Without this guard each such reference emits its own
+        # span, so a single DB call shows up as duplicate ``postgres ...`` spans.
+        emitted_otel_logger_ids: set = set()
         for callback in litellm.service_callback:
             if callback == "prometheus_system":
                 await self.init_prometheus_services_logger_if_none()
@@ -144,19 +198,18 @@ class ServiceLogging(CustomLogger):
                     end_time=end_time,
                     event_metadata=event_metadata,
                 )
-            elif callback == "otel" or isinstance(callback, OpenTelemetry):
-                _otel_logger_to_use: Optional[OpenTelemetry] = None
-                if isinstance(callback, OpenTelemetry):
-                    _otel_logger_to_use = callback
-                else:
-                    from litellm.proxy.proxy_server import open_telemetry_logger
-
-                    if open_telemetry_logger is not None and isinstance(
-                        open_telemetry_logger, OpenTelemetry
-                    ):
-                        _otel_logger_to_use = open_telemetry_logger
-
-                if _otel_logger_to_use is not None and parent_otel_span is not None:
+            else:
+                _otel_logger_to_use = self._resolve_otel_service_logger(callback)
+                # No ``parent_otel_span is not None`` gate: a background service
+                # call (no request on the stack) has no parent, and dropping it
+                # here is what hid those calls from traces entirely. The OTel
+                # logger decides what to do with a missing parent — legacy V1
+                # no-ops, V2 emits a root span (and skips metrics-only pings).
+                if (
+                    _otel_logger_to_use is not None
+                    and id(_otel_logger_to_use) not in emitted_otel_logger_ids
+                ):
+                    emitted_otel_logger_ids.add(id(_otel_logger_to_use))
                     await _otel_logger_to_use.async_service_success_hook(
                         payload=payload,
                         parent_otel_span=parent_otel_span,
@@ -238,6 +291,9 @@ class ServiceLogging(CustomLogger):
             event_metadata=event_metadata,
         )
 
+        # Dedupe OTel loggers per event — see ``async_service_success_hook`` for why
+        # the same logger can be referenced twice in ``service_callback``.
+        emitted_otel_logger_ids: set = set()
         for callback in litellm.service_callback:
             if callback == "prometheus_system":
                 await self.init_prometheus_services_logger_if_none()
@@ -255,22 +311,19 @@ class ServiceLogging(CustomLogger):
                     end_time=end_time,
                     event_metadata=event_metadata,
                 )
-            elif callback == "otel" or isinstance(callback, OpenTelemetry):
-                _otel_logger_to_use: Optional[OpenTelemetry] = None
-                if isinstance(callback, OpenTelemetry):
-                    _otel_logger_to_use = callback
-                else:
-                    from litellm.proxy.proxy_server import open_telemetry_logger
-
-                    if open_telemetry_logger is not None and isinstance(
-                        open_telemetry_logger, OpenTelemetry
-                    ):
-                        _otel_logger_to_use = open_telemetry_logger
+            else:
+                _otel_logger_to_use = self._resolve_otel_service_logger(callback)
 
                 if not isinstance(error, str):
                     error = str(error)
 
-                if _otel_logger_to_use is not None and parent_otel_span is not None:
+                # See the success hook: no parent gate, so background failures
+                # are traced too. V1 no-ops without a parent; V2 emits a root.
+                if (
+                    _otel_logger_to_use is not None
+                    and id(_otel_logger_to_use) not in emitted_otel_logger_ids
+                ):
+                    emitted_otel_logger_ids.add(id(_otel_logger_to_use))
                     await _otel_logger_to_use.async_service_failure_hook(
                         payload=payload,
                         error=error,

litellm/a2a_protocol/litellm_completion_bridge/handler.py

@@ -107,6 +107,14 @@ class A2ACompletionBridgeHandler:
             if k not in ("model", "custom_llm_provider") and k not in _AGENT_ONLY_PARAMS
         }
         completion_params.update(litellm_params_to_add)
+        # Apply forward metadata AFTER the litellm_params merge so the helper
+        # sees any agent-owner-configured ``extra_body.metadata`` and can keep
+        # those keys authoritative over the client-supplied A2A metadata.
+        A2ACompletionBridgeTransformation.apply_forward_metadata_to_completion_params(
+            completion_params=completion_params,
+            a2a_message=message,
+            params=params,
+        )
 
         # Call litellm.acompletion
         response = await litellm.acompletion(**completion_params)
@@ -214,6 +222,14 @@ class A2ACompletionBridgeHandler:
             if k not in ("model", "custom_llm_provider") and k not in _AGENT_ONLY_PARAMS
         }
         completion_params.update(litellm_params_to_add)
+        # Apply forward metadata AFTER the litellm_params merge so the helper
+        # sees any agent-owner-configured ``extra_body.metadata`` and can keep
+        # those keys authoritative over the client-supplied A2A metadata.
+        A2ACompletionBridgeTransformation.apply_forward_metadata_to_completion_params(
+            completion_params=completion_params,
+            a2a_message=message,
+            params=params,
+        )
 
         # 1. Emit initial task event (kind: "task", status: "submitted")
         task_event = A2ACompletionBridgeTransformation.create_task_event(ctx)

litellm/a2a_protocol/litellm_completion_bridge/transformation.py

@@ -45,10 +45,80 @@ class A2ACompletionBridgeTransformation:
     Static methods for transforming between A2A and OpenAI message formats.
     """
 
+    @staticmethod
+    def _extract_text_from_a2a_parts(parts: List[Dict[str, Any]]) -> str:
+        """Extract text from A2A parts (with or without explicit ``kind``)."""
+        content_parts: List[str] = []
+        for part in parts:
+            if not isinstance(part, dict):
+                continue
+            kind = part.get("kind")
+            text = part.get("text")
+            if text is None:
+                continue
+            if kind in (None, "", "text"):
+                content_parts.append(str(text))
+        return "\n".join(content_parts)
+
+    @staticmethod
+    def get_forward_metadata(
+        a2a_message: Dict[str, Any],
+        params: Optional[Dict[str, Any]] = None,
+    ) -> Optional[Dict[str, Any]]:
+        """
+        Merge A2A metadata from MessageSendParams and the message for downstream providers.
+
+        Forwarded once on the LangGraph run payload (``metadata``), not duplicated on
+        each input message — see ``apply_forward_metadata_to_completion_params``.
+        """
+        merged: Dict[str, Any] = {}
+        if params and isinstance(params.get("metadata"), dict):
+            merged.update(params["metadata"])
+        message_metadata = a2a_message.get("metadata")
+        if isinstance(message_metadata, dict):
+            merged.update(message_metadata)
+        return merged or None
+
+    @staticmethod
+    def apply_forward_metadata_to_completion_params(
+        completion_params: Dict[str, Any],
+        a2a_message: Dict[str, Any],
+        params: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Attach A2A metadata to completion kwargs for provider bridges (e.g. LangGraph).
+
+        Uses ``extra_body`` so we do not collide with LiteLLM's spend-log ``metadata`` kwarg.
+        """
+        forward_metadata = A2ACompletionBridgeTransformation.get_forward_metadata(
+            a2a_message=a2a_message,
+            params=params,
+        )
+        if not forward_metadata:
+            return
+
+        extra_body = completion_params.get("extra_body")
+        if not isinstance(extra_body, dict):
+            extra_body = {}
+        # Layer client-supplied A2A metadata under any agent-owner-configured
+        # ``extra_body.metadata`` so the configured keys remain authoritative
+        # and an A2A caller cannot overwrite server-set run metadata.
+        existing_metadata = extra_body.get("metadata")
+        existing_dict: Dict[str, Any] = (
+            existing_metadata if isinstance(existing_metadata, dict) else {}
+        )
+        merged_metadata: Dict[str, Any] = {**forward_metadata, **existing_dict}
+        extra_body = {**extra_body, "metadata": merged_metadata}
+        completion_params["extra_body"] = extra_body
+
+        verbose_logger.debug(
+            f"A2A -> completion forward metadata keys={list(forward_metadata.keys())}"
+        )
+
     @staticmethod
     def a2a_message_to_openai_messages(
         a2a_message: Dict[str, Any],
-    ) -> List[Dict[str, str]]:
+    ) -> List[Dict[str, Any]]:
         """
         Transform an A2A message to OpenAI message format.
 
@@ -70,21 +140,20 @@ class A2ACompletionBridgeTransformation:
         elif role == "system":
             openai_role = "system"
 
-        # Extract text content from parts
-        content_parts = []
-        for part in parts:
-            kind = part.get("kind", "")
-            if kind == "text":
-                text = part.get("text", "")
-                content_parts.append(text)
+        if not isinstance(parts, list):
+            parts = []
 
-        content = "\n".join(content_parts) if content_parts else ""
+        content = A2ACompletionBridgeTransformation._extract_text_from_a2a_parts(parts)
+
+        # Do not attach A2A message.metadata here — the completion bridge forwards it
+        # once at run level via extra_body.metadata (LangGraph POST /runs/wait shape).
+        openai_message: Dict[str, Any] = {"role": openai_role, "content": content}
 
         verbose_logger.debug(
             f"A2A -> OpenAI transform: role={role} -> {openai_role}, content_length={len(content)}"
         )
 
-        return [{"role": openai_role, "content": content}]
+        return [openai_message]
 
     @staticmethod
     def openai_response_to_a2a_response(
@@ -110,6 +179,7 @@ class A2ACompletionBridgeTransformation:
 
         # Build A2A message
         a2a_message = {
+            "kind": "message",
             "role": "agent",
             "parts": [{"kind": "text", "text": content}],
             "messageId": uuid4().hex,
@@ -119,9 +189,7 @@ class A2ACompletionBridgeTransformation:
         a2a_response = {
             "jsonrpc": "2.0",
             "id": request_id,
-            "result": {
-                "message": a2a_message,
-            },
+            "result": a2a_message,
         }
 
         verbose_logger.debug(f"OpenAI -> A2A transform: content_length={len(content)}")
@@ -235,50 +303,3 @@ class A2ACompletionBridgeTransformation:
                 "taskId": ctx.task_id,
             },
         }
-
-    @staticmethod
-    def openai_chunk_to_a2a_chunk(
-        chunk: Any,
-        request_id: Optional[str] = None,
-        is_final: bool = False,
-    ) -> Optional[Dict[str, Any]]:
-        """
-        Transform a LiteLLM streaming chunk to A2A streaming format.
-
-        NOTE: This method is deprecated for streaming. Use the event-based
-        methods (create_task_event, create_status_update_event,
-        create_artifact_update_event) instead for proper A2A streaming.
-
-        Args:
-            chunk: LiteLLM ModelResponse chunk
-            request_id: Original A2A request ID
-            is_final: Whether this is the final chunk
-
-        Returns:
-            A2A streaming chunk dict or None if no content
-        """
-        # Extract delta content
-        content = ""
-        if chunk is not None and hasattr(chunk, "choices") and chunk.choices:
-            choice = chunk.choices[0]
-            if hasattr(choice, "delta") and choice.delta:
-                content = choice.delta.content or ""
-
-        if not content and not is_final:
-            return None
-
-        # Build A2A streaming chunk (legacy format)
-        a2a_chunk = {
-            "jsonrpc": "2.0",
-            "id": request_id,
-            "result": {
-                "message": {
-                    "role": "agent",
-                    "parts": [{"kind": "text", "text": content}],
-                    "messageId": uuid4().hex,
-                },
-                "final": is_final,
-            },
-        }
-
-        return a2a_chunk

litellm/a2a_protocol/providers/litellm_completion/README.md

@@ -1,74 +0,0 @@
-# A2A to LiteLLM Completion Bridge
-
-Routes A2A protocol requests through `litellm.acompletion`, enabling any LiteLLM-supported provider to be invoked via A2A.
-
-## Flow
-
-```
-A2A Request → Transform → litellm.acompletion → Transform → A2A Response
-```
-
-## SDK Usage
-
-Use the existing `asend_message` and `asend_message_streaming` functions with `litellm_params`:
-
-```python
-from litellm.a2a_protocol import asend_message, asend_message_streaming
-from a2a.types import SendMessageRequest, SendStreamingMessageRequest, MessageSendParams
-from uuid import uuid4
-
-# Non-streaming
-request = SendMessageRequest(
-    id=str(uuid4()),
-    params=MessageSendParams(
-        message={"role": "user", "parts": [{"kind": "text", "text": "Hello!"}], "messageId": uuid4().hex}
-    )
-)
-response = await asend_message(
-    request=request,
-    api_base="http://localhost:2024",
-    litellm_params={"custom_llm_provider": "langgraph", "model": "agent"},
-)
-
-# Streaming
-stream_request = SendStreamingMessageRequest(
-    id=str(uuid4()),
-    params=MessageSendParams(
-        message={"role": "user", "parts": [{"kind": "text", "text": "Hello!"}], "messageId": uuid4().hex}
-    )
-)
-async for chunk in asend_message_streaming(
-    request=stream_request,
-    api_base="http://localhost:2024",
-    litellm_params={"custom_llm_provider": "langgraph", "model": "agent"},
-):
-    print(chunk)
-```
-
-## Proxy Usage
-
-Configure an agent with `custom_llm_provider` in `litellm_params`:
-
-```yaml
-agents:
-  - agent_name: my-langgraph-agent
-    agent_card_params:
-      name: "LangGraph Agent"
-      url: "http://localhost:2024"  # Used as api_base
-    litellm_params:
-      custom_llm_provider: langgraph
-      model: agent
-```
-
-When an A2A request hits `/a2a/{agent_id}/message/send`, the bridge:
-
-1. Detects `custom_llm_provider` in agent's `litellm_params`
-2. Transforms A2A message → OpenAI messages
-3. Calls `litellm.acompletion(model="langgraph/agent", api_base="http://localhost:2024")`
-4. Transforms response → A2A format
-
-## Classes
-
-- `A2ACompletionBridgeTransformation` - Static methods for message format conversion
-- `A2ACompletionBridgeHandler` - Static methods for handling requests (streaming/non-streaming)
-

litellm/a2a_protocol/providers/litellm_completion/__init__.py

@@ -1,5 +0,0 @@
-"""
-LiteLLM Completion bridge provider for A2A protocol.
-
-Routes A2A requests through litellm.acompletion based on custom_llm_provider.
-"""

litellm/a2a_protocol/providers/litellm_completion/handler.py

@@ -1,301 +0,0 @@
-"""
-Handler for A2A to LiteLLM completion bridge.
-
-Routes A2A requests through litellm.acompletion based on custom_llm_provider.
-
-A2A Streaming Events (in order):
-1. Task event (kind: "task") - Initial task creation with status "submitted"
-2. Status update (kind: "status-update") - Status change to "working"
-3. Artifact update (kind: "artifact-update") - Content/artifact delivery
-4. Status update (kind: "status-update") - Final status "completed" with final=true
-"""
-
-from typing import Any, AsyncIterator, Dict, Optional
-
-import litellm
-from litellm._logging import verbose_logger
-from litellm.a2a_protocol.litellm_completion_bridge.pydantic_ai_transformation import (
-    PydanticAITransformation,
-)
-from litellm.a2a_protocol.litellm_completion_bridge.transformation import (
-    A2ACompletionBridgeTransformation,
-    A2AStreamingContext,
-)
-
-
-class A2ACompletionBridgeHandler:
-    """
-    Static methods for handling A2A requests via LiteLLM completion.
-    """
-
-    @staticmethod
-    async def handle_non_streaming(
-        request_id: str,
-        params: Dict[str, Any],
-        litellm_params: Dict[str, Any],
-        api_base: Optional[str] = None,
-    ) -> Dict[str, Any]:
-        """
-        Handle non-streaming A2A request via litellm.acompletion.
-
-        Args:
-            request_id: A2A JSON-RPC request ID
-            params: A2A MessageSendParams containing the message
-            litellm_params: Agent's litellm_params (custom_llm_provider, model, etc.)
-            api_base: API base URL from agent_card_params
-
-        Returns:
-            A2A SendMessageResponse dict
-        """
-        # Check if this is a Pydantic AI agent request
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        if custom_llm_provider == "pydantic_ai_agents":
-            if api_base is None:
-                raise ValueError("api_base is required for Pydantic AI agents")
-
-            verbose_logger.info(
-                f"Pydantic AI: Routing to Pydantic AI agent at {api_base}"
-            )
-
-            # Send request directly to Pydantic AI agent
-            response_data = await PydanticAITransformation.send_non_streaming_request(
-                api_base=api_base,
-                request_id=request_id,
-                params=params,
-            )
-
-            return response_data
-
-        # Extract message from params
-        message = params.get("message", {})
-
-        # Transform A2A message to OpenAI format
-        openai_messages = (
-            A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
-        )
-
-        # Get completion params
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        model = litellm_params.get("model", "agent")
-
-        # Build full model string if provider specified
-        # Skip prepending if model already starts with the provider prefix
-        if custom_llm_provider and not model.startswith(f"{custom_llm_provider}/"):
-            full_model = f"{custom_llm_provider}/{model}"
-        else:
-            full_model = model
-
-        verbose_logger.info(
-            f"A2A completion bridge: model={full_model}, api_base={api_base}"
-        )
-
-        # Build completion params dict
-        completion_params = {
-            "model": full_model,
-            "messages": openai_messages,
-            "api_base": api_base,
-            "stream": False,
-        }
-        # Add litellm_params (contains api_key, client_id, client_secret, tenant_id, etc.)
-        litellm_params_to_add = {
-            k: v
-            for k, v in litellm_params.items()
-            if k not in ("model", "custom_llm_provider")
-        }
-        completion_params.update(litellm_params_to_add)
-
-        # Call litellm.acompletion
-        response = await litellm.acompletion(**completion_params)
-
-        # Transform response to A2A format
-        a2a_response = (
-            A2ACompletionBridgeTransformation.openai_response_to_a2a_response(
-                response=response,
-                request_id=request_id,
-            )
-        )
-
-        verbose_logger.info(f"A2A completion bridge completed: request_id={request_id}")
-
-        return a2a_response
-
-    @staticmethod
-    async def handle_streaming(
-        request_id: str,
-        params: Dict[str, Any],
-        litellm_params: Dict[str, Any],
-        api_base: Optional[str] = None,
-    ) -> AsyncIterator[Dict[str, Any]]:
-        """
-        Handle streaming A2A request via litellm.acompletion with stream=True.
-
-        Emits proper A2A streaming events:
-        1. Task event (kind: "task") - Initial task with status "submitted"
-        2. Status update (kind: "status-update") - Status "working"
-        3. Artifact update (kind: "artifact-update") - Content delivery
-        4. Status update (kind: "status-update") - Final "completed" status
-
-        Args:
-            request_id: A2A JSON-RPC request ID
-            params: A2A MessageSendParams containing the message
-            litellm_params: Agent's litellm_params (custom_llm_provider, model, etc.)
-            api_base: API base URL from agent_card_params
-
-        Yields:
-            A2A streaming response events
-        """
-        # Check if this is a Pydantic AI agent request
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        if custom_llm_provider == "pydantic_ai_agents":
-            if api_base is None:
-                raise ValueError("api_base is required for Pydantic AI agents")
-
-            verbose_logger.info(
-                f"Pydantic AI: Faking streaming for Pydantic AI agent at {api_base}"
-            )
-
-            # Get non-streaming response first
-            response_data = await PydanticAITransformation.send_non_streaming_request(
-                api_base=api_base,
-                request_id=request_id,
-                params=params,
-            )
-
-            # Convert to fake streaming
-            async for chunk in PydanticAITransformation.fake_streaming_from_response(
-                response_data=response_data,
-                request_id=request_id,
-            ):
-                yield chunk
-
-            return
-
-        # Extract message from params
-        message = params.get("message", {})
-
-        # Create streaming context
-        ctx = A2AStreamingContext(
-            request_id=request_id,
-            input_message=message,
-        )
-
-        # Transform A2A message to OpenAI format
-        openai_messages = (
-            A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
-        )
-
-        # Get completion params
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        model = litellm_params.get("model", "agent")
-
-        # Build full model string if provider specified
-        # Skip prepending if model already starts with the provider prefix
-        if custom_llm_provider and not model.startswith(f"{custom_llm_provider}/"):
-            full_model = f"{custom_llm_provider}/{model}"
-        else:
-            full_model = model
-
-        verbose_logger.info(
-            f"A2A completion bridge streaming: model={full_model}, api_base={api_base}"
-        )
-
-        # Build completion params dict
-        completion_params = {
-            "model": full_model,
-            "messages": openai_messages,
-            "api_base": api_base,
-            "stream": True,
-        }
-        # Add litellm_params (contains api_key, client_id, client_secret, tenant_id, etc.)
-        litellm_params_to_add = {
-            k: v
-            for k, v in litellm_params.items()
-            if k not in ("model", "custom_llm_provider")
-        }
-        completion_params.update(litellm_params_to_add)
-
-        # 1. Emit initial task event (kind: "task", status: "submitted")
-        task_event = A2ACompletionBridgeTransformation.create_task_event(ctx)
-        yield task_event
-
-        # 2. Emit status update (kind: "status-update", status: "working")
-        working_event = A2ACompletionBridgeTransformation.create_status_update_event(
-            ctx=ctx,
-            state="working",
-            final=False,
-            message_text="Processing request...",
-        )
-        yield working_event
-
-        # Call litellm.acompletion with streaming
-        response = await litellm.acompletion(**completion_params)
-
-        # 3. Accumulate content and emit artifact update
-        accumulated_text = ""
-        chunk_count = 0
-        async for chunk in response:  # type: ignore[union-attr]
-            chunk_count += 1
-
-            # Extract delta content
-            content = ""
-            if chunk is not None and hasattr(chunk, "choices") and chunk.choices:
-                choice = chunk.choices[0]
-                if hasattr(choice, "delta") and choice.delta:
-                    content = choice.delta.content or ""
-
-            if content:
-                accumulated_text += content
-
-        # Emit artifact update with accumulated content
-        if accumulated_text:
-            artifact_event = (
-                A2ACompletionBridgeTransformation.create_artifact_update_event(
-                    ctx=ctx,
-                    text=accumulated_text,
-                )
-            )
-            yield artifact_event
-
-        # 4. Emit final status update (kind: "status-update", status: "completed", final: true)
-        completed_event = A2ACompletionBridgeTransformation.create_status_update_event(
-            ctx=ctx,
-            state="completed",
-            final=True,
-        )
-        yield completed_event
-
-        verbose_logger.info(
-            f"A2A completion bridge streaming completed: request_id={request_id}, chunks={chunk_count}"
-        )
-
-
-# Convenience functions that delegate to the class methods
-async def handle_a2a_completion(
-    request_id: str,
-    params: Dict[str, Any],
-    litellm_params: Dict[str, Any],
-    api_base: Optional[str] = None,
-) -> Dict[str, Any]:
-    """Convenience function for non-streaming A2A completion."""
-    return await A2ACompletionBridgeHandler.handle_non_streaming(
-        request_id=request_id,
-        params=params,
-        litellm_params=litellm_params,
-        api_base=api_base,
-    )
-
-
-async def handle_a2a_completion_streaming(
-    request_id: str,
-    params: Dict[str, Any],
-    litellm_params: Dict[str, Any],
-    api_base: Optional[str] = None,
-) -> AsyncIterator[Dict[str, Any]]:
-    """Convenience function for streaming A2A completion."""
-    async for chunk in A2ACompletionBridgeHandler.handle_streaming(
-        request_id=request_id,
-        params=params,
-        litellm_params=litellm_params,
-        api_base=api_base,
-    ):
-        yield chunk

litellm/a2a_protocol/providers/litellm_completion/transformation.py

@@ -1,284 +0,0 @@
-"""
-Transformation utilities for A2A <-> OpenAI message format conversion.
-
-A2A Message Format:
-{
-    "role": "user",
-    "parts": [{"kind": "text", "text": "Hello!"}],
-    "messageId": "abc123"
-}
-
-OpenAI Message Format:
-{"role": "user", "content": "Hello!"}
-
-A2A Streaming Events:
-- Task event (kind: "task") - Initial task creation with status "submitted"
-- Status update (kind: "status-update") - Status changes (working, completed)
-- Artifact update (kind: "artifact-update") - Content/artifact delivery
-"""
-
-from datetime import datetime, timezone
-from typing import Any, Dict, List, Optional
-from uuid import uuid4
-
-from litellm._logging import verbose_logger
-
-
-class A2AStreamingContext:
-    """
-    Context holder for A2A streaming state.
-    Tracks task_id, context_id, and message accumulation.
-    """
-
-    def __init__(self, request_id: str, input_message: Dict[str, Any]):
-        self.request_id = request_id
-        self.task_id = str(uuid4())
-        self.context_id = str(uuid4())
-        self.input_message = input_message
-        self.accumulated_text = ""
-        self.has_emitted_task = False
-        self.has_emitted_working = False
-
-
-class A2ACompletionBridgeTransformation:
-    """
-    Static methods for transforming between A2A and OpenAI message formats.
-    """
-
-    @staticmethod
-    def a2a_message_to_openai_messages(
-        a2a_message: Dict[str, Any],
-    ) -> List[Dict[str, str]]:
-        """
-        Transform an A2A message to OpenAI message format.
-
-        Args:
-            a2a_message: A2A message with role, parts, and messageId
-
-        Returns:
-            List of OpenAI-format messages
-        """
-        role = a2a_message.get("role", "user")
-        parts = a2a_message.get("parts", [])
-
-        # Map A2A roles to OpenAI roles
-        openai_role = role
-        if role == "user":
-            openai_role = "user"
-        elif role == "assistant":
-            openai_role = "assistant"
-        elif role == "system":
-            openai_role = "system"
-
-        # Extract text content from parts
-        content_parts = []
-        for part in parts:
-            kind = part.get("kind", "")
-            if kind == "text":
-                text = part.get("text", "")
-                content_parts.append(text)
-
-        content = "\n".join(content_parts) if content_parts else ""
-
-        verbose_logger.debug(
-            f"A2A -> OpenAI transform: role={role} -> {openai_role}, content_length={len(content)}"
-        )
-
-        return [{"role": openai_role, "content": content}]
-
-    @staticmethod
-    def openai_response_to_a2a_response(
-        response: Any,
-        request_id: Optional[str] = None,
-    ) -> Dict[str, Any]:
-        """
-        Transform a LiteLLM ModelResponse to A2A SendMessageResponse format.
-
-        Args:
-            response: LiteLLM ModelResponse object
-            request_id: Original A2A request ID
-
-        Returns:
-            A2A SendMessageResponse dict
-        """
-        # Extract content from response
-        content = ""
-        if hasattr(response, "choices") and response.choices:
-            choice = response.choices[0]
-            if hasattr(choice, "message") and choice.message:
-                content = choice.message.content or ""
-
-        # Build A2A message
-        a2a_message = {
-            "role": "agent",
-            "parts": [{"kind": "text", "text": content}],
-            "messageId": uuid4().hex,
-        }
-
-        # Build A2A response
-        a2a_response = {
-            "jsonrpc": "2.0",
-            "id": request_id,
-            "result": {
-                "message": a2a_message,
-            },
-        }
-
-        verbose_logger.debug(f"OpenAI -> A2A transform: content_length={len(content)}")
-
-        return a2a_response
-
-    @staticmethod
-    def _get_timestamp() -> str:
-        """Get current timestamp in ISO format with timezone."""
-        return datetime.now(timezone.utc).isoformat()
-
-    @staticmethod
-    def create_task_event(
-        ctx: A2AStreamingContext,
-    ) -> Dict[str, Any]:
-        """
-        Create the initial task event with status 'submitted'.
-
-        This is the first event emitted in an A2A streaming response.
-        """
-        return {
-            "id": ctx.request_id,
-            "jsonrpc": "2.0",
-            "result": {
-                "contextId": ctx.context_id,
-                "history": [
-                    {
-                        "contextId": ctx.context_id,
-                        "kind": "message",
-                        "messageId": ctx.input_message.get("messageId", uuid4().hex),
-                        "parts": ctx.input_message.get("parts", []),
-                        "role": ctx.input_message.get("role", "user"),
-                        "taskId": ctx.task_id,
-                    }
-                ],
-                "id": ctx.task_id,
-                "kind": "task",
-                "status": {
-                    "state": "submitted",
-                },
-            },
-        }
-
-    @staticmethod
-    def create_status_update_event(
-        ctx: A2AStreamingContext,
-        state: str,
-        final: bool = False,
-        message_text: Optional[str] = None,
-    ) -> Dict[str, Any]:
-        """
-        Create a status update event.
-
-        Args:
-            ctx: Streaming context
-            state: Status state ('working', 'completed')
-            final: Whether this is the final event
-            message_text: Optional message text for 'working' status
-        """
-        status: Dict[str, Any] = {
-            "state": state,
-            "timestamp": A2ACompletionBridgeTransformation._get_timestamp(),
-        }
-
-        # Add message for 'working' status
-        if state == "working" and message_text:
-            status["message"] = {
-                "contextId": ctx.context_id,
-                "kind": "message",
-                "messageId": str(uuid4()),
-                "parts": [{"kind": "text", "text": message_text}],
-                "role": "agent",
-                "taskId": ctx.task_id,
-            }
-
-        return {
-            "id": ctx.request_id,
-            "jsonrpc": "2.0",
-            "result": {
-                "contextId": ctx.context_id,
-                "final": final,
-                "kind": "status-update",
-                "status": status,
-                "taskId": ctx.task_id,
-            },
-        }
-
-    @staticmethod
-    def create_artifact_update_event(
-        ctx: A2AStreamingContext,
-        text: str,
-    ) -> Dict[str, Any]:
-        """
-        Create an artifact update event with content.
-
-        Args:
-            ctx: Streaming context
-            text: The text content for the artifact
-        """
-        return {
-            "id": ctx.request_id,
-            "jsonrpc": "2.0",
-            "result": {
-                "artifact": {
-                    "artifactId": str(uuid4()),
-                    "name": "response",
-                    "parts": [{"kind": "text", "text": text}],
-                },
-                "contextId": ctx.context_id,
-                "kind": "artifact-update",
-                "taskId": ctx.task_id,
-            },
-        }
-
-    @staticmethod
-    def openai_chunk_to_a2a_chunk(
-        chunk: Any,
-        request_id: Optional[str] = None,
-        is_final: bool = False,
-    ) -> Optional[Dict[str, Any]]:
-        """
-        Transform a LiteLLM streaming chunk to A2A streaming format.
-
-        NOTE: This method is deprecated for streaming. Use the event-based
-        methods (create_task_event, create_status_update_event,
-        create_artifact_update_event) instead for proper A2A streaming.
-
-        Args:
-            chunk: LiteLLM ModelResponse chunk
-            request_id: Original A2A request ID
-            is_final: Whether this is the final chunk
-
-        Returns:
-            A2A streaming chunk dict or None if no content
-        """
-        # Extract delta content
-        content = ""
-        if chunk is not None and hasattr(chunk, "choices") and chunk.choices:
-            choice = chunk.choices[0]
-            if hasattr(choice, "delta") and choice.delta:
-                content = choice.delta.content or ""
-
-        if not content and not is_final:
-            return None
-
-        # Build A2A streaming chunk (legacy format)
-        a2a_chunk = {
-            "jsonrpc": "2.0",
-            "id": request_id,
-            "result": {
-                "message": {
-                    "role": "agent",
-                    "parts": [{"kind": "text", "text": content}],
-                    "messageId": uuid4().hex,
-                },
-                "final": is_final,
-            },
-        }
-
-        return a2a_chunk

litellm/a2a_protocol/providers/pydantic_ai_agents/transformation.py

@@ -289,16 +289,16 @@ class PydanticAITransformation:
         Transform Pydantic AI task response to standard A2A non-streaming format.
 
         Pydantic AI returns a task with history/artifacts, but the standard A2A
-        non-streaming format expects:
+        non-streaming format expects ``result`` to be the Message directly
+        (``kind="message"``), per the A2A spec / ``SendMessageResponse``:
         {
             "jsonrpc": "2.0",
             "id": "...",
             "result": {
-                "message": {
-                    "role": "agent",
-                    "parts": [{"kind": "text", "text": "..."}],
-                    "messageId": "..."
-                }
+                "kind": "message",
+                "role": "agent",
+                "parts": [{"kind": "text", "text": "..."}],
+                "messageId": "..."
             }
         }
 
@@ -316,6 +316,7 @@ class PydanticAITransformation:
 
         # Build standard A2A message
         a2a_message = {
+            "kind": "message",
             "role": "agent",
             "parts": parts if parts else [{"kind": "text", "text": full_text}],
             "messageId": message_id,
@@ -325,9 +326,7 @@ class PydanticAITransformation:
         return {
             "jsonrpc": "2.0",
             "id": request_id,
-            "result": {
-                "message": a2a_message,
-            },
+            "result": a2a_message,
         }
 
     @staticmethod

litellm/a2a_protocol/utils.py

@@ -60,6 +60,12 @@ class A2ARequestUtils:
         if not isinstance(result, dict):
             return ""
 
+        # Direct message format (A2A spec): detect by explicit kind tag only.
+        # The "parts" heuristic is too broad and would match any future result
+        # type that happens to include a "parts" field.
+        if result.get("kind") == "message":
+            return A2ARequestUtils.extract_text_from_message(result)
+
         message = result.get("message", {})
         return A2ARequestUtils.extract_text_from_message(message)
 

litellm/anthropic_interface/messages/__init__.py

@@ -10,7 +10,7 @@ This is an __init__.py file to allow the following interface
 
 """
 
-from typing import Any, AsyncIterator, Coroutine, Dict, List, Optional, Union
+from typing import Any, AsyncIterator, Coroutine, Dict, Iterator, List, Optional, Union
 
 from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
     anthropic_messages as _async_anthropic_messages,
@@ -100,8 +100,11 @@ def create(
     **kwargs,
 ) -> Union[
     AnthropicMessagesResponse,
+    Iterator[bytes],
     AsyncIterator[Any],
-    Coroutine[Any, Any, Union[AnthropicMessagesResponse, AsyncIterator[Any]]],
+    Coroutine[
+        Any, Any, Union[AnthropicMessagesResponse, AsyncIterator[Any], Iterator[bytes]]
+    ],
 ]:
     """
     Async wrapper for Anthropic's messages API

litellm/constants.py

@@ -1147,6 +1147,7 @@ BEDROCK_CONVERSE_MODELS = [
     "openai.gpt-oss-120b-1:0",
     "anthropic.claude-haiku-4-5-20251001-v1:0",
     "anthropic.claude-sonnet-4-5-20250929-v1:0",
+    "anthropic.claude-opus-4-8",
     "anthropic.claude-opus-4-7",
     "anthropic.claude-opus-4-6-v1:0",
     "anthropic.claude-opus-4-6-v1",
@@ -1408,6 +1409,13 @@ LITELLM_INTERNAL_JOBS_SERVICE_ACCOUNT_NAME = "litellm_internal_jobs"
 # Prometheus metrics, audit trails, or any other downstream consumer.
 LITELLM_PROXY_MASTER_KEY_ALIAS = "litellm_proxy_master_key"
 
+# Marker placed in ``model_call_details`` on a synthetic ``Logging`` object that
+# records a proxy-gate error (auth/rate-limit rejection) for a request that never
+# reached an upstream provider. Tracing callbacks key off it to avoid fabricating
+# an LLM-call span for a call that did not happen. See
+# ``ProxyLogging._handle_logging_proxy_only_error``.
+LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL = "litellm_no_upstream_llm_call"
+
 # Key Rotation Constants
 LITELLM_KEY_ROTATION_ENABLED = os.getenv("LITELLM_KEY_ROTATION_ENABLED", "false")
 LITELLM_KEY_ROTATION_CHECK_INTERVAL_SECONDS = int(
@@ -1443,6 +1451,12 @@ CLI_JWT_EXPIRATION_HOURS = int(
     or os.getenv("LITELLM_CLI_JWT_EXPIRATION_HOURS")
     or 24
 )
+# Comma-separated allowlisted OIDC claim map for CLI SSO polling, e.g.
+# "employment_type->acme_employment_type,org_info.department->department"
+CLI_SSO_CLAIM_MAP = (
+    os.getenv("CLI_SSO_CLAIM_MAP") or os.getenv("LITELLM_CLI_SSO_CLAIM_MAP") or ""
+)
+CLI_SSO_CLAIM_MAX_SCALAR_LENGTH = 1024
 
 ########################### UI SESSION DURATION ###########################
 # Duration for UI login session (username/password, SSO, invitation links). Format: "30s", "30m", "24h", "7d"

litellm/cost_calculator.py

@@ -24,6 +24,7 @@ from litellm.litellm_core_utils.llm_cost_calc.usage_object_transformation import
 from litellm.litellm_core_utils.llm_cost_calc.utils import (
     CostCalculatorUtils,
     _generic_cost_per_character,
+    _get_regional_uplift_multiplier,
     _get_service_tier_cost_key,
     _parse_prompt_tokens_details,
     calculate_cost_component,
@@ -132,6 +133,8 @@ _VIDEO_CALL_TYPES = frozenset(
     {
         CallTypes.create_video.value,
         CallTypes.acreate_video.value,
+        CallTypes.video_edit.value,
+        CallTypes.avideo_edit.value,
         CallTypes.video_remix.value,
         CallTypes.avideo_remix.value,
     }
@@ -312,6 +315,10 @@ def cost_per_token(  # noqa: PLR0915
     audio_transcription_file_duration: float = 0.0,  # for audio transcription calls - the file time in seconds
     ### SERVICE TIER ###
     service_tier: Optional[str] = None,  # for OpenAI service tier pricing
+    ### DATA RESIDENCY ###
+    data_residency: Optional[
+        str
+    ] = None,  # for OpenAI regional-processing uplift (e.g. "eu", "us")
     response: Optional[Any] = None,
     ### REQUEST MODEL ###
     request_model: Optional[str] = None,  # original request model for router detection
@@ -412,9 +419,36 @@ def cost_per_token(  # noqa: PLR0915
     prompt_tokens_cost_usd_dollar: float = 0
     completion_tokens_cost_usd_dollar: float = 0
     model_cost_ref = litellm.model_cost
+    # Only callers that explicitly pass `custom_llm_provider` get the
+    # dedup/prefix-join treatment. When provider is omitted, preserve legacy
+    # behavior: `model_with_provider` stays equal to the raw `model` string
+    # (provider is detected below for downstream use only).
+    caller_supplied_provider = custom_llm_provider is not None
+
+    # `model` is normally a string, but callers that mock the transport can pass
+    # non-string objects. Only run the string-based dedup/prefix-join when it is
+    # actually a string — e.g. a MagicMock's `.startswith()` is always truthy and
+    # its slices return new mocks, which would spin the dedup loop forever.
+    model_is_str = isinstance(model, str)
+
+    # Router/proxy deployments may repeat the provider segment (e.g. model_name
+    # "openai/openai/gpt-5.5"). Strip duplicated `{provider}/` chains before joining.
+    if caller_supplied_provider and model_is_str:
+        _dup_prefix = f"{custom_llm_provider}/"
+        while model.startswith(_dup_prefix):
+            _remainder = model[len(_dup_prefix) :]
+            if _remainder.startswith(_dup_prefix):
+                model = _remainder
+            else:
+                break
+
     model_with_provider = model
-    if custom_llm_provider is not None:
-        model_with_provider = custom_llm_provider + "/" + model
+    if caller_supplied_provider:
+        _prov_prefix = f"{custom_llm_provider}/"
+        if model_is_str and model.startswith(_prov_prefix):
+            model_with_provider = model
+        else:
+            model_with_provider = f"{custom_llm_provider}/{model}"
         if region_name is not None:
             model_with_provider_and_region = (
                 f"{custom_llm_provider}/{region_name}/{model}"
@@ -425,6 +459,9 @@ def cost_per_token(  # noqa: PLR0915
                 model_with_provider = model_with_provider_and_region
     else:
         _, custom_llm_provider, _, _ = litellm.get_llm_provider(model=model)
+
+    assert custom_llm_provider is not None  # caller-supplied or get_llm_provider
+
     model_without_prefix = model
     model_parts = model.split("/", 1)
     if len(model_parts) > 1:
@@ -493,6 +530,7 @@ def cost_per_token(  # noqa: PLR0915
                 usage=usage_block,
                 custom_llm_provider=custom_llm_provider,
                 service_tier=service_tier,
+                data_residency=data_residency,
             )
 
         return prompt_cost, completion_cost
@@ -521,7 +559,10 @@ def cost_per_token(  # noqa: PLR0915
         or call_type == CallTypes.retrieve_batch
     ):
         return batch_cost_calculator(
-            usage=usage_block, model=model, custom_llm_provider=custom_llm_provider
+            usage=usage_block,
+            model=model,
+            custom_llm_provider=custom_llm_provider,
+            data_residency=data_residency,
         )
     elif call_type == "atranscription" or call_type == "transcription":
         if _transcription_usage_has_token_details(usage_block):
@@ -529,6 +570,7 @@ def cost_per_token(  # noqa: PLR0915
                 model=model_without_prefix,
                 usage=usage_block,
                 service_tier=service_tier,
+                data_residency=data_residency,
             )
 
         return openai_cost_per_second(
@@ -579,7 +621,10 @@ def cost_per_token(  # noqa: PLR0915
         )
     elif custom_llm_provider == "openai":
         return openai_cost_per_token(
-            model=model, usage=usage_block, service_tier=service_tier
+            model=model,
+            usage=usage_block,
+            service_tier=service_tier,
+            data_residency=data_residency,
         )
     elif custom_llm_provider == "databricks":
         return databricks_cost_per_token(model=model, usage=usage_block)
@@ -631,6 +676,7 @@ def cost_per_token(  # noqa: PLR0915
                 usage=usage_block,
                 custom_llm_provider=custom_llm_provider,
                 service_tier=service_tier,
+                data_residency=data_residency,
             )
 
         if (
@@ -1117,6 +1163,10 @@ def completion_cost(  # noqa: PLR0915
     litellm_logging_obj: Optional[LitellmLoggingObject] = None,
     ### SERVICE TIER ###
     service_tier: Optional[str] = None,  # for OpenAI service tier pricing
+    ### DATA RESIDENCY ###
+    data_residency: Optional[
+        str
+    ] = None,  # for OpenAI regional-processing uplift (e.g. "eu", "us")
 ) -> float:
     """
     Calculate the cost of a given completion call fot GPT-3.5-turbo, llama2, any litellm supported llm.
@@ -1516,6 +1566,7 @@ def completion_cost(  # noqa: PLR0915
                         combined_usage_object=cost_per_token_usage_object,
                         custom_llm_provider=custom_llm_provider,
                         litellm_model_name=model,
+                        data_residency=data_residency,
                     )
                 elif call_type == _MCP_CALL_TYPE:
                     from litellm.proxy._experimental.mcp_server.cost_calculator import (
@@ -1600,6 +1651,7 @@ def completion_cost(  # noqa: PLR0915
                     audio_transcription_file_duration=audio_transcription_file_duration,
                     rerank_billed_units=rerank_billed_units,
                     service_tier=service_tier,
+                    data_residency=data_residency,
                     response=completion_response,
                     request_model=request_model_for_cost,
                 )
@@ -1811,6 +1863,10 @@ def response_cost_calculator(
     litellm_logging_obj: Optional[LitellmLoggingObject] = None,
     ### SERVICE TIER ###
     service_tier: Optional[str] = None,  # for OpenAI service tier pricing
+    ### DATA RESIDENCY ###
+    data_residency: Optional[
+        str
+    ] = None,  # for OpenAI regional-processing uplift (e.g. "eu", "us")
 ) -> float:
     """
     Returns
@@ -1844,6 +1900,7 @@ def response_cost_calculator(
                 router_model_id=router_model_id,
                 litellm_logging_obj=litellm_logging_obj,
                 service_tier=service_tier,
+                data_residency=data_residency,
             )
         return response_cost
     except Exception as e:
@@ -2202,6 +2259,7 @@ def batch_cost_calculator(
     model: str,
     custom_llm_provider: Optional[str] = None,
     model_info: Optional[ModelInfo] = None,
+    data_residency: Optional[str] = None,
 ) -> Tuple[float, float]:
     """
     Calculate the cost of a batch job.
@@ -2286,6 +2344,11 @@ def batch_cost_calculator(
             usage.completion_tokens * (output_cost_per_token) / 2
         )  # batch cost is usually half of the regular token cost
 
+    uplift = _get_regional_uplift_multiplier(model_info, data_residency)
+    if uplift != 1.0:
+        total_prompt_cost *= uplift
+        total_completion_cost *= uplift
+
     return total_prompt_cost, total_completion_cost
 
 
@@ -2431,6 +2494,7 @@ def handle_realtime_stream_cost_calculation(
     combined_usage_object: Usage,
     custom_llm_provider: str,
     litellm_model_name: str,
+    data_residency: Optional[str] = None,
 ) -> float:
     """
     Handles the cost calculation for realtime stream responses.
@@ -2461,6 +2525,7 @@ def handle_realtime_stream_cost_calculation(
                 model=model_name,
                 usage=combined_usage_object,
                 custom_llm_provider=custom_llm_provider,
+                data_residency=data_residency,
             )
         except Exception:
             continue

litellm/integrations/arize/arize_phoenix.py

@@ -1,5 +1,7 @@
 import os
-from typing import TYPE_CHECKING, Any, Optional, Union
+import threading
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Optional, Tuple, Union
 
 from litellm._logging import verbose_logger
 from litellm.integrations.arize import _utils
@@ -8,8 +10,10 @@ from litellm.types.integrations.arize_phoenix import ArizePhoenixConfig
 
 if TYPE_CHECKING:
     from opentelemetry.sdk.trace import TracerProvider
+    from opentelemetry.sdk.trace.export import SpanProcessor
     from opentelemetry.trace import Span as _Span
     from opentelemetry.trace import SpanKind
+    from opentelemetry.trace import Tracer
 
     from litellm.integrations.opentelemetry import OpenTelemetry as _OpenTelemetry
     from litellm.integrations.opentelemetry import (
@@ -21,20 +25,27 @@ if TYPE_CHECKING:
     OpenTelemetryConfig = _OpenTelemetryConfig
     Span = Union[_Span, Any]
     OpenTelemetry = _OpenTelemetry
+    LITELLM_TRACER_NAME: str
 else:
     Protocol = Any
     OpenTelemetryConfig = Any
     Span = Any
+    Tracer = Any
     TracerProvider = Any
     SpanKind = Any
-    # Import OpenTelemetry at runtime
+    SpanProcessor = Any
     try:
-        from litellm.integrations.opentelemetry import OpenTelemetry
+        from litellm.integrations.opentelemetry import (
+            LITELLM_TRACER_NAME,
+            OpenTelemetry,
+        )
     except ImportError:
+        LITELLM_TRACER_NAME = "litellm"
         OpenTelemetry = None  # type: ignore
 
 
 ARIZE_HOSTED_PHOENIX_ENDPOINT = "https://otlp.arize.com/v1/traces"
+_MAX_PROJECT_PROVIDERS = 64
 
 
 class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
@@ -48,37 +59,142 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
 
     def _init_tracing(self, tracer_provider):
         """
-        Override to always create a *private* TracerProvider for Arize Phoenix.
+        Override to create per-project TracerProviders (LRU-cached) for Arize Phoenix.
 
         The base ``OpenTelemetry._init_tracing`` falls back to the global
         TracerProvider when one already exists.  That causes whichever
         integration initialises second to silently reuse the first one's
         exporter, so spans only reach one destination.
-
-        By creating our own provider we guarantee Arize Phoenix always gets
-        its own exporter pipeline, regardless of initialisation order.
         """
-        from opentelemetry.sdk.trace import TracerProvider
         from opentelemetry.trace import SpanKind
 
         if tracer_provider is not None:
-            # Explicitly supplied (e.g. in tests) — honour it.
-            self.tracer = tracer_provider.get_tracer("litellm")
+            self._use_injected_tracer_provider = True
+            self._shared_span_processor = None
+            self.tracer = tracer_provider.get_tracer(LITELLM_TRACER_NAME)
             self.span_kind = SpanKind
             return
 
-        # Always create a dedicated provider — never touch the global one.
-        provider = TracerProvider(resource=self._get_litellm_resource(self.config))
-        provider.add_span_processor(self._get_span_processor())
-        self.tracer = provider.get_tracer("litellm")
+        self._use_injected_tracer_provider = False
+        self._project_providers: OrderedDict[str, TracerProvider] = OrderedDict()
+        self._project_providers_lock = threading.Lock()
+        self._shared_span_processor = self._get_span_processor()
         self.span_kind = SpanKind
+
+        default_project = self._resolve_project_name({})
+        self.tracer = self._get_tracer_for(default_project)
         verbose_logger.debug(
-            "ArizePhoenixLogger: Created dedicated TracerProvider "
-            "(endpoint=%s, exporter=%s)",
+            "ArizePhoenixLogger: Initialized per-project TracerProvider cache "
+            "(default_project=%s, endpoint=%s, exporter=%s)",
+            default_project,
             self.config.endpoint,
             self.config.exporter,
         )
 
+    def flush_tracer_providers(self) -> None:
+        """
+        Flush all cached per-project providers and the shared span processor.
+
+        Call on graceful proxy shutdown. Do not call on LRU eviction — in-flight
+        spans may still reference evicted providers.
+        """
+        if getattr(self, "_use_injected_tracer_provider", False):
+            return
+
+        shared_processor = getattr(self, "_shared_span_processor", None)
+        if shared_processor is not None:
+            try:
+                shared_processor.force_flush()
+            except Exception as e:
+                verbose_logger.debug(
+                    "ArizePhoenixLogger: shared span processor force_flush failed: %s",
+                    e,
+                )
+
+        with getattr(self, "_project_providers_lock", threading.Lock()):
+            providers = list(getattr(self, "_project_providers", {}).values())
+
+        for provider in providers:
+            try:
+                provider.force_flush()
+            except Exception as e:
+                verbose_logger.debug(
+                    "ArizePhoenixLogger: TracerProvider force_flush failed: %s", e
+                )
+
+    def _get_litellm_resource_for_project(self, project_name: str):
+        """
+        Build an OTEL Resource with project routing attrs that win over env detector.
+
+        Phoenix uses ``openinference.project.name``; Arize AX uses ``model_id`` and
+        ``service.name``. Project attrs are merged last so OTEL_RESOURCE_ATTRIBUTES
+        from init does not pin every provider to one project.
+        """
+        from opentelemetry.sdk.resources import OTELResourceDetector, Resource
+
+        project_attributes: dict[str, str] = {
+            "openinference.project.name": project_name,
+            "model_id": project_name,
+            "service.name": project_name,
+        }
+        deployment_environment = getattr(self.config, "deployment_environment", None)
+        if deployment_environment is not None:
+            project_attributes["deployment.environment"] = deployment_environment
+
+        env_resource = OTELResourceDetector().detect()
+        project_resource = Resource.create(project_attributes)  # type: ignore[arg-type]
+        return env_resource.merge(project_resource)
+
+    def _build_tracer_provider_for_project(self, project_name: str) -> TracerProvider:
+        """Create a TracerProvider for *project_name* (caller holds no cache lock)."""
+        from opentelemetry.sdk.trace import TracerProvider
+
+        provider = TracerProvider(
+            resource=self._get_litellm_resource_for_project(project_name)
+        )
+        provider.add_span_processor(self._shared_span_processor)
+        return provider
+
+    def _get_tracer_for(self, project_name: str) -> Tracer:
+        """Return a tracer for *project_name*, creating/caching a provider on miss."""
+        if getattr(self, "_use_injected_tracer_provider", False):
+            return self.tracer
+
+        with self._project_providers_lock:
+            if project_name in self._project_providers:
+                self._project_providers.move_to_end(project_name)
+                return self._project_providers[project_name].get_tracer(
+                    LITELLM_TRACER_NAME
+                )
+
+        # OTELResourceDetector().detect() is synchronous; build outside the lock so
+        # concurrent requests for other projects are not blocked on cache misses.
+        new_provider = self._build_tracer_provider_for_project(project_name)
+
+        with self._project_providers_lock:
+            if project_name in self._project_providers:
+                self._project_providers.move_to_end(project_name)
+                return self._project_providers[project_name].get_tracer(
+                    LITELLM_TRACER_NAME
+                )
+
+            if len(self._project_providers) >= _MAX_PROJECT_PROVIDERS:
+                self._project_providers.popitem(last=False)
+
+            self._project_providers[project_name] = new_provider
+            return new_provider.get_tracer(LITELLM_TRACER_NAME)
+
+    def _resolve_tracer_for_kwargs(self, kwargs: dict) -> Tuple[str, Tracer]:
+        """Resolve project name once and return the matching tracer."""
+        project_name = self._resolve_project_name(kwargs)
+        return project_name, self._get_tracer_for(project_name)
+
+    def get_tracer_to_use_for_request(self, kwargs: dict) -> Tracer:
+        """Route guardrail/raw-request spans to the same per-project tracer as the request."""
+        if getattr(self, "_use_injected_tracer_provider", False):
+            return self.tracer
+        return self._resolve_tracer_for_kwargs(kwargs)[1]
+
     def _init_otel_logger_on_litellm_proxy(self):
         """
         Override: Arize Phoenix should NOT overwrite the proxy's
@@ -93,56 +209,109 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
 
     @staticmethod
     def set_arize_phoenix_attributes(span: Span, kwargs, response_obj):
-        from litellm.integrations.opentelemetry_utils.base_otel_llm_obs_attributes import (
-            safe_set_attribute,
-        )
-
         _utils.set_attributes(span, kwargs, response_obj, ArizeOTELAttributes)
-
-        # Dynamic project name: check metadata first, then fall back to env var config
-        dynamic_project_name = ArizePhoenixLogger._get_dynamic_project_name(kwargs)
-        if dynamic_project_name:
-            safe_set_attribute(span, "openinference.project.name", dynamic_project_name)
-        else:
-            # Fall back to static config from env var
-            config = ArizePhoenixLogger.get_arize_phoenix_config()
-            if config.project_name:
-                safe_set_attribute(
-                    span, "openinference.project.name", config.project_name
-                )
-
         return
 
     @staticmethod
-    def _get_dynamic_project_name(kwargs) -> Optional[str]:
-        """
-        Retrieve dynamic Phoenix project name from request metadata.
+    def _normalize_project_name(name: Optional[str]) -> Optional[str]:
+        if name is None:
+            return None
+        normalized = str(name).strip()
+        return normalized if normalized else None
 
-        Users can set `metadata.phoenix_project_name` in their request to route
-        traces to different Phoenix projects dynamically.
-        """
-        standard_logging_payload = kwargs.get("standard_logging_object")
-        if isinstance(standard_logging_payload, dict):
-            metadata = standard_logging_payload.get("metadata")
+    @staticmethod
+    def _iter_metadata_dicts_from_kwargs(kwargs: dict):
+        """Yield request metadata dicts; standard_logging_object before litellm_params."""
+        for key in ("standard_logging_object", "litellm_params"):
+            found_key = kwargs.get(key)
+            if not isinstance(found_key, dict):
+                continue
+            metadata = found_key.get("metadata")
             if isinstance(metadata, dict):
-                project_name = metadata.get("phoenix_project_name")
-                if project_name:
-                    return str(project_name)
+                yield metadata
 
-        # Also check litellm_params.metadata for SDK usage
+    @staticmethod
+    def _is_proxy_request(kwargs: dict) -> bool:
+        """True when the call is routed through the LiteLLM proxy.
+
+        Proxy mode is determined solely by the server-set ``proxy_server_request``
+        field in ``litellm_params``.  Checking request metadata for
+        ``user_api_key_auth_metadata`` is intentionally avoided: that field is
+        user-supplied and would let an authenticated caller fake proxy-mode
+        detection to route their telemetry into arbitrary Arize/Phoenix projects.
+        """
         litellm_params = kwargs.get("litellm_params")
-        if isinstance(litellm_params, dict):
-            metadata = litellm_params.get("metadata") or {}
-        else:
-            metadata = {}
-        if isinstance(metadata, dict):
-            project_name = metadata.get("phoenix_project_name")
-            if project_name:
-                return str(project_name)
+        return isinstance(litellm_params, dict) and bool(
+            litellm_params.get("proxy_server_request")
+        )
 
+    @staticmethod
+    def _project_from_metadata_dict(
+        metadata: dict, metadata_key: str, *, proxy_mode: bool
+    ) -> Optional[str]:
+        """
+        Read a Phoenix project field from proxy/SDK metadata.
+
+        On the proxy, only ``user_api_key_auth_metadata`` (team/key config) may
+        select the project. SDK callers may still set project fields directly on
+        ``metadata``.
+        """
+        auth_metadata = metadata.get("user_api_key_auth_metadata")
+        if isinstance(auth_metadata, dict):
+            project = ArizePhoenixLogger._normalize_project_name(
+                auth_metadata.get(metadata_key)
+            )
+            if project:
+                return project
+
+        if not proxy_mode:
+            return ArizePhoenixLogger._normalize_project_name(
+                metadata.get(metadata_key)
+            )
         return None
 
-    def _get_phoenix_context(self, kwargs):
+    @staticmethod
+    def _metadata_project_from_kwargs(kwargs: dict, metadata_key: str) -> Optional[str]:
+        proxy_mode = ArizePhoenixLogger._is_proxy_request(kwargs)
+        for metadata in ArizePhoenixLogger._iter_metadata_dicts_from_kwargs(kwargs):
+            project = ArizePhoenixLogger._project_from_metadata_dict(
+                metadata, metadata_key, proxy_mode=proxy_mode
+            )
+            if project:
+                return project
+        return None
+
+    @staticmethod
+    def _resolve_project_name(kwargs: dict) -> str:
+        """
+        Resolve the target Phoenix/Arize project for this request.
+
+        Proxy priority: ``user_api_key_auth_metadata.phoenix_project_name_override``,
+        ``user_api_key_auth_metadata.phoenix_project_name``, env, then ``default``.
+        SDK priority: request metadata fields, then env, then ``default``.
+        """
+        override = ArizePhoenixLogger._metadata_project_from_kwargs(
+            kwargs, "phoenix_project_name_override"
+        )
+        if override:
+            return override
+
+        phoenix_name = ArizePhoenixLogger._metadata_project_from_kwargs(
+            kwargs, "phoenix_project_name"
+        )
+        if phoenix_name:
+            return phoenix_name
+
+        env_name = ArizePhoenixLogger._normalize_project_name(
+            os.environ.get("PHOENIX_PROJECT_NAME")
+            or os.environ.get("ARIZE_PROJECT_NAME")
+        )
+        if env_name:
+            return env_name
+
+        return "default"
+
+    def _get_phoenix_context(self, kwargs, tracer: Optional[Tracer] = None):
         """
         Build a trace context for Phoenix's dedicated TracerProvider.
 
@@ -159,11 +328,13 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
         """
         from opentelemetry import trace
 
+        if tracer is None:
+            tracer = self._resolve_tracer_for_kwargs(kwargs)[1]
+
         litellm_params = kwargs.get("litellm_params", {}) or {}
         proxy_server_request = litellm_params.get("proxy_server_request", {}) or {}
         headers = proxy_server_request.get("headers", {}) or {}
 
-        # Propagate distributed trace context if the caller sent a traceparent
         traceparent_ctx = (
             self.get_traceparent_from_header(headers=headers)
             if headers.get("traceparent")
@@ -173,10 +344,8 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
         is_proxy_mode = bool(proxy_server_request)
 
         if is_proxy_mode:
-            # Create a parent span on Phoenix's own tracer so both parent
-            # and child are exported to Phoenix.
             start_time_val = kwargs.get("start_time", kwargs.get("api_call_start_time"))
-            parent_span = self.tracer.start_span(
+            parent_span = tracer.start_span(
                 name="litellm_proxy_request",
                 start_time=(
                     self._to_ns(start_time_val) if start_time_val is not None else None
@@ -187,100 +356,77 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
             ctx = trace.set_span_in_context(parent_span)
             return ctx, parent_span
 
-        # SDK mode — no parent span needed
         return traceparent_ctx, None
 
     def _handle_success(self, kwargs, response_obj, start_time, end_time):
-        """
-        Override to always create spans on ArizePhoenixLogger's dedicated TracerProvider.
-
-        The base class's ``_get_span_context`` would find the parent span created by
-        the ``otel`` callback on the *global* TracerProvider.  That span is invisible
-        in Phoenix (different exporter pipeline), so we ignore it and build our own
-        hierarchy via ``_get_phoenix_context``.
-        """
-        from opentelemetry.trace import Status, StatusCode
-
-        verbose_logger.debug(
-            "ArizePhoenixLogger: Logging kwargs: %s, OTEL config settings=%s",
-            kwargs,
-            self.config,
+        self._handle_phoenix_trace(
+            kwargs, response_obj, start_time, end_time, success=True
         )
 
-        ctx, parent_span = self._get_phoenix_context(kwargs)
-
-        # Create litellm_request span (child of our parent when in proxy mode)
-        span = self.tracer.start_span(
-            name=self._get_span_name(kwargs),
-            start_time=self._to_ns(start_time),
-            context=ctx,
-        )
-        span.set_status(Status(StatusCode.OK))
-        self.set_attributes(span, kwargs, response_obj)
-
-        # Raw-request sub-span (if enabled) — must be created before
-        # ending the parent span so the hierarchy is valid.
-        self._maybe_log_raw_request(kwargs, response_obj, start_time, end_time, span)
-        span.end(end_time=self._to_ns(end_time))
-
-        # Guardrail span
-        self._create_guardrail_span(kwargs=kwargs, context=ctx)
-
-        # Annotate and close our proxy parent span
-        if parent_span is not None:
-            parent_span.set_status(Status(StatusCode.OK))
-            self.set_attributes(parent_span, kwargs, response_obj)
-            parent_span.end(end_time=self._to_ns(end_time))
-
-        # Metrics & cost recording
-        self._record_metrics(kwargs, response_obj, start_time, end_time)
-
-        # Semantic logs
-        if self.config.enable_events:
-            self._emit_semantic_logs(kwargs, response_obj, span)
-
     def _handle_failure(self, kwargs, response_obj, start_time, end_time):
-        """
-        Override to always create failure spans on ArizePhoenixLogger's dedicated
-        TracerProvider.  Mirrors ``_handle_success`` but sets ERROR status.
-        """
+        self._handle_phoenix_trace(
+            kwargs, response_obj, start_time, end_time, success=False
+        )
+
+    def _handle_phoenix_trace(
+        self,
+        kwargs,
+        response_obj,
+        start_time,
+        end_time,
+        *,
+        success: bool,
+    ):
         from opentelemetry.trace import Status, StatusCode
 
         verbose_logger.debug(
-            "ArizePhoenixLogger: Failure - Logging kwargs: %s, OTEL config settings=%s",
+            "ArizePhoenixLogger: %s - kwargs: %s, OTEL config settings=%s",
+            "success" if success else "failure",
             kwargs,
             self.config,
         )
 
-        ctx, parent_span = self._get_phoenix_context(kwargs)
+        _project_name, tracer = self._resolve_tracer_for_kwargs(kwargs)
+        ctx, parent_span = self._get_phoenix_context(kwargs, tracer=tracer)
 
-        # Create litellm_request span (child of our parent when in proxy mode)
-        span = self.tracer.start_span(
+        status = Status(StatusCode.OK if success else StatusCode.ERROR)
+
+        span = tracer.start_span(
             name=self._get_span_name(kwargs),
             start_time=self._to_ns(start_time),
             context=ctx,
         )
-        span.set_status(Status(StatusCode.ERROR))
+        span.set_status(status)
         self.set_attributes(span, kwargs, response_obj)
-        self._record_exception_on_span(span=span, kwargs=kwargs)
+        if not success:
+            self._record_exception_on_span(span=span, kwargs=kwargs)
+
+        if success:
+            self._maybe_log_raw_request(
+                kwargs, response_obj, start_time, end_time, span
+            )
         span.end(end_time=self._to_ns(end_time))
 
-        # Guardrail span
         self._create_guardrail_span(kwargs=kwargs, context=ctx)
 
-        # Annotate and close our proxy parent span
         if parent_span is not None:
-            parent_span.set_status(Status(StatusCode.ERROR))
+            parent_span.set_status(status)
             self.set_attributes(parent_span, kwargs, response_obj)
-            self._record_exception_on_span(span=parent_span, kwargs=kwargs)
+            if not success:
+                self._record_exception_on_span(span=parent_span, kwargs=kwargs)
             parent_span.end(end_time=self._to_ns(end_time))
 
+        if success:
+            self._record_metrics(kwargs, response_obj, start_time, end_time)
+
+            if self.config.enable_events:
+                self._emit_semantic_logs(kwargs, response_obj, span)
+
     @staticmethod
     def get_arize_phoenix_config() -> ArizePhoenixConfig:
         """
         Retrieves the Arize Phoenix configuration based on environment variables.
         Returns:
-            ArizePhoenixConfig: A Pydantic model containing Arize Phoenix configuration.
         """
         api_key = os.environ.get("PHOENIX_API_KEY", None)
 
@@ -295,18 +441,15 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
         protocol: Protocol = "otlp_http"
 
         if collector_endpoint:
-            # Parse the endpoint to determine protocol
             if collector_endpoint.startswith("grpc://") or (
                 ":4317" in collector_endpoint and "/v1/traces" not in collector_endpoint
             ):
                 endpoint = collector_endpoint
                 protocol = "otlp_grpc"
             else:
-                # Phoenix Cloud endpoints (app.phoenix.arize.com) include the space in the URL
                 if "app.phoenix.arize.com" in collector_endpoint:
                     endpoint = collector_endpoint
                     protocol = "otlp_http"
-                # For other HTTP endpoints, ensure they have the correct path
                 elif "/v1/traces" not in collector_endpoint:
                     if collector_endpoint.endswith("/v1"):
                         endpoint = collector_endpoint + "/traces"
@@ -318,7 +461,6 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
                     endpoint = collector_endpoint
                 protocol = "otlp_http"
         else:
-            # If no endpoint specified, self hosted phoenix
             endpoint = "http://localhost:6006/v1/traces"
             protocol = "otlp_http"
             verbose_logger.debug(
@@ -329,12 +471,11 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
         if api_key is not None:
             otlp_auth_headers = f"Authorization=Bearer {api_key}"
         elif "app.phoenix.arize.com" in endpoint:
-            # Phoenix Cloud requires an API key
             raise ValueError(
                 "PHOENIX_API_KEY must be set when using Phoenix Cloud (app.phoenix.arize.com)."
             )
 
-        project_name = os.environ.get("PHOENIX_PROJECT_NAME", "default")
+        project_name = os.environ.get("PHOENIX_PROJECT_NAME") or "default"
 
         return ArizePhoenixConfig(
             otlp_auth_headers=otlp_auth_headers,
@@ -343,8 +484,6 @@ class ArizePhoenixLogger(OpenTelemetry):  # type: ignore
             project_name=project_name,
         )
 
-    ## cannot suppress additional proxy server spans, removed previous methods.
-
     async def async_health_check(self):
         config = self.get_arize_phoenix_config()
 

litellm/integrations/datadog/datadog_cost_management.py

@@ -2,10 +2,17 @@ import asyncio
 import os
 import time
 from datetime import datetime
-from typing import Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple, cast
 
 from litellm._logging import verbose_logger
 from litellm.integrations.custom_batch_logger import CustomBatchLogger
+from litellm.integrations.datadog.datadog_handler import (
+    get_datadog_env,
+    get_datadog_hostname,
+    get_datadog_pod_name,
+    get_datadog_service,
+)
+from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     httpxSpecialProvider,
@@ -15,9 +22,30 @@ from litellm.types.integrations.datadog_cost_management import (
 )
 from litellm.types.utils import StandardLoggingPayload
 
+# Reserved tag keys whose values come from trusted sources (infra env, LiteLLM
+# core payload fields, or proxy-controlled auth metadata). User-supplied
+# request_tags / metadata cannot overwrite these, even when the key is
+# allowlisted via cost_tag_keys, because that would let an authenticated caller
+# spoof cost attribution (e.g. request_tags=["team:victim-team"]).
+_RESERVED_TAG_KEYS: frozenset = frozenset(
+    {
+        "env",
+        "service",
+        "host",
+        "pod_name",
+        "provider",
+        "model",
+        "model_id",
+        "team",
+        "user",
+        "model_group",
+    }
+)
+
 
 class DatadogCostManagementLogger(CustomBatchLogger):
-    def __init__(self, **kwargs):
+    def __init__(self, cost_tag_keys: Optional[List[str]] = None, **kwargs):
+        self.cost_tag_keys: List[str] = list(cost_tag_keys) if cost_tag_keys else []
         self.dd_api_key = os.getenv("DD_API_KEY")
         self.dd_app_key = os.getenv("DD_APP_KEY")
         self.dd_site = os.getenv("DD_SITE", "datadoghq.com")
@@ -68,20 +96,21 @@ class DatadogCostManagementLogger(CustomBatchLogger):
         if not self.log_queue:
             return
 
+        batch_to_send = self.log_queue[:]
+        self.log_queue = []
+
         try:
-            # Aggregate costs from the batch
-            aggregated_entries = self._aggregate_costs(self.log_queue)
-
+            aggregated_entries = self._aggregate_costs(batch_to_send)
             if not aggregated_entries:
+                verbose_logger.debug(
+                    "Datadog Cost Management: batch produced no aggregable entries; "
+                    "dropping %d log(s) from queue.",
+                    len(batch_to_send),
+                )
                 return
-
-            # Send to Datadog
             await self._upload_to_datadog(aggregated_entries)
-
-            # Clear queue only on success (or if we decide to drop on failure)
-            # CustomBatchLogger clears queue in flush_queue, so we just process here
-
         except Exception as e:
+            self.log_queue = batch_to_send + self.log_queue
             verbose_logger.exception(
                 f"Datadog Cost Management: Error in async_send_batch: {str(e)}"
             )
@@ -151,45 +180,81 @@ class DatadogCostManagementLogger(CustomBatchLogger):
         return list(aggregator.values())
 
     def _extract_tags(self, log: StandardLoggingPayload) -> Dict[str, str]:
-        from litellm.integrations.datadog.datadog_handler import (
-            get_datadog_env,
-            get_datadog_hostname,
-            get_datadog_pod_name,
-            get_datadog_service,
-        )
-
-        tags = {
+        tags: Dict[str, str] = {
             "env": get_datadog_env(),
             "service": get_datadog_service(),
             "host": get_datadog_hostname(),
             "pod_name": get_datadog_pod_name(),
         }
 
-        # Add metadata as tags
-        metadata = log.get("metadata", {})
-        if metadata:
-            # Add user info
-            # Add user info
-            if metadata.get("user_api_key_alias"):
-                tags["user"] = str(metadata["user_api_key_alias"])
+        # Always-on canonical FOCUS dimensions from top-level payload fields.
+        # Non-sensitive and required for Datadog Custom Costs per-model attribution.
+        self._add_tag(tags, "provider", log.get("custom_llm_provider"))
+        self._add_tag(tags, "model", log.get("model"))
+        self._add_tag(tags, "model_id", log.get("model_id"))
 
-            # Add Team Tag
-            team_tag = (
-                metadata.get("user_api_key_team_alias")
-                or metadata.get("team_alias")  # type: ignore
-                or metadata.get("user_api_key_team_id")
-                or metadata.get("team_id")  # type: ignore
-            )
+        # cast because StandardLoggingMetadata is a TypedDict; we iterate it
+        # as a generic mapping below.
+        metadata: Dict[str, Any] = cast(Dict[str, Any], log.get("metadata") or {})
 
-            if team_tag:
-                tags["team"] = str(team_tag)
-            # model_group is not in StandardLoggingMetadata TypedDict, so we need to access it via dict.get()
-            model_group = metadata.get("model_group")  # type: ignore[misc]
-            if model_group:
-                tags["model_group"] = str(model_group)
+        # Backwards-compat: team/user/model_group preserved regardless of allowlist.
+        if metadata.get("user_api_key_alias"):
+            tags["user"] = str(metadata["user_api_key_alias"])
+        team_tag = (
+            metadata.get("user_api_key_team_alias")
+            or metadata.get("team_alias")
+            or metadata.get("user_api_key_team_id")
+            or metadata.get("team_id")
+        )
+        if team_tag:
+            tags["team"] = str(team_tag)
+        if metadata.get("model_group"):
+            tags["model_group"] = str(metadata["model_group"])
+
+        # Allowlist-gated: request_tags (split on `:`) and arbitrary metadata.*.
+        # Reserved keys are hard-blocked here regardless of allowlist membership —
+        # see _RESERVED_TAG_KEYS for the rationale.
+        if self.cost_tag_keys:
+            allow = set(self.cost_tag_keys)
+            for rt in log.get("request_tags") or []:
+                if not isinstance(rt, str) or ":" not in rt:
+                    continue
+                k, _, v = rt.partition(":")
+                if k in allow and v:
+                    self._set_custom_tag(tags, k, v)
+            for k, v in metadata.items():
+                if k in allow and v is not None and not isinstance(v, (dict, list)):
+                    self._set_custom_tag(tags, k, str(v))
+            for nested_key in ("spend_logs_metadata", "requester_metadata"):
+                nested = metadata.get(nested_key)
+                if isinstance(nested, dict):
+                    for k, v in nested.items():
+                        if (
+                            k in allow
+                            and v is not None
+                            and not isinstance(v, (dict, list))
+                        ):
+                            self._set_custom_tag(tags, k, str(v))
 
         return tags
 
+    @staticmethod
+    def _set_custom_tag(tags: Dict[str, str], key: str, value: str) -> None:
+        if key in _RESERVED_TAG_KEYS:
+            verbose_logger.debug(
+                "Datadog Cost Management: dropping user-supplied tag %r=%r — "
+                "key is reserved for trusted cost attribution.",
+                key,
+                value,
+            )
+            return
+        tags[key] = value
+
+    @staticmethod
+    def _add_tag(tags: Dict[str, str], key: str, value: Any) -> None:
+        if value:
+            tags[key] = str(value)
+
     async def _upload_to_datadog(self, payload: List[Dict]):
         if not self.dd_api_key or not self.dd_app_key:
             return
@@ -201,8 +266,6 @@ class DatadogCostManagementLogger(CustomBatchLogger):
         }
 
         # The API endpoint expects a list of objects directly in the body (file content behavior)
-        from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
-
         data_json = safe_dumps(payload)
 
         response = await self.async_client.put(

litellm/integrations/datadog/datadog_metrics.py

@@ -144,7 +144,26 @@ class DatadogMetricsLogger(CustomBatchLogger):
             }
             self.log_queue.append(series_llm_latency)
 
-        # 3. Request Count / Status Code
+        # 3. LiteLLM Overhead Latency Metric (total - llm_api time)
+        hidden_params = log.get("hidden_params", {}) or {}
+        litellm_overhead_time_ms = hidden_params.get("litellm_overhead_time_ms")
+        if litellm_overhead_time_ms is not None:
+            overhead_tags = self._extract_tags(log)  # no status_code on latency metric
+            series_overhead: DatadogMetricSeries = {
+                "metric": "litellm.overhead.latency",
+                "type": 3,  # gauge
+                "points": [
+                    {
+                        "timestamp": timestamp,
+                        "value": litellm_overhead_time_ms
+                        / 1000,  # convert ms → seconds
+                    }
+                ],
+                "tags": overhead_tags,
+            }
+            self.log_queue.append(series_overhead)
+
+        # 4. Request Count / Status Code
         series_count: DatadogMetricSeries = {
             "metric": "litellm.llm_api.request_count",
             "type": 1,  # count

litellm/integrations/galileo.py

@@ -1,18 +1,29 @@
+import json
 import os
-from typing import Any, Dict, List, Optional
+import re
+from typing import Any, Dict, List, Optional, Tuple, cast
 
 from pydantic import BaseModel, Field
 
 import litellm
 from litellm._logging import verbose_logger
 from litellm.integrations.custom_logger import CustomLogger
+from litellm.litellm_core_utils.prompt_templates.common_utils import (
+    convert_content_list_to_str,
+    get_content_from_model_response,
+)
 from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     httpxSpecialProvider,
 )
+from litellm.types.llms.openai import AllMessageValues
+
+GALILEO_CLOUD_API_BASE_URL = "https://api.galileo.ai"
+# Cap the in-memory buffer so persistent flush failures (e.g. Galileo
+# unavailable, invalid credentials) cannot leak memory unboundedly.
+GALILEO_MAX_IN_MEMORY_RECORDS = 1000
 
 
-# from here: https://docs.rungalileo.io/galileo/gen-ai-studio-products/galileo-observe/how-to/logging-data-via-restful-apis#structuring-your-records
 class LLMResponse(BaseModel):
     latency_ms: int
     status_code: int
@@ -37,65 +48,190 @@ class GalileoObserve(CustomLogger):
     def __init__(self) -> None:
         self.in_memory_records: List[dict] = []
         self.batch_size = 1
-        self.base_url = os.getenv("GALILEO_BASE_URL", None)
-        self.project_id = os.getenv("GALILEO_PROJECT_ID", None)
+        self.api_key = os.getenv("GALILEO_API_KEY")
+        self.project_id = os.getenv("GALILEO_PROJECT_ID")
+        self.log_stream_id = os.getenv("GALILEO_LOG_STREAM_ID")
+        self.username = os.getenv("GALILEO_USERNAME")
+        self.password = os.getenv("GALILEO_PASSWORD")
+        self.base_url = self._normalize_base_url(os.getenv("GALILEO_BASE_URL"))
+        if self.api_key and not self.base_url:
+            self.base_url = GALILEO_CLOUD_API_BASE_URL
+        self.use_v2_api = bool(self.api_key)
         self.headers: Optional[Dict[str, str]] = None
         self.async_httpx_handler = get_async_httpx_client(
             llm_provider=httpxSpecialProvider.LoggingCallback
         )
-        pass
 
-    def set_galileo_headers(self):
-        # following https://docs.rungalileo.io/galileo/gen-ai-studio-products/galileo-observe/how-to/logging-data-via-restful-apis#logging-your-records
+    @staticmethod
+    def _normalize_base_url(base_url: Optional[str]) -> Optional[str]:
+        if base_url:
+            return base_url.rstrip("/")
+        return None
 
-        headers = {
-            "accept": "application/json",
-            "Content-Type": "application/x-www-form-urlencoded",
-        }
-        galileo_login_response = litellm.module_level_client.post(
+    def _is_configured(self) -> bool:
+        if not self.project_id or not self.base_url:
+            return False
+        if self.use_v2_api:
+            return bool(self.api_key)
+        return bool(self.username and self.password)
+
+    async def async_set_galileo_headers(self) -> None:
+        galileo_login_response = await self.async_httpx_handler.post(
             url=f"{self.base_url}/login",
-            headers=headers,
+            headers={
+                "accept": "application/json",
+                "Content-Type": "application/x-www-form-urlencoded",
+            },
             data={
-                "username": os.getenv("GALILEO_USERNAME"),
-                "password": os.getenv("GALILEO_PASSWORD"),
+                "username": self.username,
+                "password": self.password,
             },
         )
-
+        galileo_login_response.raise_for_status()
         access_token = galileo_login_response.json()["access_token"]
-
         self.headers = {
             "accept": "application/json",
             "Content-Type": "application/json",
             "Authorization": f"Bearer {access_token}",
         }
 
-    def get_output_str_from_response(self, response_obj, kwargs):
-        output = None
-        if response_obj is not None and (
-            kwargs.get("call_type", None) == "embedding"
-            or isinstance(response_obj, litellm.EmbeddingResponse)
-        ):
-            output = None
-        elif response_obj is not None and isinstance(
-            response_obj, litellm.ModelResponse
-        ):
-            output = response_obj["choices"][0]["message"].json()
-        elif response_obj is not None and isinstance(
-            response_obj, litellm.TextCompletionResponse
-        ):
-            output = response_obj.choices[0].text
-        elif response_obj is not None and isinstance(
-            response_obj, litellm.ImageResponse
-        ):
-            output = response_obj["data"]
+    async def _ensure_headers(self) -> bool:
+        if self.headers is not None:
+            return True
 
-        return output
+        if self.use_v2_api:
+            if not self.api_key:
+                return False
+            self.headers = {
+                "accept": "application/json",
+                "Content-Type": "application/json",
+                "Galileo-API-Key": self.api_key,
+            }
+            return True
+
+        if not (self.username and self.password and self.base_url):
+            return False
+
+        try:
+            await self.async_set_galileo_headers()
+            return True
+        except Exception as e:
+            verbose_logger.debug("Galileo Logger: failed to authenticate: %s", e)
+            return False
+
+    @staticmethod
+    def _galileo_input_messages(
+        messages: Optional[List[Any]], input_text: str
+    ) -> List[Dict[str, str]]:
+        if not messages:
+            return [{"role": "user", "content": input_text}]
+
+        galileo_messages: List[Dict[str, str]] = []
+        for message in messages:
+            if not isinstance(message, dict):
+                continue
+            role = message.get("role")
+            if not role:
+                continue
+            galileo_messages.append(
+                {
+                    "role": str(role),
+                    "content": convert_content_list_to_str(
+                        message=cast(AllMessageValues, message)
+                    ),
+                }
+            )
+
+        if galileo_messages:
+            return galileo_messages
+        return [{"role": "user", "content": input_text}]
+
+    @staticmethod
+    def _record_to_v2_span(record: Dict[str, Any]) -> Dict[str, Any]:
+        created_at = record.get("created_at", "")
+        if created_at and not re.search(r"(Z|[+-]\d{2}:?\d{2})$", created_at):
+            created_at = f"{created_at}Z"
+
+        span: Dict[str, Any] = {
+            "type": "llm",
+            "name": record.get("node_type", "litellm"),
+            "created_at": created_at,
+            "input": GalileoObserve._galileo_input_messages(
+                record.get("messages"), record.get("input_text", "")
+            ),
+            "output": {
+                "role": "assistant",
+                "content": record.get("output_text", ""),
+            },
+            "status_code": record.get("status_code", 200),
+            "model": record.get("model"),
+            "metrics": {
+                "duration_ns": int(record.get("latency_ms", 0)) * 1_000_000,
+                "num_input_tokens": record.get("num_input_tokens"),
+                "num_output_tokens": record.get("num_output_tokens"),
+            },
+        }
+        if record.get("tags"):
+            span["tags"] = record["tags"]
+        return span
+
+    def _get_ingest_request(self) -> Optional[Tuple[str, Dict[str, Any]]]:
+        if not self.base_url or not self.project_id:
+            return None
+
+        # Snapshot the records to be sent into a new list so concurrent appends
+        # during the network round-trip (across the await points in
+        # flush_in_memory_records) aren't silently dropped when we later clear
+        # the in-memory buffer.
+        records = list(self.in_memory_records)
+
+        if self.use_v2_api:
+            payload: Dict[str, Any] = {
+                "spans": [self._record_to_v2_span(record) for record in records],
+                "reliable": False,
+            }
+            if self.log_stream_id:
+                payload["log_stream_id"] = self.log_stream_id
+            return (
+                f"{self.base_url}/v2/projects/{self.project_id}/spans",
+                payload,
+            )
+
+        return (
+            f"{self.base_url}/projects/{self.project_id}/observe/ingest",
+            {"records": records},
+        )
+
+    def get_output_str_from_response(
+        self, response_obj: Any, kwargs: Dict[str, Any]
+    ) -> Optional[str]:
+        if response_obj is None:
+            return None
+        if kwargs.get("call_type", None) == "embedding" or isinstance(
+            response_obj, litellm.EmbeddingResponse
+        ):
+            return None
+        if isinstance(response_obj, litellm.TextCompletionResponse):
+            return response_obj.choices[0].text
+        if isinstance(response_obj, litellm.ImageResponse):
+            return json.dumps(response_obj["data"], default=str)
+        if isinstance(response_obj, (litellm.ModelResponse, dict)):
+            return get_content_from_model_response(response_obj)
+        return None
 
     async def async_log_success_event(
         self, kwargs: Any, response_obj: Any, start_time: Any, end_time: Any
     ):
         verbose_logger.debug("On Async Success")
 
+        if not self._is_configured():
+            verbose_logger.debug(
+                "Galileo Logger: skipping flush — set GALILEO_PROJECT_ID and "
+                "either GALILEO_API_KEY (hosted) or GALILEO_USERNAME/GALILEO_PASSWORD "
+                "(enterprise Observe)."
+            )
+            return
+
         _latency_ms = int((end_time - start_time).total_seconds() * 1000)
         _call_type = kwargs.get("call_type", "litellm")
         input_text = litellm.utils.get_formatted_prompt(
@@ -125,26 +261,69 @@ class GalileoObserve(CustomLogger):
                 ),  # timestamp str constructed in "%Y-%m-%dT%H:%M:%S" format
             )
 
-            # dump to dict
             request_dict = request_record.model_dump()
+            messages = kwargs.get("messages")
+            if messages:
+                request_dict["messages"] = messages
             self.in_memory_records.append(request_dict)
 
+            # Bound the buffer so persistent flush failures cannot grow it
+            # without limit. Drop the oldest records once we exceed the cap.
+            if len(self.in_memory_records) > GALILEO_MAX_IN_MEMORY_RECORDS:
+                dropped = len(self.in_memory_records) - GALILEO_MAX_IN_MEMORY_RECORDS
+                self.in_memory_records = self.in_memory_records[
+                    -GALILEO_MAX_IN_MEMORY_RECORDS:
+                ]
+                verbose_logger.warning(
+                    "Galileo Logger: in-memory buffer exceeded %s records; "
+                    "dropped %s oldest record(s). Check Galileo connectivity/credentials.",
+                    GALILEO_MAX_IN_MEMORY_RECORDS,
+                    dropped,
+                )
+
             if len(self.in_memory_records) >= self.batch_size:
                 await self.flush_in_memory_records()
 
     async def flush_in_memory_records(self):
-        verbose_logger.debug("flushing in memory records")
-        response = await self.async_httpx_handler.post(
-            url=f"{self.base_url}/projects/{self.project_id}/observe/ingest",
-            headers=self.headers,
-            json={"records": self.in_memory_records},
-        )
+        if not self.in_memory_records:
+            return
 
-        if response.status_code == 200:
+        # Capture the number of records that will be sent BEFORE any await so
+        # that concurrent appends made by other asyncio tasks during the
+        # network round-trip aren't silently dropped on the success-clear.
+        records_in_payload = len(self.in_memory_records)
+
+        ingest_request = self._get_ingest_request()
+        if ingest_request is None:
             verbose_logger.debug(
-                "Galileo Logger:successfully flushed in memory records"
+                "Galileo Logger: missing GALILEO_BASE_URL or GALILEO_PROJECT_ID"
             )
-            self.in_memory_records = []
+            return
+
+        if not await self._ensure_headers():
+            verbose_logger.debug("Galileo Logger: could not set request headers")
+            return
+
+        url, payload = ingest_request
+        verbose_logger.debug("flushing in memory records to %s", url)
+
+        try:
+            response = await self.async_httpx_handler.post(
+                url=url,
+                headers=self.headers,
+                json=payload,
+            )
+        except Exception as e:
+            verbose_logger.debug(
+                "Galileo Logger: failed to flush in memory records: %s", e
+            )
+            return
+
+        if response.is_success:
+            verbose_logger.debug(
+                "Galileo Logger: successfully flushed in memory records"
+            )
+            del self.in_memory_records[:records_in_payload]
         else:
             verbose_logger.debug("Galileo Logger: failed to flush in memory records")
             verbose_logger.debug(
@@ -152,6 +331,13 @@ class GalileoObserve(CustomLogger):
                 response.text,
                 response.status_code,
             )
+            # Legacy enterprise auth caches a bearer token obtained from
+            # /login. If the request was rejected for auth reasons, drop the
+            # cached headers so the next flush re-authenticates instead of
+            # silently failing forever on a stale token. The v2 API key path
+            # uses a long-lived static key, so leave its headers in place.
+            if not self.use_v2_api and response.status_code in (401, 403):
+                self.headers = None
 
     async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time):
         verbose_logger.debug("On Async Failure")

litellm/integrations/opentelemetry.py

@@ -64,6 +64,9 @@ HTTP_RESPONSE_STATUS_CODE_ATTRIBUTE = "http.response.status_code"
 HTTP_ROUTE_ATTRIBUTE = "http.route"
 URL_PATH_ATTRIBUTE = "url.path"
 PREPROCESSING_DURATION_MS_ATTRIBUTE = "litellm.preprocessing.duration_ms"
+TEAM_METADATA_ATTRIBUTE = "litellm.team.metadata"
+MODEL_GROUP_ATTRIBUTE = "litellm.model_group"
+PROVIDER_MODEL_ATTRIBUTE = "litellm.provider.model"
 # Remove the hardcoded LITELLM_RESOURCE dictionary - we'll create it properly later
 RAW_REQUEST_SPAN_NAME = "raw_gen_ai_request"
 LITELLM_REQUEST_SPAN_NAME = "litellm_request"
@@ -702,6 +705,14 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
                 },
             )
 
+            # _record_exception_on_span only stamps when error_code is set;
+            # bare TypeError etc. has none, and the span is about to be ended.
+            error_code = (
+                error_information.get("error_code") if error_information else None
+            )
+            if not error_code:
+                self.set_response_status_code_attribute(parent_otel_span, 500)
+
             # Pre-request latency (request_data carries the propagated
             # metadata on the failure path; omitted if it failed before handoff).
             self.set_preprocessing_duration_attribute(parent_otel_span, request_data)
@@ -798,11 +809,6 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             # Pre-request latency on the SERVER span (success path).
             self.set_preprocessing_duration_attribute(parent_span, kwargs)
 
-            # http.response.status_code on the SERVER span (success path).
-            # A successful proxy response is HTTP 200; the failure path sets
-            # this from the error code in _record_exception_on_span.
-            self.set_response_status_code_attribute(parent_span, 200)
-
             # 3. Guardrail span
             self._create_guardrail_span(kwargs=kwargs, context=ctx)
 
@@ -985,7 +991,15 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             and hasattr(proxy_span, "is_recording")
             and proxy_span.is_recording()
         ):
-            proxy_span.end(end_time=self._to_ns(end_time))
+            self._close_proxy_span_ok(proxy_span, end_time)
+
+    def _close_proxy_span_ok(self, span: Span, end_time) -> None:
+        """Stamp http.response.status_code=200 + status=OK, then end the span."""
+        from opentelemetry.trace import Status, StatusCode
+
+        self.set_response_status_code_attribute(span, 200)
+        span.set_status(Status(StatusCode.OK))
+        span.end(end_time=self._to_ns(end_time))
 
     def _handle_success(self, kwargs, response_obj, start_time, end_time):
         """Create the litellm_request span then close the proxy span."""
@@ -1071,8 +1085,10 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             parent_span is not None
             and hasattr(parent_span, "name")
             and parent_span.name == LITELLM_PROXY_REQUEST_SPAN_NAME
+            and hasattr(parent_span, "is_recording")
+            and parent_span.is_recording()
         ):
-            parent_span.end(end_time=self._to_ns(end_time))
+            self._close_proxy_span_ok(parent_span, end_time)
 
         # Stamp team attributes onto the SERVER (root) span before it is
         # closed, so the trace root carries them like every child span.
@@ -1200,6 +1216,68 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
         ):
             self._set_team_attributes_from_kwargs(proxy_span, kwargs)
 
+    def _set_inference_identity_attributes(
+        self,
+        span: Span,
+        standard_logging_payload: StandardLoggingPayload,
+        litellm_params: dict,
+    ) -> None:
+        """Stamp request-identity attributes onto an inference span so every
+        LLM-call span is filterable by the route it came in on, the team's
+        metadata, and both the user-facing (model_group alias) and the
+        dispatched (provider) model names. Empty/absent values are skipped.
+        """
+        metadata = standard_logging_payload.get("metadata") or {}
+
+        http_route = metadata.get("user_api_key_request_route")
+        if http_route:
+            self.safe_set_attribute(
+                span=span, key=HTTP_ROUTE_ATTRIBUTE, value=http_route
+            )
+
+        # ``user_api_key_team_metadata`` is dropped from the standard logging
+        # payload metadata, so read it from the raw request metadata in kwargs.
+        # ``metadata`` and ``litellm_metadata`` are alternate names for the same
+        # full metadata dict (the name varies by endpoint), so first-truthy wins.
+        raw_metadata = (
+            litellm_params.get("metadata")
+            or litellm_params.get("litellm_metadata")
+            or {}
+        )
+        team_metadata = self._team_metadata_json(
+            raw_metadata.get("user_api_key_team_metadata")
+        )
+        if team_metadata:
+            self.safe_set_attribute(
+                span=span, key=TEAM_METADATA_ATTRIBUTE, value=team_metadata
+            )
+
+        model_group = standard_logging_payload.get("model_group")
+        if model_group:
+            self.safe_set_attribute(
+                span=span, key=MODEL_GROUP_ATTRIBUTE, value=model_group
+            )
+
+        hidden_params = standard_logging_payload.get("hidden_params") or {}
+        provider_model = hidden_params.get(
+            "litellm_model_name"
+        ) or standard_logging_payload.get("model")
+        if provider_model:
+            self.safe_set_attribute(
+                span=span, key=PROVIDER_MODEL_ATTRIBUTE, value=provider_model
+            )
+
+    @staticmethod
+    def _team_metadata_json(value: Any) -> Optional[str]:
+        """JSON-serialize a team's metadata dict for a single span attribute.
+
+        Returns ``None`` for a missing, non-dict, or empty mapping so the
+        empty case is dropped rather than stamping a useless ``"{}"``.
+        """
+        if not isinstance(value, dict) or not value:
+            return None
+        return safe_dumps(value)
+
     def _record_metrics(self, kwargs, response_obj, start_time, end_time):
         duration_s = (end_time - start_time).total_seconds()
         params = kwargs.get("litellm_params") or {}
@@ -2010,6 +2088,12 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
                     key="hidden_params",
                     value=safe_dumps(hidden_params),
                 )
+
+            self._set_inference_identity_attributes(
+                span=span,
+                standard_logging_payload=standard_logging_payload,
+                litellm_params=litellm_params,
+            )
             # Cost breakdown tracking
             cost_breakdown: Optional[CostBreakdown] = standard_logging_payload.get(
                 "cost_breakdown"
@@ -3041,6 +3125,11 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             management_endpoint_span.set_status(Status(StatusCode.OK))
             management_endpoint_span.end(end_time=_end_time_ns)
 
+            # The management wrapper has no other hook that closes the SERVER span.
+            self.set_response_status_code_attribute(parent_otel_span, 200)
+            parent_otel_span.set_status(Status(StatusCode.OK))
+            parent_otel_span.end(end_time=_end_time_ns)
+
     async def async_management_endpoint_failure_hook(
         self,
         logging_payload: ManagementEndpointLoggingPayload,
@@ -3091,6 +3180,24 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             management_endpoint_span.set_status(Status(StatusCode.ERROR))
             management_endpoint_span.end(end_time=_end_time_ns)
 
+            # The management wrapper has no other hook that closes the SERVER span.
+            from litellm.litellm_core_utils.litellm_logging import (
+                StandardLoggingPayloadSetup,
+            )
+
+            error_information = StandardLoggingPayloadSetup.get_error_information(
+                original_exception=_exception,
+            )
+            parent_otel_span.set_status(Status(StatusCode.ERROR))
+            self._record_exception_on_span(
+                span=parent_otel_span,
+                kwargs={
+                    "exception": _exception,
+                    "standard_logging_object": {"error_information": error_information},
+                },
+            )
+            parent_otel_span.end(end_time=_end_time_ns)
+
     def create_litellm_proxy_request_started_span(
         self,
         start_time: datetime,

litellm/integrations/otel/README.md

@@ -0,0 +1,258 @@
+# OpenTelemetry instrumentation
+
+This package produces OpenTelemetry traces for LiteLLM. It is enabled by the
+`LITELLM_OTEL_V2` environment variable (`is_otel_v2_enabled()` in
+[`config.py`](./model/config.py)); when unset, nothing in this package runs.
+
+## What gets traced
+
+A traced proxy request produces one trace with two kinds of spans:
+
+```
+SERVER span  "POST /v1/chat/completions"        ← FastAPI instrumentation
+├── INTERNAL span  "auth /v1/chat/completions"   ← auth phase     ┐
+│   ├── CLIENT span  "postgres get_key_object"    ← datastore call │
+│   └── CLIENT span  "postgres get_team_membership"                │
+├── INTERNAL span  "execute_guardrail …"         ← guardrail       │ this package
+├── CLIENT span    "chat gpt-4o"                  ← LLM call        │
+└── CLIENT span    "batch_write_to_db …"          ← spend write    ┘
+```
+
+The gen-ai spans are siblings under the server span. In particular the guardrail
+span is a sibling of the LLM call, not a child of it: pre/during/post-call
+guardrail hooks are part of the request lifecycle (a pre-call guardrail runs
+before the LLM call even starts), so they belong directly under the server span,
+alongside the LLM call.
+
+Request-level spans (LLM call, guardrail) parent to the server span via an
+**explicit anchor** — `context.set_request_root_span` captures the server span
+once at request entry, and `resolve_request_span_context` reads it — rather than
+to whatever span is momentarily active. Ambient-only parenting was wrong at two
+boundaries: inside the live `auth` phase span the active span is `auth` (so the
+span would nest under auth), and a pass-through request closes its span from a
+detached `asyncio.create_task` where the server span is no longer active (so the
+span orphaned into its own trace). The anchor — a contextvar inherited by those
+child tasks — gives a stable parent in both cases. DB/service spans keep ambient
+parenting so an auth DB lookup still nests under `auth`.
+
+**Which service calls become spans (`spans.span_role_for_service`).** LiteLLM's
+service-logging layer instruments many internal functions, but only some are
+traceable units of work:
+
+- **`DB_CALL` (CLIENT)** — outbound datastore calls (redis, postgres,
+  `batch_write_to_db`), carrying `db.system.name` / `db.operation.name` semconv.
+- **`SERVICE` (INTERNAL)** — genuine internal work worth a span (background
+  budget/reset jobs, pod-lock manager).
+- **metrics-only (no span)** — `self` (the `track_llm_api_timing` wrapper, which
+  duplicates the LLM-call span), `router` (duplicates the request), and
+  `proxy_pre_call` (a guardrail's real span is `execute_guardrail …`). These
+  still feed Prometheus/Datadog through their own hooks; they just never enter
+  the trace. `auth` is also excluded here because it gets a **live phase span**
+  instead (see below).
+
+Spans are named `"{service} {call_type}"` (e.g. `"redis set"`) so repeated calls
+to one service stay distinguishable. Like every other span they parent to the
+**ambient** context, falling back to the threaded `litellm_parent_otel_span` only
+when ambient has no live span; a background job with neither starts its own root
+trace. Caller-supplied `event_metadata` is **sanitized** before it reaches a span
+(primitives only, no live objects, no secrets/headers, bounded) — see
+`payloads.sanitize_event_metadata`.
+
+**Live phase spans.** `auth` is wrapped in a real, active span
+(`logger.phase_span`) for the duration of authentication, so the DB lookups it
+triggers nest **under** it instead of flattening onto the server span. Identity
+Baggage (team/key/user) is seeded once the key resolves, so every post-auth span
+inherits it; auth-internal DB lookups that run before the key is known stay
+unlabeled, which is correct.
+
+**Status.** On success a span's status is left `UNSET` (the semconv default,
+matching the FastAPI server span); only a genuine error sets `ERROR`.
+
+- **Server spans** (one per HTTP route) are created by the
+  `opentelemetry-instrumentation-fastapi` package. It stamps `http.*` attributes
+  and extracts inbound `traceparent` headers. This package does **not** create
+  or modify server spans — request routes never touch spans.
+- **Gen-AI spans** (LLM calls, guardrails, internal service calls) are created
+  by this package from LiteLLM's logging callbacks. Request-level spans parent to
+  the server span via the captured anchor; DB/service spans parent to the active
+  span (ambient) so they nest under the request phase that triggered them.
+
+Both kinds share a single `TracerProvider`, so they belong to the same trace
+and export through the same configured exporters. FastAPI middleware can only be
+added before the app starts serving, so the app is instrumented at
+import time **without** a provider — it binds to the OTel global
+`ProxyTracerProvider`. Once config (and the callbacks) is loaded, the proxy
+publishes the chosen logger's `TracerProvider` as the global via
+`trace.set_tracer_provider(...)`, and the server spans delegate to it. When a
+preset callback (`arize`, `langfuse_otel`, …) is configured, its provider
+becomes the global, so server spans export to that backend too.
+
+## How a request flows
+
+1. **App creation** (`proxy_server` import): when the gate is on,
+   `mount.instrument_fastapi_app(app)` calls `FastAPIInstrumentor.instrument_app`
+   with no provider (the middleware stack is frozen once the app serves, so this
+   can't wait for startup). It binds to the OTel global `ProxyTracerProvider`. Noisy
+   non-LLM routes are excluded by default (`mount._DEFAULT_EXCLUDED_ROUTES`): health
+   checks (`/health*`), the Prometheus scrape (`/metrics`), and static UI/docs assets
+   (`/litellm-asset-prefix`, `/_next`, `/ui`, `/swagger`, `/docs`, `/redoc`,
+   `/openapi.json`, favicons, `/.well-known`) — so load-balancer polling, metric
+   scrapes, and asset fetches don't flood traces. Entries are substring-matched, so
+   `/metrics` also drops the `/model/metrics` admin-analytics spans. Set
+   `OTEL_PYTHON_FASTAPI_EXCLUDED_URLS` to override the whole set (e.g. `""` to trace
+   everything, or your own comma-separated path list).
+2. **Startup** (`proxy_server.proxy_startup_event`): after the config (and
+   callbacks) is loaded, the already-registered preset `OpenTelemetryV2` logger
+   is reused — or a generic one reading `OTEL_*` envs is built when no preset is
+   configured — and its `TracerProvider` is published as the OTel global with
+   `trace.set_tracer_provider(...)`. The proxy tracer then delegates to it, so
+   server spans and gen-ai spans share one provider and the same trace.
+3. **Request**: the FastAPI instrumentation starts the server span and makes it
+   the active context for the request task. The proxy's first call into the V2
+   logger (`create_litellm_proxy_request_started_span`, at the auth boundary)
+   **captures it as the request anchor** (`set_request_root_span`), so every later
+   request-level span has a stable explicit parent regardless of what is active
+   when it emits.
+4. **LLM call span (born at the boundary)**: `OpenTelemetryV2.log_pre_api_call`
+   runs synchronously in the request task, just before the upstream call, and
+   **opens** the LLM-call span there, parented to the anchored server span
+   (`resolve_request_span_context`). The open span is held in a bounded cache keyed
+   by `litellm_call_id` (a primitive the callback kwargs carry at both `pre_call`
+   and close), so no live `Span` ever travels through a `litellm_params` metadata
+   dict. For the boundary hook to fire at all, the logger is registered into
+   `litellm.input_callback` — the list `Logging.pre_call` iterates. The async
+   success/failure callback later
+   **closes** it: it builds an `LLMCallSpanData` from the typed
+   `standard_logging_object` (token usage and cost are computed only by then),
+   stamps the attributes, sets status, and ends the span. The sync callback is a
+   no-op (closing is async-only). When `pre_call` runs off the request task — a
+   sync-only provider driven through a thread pool, where contextvars (and so the
+   anchor) don't follow — no parent is visible there, so creation is **deferred**
+   to the async callback, whose worker context was copied from the request task at
+   enqueue and so still carries the anchor. **Pass-through** endpoints call
+   `logging_obj.pre_call` in the request task too, then close from a detached
+   `asyncio.create_task`; the anchor (not the by-then-inactive server span) keeps
+   their LLM-call span in the request's trace. `pre_call` is litellm's generic
+   "log the attempt" hook, so it also fires for synthetic proxy-gate error logs
+   (auth/rate-limit rejections); those carry `LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL`
+   and are skipped, so a request rejected before reaching a provider never produces
+   a phantom CLIENT span.
+5. **Guardrails / services**: the post-call and service hooks emit guardrail and
+   service spans the same way — typed data → engine → span. Service spans
+   (Redis/Postgres) are dispatched by `litellm/_service_logger.py`, which
+   recognizes the V2 `OpenTelemetryV2` logger (a plain `CustomLogger`, not a
+   subclass of the legacy `OpenTelemetry`). It hands every service call to the
+   logger — including calls with no parent span — and the V2 adapter decides the
+   role (`DB_CALL` vs `SERVICE`), the parent (ambient → threaded → root), and
+   whether the call is a traceable operation or a metrics-only ping. Guardrail
+   span data is built from the typed, provider-agnostic
+   `StandardLoggingGuardrailInformation` — no single provider's field shape is
+   assumed.
+6. **Export**: each span ends and is handed to the provider's span processors,
+   which export to the configured backends (OTLP, console, in-memory, …).
+
+## Components
+
+### Sources of truth (`model/`, no OpenTelemetry import)
+
+These define the shape of a span without depending on the OTel SDK, so they can
+be imported anywhere. They live in [`model/`](./model) and form a closed set —
+nothing here imports outside it:
+
+- [`semconv.py`](./model/semconv.py) — attribute-key constants (`gen_ai.*`, `http.*`,
+  `litellm.*`), the GenAI operation/provider enums, and the functions that map
+  LiteLLM provider/call-type strings onto convention values.
+- [`spans.py`](./model/spans.py) — the span registry: every span role, its OTel span
+  kind, its place in the hierarchy, and its name builder.
+- [`payloads.py`](./model/payloads.py) — frozen dataclasses (`LLMCallSpanData`,
+  `GuardrailSpanData`, `ServiceSpanData`, …) built from heterogeneous logging
+  payloads via `from_*` classmethods.
+- [`config.py`](./model/config.py) — `OpenTelemetryV2Config`, a pydantic-settings
+  model that reads `OTEL_*` / `LITELLM_OTEL_*` env vars, plus the feature gate.
+  `capture_span_content` gates whether prompt/response bodies may be written as
+  span attributes; it defaults **off** (`no_content`). The Baggage allowlists are
+  configurable, not hard-coded: set `LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS` /
+  `LITELLM_OTEL_BAGGAGE_METADATA_KEYS` (comma-separated) as env vars, or
+  `baggage_promoted_keys` / `baggage_metadata_keys` (YAML lists) under
+  `callback_settings.otel` in `config.yaml` — the latter reach the config through
+  the logger's constructor kwargs.
+- [`baggage.py`](./model/baggage.py) — the single definition of which request-identity
+  values are promoted into Baggage (so child spans inherit them) and under which
+  attribute keys.
+- [`utils.py`](./model/utils.py) — value coercion, JSON serialization, and
+  extractor-table application, shared across the package.
+
+### Engine
+
+- [`emitter.py`](./emitter.py) — `SpanEmitter.emit(role, data)`: dedupe → start
+  the span → run the mapper chain to stamp attributes → set status → end. It
+  owns no attribute keys. The dedupe set (which coalesces the sync+async firing
+  of one request) is a bounded LRU so it can't grow without limit.
+- [`mappers/`](./mappers) — each mapper turns typed span data into a flat
+  `{attribute key: value}` dict. They compose: listing several mapper names in
+  the config layers multiple attribute vocabularies onto the same span.
+  - `genai` — the canonical OpenTelemetry GenAI vocabulary, always present.
+  - `legacy` — an additional vocabulary using the older semconv-ai / Traceloop
+    attribute key names, for backends that read those.
+  - `openinference`, `langfuse`, `weave`, `langtrace` — vendor vocabularies.
+  - `resolve_mappers(names)` turns config names into mapper instances.
+
+### Plumbing (`plumbing/`)
+
+The OTel-SDK wiring. Everything here imports only `model/` and each other; it
+lives in [`plumbing/`](./plumbing):
+
+- [`providers.py`](./plumbing/providers.py) — builds the `TracerProvider`, its exporters
+  (from `ExporterSpec`s), and the span processor that copies allowlisted Baggage
+  entries onto every span. `register_exporter_factory(kind, factory)` lets a
+  preset contribute a custom exporter `kind` (e.g. one that fetches an auth
+  token lazily) without coupling this module to any vendor.
+- [`context.py`](./plumbing/context.py) — trace-context and Baggage read/write helpers.
+- [`routing.py`](./plumbing/routing.py) — `TenantTracerCache`: when a request carries
+  team/key-scoped vendor credentials, route its spans through a credential-keyed
+  `TracerProvider` so one logger serves many tenants. The cache is a bounded LRU
+  that flushes + shuts down evicted providers, since the key derives from
+  request-supplied credentials and must not grow (or leak threads) without limit.
+- [`metrics.py`](./plumbing/metrics.py) — GenAI client metric instruments.
+
+### Adapter
+
+- [`logger.py`](./logger.py) — `OpenTelemetryV2`, a `CustomLogger` that
+  translates LiteLLM's logging callbacks into typed span data and hands them to
+  the engine. The LLM-call span is opened at the `log_pre_api_call` boundary
+  (parented to the live server span via ambient context) and closed at the async
+  success/failure callback; the open span is held in a bounded cache keyed by
+  `litellm_call_id`, never threaded through a metadata dict. The logger registers
+  itself into `litellm.input_callback` so `Logging.pre_call` fires the boundary
+  hook.
+- [`mount.py`](./mount.py) — `instrument_fastapi_app(app)`, the single call site
+  that attaches `opentelemetry-instrumentation-fastapi` for SERVER spans. It owns
+  the health-check exclusion default (`OTEL_PYTHON_FASTAPI_EXCLUDED_URLS`) and the
+  passthrough span-naming hook (`PASSTHROUGH_PREFIXES`) so `proxy_server` carries
+  no OTel detail. A safe no-op when the gate is off or the instrumentation package
+  is absent; must be called at app-creation time (the middleware stack freezes
+  once the app serves).
+
+### Presets
+
+- [`presets/`](./presets) — each preset reads one integration's env vars and
+  returns an `OpenTelemetryV2Config` (exporter destination + mapper vocabularies
+  + resource attributes). `PRESET_BY_CALLBACK` maps a callback name (`"arize"`,
+  `"langfuse_otel"`, …) to its preset. Integrations that support team/key-scoped
+  credentials also provide a per-request OTLP header builder
+  (`DYNAMIC_HEADERS_BY_CALLBACK`). Presets do **no** network I/O at build time:
+  AgentOps, for example, mints its JWT lazily inside a custom exporter on the
+  first export (in the `BatchSpanProcessor` worker thread), never on the event
+  loop.
+
+## Extending
+
+- **A new attribute vocabulary for a backend**: add a mapper in `mappers/`
+  (a class with a `map(data) -> AttributeMap` method, typically built from
+  `key -> extractor` tables) and register it in `mappers/__init__._MAPPER_BY_NAME`.
+- **A new integration**: add a preset in `presets/` that returns an
+  `OpenTelemetryV2Config`, and register it in `presets/__init__.PRESET_BY_CALLBACK`.
+  If it supports dynamic credentials, add a header builder to
+  `DYNAMIC_HEADERS_BY_CALLBACK`.
+- **A new span kind**: add a role to `spans.py` (registry entry + name builder),
+  a payload dataclass in `payloads.py`, and a branch in the relevant mapper(s).

litellm/integrations/otel/__init__.py

@@ -0,0 +1,102 @@
+"""Typed, semconv-aligned OpenTelemetry instrumentation for LiteLLM.
+
+The three sources of truth — attribute keys (:mod:`semconv`), the span and
+hierarchy registry (:mod:`spans`), and the typed span-data inputs
+(:mod:`payloads`) — plus :mod:`config` are exported here and are free of any
+``opentelemetry`` import. The engine layer (``emitter``, ``providers``,
+``context``, ``metrics``) and the ``CustomLogger`` adapter (``logger``) are
+reached via their submodule paths so that importing this package never
+requires the OTel SDK.
+
+The ``LITELLM_OTEL_V2`` env var gates whether the factory in
+``litellm_core_utils.litellm_logging`` constructs the ``OpenTelemetryV2``
+class (from :mod:`logger`).
+"""
+
+from litellm.integrations.otel.model.config import (
+    OTEL_V2_ENV,
+    OpenTelemetryV2Config,
+    is_otel_v2_enabled,
+)
+from litellm.integrations.otel.model.baggage import (
+    BAGGAGE_PROMOTED_KEYS,
+    DEFAULT_BAGGAGE_METADATA_KEYS,
+    promoted_baggage,
+)
+from litellm.integrations.otel.model.metadata import (
+    RequestContext,
+    RequestIdentity,
+)
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    LLMRequestParams,
+    LLMUsage,
+    ProxyRequestSpanData,
+    ServerInfo,
+    ServiceSpanData,
+    SpanError,
+)
+from litellm.integrations.otel.model.semconv import (
+    DB,
+    Error,
+    GenAI,
+    GenAIOperation,
+    GenAIProvider,
+    HTTP,
+    LiteLLM,
+    Metric,
+    Server,
+    resolve_operation,
+    resolve_provider,
+)
+from litellm.integrations.otel.model.spans import (
+    SPAN_REGISTRY,
+    LiteLLMSpanKind,
+    SpanRole,
+    SpanSpec,
+    db_system,
+    span_role_for_service,
+    validate_registry,
+)
+
+__all__ = [
+    # config
+    "OTEL_V2_ENV",
+    "OpenTelemetryV2Config",
+    "is_otel_v2_enabled",
+    # semconv
+    "BAGGAGE_PROMOTED_KEYS",
+    "DB",
+    "DEFAULT_BAGGAGE_METADATA_KEYS",
+    "Error",
+    "GenAI",
+    "GenAIOperation",
+    "GenAIProvider",
+    "HTTP",
+    "LiteLLM",
+    "Metric",
+    "Server",
+    "resolve_operation",
+    "resolve_provider",
+    # spans
+    "SPAN_REGISTRY",
+    "LiteLLMSpanKind",
+    "SpanRole",
+    "SpanSpec",
+    "db_system",
+    "span_role_for_service",
+    "validate_registry",
+    # payloads
+    "GuardrailSpanData",
+    "LLMCallSpanData",
+    "LLMRequestParams",
+    "LLMUsage",
+    "ProxyRequestSpanData",
+    "RequestContext",
+    "RequestIdentity",
+    "ServerInfo",
+    "ServiceSpanData",
+    "SpanError",
+    "promoted_baggage",
+]

litellm/integrations/otel/emitter.py

@@ -0,0 +1,175 @@
+"""The span engine: dedup, start, run the mapper chain, set status, end."""
+
+from collections import OrderedDict
+from typing import Callable, Sequence
+
+from opentelemetry.context import Context
+from opentelemetry.trace import Span, Tracer
+from opentelemetry.trace.status import Status, StatusCode
+
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.mappers import resolve_mappers
+from litellm.integrations.otel.mappers.base import AttributeMapper, SpanData
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    ServiceSpanData,
+)
+from litellm.integrations.otel.plumbing.providers import to_otel_span_kind
+from litellm.integrations.otel.model.semconv import Error
+from litellm.integrations.otel.model.spans import (
+    SPAN_REGISTRY,
+    SpanRole,
+    guardrail_span_name,
+    llm_call_span_name,
+    service_span_name,
+)
+
+# Roles emit() knows how to name and emit. PROXY_REQUEST and the management
+# routes are SERVER spans owned by the mounted FastAPI instrumentor, so they
+# have no builder here.
+_NAME_BUILDERS: dict[SpanRole, Callable[..., str]] = {
+    SpanRole.LLM_CALL: llm_call_span_name,
+    SpanRole.GUARDRAIL: guardrail_span_name,
+    # DB_CALL and SERVICE are both built from ServiceSpanData; they differ only in
+    # span kind (CLIENT vs INTERNAL) and attribute vocabulary, not in naming.
+    SpanRole.DB_CALL: service_span_name,
+    SpanRole.SERVICE: service_span_name,
+}
+
+# Cap on the dedup cache. It only needs to coalesce the sync+async firing window
+# of a single in-flight request, so a bounded LRU keeps memory flat on a
+# long-running proxy while still covering every concurrently-open call.
+_DEDUP_CACHE_MAX = 10_000
+
+
+class SpanEmitter:
+    def __init__(
+        self,
+        tracer: Tracer,
+        config: OpenTelemetryV2Config,
+        mappers: Sequence[AttributeMapper] | None = None,
+    ) -> None:
+        self._tracer = tracer
+        self._config = config
+        # The mapper chain is the sole source of span attributes. When not
+        # passed in, resolve it from the config so there's one source of truth.
+        self._mappers: list[AttributeMapper] = (
+            list(mappers)
+            if mappers is not None
+            else resolve_mappers(config.mapper_names)
+        )
+        # Bounded LRU (ordered by insertion / most-recent touch). Storing keys
+        # only — the value is unused — so it behaves like a capped set.
+        self._emitted: "OrderedDict[tuple[str, SpanRole], None]" = OrderedDict()
+
+    # -- low-level helpers --------------------------------------------------- #
+
+    def start_span(
+        self,
+        role: SpanRole,
+        name: str,
+        parent_context: Context | None = None,
+        start_time_ns: int | None = None,
+        *,
+        tracer: Tracer | None = None,
+    ) -> Span:
+        """Start a span for ``role`` without dedup or attribute mapping.
+
+        For callers that own and manage their own span lifecycle. ``tracer``
+        overrides the bound tracer for this span only, used for per-request
+        multi-tenant credential routing.
+        """
+        return (tracer or self._tracer).start_span(
+            name,
+            context=parent_context,
+            kind=to_otel_span_kind(SPAN_REGISTRY[role].kind),
+            start_time=start_time_ns,
+        )
+
+    def _seen(self, dedup_key: str | None, role: SpanRole) -> bool:
+        """Return True once a ``(dedup_key, role)`` pair has been emitted.
+
+        Guards against emitting the same span twice when a streaming call
+        fires both a sync and an async logging callback.
+        """
+        if not dedup_key:
+            return False
+        marker = (dedup_key, role)
+        if marker in self._emitted:
+            self._emitted.move_to_end(marker)
+            return True
+        self._emitted[marker] = None
+        if len(self._emitted) > _DEDUP_CACHE_MAX:
+            self._emitted.popitem(last=False)  # evict least-recently-used
+        return False
+
+    # -- the engine ---------------------------------------------------------- #
+
+    def emit(
+        self,
+        role: SpanRole,
+        data: SpanData,
+        parent_context: Context | None = None,
+        *,
+        start_time_ns: int | None = None,
+        end_time_ns: int | None = None,
+        tracer: Tracer | None = None,
+    ) -> Span | None:
+        """Emit one complete span: dedup, start, map attributes, status, end.
+
+        Return the span, or ``None`` if it was deduplicated away. ``tracer``
+        overrides the bound tracer for this span, used for per-request routing.
+        """
+        # Only LLM-call spans carry a dedup key; LLM-call and service spans
+        # carry an ``error`` field. ``isinstance`` narrows the type for mypy and
+        # keeps the engine free of duck-typed attribute reads.
+        dedup_key = data.identity.call_id if isinstance(data, LLMCallSpanData) else None
+        if self._seen(dedup_key, role):
+            return None
+        span = self.start_span(
+            role,
+            _NAME_BUILDERS[role](data),
+            parent_context=parent_context,
+            start_time_ns=start_time_ns,
+            tracer=tracer,
+        )
+        self.finish_span(role, span, data, end_time_ns=end_time_ns)
+        return span
+
+    def finish_span(
+        self,
+        role: SpanRole,
+        span: Span,
+        data: SpanData,
+        *,
+        end_time_ns: int | None = None,
+    ) -> None:
+        """Stamp attributes + status on an already-started ``span`` and end it.
+
+        The counterpart to :meth:`start_span` for callers that own a span's
+        lifecycle — the LLM-call span is opened at the request's ``pre_call``
+        boundary (so it parents to the live server span via real ambient context,
+        never a span threaded through a metadata dict) and closed here once the
+        typed payload is available. The span name is (re)built from the now-known
+        data, since the boundary opener only has a provisional name.
+        """
+        span.update_name(_NAME_BUILDERS[role](data))
+        for mapper in self._mappers:
+            for key, value in mapper.map(data).items():
+                span.set_attribute(key, value)
+        error = (
+            data.error
+            if isinstance(data, (LLMCallSpanData, ServiceSpanData, GuardrailSpanData))
+            else None
+        )
+        if error and (error.error_type or error.message):
+            span.set_attribute(Error.TYPE, error.error_type or "error")
+            span.set_status(
+                Status(StatusCode.ERROR, error.message or error.error_type or "error")
+            )
+        # On success leave the status UNSET (the semconv default) rather than
+        # forcing OK — that matches the FastAPI server span and avoids implying a
+        # span-level health signal litellm doesn't actually evaluate. Only a
+        # genuine error sets a status.
+        span.end(end_time=end_time_ns)

litellm/integrations/otel/logger.py

@@ -0,0 +1,495 @@
+"""``CustomLogger`` adapter on the OpenTelemetry span engine."""
+
+from collections import OrderedDict
+from contextlib import contextmanager
+from datetime import datetime
+from typing import TYPE_CHECKING, Any, Iterator, Mapping, cast
+
+from opentelemetry.context import attach, get_current
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.trace import Span, Tracer, get_current_span, use_span
+
+import litellm
+from litellm.integrations.custom_logger import CustomLogger
+from litellm.integrations.otel.model.baggage import promoted_baggage
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.plumbing.context import (
+    is_recordable_span,
+    resolve_parent_context,
+    resolve_request_span_context,
+    set_request_baggage,
+    set_request_root_span,
+)
+from litellm.integrations.otel.emitter import SpanEmitter
+from litellm.integrations.otel.mappers import resolve_mappers
+from litellm.integrations.otel.model.metadata import (
+    LLMCallEvent,
+    RequestIdentity,
+    guardrail_entries_from_request_data,
+    model_from_request_data,
+)
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    ServiceSpanData,
+    SpanError,
+)
+from litellm.integrations.otel.plumbing.providers import (
+    build_tracer_provider,
+    get_tracer,
+)
+from litellm.integrations.otel.plumbing.routing import TenantTracerCache
+from litellm.integrations.otel.model.spans import SpanRole, span_role_for_service
+from litellm.integrations.otel.model.utils import to_ns
+
+if TYPE_CHECKING:
+    from litellm.types.utils import StandardLoggingGuardrailInformation
+
+LITELLM_TRACER_NAME = "litellm"
+
+# Any callback whose class belongs to one of these modules is "the OTel
+# callback" for proxy-global-registration purposes.
+_OTEL_MODULES = (
+    "litellm.integrations.otel",
+    "litellm.integrations.opentelemetry",
+)
+
+
+# Cap on the open-call carrier map. A span opened at ``pre_call`` that never
+# reaches a success/failure callback (e.g. a stream that only fires stream
+# events) would otherwise linger; bounding the map evicts the oldest so memory
+# stays flat on a long-running proxy while covering every concurrent in-flight
+# call.
+_OPEN_CALLS_MAX = 10_000
+
+
+class _LLMCallSpan:
+    """The state carried from the ``pre_call`` boundary to span close.
+
+    ``span`` is the live span when it could be opened at the boundary (the server
+    span was ambient), or ``None`` when creation was deferred because no ambient
+    parent was visible — in which case the async callback creates it against its
+    own (worker-copied) ambient context using ``start_time_ns``. The presence of
+    a carrier for a call at all is the proof that ``pre_call`` ran, i.e. that an
+    upstream call was actually attempted.
+    """
+
+    __slots__ = ("span", "start_time_ns")
+
+    def __init__(self, span: "Span | None", start_time_ns: int | None) -> None:
+        self.span = span
+        self.start_time_ns = start_time_ns
+
+
+class OpenTelemetryV2(CustomLogger):
+    """The ``CustomLogger`` for OpenTelemetry."""
+
+    def __init__(
+        self,
+        config: OpenTelemetryV2Config | None = None,
+        callback_name: str | None = None,
+        tracer_provider: TracerProvider | None = None,
+        logger_provider: Any | None = None,  # reserved for OTel logs
+        meter_provider: Any | None = None,  # reserved for metrics
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.config: OpenTelemetryV2Config = config or OpenTelemetryV2Config(**kwargs)
+        self.callback_name = callback_name
+        self._tracer_provider: TracerProvider = (
+            tracer_provider
+            if tracer_provider is not None
+            else build_tracer_provider(self.config)
+        )
+        self.tracer: Tracer = get_tracer(self._tracer_provider, LITELLM_TRACER_NAME)
+        self._emitter = SpanEmitter(
+            self.tracer, self.config, mappers=resolve_mappers(self.config.mapper_names)
+        )
+        self._tenant_tracers = TenantTracerCache(
+            self.config, callback_name, LITELLM_TRACER_NAME
+        )
+        self._open_llm_calls: "OrderedDict[str, _LLMCallSpan]" = OrderedDict()
+        self._init_otel_logger_on_litellm_proxy()
+
+    # ====================================================================== #
+    #  Proxy global registration
+    # ====================================================================== #
+
+    def _register_in_callback_list(self, callbacks: list) -> None:
+        already_otel = any(
+            cb.__class__.__module__.startswith(_OTEL_MODULES)
+            for cb in callbacks
+            if hasattr(cb, "__class__")
+        )
+        if not already_otel:
+            callbacks.append(self)
+
+    def _init_otel_logger_on_litellm_proxy(self) -> None:
+        try:
+            from litellm.proxy import proxy_server
+        except Exception:
+            return
+        try:
+            self._register_in_callback_list(litellm.service_callback)
+            self._register_in_callback_list(litellm.input_callback)
+            self._register_in_callback_list(litellm._async_success_callback)
+            self._register_in_callback_list(litellm._async_failure_callback)
+        except Exception:
+            pass
+        if getattr(proxy_server, "open_telemetry_logger", None) is None:
+            setattr(proxy_server, "open_telemetry_logger", self)
+
+    # ====================================================================== #
+    #  LLM-call callbacks — the span is opened at the ``pre_call`` boundary and
+    #  closed here. See ``log_pre_api_call``.
+    # ====================================================================== #
+
+    def log_pre_api_call(self, model, messages, kwargs):
+        """Open the LLM-call span at the call boundary.
+
+        Runs synchronously inside the request task, before the upstream call —
+        the one place where the live server span is genuinely the ambient OTel
+        context — so the span parents to it natively, with no span threaded
+        through a metadata dict. The open span is stashed on the per-request
+        ``LiteLLMLoggingObj`` (a typed object) and closed in the async callback.
+
+        When no recordable parent is visible (``pre_call`` was driven from a thread
+        pool for a sync-only provider, where contextvars — and so the anchor —
+        don't follow), creation is deferred: only the start time is recorded, and
+        the async callback — whose worker context was copied from the request task
+        and so still carries the anchor — creates the span then.
+
+        Synthetic proxy-gate error logs (auth/rate-limit rejections) also fire this
+        hook but never made an upstream call; they are tagged and skipped so no
+        phantom LLM-call span is produced.
+        """
+        call = LLMCallEvent.from_dict(kwargs)
+        if call.is_no_upstream_call:
+            return
+        call_id = call.call_id
+        if call_id is None:
+            return
+        # Idempotent: a retried call may re-enter ``pre_call`` with the same
+        # call id; keep the first span so its start time is the true one.
+        if call_id in self._open_llm_calls:
+            return
+        start_time_ns = to_ns(datetime.now())
+        span: Span | None = None
+        # Parent to the request's anchored root span (stable across the request),
+        # falling back to ambient on the SDK path. Open the span live only when
+        # that resolves to a recordable parent; otherwise defer to the close
+        # callback (the thread-pool case, where the anchor isn't visible here).
+        parent_context = resolve_request_span_context()
+        if is_recordable_span(get_current_span(parent_context)):
+            span = self._emitter.start_span(
+                SpanRole.LLM_CALL,
+                call.provisional_span_name,
+                parent_context=parent_context,
+                start_time_ns=start_time_ns,
+                tracer=self._tenant_tracers.tracer_for(
+                    self.tracer, call.dynamic_params
+                ),
+            )
+        self._open_llm_calls[call_id] = _LLMCallSpan(
+            span=span, start_time_ns=start_time_ns
+        )
+        # Evict the oldest open call if the map is over budget. A call that opens
+        # but never closes (a stream that only fires stream events) would linger
+        # otherwise; the evicted span is simply dropped (never exported).
+        if len(self._open_llm_calls) > _OPEN_CALLS_MAX:
+            self._open_llm_calls.popitem(last=False)
+
+    async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
+        self._close_llm_call(kwargs, start_time, end_time)
+
+    async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time):
+        self._close_llm_call(kwargs, start_time, end_time)
+
+    def _close_llm_call(
+        self,
+        kwargs: Mapping[str, Any],
+        start_time: datetime | float | None,
+        end_time: datetime | float | None,
+    ) -> Span | None:
+        """Finish the LLM-call span opened at ``pre_call`` (or create it deferred).
+
+        No carrier for this call id means ``pre_call`` never ran — the request was
+        rejected at the gate or blocked by a pre-call guardrail before any upstream
+        call — so there is nothing to record and no phantom span.
+        """
+        call = LLMCallEvent.from_dict(kwargs)
+        call_id = call.call_id
+        # ``pop`` is the dedup: this method runs from both the success and failure
+        # paths, and whichever fires first removes the carrier and closes the span.
+        carrier = self._open_llm_calls.pop(call_id, None) if call_id else None
+        if carrier is None:
+            return None
+        payload = call.payload
+        if payload is None:
+            if carrier.span is not None:
+                # Opened at the boundary but the payload never materialized — end
+                # it (named provisionally) so it isn't leaked as an open span.
+                carrier.span.end(end_time=to_ns(end_time))
+            return None
+        data = LLMCallSpanData.from_standard_logging_payload(
+            payload, capture_content=self.config.capture_span_content
+        )
+        end_time_ns = to_ns(end_time)
+        if carrier.span is not None:
+            # Born at the boundary: stamp attributes from the typed payload, set
+            # status, and end it. Its parent (the server span) was captured at
+            # creation from real ambient context.
+            self._emitter.finish_span(
+                SpanRole.LLM_CALL, carrier.span, data, end_time_ns=end_time_ns
+            )
+            return carrier.span
+        # Deferred: ``pre_call`` saw no recordable parent, so create the span now.
+        # The worker copied the request task's context, which carries the anchored
+        # root span — parent to it (ambient fallback on the SDK path). Seed identity
+        # Baggage so the span — and the SDK path, which has none — is labeled
+        # consistently.
+        parent_ctx = resolve_request_span_context()
+        bag = promoted_baggage(
+            data.identity,
+            data.request_model,
+            promoted_keys=tuple(self.config.baggage_promoted_keys),
+            metadata_keys=tuple(self.config.baggage_metadata_keys),
+        )
+        if bag:
+            parent_ctx = set_request_baggage(bag, context=parent_ctx)
+        return self._emitter.emit(
+            SpanRole.LLM_CALL,
+            data,
+            parent_context=parent_ctx,
+            start_time_ns=carrier.start_time_ns,
+            end_time_ns=end_time_ns,
+            tracer=self._tenant_tracers.tracer_for(self.tracer, call.dynamic_params),
+        )
+
+    # ====================================================================== #
+    #  Service hooks
+    # ====================================================================== #
+
+    async def async_service_success_hook(
+        self,
+        payload: Any,
+        parent_otel_span: Span | None = None,
+        start_time: datetime | float | None = None,
+        end_time: datetime | float | None = None,
+        event_metadata: dict | None = None,
+    ) -> None:
+        self._emit_service(
+            payload,
+            parent_otel_span=parent_otel_span,
+            start_time=start_time,
+            end_time=end_time,
+            event_metadata=event_metadata,
+            error_override=None,
+        )
+
+    async def async_service_failure_hook(
+        self,
+        payload: Any,
+        error: str | None = "",
+        parent_otel_span: Span | None = None,
+        start_time: datetime | float | None = None,
+        end_time: datetime | float | None = None,
+        event_metadata: dict | None = None,
+    ) -> None:
+        self._emit_service(
+            payload,
+            parent_otel_span=parent_otel_span,
+            start_time=start_time,
+            end_time=end_time,
+            event_metadata=event_metadata,
+            error_override=error or "error",
+        )
+
+    def _emit_service(
+        self,
+        payload: Any,
+        *,
+        parent_otel_span: Span | None,
+        start_time: datetime | float | None,
+        end_time: datetime | float | None,
+        event_metadata: dict | None,
+        error_override: str | None,
+    ) -> Span | None:
+        data = ServiceSpanData.from_payload(payload, event_metadata=event_metadata)
+        # Decide whether this service call is a span at all, and of what kind.
+        # ``None`` means metrics-only (framework instrumentation that duplicates a
+        # gen-AI span — ``self``/``router``/``proxy_pre_call`` — or ``auth``, which
+        # gets a live phase span instead). Those still feed Prometheus/Datadog via
+        # their own hooks; they just never enter the trace.
+        role = span_role_for_service(data.service_name)
+        if role is None:
+            return None
+        # A metrics-only ping with neither timing nor a parent (in-memory queue
+        # gauges) is not a traceable operation; a span for it would be a
+        # zero-duration root with no context, so skip it. Real background work
+        # (budget/reset jobs, spend flush) passes start/end times and still emits
+        # as a root; anything with a parent emits regardless.
+        if (
+            error_override is None
+            and start_time is None
+            and end_time is None
+            and parent_otel_span is None
+        ):
+            return None
+        if error_override is not None and data.error is None:
+            data = ServiceSpanData(
+                service_name=data.service_name,
+                call_type=data.call_type,
+                error=SpanError(message=error_override),
+                event_metadata=data.event_metadata,
+            )
+        # Parent like every other span: ambient context first (so identity Baggage
+        # rides along and the call nests under whatever request phase is active —
+        # e.g. a DB lookup under the live ``auth`` span), falling back to the
+        # server span the proxy threaded as ``parent_otel_span``. A background
+        # service call has neither, so it starts its own root trace.
+        parent_context = resolve_parent_context(threaded=parent_otel_span)
+        return self._emitter.emit(
+            role,
+            data,
+            parent_context=parent_context,
+            start_time_ns=to_ns(start_time),
+            end_time_ns=to_ns(end_time),
+        )
+
+    # ====================================================================== #
+    #  async_post_call_* hooks — emit guardrail spans. The server span's status
+    #  / errors are the FastAPI instrumentor's job, so we don't touch it here.
+    # ====================================================================== #
+
+    def seed_request_identity(self, user_api_key_dict: Any, model: Any = None) -> None:
+        """Attach request-identity Baggage to the current context + server span.
+
+        Seeding identity into Baggage makes **every** span emitted afterwards for
+        this request — LLM call, guardrail, DB call — inherit it via
+        ``LiteLLMBaggageSpanProcessor``. Called once at the auth boundary (as soon
+        as the key resolves) so post-auth spans are labeled consistently; the
+        Baggage rides the request task's contextvar from there on. Auth-internal
+        DB lookups that run before the key is known stay unlabeled — identity
+        isn't determined yet, which is correct.
+        """
+        try:
+            identity = RequestIdentity.from_user_api_key_auth(user_api_key_dict)
+            bag = promoted_baggage(
+                identity,
+                model,
+                promoted_keys=tuple(self.config.baggage_promoted_keys),
+                metadata_keys=tuple(self.config.baggage_metadata_keys),
+            )
+            if bag:
+                # Attach (no detach): the contextvar is scoped to this request's
+                # asyncio task and is reclaimed when the task ends.
+                attach(set_request_baggage(bag, context=get_current()))
+                # The server span was started by the instrumentor before this ran,
+                # so the Baggage processor (which only fires at span start) won't
+                # backfill it — stamp identity on it directly.
+                server_span = get_current_span()
+                if is_recordable_span(server_span):
+                    # Re-capture the anchor here too: this runs post-auth with the
+                    # server span active and covers entrypoints that bypass
+                    # ``create_litellm_proxy_request_started_span`` (e.g. the SDK
+                    # path's ``async_pre_call_hook``). Idempotent.
+                    set_request_root_span(server_span)
+                    for key, value in bag.items():
+                        server_span.set_attribute(key, value)
+        except Exception:
+            pass
+
+    @contextmanager
+    def start_phase_span(self, name: str) -> "Iterator[Span]":
+        span = self._emitter.start_span(SpanRole.SERVICE, name)
+        with use_span(span, end_on_exit=True):
+            yield span
+
+    async def async_pre_call_hook(
+        self,
+        user_api_key_dict: Any,
+        cache: Any,
+        data: dict,
+        call_type: Any,
+    ) -> dict:
+        self.seed_request_identity(
+            user_api_key_dict,
+            model=model_from_request_data(data),
+        )
+        return data
+
+    async def async_post_call_success_hook(
+        self,
+        data: Mapping[str, Any],
+        user_api_key_dict: Any,
+        response: Any,
+    ) -> Any:
+        self._emit_guardrail_spans(data)
+        return response
+
+    async def async_post_call_failure_hook(
+        self,
+        request_data: Mapping[str, Any],
+        original_exception: BaseException | None,
+        user_api_key_dict: Any,
+        traceback_str: str | None = None,
+    ) -> None:
+        self._emit_guardrail_spans(request_data)
+
+    def _emit_guardrail_spans(self, request_data: Mapping[str, Any]) -> None:
+        # A guardrail is a sibling of the LLM call under the request's root span,
+        # so parent it to the explicit anchor — not the active span, which on the
+        # failure path can be the live ``auth`` phase span (post-call failure hooks
+        # run from inside it on an auth rejection). Emit with the guardrail's actual
+        # execution window so a pre_call guardrail is placed before the LLM call
+        # rather than at post-call emission time.
+        guardrails = guardrail_entries_from_request_data(request_data)
+        if not guardrails:
+            return
+        parent_ctx = resolve_request_span_context()
+        for entry in guardrails:
+            data = GuardrailSpanData.from_logging_entry(
+                cast("StandardLoggingGuardrailInformation", entry)
+            )
+            self._emitter.emit(
+                SpanRole.GUARDRAIL,
+                data,
+                parent_context=parent_ctx,
+                start_time_ns=to_ns(data.start_time),
+                end_time_ns=to_ns(data.end_time),
+            )
+
+    def create_litellm_proxy_request_started_span(
+        self, start_time: datetime, headers: Mapping[str, str] | None
+    ) -> Span | None:
+        span = get_current_span()
+        if not is_recordable_span(span):
+            return None
+        set_request_root_span(span)
+        return span
+
+
+def _registered_v2_logger() -> "OpenTelemetryV2 | None":
+    try:
+        from litellm.proxy import proxy_server
+    except Exception:
+        return None
+    logger = getattr(proxy_server, "open_telemetry_logger", None)
+    return logger if isinstance(logger, OpenTelemetryV2) else None
+
+
+def seed_request_identity(user_api_key_dict: Any, model: Any = None) -> None:
+    logger = _registered_v2_logger()
+    if logger is not None:
+        logger.seed_request_identity(user_api_key_dict, model=model)
+
+
+@contextmanager
+def phase_span(name: str) -> "Iterator[Span | None]":
+    logger = _registered_v2_logger()
+    if logger is None:
+        yield None
+        return
+    with logger.start_phase_span(name) as span:
+        yield span

litellm/integrations/otel/mappers/__init__.py

@@ -0,0 +1,58 @@
+"""Attribute mappers: pure ``LLMCallSpanData -> {attribute key: value}`` functions.
+
+Composition over inheritance: vocabularies layer onto the same span. Listing
+``["genai", "openinference"]`` in ``config.mapper_names`` makes every span
+carry both the canonical ``gen_ai.*`` keys and the OpenInference (Arize +
+Phoenix) keys. Add ``"langfuse"`` and it works for all three backends at once.
+"""
+
+from typing import Callable, Iterable
+
+from litellm.integrations.otel.mappers.base import (
+    AttributeMap,
+    AttributeMapper,
+    AttrValue,
+)
+from litellm.integrations.otel.mappers.genai import GenAIMapper
+from litellm.integrations.otel.mappers.langfuse import LangfuseMapper
+from litellm.integrations.otel.mappers.langtrace import LangtraceMapper
+from litellm.integrations.otel.mappers.legacy import LegacyMapper
+from litellm.integrations.otel.mappers.openinference import OpenInferenceMapper
+from litellm.integrations.otel.mappers.weave import WeaveMapper
+
+# Registry keyed by ``config.mapper_names`` entries.
+_MAPPER_BY_NAME: dict[str, Callable[[], AttributeMapper]] = {
+    "genai": GenAIMapper,
+    "legacy": LegacyMapper,
+    "openinference": OpenInferenceMapper,
+    "langfuse": LangfuseMapper,
+    "weave": WeaveMapper,
+    "langtrace": LangtraceMapper,
+}
+
+
+def resolve_mappers(names: Iterable[str]) -> list[AttributeMapper]:
+    """Resolve mapper names to instances. Unknown names raise ``ValueError``."""
+    out: list[AttributeMapper] = []
+    for name in names:
+        factory = _MAPPER_BY_NAME.get(name)
+        if factory is None:
+            raise ValueError(
+                f"unknown mapper name {name!r}; known: " f"{sorted(_MAPPER_BY_NAME)}"
+            )
+        out.append(factory())
+    return out
+
+
+__all__ = [
+    "AttributeMap",
+    "AttributeMapper",
+    "AttrValue",
+    "GenAIMapper",
+    "LangfuseMapper",
+    "LangtraceMapper",
+    "LegacyMapper",
+    "OpenInferenceMapper",
+    "WeaveMapper",
+    "resolve_mappers",
+]

litellm/integrations/otel/mappers/base.py

@@ -0,0 +1,36 @@
+"""Mapper protocol and attribute value types."""
+
+from typing import Sequence
+
+from typing_extensions import Protocol, runtime_checkable
+
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    ServiceSpanData,
+)
+
+AttrScalar = str | bool | int | float
+# Mirrors ``opentelemetry.util.types.AttributeValue`` (homogeneous sequences)
+# without importing the SDK, so mappers stay OTel-free.
+AttrValue = (
+    AttrScalar | Sequence[str] | Sequence[bool] | Sequence[int] | Sequence[float]
+)
+AttributeMap = dict[str, AttrValue]
+
+# The closed set of span-data types the engine routes through the mapper chain.
+# Server spans (PROXY_REQUEST + management routes) belong to the mounted FastAPI
+# instrumentor, not the mapper chain.
+SpanData = LLMCallSpanData | GuardrailSpanData | ServiceSpanData
+
+
+@runtime_checkable
+class AttributeMapper(Protocol):
+    """Maps a typed span input to a flat dict of OTel span attributes.
+
+    One method per mapper, dispatched internally on the ``data`` type. The
+    engine calls this uniformly for every span kind — mappers that don't speak
+    a given type return ``{}``. This is why the engine contains no attribute keys.
+    """
+
+    def map(self, data: SpanData) -> AttributeMap: ...

litellm/integrations/otel/mappers/genai.py

@@ -0,0 +1,137 @@
+"""Canonical OpenTelemetry GenAI semantic-convention mapper (always active).
+
+Owns the attribute schema for every span kind the engine emits — LLM call,
+guardrail, and service — so the engine itself never references attribute keys.
+
+Each span kind declares its schema as a flat ``attribute key -> extractor``
+table: one lambda per mapping operation, applied against the typed span data.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import collect, drop_none
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    ServiceSpanData,
+    ToolDefinition,
+)
+from litellm.integrations.otel.model.semconv import DB, Error, GenAI, LiteLLM, Server
+from litellm.integrations.otel.model.spans import db_system
+
+
+class GenAIMapper:
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        GenAI.OPERATION_NAME: lambda d: d.operation.value,
+        GenAI.PROVIDER_NAME: lambda d: d.provider or None,
+        GenAI.REQUEST_MODEL: lambda d: d.request_model or None,
+        GenAI.REQUEST_TEMPERATURE: lambda d: d.request_params.temperature,
+        GenAI.REQUEST_TOP_P: lambda d: d.request_params.top_p,
+        GenAI.REQUEST_TOP_K: lambda d: d.request_params.top_k,
+        GenAI.REQUEST_MAX_TOKENS: lambda d: d.request_params.max_tokens,
+        GenAI.REQUEST_FREQUENCY_PENALTY: lambda d: d.request_params.frequency_penalty,
+        GenAI.REQUEST_PRESENCE_PENALTY: lambda d: d.request_params.presence_penalty,
+        GenAI.REQUEST_STOP_SEQUENCES: lambda d: (
+            list(d.request_params.stop_sequences)
+            if d.request_params.stop_sequences
+            else None
+        ),
+        GenAI.REQUEST_SEED: lambda d: d.request_params.seed,
+        GenAI.RESPONSE_MODEL: lambda d: d.response_model,
+        GenAI.RESPONSE_ID: lambda d: d.response_id,
+        GenAI.RESPONSE_FINISH_REASONS: lambda d: (
+            list(d.finish_reasons) if d.finish_reasons else None
+        ),
+        GenAI.USAGE_INPUT_TOKENS: lambda d: d.usage.input_tokens,
+        GenAI.USAGE_OUTPUT_TOKENS: lambda d: d.usage.output_tokens,
+        Error.TYPE: lambda d: d.error.error_type if d.error else None,
+        Server.ADDRESS: lambda d: d.server.address if d.server else None,
+        Server.PORT: lambda d: d.server.port if d.server else None,
+        LiteLLM.CALL_ID: lambda d: d.identity.call_id or None,
+        # The provider/underlying model is only known once routing has picked a
+        # deployment, so it can't ride identity Baggage (seeded at auth, before
+        # routing) onto the boundary-born LLM span — stamp it directly here.
+        LiteLLM.PROVIDER_MODEL: lambda d: d.identity.provider_model or None,
+        f"{LiteLLM.COST_PREFIX}total": lambda d: d.response_cost,
+        LiteLLM.REQUEST_STREAMING: lambda d: d.is_streaming,
+    }
+
+    _TOOL_ATTRS: dict[str, Callable[[ToolDefinition], AttrValue | None]] = {
+        "name": lambda t: t.name,
+        "description": lambda t: t.description or None,
+        "parameters": lambda t: t.parameters_json or None,
+    }
+
+    _GUARDRAIL_ATTRS: dict[str, Callable[[GuardrailSpanData], AttrValue | None]] = {
+        LiteLLM.GUARDRAIL_NAME: lambda d: d.guardrail_name,
+        LiteLLM.GUARDRAIL_MODE: lambda d: d.mode,
+        LiteLLM.GUARDRAIL_STATUS: lambda d: d.status,
+        LiteLLM.GUARDRAIL_PROVIDER: lambda d: d.provider,
+        LiteLLM.GUARDRAIL_ACTION: lambda d: d.action,
+        LiteLLM.GUARDRAIL_RESPONSE: lambda d: d.response_json,
+        LiteLLM.GUARDRAIL_VIOLATION_CATEGORIES: lambda d: (
+            list(d.violation_categories) if d.violation_categories else None
+        ),
+        LiteLLM.GUARDRAIL_CONFIDENCE_SCORE: lambda d: d.confidence_score,
+        LiteLLM.GUARDRAIL_RISK_SCORE: lambda d: d.risk_score,
+        LiteLLM.GUARDRAIL_MASKED_ENTITY_COUNT: lambda d: d.masked_entity_count,
+        LiteLLM.GUARDRAIL_DURATION: lambda d: d.duration,
+        LiteLLM.GUARDRAIL_ID: lambda d: d.guardrail_id,
+        LiteLLM.GUARDRAIL_POLICY_TEMPLATE: lambda d: d.policy_template,
+        LiteLLM.GUARDRAIL_DETECTION_METHOD: lambda d: d.detection_method,
+    }
+
+    _SERVICE_ATTRS: dict[str, Callable[[ServiceSpanData], AttrValue | None]] = {
+        LiteLLM.SERVICE_NAME: lambda d: d.service_name,
+        LiteLLM.SERVICE_CALL_TYPE: lambda d: d.call_type,
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case GuardrailSpanData():
+                return self._guardrail(data)
+            case ServiceSpanData():
+                return self._service(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        attrs = collect(cls._LLM_CALL_ATTRS, data)
+        attrs.update(
+            drop_none(
+                {
+                    f"gen_ai.tool.{idx}.{suffix}": extract(tool)
+                    for idx, tool in enumerate(data.tools)
+                    for suffix, extract in cls._TOOL_ATTRS.items()
+                }
+            )
+        )
+        return attrs
+
+    @classmethod
+    def _guardrail(cls, data: GuardrailSpanData) -> AttributeMap:
+        return collect(cls._GUARDRAIL_ATTRS, data)
+
+    @classmethod
+    def _service(cls, data: ServiceSpanData) -> AttributeMap:
+        attrs = collect(cls._SERVICE_ATTRS, data)
+        # An outbound datastore call (DB_CALL / CLIENT span) also carries db.*
+        # semconv. Internal services (router, budget jobs, …) have no db.system,
+        # so they get only the litellm.service.* keys above.
+        system = db_system(data.service_name)
+        if system is not None:
+            attrs[DB.SYSTEM_NAME] = system
+            if data.call_type:
+                attrs[DB.OPERATION_NAME] = data.call_type
+        attrs.update(
+            {
+                f"{LiteLLM.METADATA_PREFIX}{key}": value
+                for key, value in data.event_metadata.items()
+            }
+        )
+        return attrs

litellm/integrations/otel/mappers/langfuse.py

@@ -0,0 +1,84 @@
+"""Langfuse OTLP attribute mapper.
+
+Langfuse ingests OTLP spans and reads from its own vendor namespace
+(``langfuse.observation.*``, ``langfuse.trace.*``). Compose this mapper after
+``GenAIMapper`` to send canonical + Langfuse-flavored spans simultaneously.
+
+Every attribute is declared as a ``key -> extractor`` table entry (one callable
+per mapping operation): ``_LLM_CALL_ATTRS`` for scalars and ``_BLOB_ATTRS`` for
+the JSON-serialized payloads. ``_llm_call`` just applies both tables.
+"""
+
+import json
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import (
+    collect,
+    json_if,
+    output_messages,
+    serialize_messages,
+)
+from litellm.integrations.otel.model.payloads import (
+    LLMCallSpanData,
+    LLMRequestParams,
+    LLMUsage,
+)
+
+
+class LangfuseMapper:
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "langfuse.observation.type": lambda d: "generation",
+        "langfuse.observation.model.name": lambda d: d.request_model or None,
+        "langfuse.observation.metadata.provider": lambda d: d.provider or None,
+        "langfuse.observation.id": lambda d: d.identity.call_id or None,
+        "langfuse.trace.metadata.team_id": lambda d: d.identity.team_id or None,
+        "langfuse.trace.metadata.team_alias": lambda d: d.identity.team_alias or None,
+    }
+
+    # Sub-tables folded into their respective JSON blobs.
+    _MODEL_PARAMS: dict[str, Callable[[LLMRequestParams], AttrValue | None]] = {
+        "temperature": lambda rp: rp.temperature,
+        "top_p": lambda rp: rp.top_p,
+        "max_tokens": lambda rp: rp.max_tokens,
+        "frequency_penalty": lambda rp: rp.frequency_penalty,
+        "presence_penalty": lambda rp: rp.presence_penalty,
+        "seed": lambda rp: rp.seed,
+    }
+    _USAGE_FIELDS: dict[str, Callable[[LLMUsage], AttrValue | None]] = {
+        "input": lambda u: u.input_tokens,
+        "output": lambda u: u.output_tokens,
+        "total": lambda u: u.total_tokens,
+    }
+
+    # JSON-payload attributes: each builder returns the serialized blob or None.
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "langfuse.observation.model.parameters": lambda d: json_if(
+            collect(LangfuseMapper._MODEL_PARAMS, d.request_params)
+        ),
+        "langfuse.observation.input": lambda d: serialize_messages(d.messages_in),
+        "langfuse.observation.output": lambda d: serialize_messages(output_messages(d)),
+        "langfuse.observation.usage_details": lambda d: json_if(
+            collect(LangfuseMapper._USAGE_FIELDS, d.usage)
+        ),
+        "langfuse.observation.cost_details": lambda d: (
+            json.dumps({"total": d.response_cost})
+            if d.response_cost is not None
+            else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+        }

litellm/integrations/otel/mappers/langtrace.py

@@ -0,0 +1,64 @@
+"""Langtrace attribute mapper.
+
+Produces Langtrace's attribute vocabulary so a span can be ingested by a
+Langtrace backend. Compose it alongside other mappers like any other
+vocabulary.
+
+Scalar attributes are declared as a flat ``key -> extractor`` table (one lambda
+per mapping operation); the prompt/completion blobs are serialized as a tail.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import (
+    collect,
+    json_or_none,
+    output_messages,
+)
+from litellm.integrations.otel.model.payloads import LLMCallSpanData
+
+
+class LangtraceMapper:
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "gen_ai.operation.name": lambda d: "chat",
+        "langtrace.service.name": lambda d: d.provider or None,
+        "llm.model": lambda d: d.request_model or None,
+        "gen_ai.response.model": lambda d: d.response_model or None,
+        "gen_ai.response_id": lambda d: d.response_id or None,
+        "gen_ai.system_fingerprint": lambda d: d.system_fingerprint or None,
+        "llm.temperature": lambda d: d.request_params.temperature,
+        "llm.top_p": lambda d: d.request_params.top_p,
+        "llm.top_k": lambda d: d.request_params.top_k,
+        "llm.max_tokens": lambda d: d.request_params.max_tokens,
+        "llm.frequency_penalty": lambda d: d.request_params.frequency_penalty,
+        "llm.presence_penalty": lambda d: d.request_params.presence_penalty,
+        "llm.stream": lambda d: d.is_streaming,
+        "llm.token.counts.prompt": lambda d: d.usage.input_tokens,
+        "llm.token.counts.completion": lambda d: d.usage.output_tokens,
+        "llm.token.counts.total": lambda d: d.usage.total_tokens,
+    }
+
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "llm.prompts": lambda d: (
+            json_or_none(list(d.messages_in)) if d.messages_in else None
+        ),
+        "llm.completions": lambda d: (
+            json_or_none(output_messages(d)) if d.choices_out else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+        }

litellm/integrations/otel/mappers/legacy.py

@@ -0,0 +1,97 @@
+"""Mapper for the older semantic-convention attribute vocabulary.
+
+Emits attributes under the semconv-ai / Traceloop key names (e.g.
+``gen_ai.system``, ``gen_ai.usage.prompt_tokens``, ``llm.is_streaming``) plus a
+few bare, unprefixed service keys (``service``, ``call_type``, ``error``), for
+backends that consume those names.
+
+Like ``GenAIMapper``, each span kind declares its schema as a flat
+``attribute key -> extractor`` table: one lambda per mapping operation.
+"""
+
+from typing import Callable, Final
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import collect, drop_none
+from litellm.integrations.otel.model.payloads import (
+    LLMCallSpanData,
+    ServiceSpanData,
+    ToolDefinition,
+)
+
+# Attribute keys in the semconv-ai / Traceloop vocabulary.
+_LEGACY_SYSTEM: Final = "gen_ai.system"
+_LEGACY_PROMPT_TOKENS: Final = "gen_ai.usage.prompt_tokens"
+_LEGACY_COMPLETION_TOKENS: Final = "gen_ai.usage.completion_tokens"
+_LEGACY_TOTAL_TOKENS: Final = "gen_ai.usage.total_tokens"
+_LEGACY_IS_STREAMING: Final = "llm.is_streaming"
+_LEGACY_TOP_K: Final = "llm.top_k"
+_LEGACY_FREQUENCY_PENALTY: Final = "llm.frequency_penalty"
+_LEGACY_PRESENCE_PENALTY: Final = "llm.presence_penalty"
+_LEGACY_STOP_SEQUENCES: Final = "llm.chat.stop_sequences"
+_LEGACY_SERVICE: Final = "service"
+_LEGACY_CALL_TYPE: Final = "call_type"
+_LEGACY_ERROR: Final = "error"
+
+
+class LegacyMapper:
+    """Emits LLM-call and service attributes under the older key names."""
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        _LEGACY_SYSTEM: lambda d: d.provider or None,
+        _LEGACY_PROMPT_TOKENS: lambda d: d.usage.input_tokens,
+        _LEGACY_COMPLETION_TOKENS: lambda d: d.usage.output_tokens,
+        _LEGACY_TOTAL_TOKENS: lambda d: d.usage.total_tokens,
+        _LEGACY_IS_STREAMING: lambda d: d.is_streaming,
+        _LEGACY_TOP_K: lambda d: d.request_params.top_k,
+        _LEGACY_FREQUENCY_PENALTY: lambda d: d.request_params.frequency_penalty,
+        _LEGACY_PRESENCE_PENALTY: lambda d: d.request_params.presence_penalty,
+        _LEGACY_STOP_SEQUENCES: lambda d: (
+            list(d.request_params.stop_sequences)
+            if d.request_params.stop_sequences
+            else None
+        ),
+    }
+
+    _TOOL_ATTRS: dict[str, Callable[[ToolDefinition], AttrValue | None]] = {
+        "name": lambda t: t.name,
+        "description": lambda t: t.description or None,
+        "parameters": lambda t: t.parameters_json or None,
+    }
+
+    _SERVICE_ATTRS: dict[str, Callable[[ServiceSpanData], AttrValue | None]] = {
+        _LEGACY_SERVICE: lambda d: d.service_name,
+        _LEGACY_CALL_TYPE: lambda d: d.call_type,
+        _LEGACY_ERROR: lambda d: (
+            d.error.message if d.error is not None and d.error.message else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case ServiceSpanData():
+                return self._service(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        attrs = collect(cls._LLM_CALL_ATTRS, data)
+        attrs.update(
+            drop_none(
+                {
+                    f"llm.request.functions.{idx}.{suffix}": extract(tool)
+                    for idx, tool in enumerate(data.tools)
+                    for suffix, extract in cls._TOOL_ATTRS.items()
+                }
+            )
+        )
+        return attrs
+
+    @classmethod
+    def _service(cls, data: ServiceSpanData) -> AttributeMap:
+        attrs = collect(cls._SERVICE_ATTRS, data)
+        attrs.update(dict(data.event_metadata))
+        return attrs

litellm/integrations/otel/mappers/openinference.py

@@ -0,0 +1,128 @@
+"""OpenInference attribute mapper (Arize + Arize-Phoenix shared vocabulary).
+
+Spec: https://github.com/Arize-ai/openinference/tree/main/spec — the standard
+both Arize and Phoenix consume. Composing this mapper after ``GenAIMapper``
+gives the same span both vocabularies, so a single trace lights up Arize +
+Phoenix + any other OpenInference-aware backend simultaneously.
+"""
+
+import json
+from typing import Callable, Sequence
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import (
+    collect,
+    drop_none,
+    json_if,
+    message_content,
+    output_messages,
+)
+from litellm.integrations.otel.model.payloads import (
+    LLMCallSpanData,
+    LLMRequestParams,
+    ToolDefinition,
+)
+
+
+class OpenInferenceMapper:
+    """Emits OpenInference attributes for LLM_CALL spans.
+
+    Key families (per the OpenInference spec):
+    - ``openinference.span.kind`` — discriminator (``"LLM"`` here)
+    - ``llm.model_name`` / ``llm.provider`` / ``llm.invocation_parameters``
+    - ``llm.input_messages.{i}.message.role`` / ``...content``
+    - ``llm.output_messages.{i}.message.role`` / ``...content``
+    - ``llm.token_count.prompt`` / ``...completion`` / ``...total``
+    - ``input.value`` / ``output.value`` — JSON-serialized request / response
+    """
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "openinference.span.kind": lambda d: "LLM",
+        "llm.model_name": lambda d: d.request_model or None,
+        "llm.provider": lambda d: d.provider or None,
+        "llm.token_count.prompt": lambda d: d.usage.input_tokens,
+        "llm.token_count.completion": lambda d: d.usage.output_tokens,
+        "llm.token_count.total": lambda d: d.usage.total_tokens,
+    }
+
+    # Folded into the ``llm.invocation_parameters`` JSON blob.
+    _INVOCATION_PARAMS: dict[str, Callable[[LLMRequestParams], AttrValue | None]] = {
+        "temperature": lambda rp: rp.temperature,
+        "top_p": lambda rp: rp.top_p,
+        "top_k": lambda rp: rp.top_k,
+        "max_tokens": lambda rp: rp.max_tokens,
+        "frequency_penalty": lambda rp: rp.frequency_penalty,
+        "presence_penalty": lambda rp: rp.presence_penalty,
+        "seed": lambda rp: rp.seed,
+    }
+
+    # Per-tool extractors, keyed by the ``llm.tools.{idx}.*`` suffix.
+    _TOOL_ATTRS: dict[str, Callable[[ToolDefinition], AttrValue | None]] = {
+        "tool.name": lambda t: t.name,
+        "tool.description": lambda t: t.description or None,
+        "tool.json_schema": lambda t: t.parameters_json or None,
+    }
+
+    # JSON-payload attributes: each builder returns the serialized blob or None.
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "llm.invocation_parameters": lambda d: json_if(
+            collect(OpenInferenceMapper._INVOCATION_PARAMS, d.request_params)
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+            **cls._messages("llm.input_messages", "input.value", data.messages_in),
+            **cls._messages(
+                "llm.output_messages", "output.value", output_messages(data)
+            ),
+            **cls._tools(data),
+        }
+
+    @staticmethod
+    def _messages(
+        prefix: str, value_key: str, messages: Sequence[object]
+    ) -> AttributeMap:
+        """Per-message ``{prefix}.{idx}.message.*`` keys + the ``value_key`` blob."""
+        parsed = [
+            (m.get("role") if isinstance(m, dict) else None, message_content(m))
+            for m in messages
+        ]
+        attrs = drop_none(
+            {
+                key: value
+                for idx, (role, content) in enumerate(parsed)
+                for key, value in (
+                    (
+                        f"{prefix}.{idx}.message.role",
+                        role if isinstance(role, str) else None,
+                    ),
+                    (f"{prefix}.{idx}.message.content", content),
+                )
+            }
+        )
+        if parsed:
+            attrs[value_key] = json.dumps(
+                [{"role": role, "content": content} for role, content in parsed]
+            )
+        return attrs
+
+    @classmethod
+    def _tools(cls, data: LLMCallSpanData) -> AttributeMap:
+        return drop_none(
+            {
+                f"llm.tools.{idx}.{suffix}": extract(tool)
+                for idx, tool in enumerate(data.tools)
+                for suffix, extract in cls._TOOL_ATTRS.items()
+            }
+        )

litellm/integrations/otel/mappers/utils.py

@@ -0,0 +1,76 @@
+"""Shared helpers for the attribute mappers.
+
+Small, mapper-agnostic utilities — JSON serialization, message extraction, and
+extractor-table application — pulled out of the individual mapper modules so
+they live in one place.
+"""
+
+import json
+from typing import Callable, Mapping, Sequence
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue
+from litellm.integrations.otel.model.payloads import LLMCallSpanData
+
+
+def drop_none(values: Mapping[str, AttrValue | None]) -> AttributeMap:
+    """Return ``values`` with ``None``-valued entries removed."""
+    return {k: v for k, v in values.items() if v is not None}
+
+
+def collect(table: Mapping[str, Callable], source: object) -> AttributeMap:
+    """Apply an extractor table to ``source``, dropping ``None`` results."""
+    return drop_none({key: extract(source) for key, extract in table.items()})
+
+
+def json_if(payload: Mapping[str, object]) -> str | None:
+    """JSON-serialize ``payload`` only when it's non-empty; else ``None``."""
+    return json.dumps(payload) if payload else None
+
+
+def json_or_none(value: object) -> str | None:
+    """JSON-serialize ``value`` (falling back to ``str``); ``None`` on failure."""
+    try:
+        return json.dumps(value, default=str)
+    except Exception:
+        return None
+
+
+def stringify_message(message: object) -> str | None:
+    """JSON-serialize a chat message dict; ``None`` if not a dict or on failure."""
+    if not isinstance(message, dict):
+        return None
+    try:
+        return json.dumps(message, default=str)
+    except Exception:
+        return None
+
+
+def serialize_messages(messages: Sequence[object]) -> str | None:
+    """Round-trip a sequence of message dicts through ``stringify_message``."""
+    serialized = [
+        json.loads(s) for s in (stringify_message(m) for m in messages) if s is not None
+    ]
+    return json.dumps(serialized) if serialized else None
+
+
+def message_content(message: object) -> str | None:
+    """Extract the textual ``content`` from a chat message dict."""
+    if not isinstance(message, dict):
+        return None
+    content = message.get("content")
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        # multimodal: concatenate text parts only
+        parts = [
+            part.get("text", "")
+            for part in content
+            if isinstance(part, dict) and part.get("type") == "text"
+        ]
+        return "".join(p for p in parts if isinstance(p, str)) or None
+    return None
+
+
+def output_messages(data: LLMCallSpanData) -> list:
+    """The ``message`` payload of each response choice."""
+    return [c.get("message") for c in data.choices_out if isinstance(c, dict)]

litellm/integrations/otel/mappers/weave.py

@@ -0,0 +1,48 @@
+"""Weave (W&B) attribute mapper.
+
+Weave consumes OpenInference + a small set of Weave-specific keys (display
+name, thread id, output value). This mapper layers the latter on top of
+OpenInference's vocabulary — compose ``["genai", "openinference", "weave"]``
+to feed a Weave backend.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import collect, json_or_none
+from litellm.integrations.otel.model.payloads import LLMCallSpanData
+
+
+class WeaveMapper:
+    """Maps ``LLMCallSpanData`` to Weave's vendor attributes."""
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        # ``display_name`` has the form ``"{operation} {model}"``. The span
+        # name already covers that, but Weave reads this attribute too.
+        "weave.display_name": lambda d: (
+            f"{d.operation.value} {d.request_model}" if d.request_model else None
+        ),
+        "weave.call_id": lambda d: d.identity.call_id or None,
+    }
+
+    # JSON-payload attributes: each builder returns the serialized blob or None.
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        # Weave treats the response choices as the "output" payload.
+        "weave.output": lambda d: (
+            json_or_none(list(d.choices_out)) if d.choices_out else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+        }

litellm/integrations/otel/model/baggage.py

@@ -0,0 +1,76 @@
+"""Baggage promotion: request-identity values carried across child spans.
+
+A bounded set of identity values is written into OpenTelemetry Baggage on the
+LLM-call span so that child spans (guardrail, service) inherit them.
+``providers.LiteLLMBaggageSpanProcessor`` reads Baggage at span start and stamps
+the allowlisted keys onto every span.
+
+This module is the single place baggage is defined: ``_PROMOTABLE`` maps each
+promotable attribute key to how its value is read, and the two ``*_KEYS``
+defaults select what is promoted unless the config overrides them.
+"""
+
+from collections.abc import Callable
+from typing import Final
+
+from litellm.integrations.otel.model.metadata import RequestIdentity
+from litellm.integrations.otel.model.semconv import GenAI, LiteLLM
+
+# Attribute key -> value extractor over (identity, request_model). The single
+# definition of what may be promoted and under which key.
+_PROMOTABLE: Final[dict[str, Callable[[RequestIdentity, str | None], str | None]]] = {
+    LiteLLM.TEAM_ID: lambda identity, model: identity.team_id,
+    LiteLLM.TEAM_ALIAS: lambda identity, model: identity.team_alias,
+    LiteLLM.TEAM_METADATA: lambda identity, model: identity.team_metadata,
+    LiteLLM.KEY_HASH: lambda identity, model: identity.key_hash,
+    LiteLLM.END_USER: lambda identity, model: identity.end_user,
+    GenAI.REQUEST_MODEL: lambda identity, model: model,
+    LiteLLM.PROVIDER_MODEL: lambda identity, model: identity.provider_model,
+}
+
+# Keys promoted by default (a subset of ``_PROMOTABLE``). ``END_USER`` is
+# promotable but off by default — it identifies an individual user, so stamping
+# it onto every span is opt-in via ``config.baggage_promoted_keys``.
+BAGGAGE_PROMOTED_KEYS: Final[tuple[str, ...]] = (
+    LiteLLM.TEAM_ID,
+    LiteLLM.TEAM_ALIAS,
+    LiteLLM.TEAM_METADATA,
+    LiteLLM.KEY_HASH,
+    GenAI.REQUEST_MODEL,
+    LiteLLM.PROVIDER_MODEL,
+)
+
+# Metadata sub-keys eligible for promotion under the ``litellm.metadata.*``
+# namespace. The full metadata blob is never promoted; only this allowlist is.
+DEFAULT_BAGGAGE_METADATA_KEYS: Final[tuple[str, ...]] = (
+    "user_api_key_org_id",
+    "user_api_key_user_id",
+    "user_api_key_alias",
+    "user_api_key_end_user_id",
+    "requester_ip_address",
+)
+
+
+def promoted_baggage(
+    identity: RequestIdentity,
+    request_model: str | None,
+    promoted_keys: tuple[str, ...],
+    metadata_keys: tuple[str, ...] = DEFAULT_BAGGAGE_METADATA_KEYS,
+) -> dict[str, str]:
+    """Identity values to write into Baggage, filtered to ``promoted_keys``.
+
+    ``promoted_keys`` selects from ``_PROMOTABLE``; ``metadata_keys`` selects
+    sub-keys of ``identity.metadata`` to promote under ``litellm.metadata.*``.
+    Empty values are dropped.
+    """
+    out: dict[str, str] = {}
+    for key, extract in _PROMOTABLE.items():
+        if key in promoted_keys:
+            value = extract(identity, request_model)
+            if value:
+                out[key] = value
+    for meta_key in metadata_keys:
+        value = identity.metadata.get(meta_key)
+        if value:
+            out[f"{LiteLLM.METADATA_PREFIX}{meta_key}"] = value
+    return out

litellm/integrations/otel/model/config.py

@@ -0,0 +1,236 @@
+"""Typed configuration for the OpenTelemetry instrumentation."""
+
+from typing import Any, List
+
+from pydantic import AliasChoices, BaseModel, Field, field_validator, model_validator
+from pydantic_settings import BaseSettings, NoDecode, SettingsConfigDict
+from typing_extensions import Annotated
+
+from litellm.integrations.otel.model.baggage import (
+    BAGGAGE_PROMOTED_KEYS,
+    DEFAULT_BAGGAGE_METADATA_KEYS,
+)
+
+#: Master feature-flag env var. The logger is inert until this is truthy.
+OTEL_V2_ENV = "LITELLM_OTEL_V2"
+
+
+class CaptureMessageContent(str):
+    NO_CONTENT = "no_content"
+    SPAN_ONLY = "span_only"
+    EVENT_ONLY = "event_only"
+    SPAN_AND_EVENT = "span_and_event"
+
+
+class _OTelV2Flag(BaseSettings):
+    model_config = SettingsConfigDict(extra="ignore")
+
+    enabled: bool = Field(default=False, validation_alias=AliasChoices(OTEL_V2_ENV))
+
+
+def is_otel_v2_enabled() -> bool:
+    return _OTelV2Flag().enabled
+
+
+class ExporterSpec(BaseModel):
+    """One span-export destination.
+
+    The shared ``TracerProvider`` attaches one ``SpanProcessor`` per spec, so
+    listing several specs sends every span to all of them at once (e.g. Arize +
+    Phoenix + your own Honeycomb).
+    """
+
+    model_config = {"extra": "forbid"}
+
+    kind: str = Field(
+        default="console",
+        description="console | in_memory | otlp_http | otlp_grpc | <factory kind>",
+    )
+    endpoint: str | None = None
+    headers: str | None = None
+    options: dict[str, str] | None = Field(
+        default=None,
+        description=(
+            "Factory-specific configuration for a custom exporter ``kind`` "
+            "registered via ``providers.register_exporter_factory`` (e.g. an "
+            "API key a lazy-auth exporter fetches a token with). Ignored by the "
+            "built-in console/in_memory/otlp exporters."
+        ),
+    )
+    use_simple_processor: bool | None = Field(
+        default=None,
+        description=(
+            "Force SimpleSpanProcessor regardless of exporter kind. Default: "
+            "auto (Simple for console/in_memory, Batch otherwise)."
+        ),
+    )
+
+
+class OpenTelemetryV2Config(BaseSettings):
+    model_config = SettingsConfigDict(populate_by_name=True, extra="ignore")
+
+    # ----- single-destination shorthand, read from standard OTEL_* envs ----- #
+    exporter: str = Field(
+        default="console",
+        validation_alias=AliasChoices("OTEL_EXPORTER", "OTEL_EXPORTER_OTLP_PROTOCOL"),
+        description=(
+            "Exporter kind for the single-destination shorthand. The model "
+            "validator folds this (with ``endpoint`` / ``headers``) into a "
+            "one-entry ``exporters`` list when ``exporters`` is empty; set "
+            "``exporters`` directly for multiple destinations."
+        ),
+    )
+    endpoint: str | None = Field(
+        default=None,
+        validation_alias=AliasChoices("OTEL_ENDPOINT", "OTEL_EXPORTER_OTLP_ENDPOINT"),
+    )
+    headers: str | None = Field(
+        default=None,
+        validation_alias=AliasChoices("OTEL_HEADERS", "OTEL_EXPORTER_OTLP_HEADERS"),
+    )
+    service_name: str = Field(
+        default="litellm", validation_alias=AliasChoices("OTEL_SERVICE_NAME")
+    )
+    deployment_environment: str | None = Field(
+        default=None, validation_alias=AliasChoices("OTEL_ENVIRONMENT_NAME")
+    )
+
+    enable_metrics: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("LITELLM_OTEL_INTEGRATION_ENABLE_METRICS"),
+    )
+    enable_events: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("LITELLM_OTEL_INTEGRATION_ENABLE_EVENTS"),
+    )
+    capture_message_content: str = Field(
+        default=CaptureMessageContent.NO_CONTENT,
+        validation_alias=AliasChoices(
+            "OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT"
+        ),
+    )
+    legacy_compat: bool = Field(
+        default=True, validation_alias=AliasChoices("LITELLM_OTEL_LEGACY_COMPAT")
+    )
+
+    # ----- explicit multi-destination / vocabulary configuration ------------ #
+
+    exporters: list[ExporterSpec] = Field(
+        default_factory=list,
+        description=(
+            "One destination per spec. The shared TracerProvider attaches a "
+            "SpanProcessor per entry. When empty, the model validator folds "
+            "the ``exporter`` / ``endpoint`` / ``headers`` shorthand into a "
+            "single spec so there is always at least one destination."
+        ),
+    )
+
+    mapper_names: Annotated[List[str], NoDecode] = Field(
+        default_factory=lambda: ["genai"],
+        description=(
+            "Ordered attribute vocabularies to emit. ``genai`` is the "
+            "canonical OTel GenAI vocabulary and is always placed first. "
+            "Vendor names: ``openinference`` (Arize + Phoenix), ``langfuse``, "
+            "``weave``, ``langtrace``."
+        ),
+    )
+
+    resource_attributes: dict[str, str] = Field(
+        default_factory=dict,
+        description=(
+            "Extra Resource attributes beyond ``service.name`` and "
+            "``deployment.environment`` (e.g. integration-specific markers)."
+        ),
+    )
+
+    baggage_promoted_keys: Annotated[List[str], NoDecode] = Field(
+        default_factory=lambda: list(BAGGAGE_PROMOTED_KEYS),
+        validation_alias=AliasChoices(
+            "baggage_promoted_keys", "LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS"
+        ),
+        description=(
+            "Identity attribute keys written into Baggage and stamped on every "
+            "child span (e.g. ``litellm.team.id``). Configure via the "
+            "``LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS`` env var (comma-separated) or "
+            "``callback_settings.otel.baggage_promoted_keys`` in config.yaml (a "
+            "YAML list)."
+        ),
+    )
+    baggage_metadata_keys: Annotated[List[str], NoDecode] = Field(
+        default_factory=lambda: list(DEFAULT_BAGGAGE_METADATA_KEYS),
+        validation_alias=AliasChoices(
+            "baggage_metadata_keys", "LITELLM_OTEL_BAGGAGE_METADATA_KEYS"
+        ),
+        description=(
+            "Metadata sub-keys promoted under the ``litellm.metadata.*`` "
+            "namespace. Configure via the ``LITELLM_OTEL_BAGGAGE_METADATA_KEYS`` "
+            "env var (comma-separated) or "
+            "``callback_settings.otel.baggage_metadata_keys`` in config.yaml."
+        ),
+    )
+
+    @field_validator(
+        "baggage_promoted_keys",
+        "baggage_metadata_keys",
+        "mapper_names",
+        mode="before",
+    )
+    @classmethod
+    def _split_csv(cls, value: Any) -> Any:
+        """Accept a comma-separated string for list fields.
+
+        Env vars are strings, but these fields are lists. Pydantic-settings would
+        otherwise require JSON for a list env var; splitting on commas here lets
+        an operator write ``LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS=litellm.team.id,litellm.api_key.hash``.
+        YAML lists (from ``callback_settings.otel.*``) and real lists pass through
+        unchanged.
+        """
+        if isinstance(value, str):
+            return [item.strip() for item in value.split(",") if item.strip()]
+        return value
+
+    @model_validator(mode="after")
+    def _normalize(self) -> "OpenTelemetryV2Config":
+        # An endpoint with the default exporter kind implies OTLP/HTTP.
+        if self.endpoint and self.exporter == "console":
+            self.exporter = "otlp_http"
+        # When no explicit destinations are given, fold the single-destination
+        # shorthand into one spec so the provider always has a destination.
+        if not self.exporters:
+            self.exporters = [
+                ExporterSpec(
+                    kind=self.exporter,
+                    endpoint=self.endpoint,
+                    headers=self.headers,
+                )
+            ]
+        # Ensure ``genai`` is always present and first.
+        names = list(self.mapper_names)
+        if "genai" in names:
+            names = ["genai"] + [n for n in names if n != "genai"]
+        else:
+            names = ["genai"] + names
+        # When enabled, also emit attribute keys under their semconv-ai /
+        # Traceloop names via the ``legacy`` mapper. Append it at the tail so
+        # the canonical ``genai`` keys win on any conflict.
+        if self.legacy_compat and "legacy" not in names:
+            names.append("legacy")
+        self.mapper_names = names
+        return self
+
+    @property
+    def capture_span_content(self) -> bool:
+        """Whether prompt/response content may be stamped as span attributes.
+
+        Defaults off (``no_content``): an operator must opt in before message
+        bodies leave the process, so a user request can never force its prompt
+        or completion into the configured backend while capture is disabled.
+        """
+        return self.capture_message_content in (
+            CaptureMessageContent.SPAN_ONLY,
+            CaptureMessageContent.SPAN_AND_EVENT,
+        )
+
+    @classmethod
+    def from_env(cls) -> "OpenTelemetryV2Config":
+        return cls()

litellm/integrations/otel/model/metadata.py

@@ -0,0 +1,315 @@
+"""The single translation layer between a request's metadata and the spans.
+
+Every relevant field litellm exposes about a request — the user-facing model,
+the model actually dispatched to the provider, the deployment, and the caller's
+identity (team, key, end-user) — is parsed **once**, here, out of the
+``StandardLoggingPayload`` (or a ``UserAPIKeyAuth`` at the auth boundary). Span
+data, baggage promotion, and the mappers then read these typed fields instead of
+each digging into the raw ``metadata`` / ``hidden_params`` dicts.
+
+Two models live here because a request's identity is known *before* its model
+resolution is:
+
+* :class:`RequestIdentity` — team / key / end-user, seeded into Baggage at the
+  auth boundary (``from_user_api_key_auth``), before routing has picked a
+  deployment. ``provider_model`` is therefore absent from that early seed and is
+  only filled in from the payload once the call closes.
+* :class:`RequestContext` — the full picture available at close: the resolved
+  request vs. provider model split, plus the response model, model group, model
+  id, and api base, wrapping the :class:`RequestIdentity`.
+
+The request-vs-provider model split is the subtle part. On the proxy a caller
+asks for a *model group* (e.g. ``gpt-4o``) that routes to a concrete deployment
+(e.g. ``azure/my-deployment``); the two are distinct and both worth recording.
+``StandardLoggingPayload`` exposes them as:
+
+* ``model_group`` — the user-facing name the caller requested.
+* ``model`` — already reconstructed (see ``reconstruct_model_name``) to the name
+  litellm dispatched to the provider (the deployment, provider-prefixed).
+* ``hidden_params.litellm_model_name`` — a secondary source for the dispatched
+  model (populated only on some call paths, e.g. files).
+
+So ``gen_ai.request.model`` is the *group* (falling back to the call model on the
+SDK path, which has no group), and ``litellm.provider.model`` is the *dispatched*
+model. They coincide on the SDK path, which is correct.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any, Mapping, cast
+
+from litellm.constants import LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL
+from litellm.integrations.otel.model.semconv import resolve_operation
+from litellm.integrations.otel.model.utils import as_str
+
+if TYPE_CHECKING:
+    from litellm.types.utils import StandardLoggingPayload
+
+
+@dataclass(frozen=True)
+class RequestIdentity:
+    call_id: str | None = None
+    team_id: str | None = None
+    team_alias: str | None = None
+    # The team's free-form metadata dict, JSON-serialized (empty/missing -> None).
+    team_metadata: str | None = None
+    key_hash: str | None = None
+    end_user: str | None = None
+    # The model litellm dispatched to the provider. Only known once the call
+    # completes (routing has picked a deployment), so it's absent from the
+    # auth-time seed and filled only from the payload.
+    provider_model: str | None = None
+    metadata: Mapping[str, str] = field(default_factory=dict)
+
+    @classmethod
+    def from_payload(cls, payload: "StandardLoggingPayload") -> "RequestIdentity":
+        """Parse caller identity out of a closed request's payload metadata.
+
+        ``provider_model`` is resolved here too (see :func:`resolve_provider_model`)
+        so the identity carried into Baggage labels every span with the dispatched
+        model, not just the user-facing one.
+        """
+        raw_meta = cast(Mapping[str, object], payload.get("metadata") or {})
+        metadata = {
+            key: str(value)
+            for key, value in raw_meta.items()
+            if isinstance(value, (str, bool, int, float))
+        }
+        return cls(
+            call_id=as_str(payload.get("litellm_call_id")) or as_str(payload.get("id")),
+            # StandardLoggingMetadata's canonical key is ``user_api_key_team_id``;
+            # the bare ``team_id`` is a legacy alias and is often empty, so prefer
+            # the canonical key and fall back to the alias.
+            team_id=as_str(raw_meta.get("user_api_key_team_id"))
+            or as_str(raw_meta.get("team_id")),
+            team_alias=as_str(raw_meta.get("user_api_key_team_alias"))
+            or as_str(raw_meta.get("team_alias")),
+            team_metadata=_team_metadata_json(
+                raw_meta.get("user_api_key_team_metadata")
+            ),
+            key_hash=as_str(raw_meta.get("user_api_key_hash")),
+            end_user=as_str(payload.get("end_user"))
+            or as_str(raw_meta.get("user_api_key_end_user_id")),
+            provider_model=resolve_provider_model(payload),
+            metadata=metadata,
+        )
+
+    @classmethod
+    def from_user_api_key_auth(cls, auth: object) -> "RequestIdentity":
+        """Identity from a ``UserAPIKeyAuth`` (duck-typed to keep this module
+        free of a proxy import).
+
+        Used in the pre-call hook to seed Baggage early — before any LLM,
+        guardrail, or service span is created — so the whole request's spans
+        inherit identity, not just the LLM-call span. Metadata sub-keys use the
+        ``user_api_key_*`` names that ``baggage.DEFAULT_BAGGAGE_METADATA_KEYS``
+        promotes.
+        """
+        get = lambda name: getattr(auth, name, None)  # noqa: E731
+        metadata = {
+            meta_key: str(value)
+            for meta_key, attr in (
+                ("user_api_key_user_id", "user_id"),
+                ("user_api_key_org_id", "org_id"),
+                ("user_api_key_alias", "key_alias"),
+                ("user_api_key_end_user_id", "end_user_id"),
+            )
+            if (value := get(attr))
+        }
+        return cls(
+            team_id=as_str(get("team_id")),
+            team_alias=as_str(get("team_alias")),
+            team_metadata=_team_metadata_json(get("team_metadata")),
+            key_hash=as_str(get("api_key")),
+            end_user=as_str(get("end_user_id")),
+            # ``provider_model`` is unknown at the auth boundary — routing hasn't
+            # picked a deployment yet — so it's only populated from the payload.
+            metadata=metadata,
+        )
+
+
+@dataclass(frozen=True)
+class RequestContext:
+    """The fully-resolved view of a closed request, parsed once from the payload.
+
+    ``request_model`` is the user-facing requested model and ``provider_model``
+    (on :attr:`identity`) is the model litellm dispatched to the provider; the two
+    differ on the proxy (group vs. deployment) and coincide on the SDK path.
+    """
+
+    request_model: str
+    response_model: str | None
+    model_group: str | None
+    model_id: str | None
+    api_base: str | None
+    identity: RequestIdentity
+
+    @property
+    def provider_model(self) -> str | None:
+        """The dispatched-model name, carried on the identity for Baggage."""
+        return self.identity.provider_model
+
+    @classmethod
+    def from_standard_logging_payload(
+        cls, payload: "StandardLoggingPayload"
+    ) -> "RequestContext":
+        raw_meta = cast(Mapping[str, object], payload.get("metadata") or {})
+        hidden = cast(Mapping[str, object], payload.get("hidden_params") or {})
+        raw_response = payload.get("response")
+        response = cast(
+            Mapping[str, object], raw_response if isinstance(raw_response, dict) else {}
+        )
+        model_group = as_str(payload.get("model_group")) or as_str(
+            raw_meta.get("model_group")
+        )
+        return cls(
+            # The user asked for the group; fall back to the call model on the SDK
+            # path, which has no group. Empty string (never None) so the span name
+            # builder and the mapper see a plain string.
+            request_model=model_group or as_str(payload.get("model")) or "",
+            response_model=as_str(response.get("model")),
+            model_group=model_group,
+            model_id=as_str(payload.get("model_id"))
+            or _model_info_id(raw_meta.get("model_info")),
+            api_base=as_str(payload.get("api_base")) or as_str(hidden.get("api_base")),
+            identity=RequestIdentity.from_payload(payload),
+        )
+
+
+# --- live-callback kwargs parsing ------------------------------------------- #
+#
+# The model and helpers below parse the *live* callback ``kwargs`` god object (and
+# the raw pre/post-call ``data`` dicts) — the untyped request state that reaches a
+# ``CustomLogger`` before, or instead of, a ``StandardLoggingPayload``. They live
+# here, with the payload/auth parsers, so every read out of a request's raw dicts
+# is in one place rather than scattered across the ``CustomLogger``.
+
+
+@dataclass(frozen=True)
+class LLMCallEvent:
+    """The typed view of the live callback ``kwargs`` (``model_call_details``).
+
+    litellm hands every callback an untyped ``kwargs`` god object. The fields the
+    OTel logger needs out of it are parsed **once**, here, so the ``CustomLogger``
+    reads typed attributes instead of digging into the dict at each boundary.
+    """
+
+    # The ``litellm_call_id`` correlating ``pre_call`` with the close callback.
+    # Present in ``model_call_details`` at ``pre_call`` and in both the kwargs and
+    # the ``standard_logging_object`` at success/failure, so it's a stable key for
+    # the open-call carrier — no back-reference to the logging object required (the
+    # object isn't reachable from the callback kwargs at ``pre_call`` time).
+    call_id: str | None
+    # The ``StandardLoggingPayload`` carried on a success/failure callback; ``None``
+    # at ``pre_call``, or when the call closed before any payload materialized (so
+    # there is nothing to stamp on the span).
+    payload: "StandardLoggingPayload | None"
+    # The ``standard_callback_dynamic_params`` routing the call to a per-tenant
+    # tracer (its own exporter/endpoint), or ``None`` when the call isn't scoped.
+    dynamic_params: Any
+    # True for synthetic proxy-gate logs (auth / rate-limit rejections): they fire
+    # the ``pre_call`` hook but never made an upstream call, so they get no span.
+    is_no_upstream_call: bool
+    # A best-effort ``"{operation} {model}"`` name known at ``pre_call`` time. The
+    # span is renamed from the typed payload at close (``finish_span``); this only
+    # needs to be reasonable for a span that never gets closed (a leak).
+    provisional_span_name: str
+
+    @classmethod
+    def from_dict(cls, kwargs: Mapping[str, Any]) -> "LLMCallEvent":
+        raw_payload = kwargs.get("standard_logging_object")
+        payload = cast("StandardLoggingPayload", raw_payload) if raw_payload else None
+        operation = resolve_operation(as_str(kwargs.get("call_type")))
+        model = as_str(kwargs.get("model")) or ""
+        return cls(
+            call_id=_call_id(payload, kwargs),
+            payload=payload,
+            dynamic_params=kwargs.get("standard_callback_dynamic_params"),
+            is_no_upstream_call=bool(kwargs.get(LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL)),
+            provisional_span_name=f"{operation.value} {model}".strip(),
+        )
+
+
+def _call_id(
+    payload: "StandardLoggingPayload | None", kwargs: Mapping[str, Any]
+) -> str | None:
+    """The call id from the payload (when closed) or the bare kwargs (at pre_call)."""
+    if payload is not None:
+        call_id = as_str(payload.get("litellm_call_id")) or as_str(payload.get("id"))
+        if call_id:
+            return call_id
+    return as_str(kwargs.get("litellm_call_id"))
+
+
+def model_from_request_data(data: object) -> str | None:
+    """The user-facing ``model`` from a pre-call ``data`` dict (``None`` if absent).
+
+    Read at the auth boundary to label early Baggage before routing has resolved
+    a deployment; ``data`` is duck-typed since it arrives untyped from the proxy.
+    """
+    if isinstance(data, Mapping):
+        return as_str(data.get("model"))
+    return None
+
+
+def guardrail_entries_from_request_data(
+    request_data: Mapping[str, Any],
+) -> list[dict]:
+    """The guardrail-information dicts buried in ``metadata`` of a post-call dict.
+
+    ``standard_logging_guardrail_information`` is stored as either a single dict
+    or a list of them; normalize to a list of dicts (dropping non-dict noise) so
+    the caller just iterates. Empty list when none are present.
+    """
+    metadata = request_data.get("metadata")
+    if not isinstance(metadata, Mapping):
+        return []
+    info = metadata.get("standard_logging_guardrail_information")
+    if isinstance(info, Mapping):
+        return [cast(dict, info)]
+    if isinstance(info, list):
+        return [entry for entry in info if isinstance(entry, dict)]
+    return []
+
+
+def resolve_provider_model(payload: "StandardLoggingPayload") -> str | None:
+    """The model litellm dispatched to the provider, from the payload.
+
+    Prefers the explicit ``hidden_params.litellm_model_name`` (set on call paths
+    that know it, e.g. files), then the top-level ``model`` — which
+    ``reconstruct_model_name`` has already resolved to the deployment's
+    provider-prefixed name. Returns ``None`` only when neither is present.
+    """
+    raw_meta = cast(Mapping[str, object], payload.get("metadata") or {})
+    hidden = cast(Mapping[str, object], payload.get("hidden_params") or {})
+    return (
+        # ``deployment`` survives only on paths that don't strip it from metadata;
+        # harmless (and most precise) to prefer it when present.
+        as_str(raw_meta.get("deployment"))
+        or as_str(hidden.get("litellm_model_name"))
+        or as_str(payload.get("model"))
+    )
+
+
+def _model_info_id(model_info: object) -> str | None:
+    """The deployment id from a ``metadata.model_info`` sub-dict, if present."""
+    if isinstance(model_info, Mapping):
+        return as_str(model_info.get("id"))
+    return None
+
+
+def _team_metadata_json(value: object) -> str | None:
+    """JSON-serialize a team's metadata dict for a single Baggage value.
+
+    Returns ``None`` for a missing, non-dict, or empty mapping so the empty case
+    is dropped rather than promoting a useless ``"{}"``. Keys are sorted for a
+    stable, diff-friendly serialization.
+    """
+    if not isinstance(value, Mapping) or not value:
+        return None
+    try:
+        return json.dumps(value, default=str, sort_keys=True)
+    except Exception:
+        return None

litellm/integrations/otel/model/payloads.py

@@ -0,0 +1,468 @@
+"""Typed span-data inputs: frozen dataclasses the engine and mappers consume."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import TYPE_CHECKING, ClassVar, Mapping, cast
+from urllib.parse import urlsplit
+
+from litellm.integrations.otel.model.metadata import (
+    RequestContext,
+    RequestIdentity,
+)
+from litellm.integrations.otel.model.semconv import (
+    GenAIOperation,
+    resolve_operation,
+    resolve_provider,
+)
+from litellm.integrations.otel.model.utils import (
+    as_bool,
+    as_float,
+    as_int,
+    as_str,
+    as_str_tuple,
+)
+
+# ``RequestIdentity`` and the request-metadata translation now live in
+# :mod:`metadata`; re-exported here so existing ``model.payloads`` imports keep
+# resolving it.
+__all__ = [
+    "RequestContext",
+    "RequestIdentity",
+    "GuardrailSpanData",
+    "LLMCallSpanData",
+    "LLMRequestParams",
+    "LLMUsage",
+    "ProxyRequestSpanData",
+    "ServerInfo",
+    "ServiceSpanData",
+    "SpanError",
+    "ToolDefinition",
+]
+
+if TYPE_CHECKING:
+    from litellm.types.services import ServiceLoggerPayload
+    from litellm.types.utils import (
+        StandardLoggingGuardrailInformation,
+        StandardLoggingPayload,
+    )
+
+
+# --- typed sub-structures ---------------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class LLMRequestParams:
+    temperature: float | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    max_tokens: int | None = None
+    frequency_penalty: float | None = None
+    presence_penalty: float | None = None
+    stop_sequences: tuple[str, ...] | None = None
+    seed: int | None = None
+
+    @classmethod
+    def from_model_parameters(cls, params: Mapping[str, object]) -> "LLMRequestParams":
+        max_tokens = as_int(params.get("max_tokens"))
+        if max_tokens is None:
+            max_tokens = as_int(params.get("max_completion_tokens"))
+        return cls(
+            temperature=as_float(params.get("temperature")),
+            top_p=as_float(params.get("top_p")),
+            top_k=as_int(params.get("top_k")),
+            max_tokens=max_tokens,
+            frequency_penalty=as_float(params.get("frequency_penalty")),
+            presence_penalty=as_float(params.get("presence_penalty")),
+            stop_sequences=as_str_tuple(params.get("stop")),
+            seed=as_int(params.get("seed")),
+        )
+
+
+@dataclass(frozen=True)
+class LLMUsage:
+    input_tokens: int | None = None
+    output_tokens: int | None = None
+    total_tokens: int | None = None
+
+
+@dataclass(frozen=True)
+class SpanError:
+    error_type: str | None = None
+    message: str | None = None
+
+
+@dataclass(frozen=True)
+class ServerInfo:
+    address: str | None = None
+    port: int | None = None
+
+    @classmethod
+    def from_api_base(cls, api_base: str | None) -> ServerInfo | None:
+        if not api_base:
+            return None
+        parsed = urlsplit(api_base if "://" in api_base else f"//{api_base}")
+        if not parsed.hostname:
+            return None
+        return cls(address=parsed.hostname, port=parsed.port)
+
+
+@dataclass(frozen=True)
+class GuardrailSpanData:
+    guardrail_name: str
+    mode: str | None = None
+    status: str | None = None
+    masked_entity_count: int | None = None
+    provider: str | None = None
+    action: str | None = None
+    # The guardrail verdict / provider response (e.g. the moderation result),
+    # JSON-serialized. This is the detail that belongs on the guardrail span.
+    response_json: str | None = None
+    violation_categories: tuple[str, ...] = ()
+    confidence_score: float | None = None
+    risk_score: float | None = None
+    duration: float | None = None
+    # Actual execution window (epoch seconds) from the logging entry, so the span
+    # is placed when the guardrail really ran — a pre_call guardrail before the
+    # LLM call — rather than at post-call emission time.
+    start_time: float | None = None
+    end_time: float | None = None
+    # Provider-agnostic configuration/detection metadata (see
+    # ``StandardLoggingGuardrailInformation``). Present for any guardrail that
+    # populates them, not just one provider's shape.
+    guardrail_id: str | None = None
+    policy_template: str | None = None
+    detection_method: str | None = None
+    # Set when the guardrail intervened/blocked or failed, so the emitter marks
+    # the span ERROR — a blocking guardrail is an error outcome for that span.
+    error: SpanError | None = None
+
+    # Guardrail statuses that mean the guardrail did not pass the request through.
+    _ERROR_STATUSES: ClassVar[frozenset[str]] = frozenset(
+        {"guardrail_intervened", "guardrail_failed_to_respond"}
+    )
+
+    @classmethod
+    def from_logging_entry(
+        cls, entry: "StandardLoggingGuardrailInformation"
+    ) -> "GuardrailSpanData":
+        """Build from one ``standard_logging_guardrail_information`` entry.
+
+        Reads the canonical, provider-agnostic ``StandardLoggingGuardrailInformation``
+        keys only — no guessing at a single provider's field names. Values that are
+        typed as enums or lists (e.g. ``guardrail_mode``) are normalized to a
+        stable string rather than assumed to already be plain strings.
+        """
+        get = cast(Mapping[str, object], entry).get
+        status = as_str(get("guardrail_status"))
+        response = get("guardrail_response")
+        error = (
+            SpanError(error_type=status, message=as_str(get("guardrail_action")))
+            if status in cls._ERROR_STATUSES
+            else None
+        )
+        return cls(
+            guardrail_name=as_str(get("guardrail_name")) or "guardrail",
+            mode=_guardrail_mode_str(get("guardrail_mode")),
+            status=status,
+            masked_entity_count=_total_masked_entities(get("masked_entity_count")),
+            provider=as_str(get("guardrail_provider")),
+            action=as_str(get("guardrail_action")),
+            response_json=_json_or_none(response) if response is not None else None,
+            violation_categories=as_str_tuple(get("violation_categories")) or (),
+            confidence_score=as_float(get("confidence_score")),
+            risk_score=as_float(get("risk_score")),
+            duration=as_float(get("duration")),
+            start_time=as_float(get("start_time")),
+            end_time=as_float(get("end_time")),
+            guardrail_id=as_str(get("guardrail_id")),
+            policy_template=as_str(get("policy_template")),
+            detection_method=as_str(get("detection_method")),
+            error=error,
+        )
+
+
+@dataclass(frozen=True)
+class ServiceSpanData:
+    service_name: str
+    call_type: str | None = None
+    error: SpanError | None = None
+    # Caller-supplied attributes to stamp on the service span, passed through
+    # from ``async_service_*_hook(event_metadata=...)``. The mapper owns how
+    # these are namespaced: the canonical vocabulary uses ``litellm.metadata.*``
+    # keys, the semconv-ai / Traceloop vocabulary uses the bare key names.
+    event_metadata: Mapping[str, str] = field(default_factory=dict)
+
+    @classmethod
+    def from_payload(
+        cls,
+        payload: "ServiceLoggerPayload",
+        event_metadata: Mapping[str, object] | None = None,
+    ) -> "ServiceSpanData":
+        # ``payload.service`` is a ``ServiceTypes(str, Enum)`` and ``error`` is
+        # ``Optional[str]`` on the Pydantic model — no defensive reads needed.
+        # ``event_metadata`` is sanitized: the legacy service decorators pass raw
+        # call-site data (live objects, full request metadata, response headers),
+        # none of which belongs on a span.
+        return cls(
+            service_name=payload.service.value,
+            call_type=payload.call_type,
+            error=SpanError(message=payload.error) if payload.error else None,
+            event_metadata=sanitize_event_metadata(event_metadata),
+        )
+
+
+@dataclass(frozen=True)
+class ProxyRequestSpanData:
+    http_method: str
+    route: str
+    url_path: str | None = None
+    status_code: int | None = None
+    identity: RequestIdentity | None = None
+
+
+# --- the primary LLM-call model ---------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class ToolDefinition:
+    """A single function/tool declared on a chat-completion request."""
+
+    name: str
+    description: str | None = None
+    parameters_json: str | None = (
+        None  # JSON-serialized schema (str so it's an AttrValue)
+    )
+
+
+@dataclass(frozen=True)
+class LLMCallSpanData:
+    operation: GenAIOperation
+    provider: str
+    request_model: str
+    response_model: str | None
+    response_id: str | None
+    request_params: LLMRequestParams
+    usage: LLMUsage
+    finish_reasons: tuple[str, ...]
+    error: SpanError | None
+    response_cost: float | None
+    server: ServerInfo | None
+    identity: RequestIdentity
+    is_streaming: bool | None = None
+    tools: tuple[ToolDefinition, ...] = ()
+    # Raw messages and response, needed by vendor mappers (OpenInference,
+    # Langfuse, Weave) that stamp message-level attributes. ``messages_in`` is
+    # the request payload; ``choices_out`` mirrors ``response.choices`` from
+    # the StandardLoggingPayload. Both are tuples of immutable mappings so the
+    # dataclass stays hashable and frozen.
+    messages_in: tuple[Mapping[str, object], ...] = ()
+    choices_out: tuple[Mapping[str, object], ...] = ()
+    system_fingerprint: str | None = None
+
+    @classmethod
+    def from_standard_logging_payload(
+        cls, payload: "StandardLoggingPayload", capture_content: bool = False
+    ) -> "LLMCallSpanData":
+        params = cast(Mapping[str, object], payload.get("model_parameters") or {})
+        # The single parse of the request's metadata — the request-vs-provider
+        # model split, the response model, api base, and identity all come from
+        # here rather than being re-derived from the raw payload dicts.
+        context = RequestContext.from_standard_logging_payload(payload)
+        # Normalize ``response`` to a dict once so the content/id reads below are a
+        # plain ``.get`` — no repeated ``isinstance`` guards.
+        raw_response = payload.get("response")
+        response = cast(
+            Mapping[str, object], raw_response if isinstance(raw_response, dict) else {}
+        )
+        choices_out = _dicts(response.get("choices"))
+        # ``finish_reasons`` is metadata, not content, so derive it from
+        # ``choices_out`` before gating. The raw message/choice bodies are only
+        # retained when content capture is enabled (see ``capture_span_content``);
+        # otherwise the content-bearing mappers receive empty sequences and emit
+        # no prompt/response text.
+        finish_reasons = _finish_reasons(choices_out)
+        return cls(
+            operation=resolve_operation(as_str(payload.get("call_type"))),
+            provider=resolve_provider(as_str(payload.get("custom_llm_provider"))),
+            request_model=context.request_model,
+            response_model=context.response_model,
+            response_id=as_str(response.get("id")),
+            request_params=LLMRequestParams.from_model_parameters(params),
+            usage=LLMUsage(
+                input_tokens=as_int(payload.get("prompt_tokens")),
+                output_tokens=as_int(payload.get("completion_tokens")),
+                total_tokens=as_int(payload.get("total_tokens")),
+            ),
+            finish_reasons=finish_reasons,
+            error=_parse_error(payload),
+            response_cost=as_float(payload.get("response_cost")),
+            server=ServerInfo.from_api_base(context.api_base),
+            identity=context.identity,
+            is_streaming=as_bool(payload.get("stream")),
+            tools=_extract_tools(params),
+            messages_in=_dicts(payload.get("messages")) if capture_content else (),
+            choices_out=choices_out if capture_content else (),
+            system_fingerprint=as_str(response.get("system_fingerprint")),
+        )
+
+
+# --- service event_metadata sanitization ------------------------------------ #
+
+# Substrings (case-insensitive) of keys that must never reach a span: secrets,
+# tokens, and raw request/response dumps the legacy service decorators pass.
+_SENSITIVE_METADATA_SUBSTRINGS: tuple[str, ...] = (
+    "api_key",
+    "token",
+    "secret",
+    "password",
+    "cookie",
+    "authorization",
+    "header",
+    "hidden_params",
+)
+# Keys that carry raw call-site internals — live objects, full kwargs/args. The
+# operation name is already the span's ``call_type``, so ``function_name`` is
+# redundant.
+_DROP_METADATA_KEYS: frozenset = frozenset(
+    {"function_kwargs", "function_args", "function_name"}
+)
+_MAX_METADATA_VALUE_LEN = 1024
+_MAX_METADATA_ITEMS = 32
+
+
+def sanitize_event_metadata(
+    event_metadata: Mapping[str, object] | None,
+) -> dict[str, str]:
+    """Reduce caller-supplied ``event_metadata`` to span-safe string attributes.
+
+    Keeps only primitive values (str/int/float/bool) under non-sensitive keys —
+    never ``repr()``-ing objects, dicts, or lists, never stamping secrets/headers,
+    and bounding the count and per-value length. This is the single chokepoint:
+    both the GenAI and legacy mappers read the cleaned result.
+    """
+    if not event_metadata:
+        return {}
+    clean: dict[str, str] = {}
+    for key, value in event_metadata.items():
+        if len(clean) >= _MAX_METADATA_ITEMS:
+            break
+        if not isinstance(key, str) or key in _DROP_METADATA_KEYS:
+            continue
+        lowered = key.lower()
+        if any(token in lowered for token in _SENSITIVE_METADATA_SUBSTRINGS):
+            continue
+        # ``bool`` is a subclass of ``int``, so it's covered. Non-primitive values
+        # (objects, dicts, lists) are dropped rather than stringified.
+        if isinstance(value, (str, int, float)):
+            clean[key] = str(value)[:_MAX_METADATA_VALUE_LEN]
+    return clean
+
+
+def _json_or_none(value: object) -> str | None:
+    """JSON-serialize ``value`` (already-string values pass through). ``None`` on failure."""
+    if isinstance(value, str):
+        return value
+    try:
+        return json.dumps(value, default=str)
+    except Exception:
+        return None
+
+
+def _guardrail_mode_str(value: object) -> str | None:
+    """Normalize ``guardrail_mode`` to a stable string.
+
+    ``guardrail_mode`` is typed as a ``GuardrailEventHooks`` enum, a list of them,
+    or a ``GuardrailMode`` — not a plain string. Emit the enum *value* (e.g.
+    ``"pre_call"``) rather than ``str(enum)`` (``"GuardrailEventHooks.pre_call"``),
+    and join a list of modes so a guardrail that runs at multiple hooks is
+    represented faithfully.
+    """
+    if value is None:
+        return None
+    if isinstance(value, (list, tuple)):
+        parts: list[str] = []
+        for item in value:
+            if item is None:
+                continue
+            part = as_str(item.value) if isinstance(item, Enum) else as_str(item)
+            if part:
+                parts.append(part)
+        return ",".join(parts) or None
+    if isinstance(value, Enum):
+        return as_str(value.value)
+    return as_str(value)
+
+
+def _total_masked_entities(value: object) -> int | None:
+    """``masked_entity_count`` is a ``{entity_type: count}`` map — sum to a total."""
+    if isinstance(value, Mapping):
+        total = sum(v for v in value.values() if isinstance(v, int))
+        return total or None
+    return as_int(value)
+
+
+def _dicts(value: object) -> tuple[Mapping[str, object], ...]:
+    """The dict items of ``value`` (when it's a list), as a tuple. Else empty."""
+    if not isinstance(value, list):
+        return ()
+    return tuple(item for item in value if isinstance(item, dict))
+
+
+def _finish_reasons(choices: tuple[Mapping[str, object], ...]) -> tuple[str, ...]:
+    """Non-empty ``finish_reason`` of each response choice."""
+    return tuple(r for c in choices if (r := as_str(c.get("finish_reason"))))
+
+
+def _parse_error(payload: "StandardLoggingPayload") -> SpanError | None:
+    """A ``SpanError`` for a failed request, or ``None`` on success."""
+    if payload.get("status") != "failure":
+        return None
+    info = cast(Mapping[str, object], payload.get("error_information") or {})
+    return SpanError(
+        error_type=as_str(info.get("error_class")) or as_str(info.get("error_code")),
+        message=as_str(info.get("error_message")) or as_str(payload.get("error_str")),
+    )
+
+
+def _tool_from_entry(entry: object) -> ToolDefinition | None:
+    """One ``tools``/``functions`` entry → ``ToolDefinition``, or ``None`` if unusable."""
+    if not isinstance(entry, dict):
+        return None
+    fn = entry.get("function") if "function" in entry else entry
+    if not isinstance(fn, dict):
+        return None
+    name = as_str(fn.get("name"))
+    if not name:
+        return None
+    params = fn.get("parameters")
+    parameters_json: str | None = None
+    if params is not None:
+        try:
+            parameters_json = json.dumps(params, default=str)
+        except Exception:
+            parameters_json = None
+    return ToolDefinition(
+        name=name,
+        description=as_str(fn.get("description")),
+        parameters_json=parameters_json,
+    )
+
+
+def _extract_tools(
+    model_parameters: Mapping[str, object],
+) -> tuple[ToolDefinition, ...]:
+    """Pull declared tools from request params (OpenAI / Anthropic shape).
+
+    Accepts the chat-completion ``tools=[{"type":"function", "function":
+    {...}}, ...]`` shape, and falls back to the ``functions=[...]`` shape.
+    Returns an empty tuple when neither is present.
+    """
+    raw_tools = model_parameters.get("tools")
+    if not isinstance(raw_tools, list):
+        raw_tools = model_parameters.get("functions")  # ``functions`` shape
+    if not isinstance(raw_tools, list):
+        return ()
+    return tuple(t for entry in raw_tools if (t := _tool_from_entry(entry)) is not None)

litellm/integrations/otel/model/semconv.py

@@ -0,0 +1,201 @@
+"""
+Keys follow the OpenTelemetry GenAI semantic conventions (experimental). Anything
+without a semconv equivalent lives under the ``litellm.*`` vendor namespace.
+"""
+
+from enum import Enum
+from typing import Final
+
+
+class GenAIOperation(str, Enum):
+    """Values for ``gen_ai.operation.name``."""
+
+    CHAT = "chat"
+    TEXT_COMPLETION = "text_completion"
+    EMBEDDINGS = "embeddings"
+    GENERATE_CONTENT = "generate_content"
+    CREATE_AGENT = "create_agent"  # reserved for future agent spans
+    INVOKE_AGENT = "invoke_agent"  # reserved for future agent spans
+    EXECUTE_TOOL = "execute_tool"  # reserved for future tool spans
+
+
+class GenAIProvider(str, Enum):
+    """Common values for the ``gen_ai.provider.name`` attribute."""
+
+    OPENAI = "openai"
+    ANTHROPIC = "anthropic"
+    AWS_BEDROCK = "aws.bedrock"
+    AZURE_AI_OPENAI = "azure.ai.openai"
+    AZURE_AI_INFERENCE = "azure.ai.inference"
+    GCP_GEMINI = "gcp.gemini"
+    GCP_VERTEX_AI = "gcp.vertex_ai"
+    COHERE = "cohere"
+    MISTRAL_AI = "mistral_ai"
+    DEEPSEEK = "deepseek"
+    GROQ = "groq"
+    PERPLEXITY = "perplexity"
+    X_AI = "x_ai"
+    IBM_WATSONX_AI = "ibm.watsonx.ai"
+
+
+class GenAI:
+    """Canonical OTel GenAI span-attribute keys."""
+
+    # request
+    OPERATION_NAME: Final = "gen_ai.operation.name"
+    PROVIDER_NAME: Final = "gen_ai.provider.name"
+    REQUEST_MODEL: Final = "gen_ai.request.model"
+    REQUEST_TEMPERATURE: Final = "gen_ai.request.temperature"
+    REQUEST_TOP_P: Final = "gen_ai.request.top_p"
+    REQUEST_TOP_K: Final = "gen_ai.request.top_k"
+    REQUEST_MAX_TOKENS: Final = "gen_ai.request.max_tokens"
+    REQUEST_FREQUENCY_PENALTY: Final = "gen_ai.request.frequency_penalty"
+    REQUEST_PRESENCE_PENALTY: Final = "gen_ai.request.presence_penalty"
+    REQUEST_STOP_SEQUENCES: Final = "gen_ai.request.stop_sequences"
+    REQUEST_SEED: Final = "gen_ai.request.seed"
+    REQUEST_CHOICE_COUNT: Final = "gen_ai.request.choice.count"
+    REQUEST_ENCODING_FORMATS: Final = "gen_ai.request.encoding_formats"
+    # response
+    RESPONSE_ID: Final = "gen_ai.response.id"
+    RESPONSE_MODEL: Final = "gen_ai.response.model"
+    RESPONSE_FINISH_REASONS: Final = "gen_ai.response.finish_reasons"
+    # usage
+    USAGE_INPUT_TOKENS: Final = "gen_ai.usage.input_tokens"
+    USAGE_OUTPUT_TOKENS: Final = "gen_ai.usage.output_tokens"
+    # content (opt-in, gated by capture mode)
+    INPUT_MESSAGES: Final = "gen_ai.input.messages"
+    OUTPUT_MESSAGES: Final = "gen_ai.output.messages"
+    SYSTEM_INSTRUCTIONS: Final = "gen_ai.system_instructions"
+    OUTPUT_TYPE: Final = "gen_ai.output.type"
+    CONVERSATION_ID: Final = "gen_ai.conversation.id"
+    # agent / tool (reserved)
+    AGENT_ID: Final = "gen_ai.agent.id"
+    AGENT_NAME: Final = "gen_ai.agent.name"
+    TOOL_NAME: Final = "gen_ai.tool.name"
+    TOOL_CALL_ID: Final = "gen_ai.tool.call.id"
+
+
+class Error:
+    TYPE: Final = "error.type"
+
+
+class Server:
+    ADDRESS: Final = "server.address"
+    PORT: Final = "server.port"
+
+
+class DB:
+    """Database / cache client-span keys (OTel ``db.*`` semconv).
+
+    Stamped on ``DB_CALL`` spans (redis / postgres), which are CLIENT spans for
+    outbound datastore calls — not on the INTERNAL ``SERVICE`` spans.
+    """
+
+    SYSTEM_NAME: Final = "db.system.name"
+    OPERATION_NAME: Final = "db.operation.name"
+
+
+class HTTP:
+    """HTTP server-span keys. Belong on the SERVER span only (never promoted)."""
+
+    REQUEST_METHOD: Final = "http.request.method"
+    ROUTE: Final = "http.route"
+    RESPONSE_STATUS_CODE: Final = "http.response.status_code"
+    URL_PATH: Final = "url.path"
+
+
+class LiteLLM:
+    """Vendor-extension keys (no semconv equivalent). Always ``litellm.*``."""
+
+    CALL_ID: Final = "litellm.call_id"
+    COST_PREFIX: Final = "litellm.cost."
+    METADATA_PREFIX: Final = "litellm.metadata."
+    TEAM_ID: Final = "litellm.team.id"
+    TEAM_ALIAS: Final = "litellm.team.alias"
+    # The team's free-form metadata dict, JSON-serialized into a single value.
+    TEAM_METADATA: Final = "litellm.team.metadata"
+    KEY_HASH: Final = "litellm.api_key.hash"
+    END_USER: Final = "litellm.end_user.id"
+    # The model string litellm actually sent to the provider (the deployment's
+    # ``litellm_params.model``), distinct from the user-facing ``gen_ai.request.model``.
+    PROVIDER_MODEL: Final = "litellm.provider.model"
+    REQUEST_STREAMING: Final = "litellm.request.streaming"
+    GUARDRAIL_NAME: Final = "litellm.guardrail.name"
+    GUARDRAIL_MODE: Final = "litellm.guardrail.mode"
+    GUARDRAIL_STATUS: Final = "litellm.guardrail.status"
+    GUARDRAIL_PROVIDER: Final = "litellm.guardrail.provider"
+    GUARDRAIL_ACTION: Final = "litellm.guardrail.action"
+    GUARDRAIL_RESPONSE: Final = "litellm.guardrail.response"
+    GUARDRAIL_VIOLATION_CATEGORIES: Final = "litellm.guardrail.violation_categories"
+    GUARDRAIL_CONFIDENCE_SCORE: Final = "litellm.guardrail.confidence_score"
+    GUARDRAIL_RISK_SCORE: Final = "litellm.guardrail.risk_score"
+    GUARDRAIL_MASKED_ENTITY_COUNT: Final = "litellm.guardrail.masked_entity_count"
+    GUARDRAIL_DURATION: Final = "litellm.guardrail.duration"
+    GUARDRAIL_ID: Final = "litellm.guardrail.id"
+    GUARDRAIL_POLICY_TEMPLATE: Final = "litellm.guardrail.policy_template"
+    GUARDRAIL_DETECTION_METHOD: Final = "litellm.guardrail.detection_method"
+    SERVICE_NAME: Final = "litellm.service.name"
+    SERVICE_CALL_TYPE: Final = "litellm.service.call_type"
+    PREPROCESSING_MS: Final = "litellm.preprocessing.duration_ms"
+
+
+class Metric:
+    """GenAI metric instrument names."""
+
+    TOKEN_USAGE: Final = "gen_ai.client.token.usage"
+    OPERATION_DURATION: Final = "gen_ai.client.operation.duration"
+
+
+# litellm ``custom_llm_provider`` -> ``gen_ai.provider.name`` value.
+_PROVIDER_BY_LITELLM: dict[str, GenAIProvider] = {
+    "openai": GenAIProvider.OPENAI,
+    "text-completion-openai": GenAIProvider.OPENAI,
+    "azure": GenAIProvider.AZURE_AI_OPENAI,
+    "azure_ai": GenAIProvider.AZURE_AI_INFERENCE,
+    "anthropic": GenAIProvider.ANTHROPIC,
+    "bedrock": GenAIProvider.AWS_BEDROCK,
+    "bedrock_converse": GenAIProvider.AWS_BEDROCK,
+    "vertex_ai": GenAIProvider.GCP_VERTEX_AI,
+    "vertex_ai_beta": GenAIProvider.GCP_VERTEX_AI,
+    "gemini": GenAIProvider.GCP_GEMINI,
+    "cohere": GenAIProvider.COHERE,
+    "cohere_chat": GenAIProvider.COHERE,
+    "mistral": GenAIProvider.MISTRAL_AI,
+    "deepseek": GenAIProvider.DEEPSEEK,
+    "groq": GenAIProvider.GROQ,
+    "perplexity": GenAIProvider.PERPLEXITY,
+    "xai": GenAIProvider.X_AI,
+    "watsonx": GenAIProvider.IBM_WATSONX_AI,
+}
+
+# litellm ``call_type`` -> ``gen_ai.operation.name``.
+_OPERATION_BY_CALL_TYPE: dict[str, GenAIOperation] = {
+    "completion": GenAIOperation.CHAT,
+    "acompletion": GenAIOperation.CHAT,
+    "completion_with_retries": GenAIOperation.CHAT,
+    "text_completion": GenAIOperation.TEXT_COMPLETION,
+    "atext_completion": GenAIOperation.TEXT_COMPLETION,
+    "embedding": GenAIOperation.EMBEDDINGS,
+    "aembedding": GenAIOperation.EMBEDDINGS,
+    "responses": GenAIOperation.CHAT,
+    "aresponses": GenAIOperation.CHAT,
+}
+
+
+def resolve_provider(custom_llm_provider: str | None) -> str:
+    """Map a litellm provider string to a ``gen_ai.provider.name`` value.
+
+    Unknown providers pass through verbatim — the convention explicitly allows
+    provider-specific values, so an unmapped name is still valid.
+    """
+    if not custom_llm_provider:
+        return ""
+    mapped = _PROVIDER_BY_LITELLM.get(custom_llm_provider.lower())
+    return mapped.value if mapped is not None else custom_llm_provider
+
+
+def resolve_operation(call_type: str | None) -> GenAIOperation:
+    """Map a litellm ``call_type`` to a ``gen_ai.operation.name`` value."""
+    if not call_type:
+        return GenAIOperation.CHAT
+    return _OPERATION_BY_CALL_TYPE.get(call_type.lower(), GenAIOperation.CHAT)

litellm/integrations/otel/model/spans.py

@@ -0,0 +1,203 @@
+"""
+This module declares every span the instrumentation can emit and the hierarchy.
+
+Span-name patterns live here as typed builder functions.
+
+Canonical hierarchy::
+
+    PROXY_REQUEST  (SERVER, root)     # owned by the FastAPI instrumentor
+    ├── SERVICE    (INTERNAL)         # auth phase span (live; see logger.phase_span)
+    │   └── DB_CALL (CLIENT)          #   its key/user/team lookups nest here
+    ├── GUARDRAIL  (INTERNAL)         # request-lifecycle hook, sibling of LLM_CALL
+    ├── LLM_CALL   (CLIENT)
+    └── DB_CALL    (CLIENT)           # e.g. the spend-log write
+
+Guardrails parent to PROXY_REQUEST, not LLM_CALL: pre/during/post-call guardrail
+hooks are orchestrated by the request lifecycle (a pre-call guardrail runs
+before the LLM call even starts), so a guardrail is a sibling of the LLM call,
+not a child of it. The emitter parents every span to the ambient OTel context
+(the active server span), which matches this.
+
+Not every service call becomes a span — :func:`span_role_for_service` decides:
+
+- ``DB_CALL`` (CLIENT) — outbound datastores (redis, postgres,
+  ``batch_write_to_db``), carrying ``db.*`` semconv.
+- ``SERVICE`` (INTERNAL) — genuine internal work worth a span (background
+  budget/reset jobs, pod-lock manager).
+- ``None`` (metrics-only) — framework instrumentation that duplicates a gen-AI
+  span (``self`` = the ``track_llm_api_timing`` wrapper, ``router``,
+  ``proxy_pre_call``) or ``auth`` (which gets a live phase span instead). These
+  still feed Prometheus/Datadog; they just never enter the trace.
+
+``DB_CALL`` and ``SERVICE`` are built from the same ``ServiceSpanData``; only the
+role (hence span kind and attribute vocabulary) differs. A service call can fire
+outside any request (a background job), in which case it parents to no server
+span and starts its own root trace rather than being dropped.
+
+Management/admin endpoints are ordinary FastAPI routes — their SERVER spans are
+owned by the instrumentor too, so they don't appear as a role here.
+"""
+
+from dataclasses import dataclass
+from enum import Enum
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from litellm.integrations.otel.model.payloads import (
+        GuardrailSpanData,
+        LLMCallSpanData,
+        ProxyRequestSpanData,
+        ServiceSpanData,
+    )
+
+
+class SpanRole(str, Enum):
+    PROXY_REQUEST = "proxy_request"
+    LLM_CALL = "llm_call"
+    GUARDRAIL = "guardrail"
+    DB_CALL = "db_call"
+    SERVICE = "service"
+
+
+class LiteLLMSpanKind(str, Enum):
+    SERVER = "server"
+    CLIENT = "client"
+    INTERNAL = "internal"
+    PRODUCER = "producer"
+    CONSUMER = "consumer"
+
+
+@dataclass(frozen=True)
+class SpanSpec:
+    role: SpanRole
+    kind: LiteLLMSpanKind
+    parent: SpanRole | None
+
+
+SPAN_REGISTRY: dict[SpanRole, SpanSpec] = {
+    SpanRole.PROXY_REQUEST: SpanSpec(
+        SpanRole.PROXY_REQUEST, LiteLLMSpanKind.SERVER, parent=None
+    ),
+    SpanRole.LLM_CALL: SpanSpec(
+        SpanRole.LLM_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST
+    ),
+    SpanRole.GUARDRAIL: SpanSpec(
+        SpanRole.GUARDRAIL, LiteLLMSpanKind.INTERNAL, parent=SpanRole.PROXY_REQUEST
+    ),
+    SpanRole.DB_CALL: SpanSpec(
+        SpanRole.DB_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST
+    ),
+    SpanRole.SERVICE: SpanSpec(
+        SpanRole.SERVICE, LiteLLMSpanKind.INTERNAL, parent=SpanRole.PROXY_REQUEST
+    ),
+}
+
+
+# ``ServiceTypes`` value -> ``db.system.name``. These are outbound datastore
+# calls and become CLIENT ``DB_CALL`` spans; ``redis_``-prefixed names cover the
+# redis-backed spend queues. Any service not mapped here is litellm-internal work
+# and stays an INTERNAL ``SERVICE`` span. This table is the single source of
+# datastore knowledge — both the role classifier and the mapper read it.
+_DB_SYSTEM_BY_SERVICE: dict[str, str] = {
+    "redis": "redis",
+    "postgres": "postgresql",
+    "batch_write_to_db": "postgresql",
+}
+
+
+def db_system(service_name: str) -> str | None:
+    """The ``db.system.name`` for a datastore service, else ``None``.
+
+    ``None`` means the service is not an outbound datastore call. Redis-backed
+    spend queues (``redis_*``) map to ``redis``.
+    """
+    if service_name in _DB_SYSTEM_BY_SERVICE:
+        return _DB_SYSTEM_BY_SERVICE[service_name]
+    if service_name.startswith("redis_"):
+        return "redis"
+    return None
+
+
+# ``ServiceTypes`` values that are NOT emitted as spans — they are framework
+# instrumentation that either duplicates a gen-AI span or has a better home as a
+# Prometheus/Datadog metric. They still flow to those metric backends via their
+# own hooks; the v2 logger just does not put them in the trace:
+#
+#   - ``self``           — ``track_llm_api_timing`` wraps the LLM call; the
+#                          ``chat {model}`` CLIENT span already represents it.
+#   - ``router``         — wraps the whole request; duplicates the server span.
+#   - ``proxy_pre_call`` — per-callback pre-call timing; a guardrail's real span
+#                          is ``execute_guardrail {name}``.
+#   - ``auth``           — emitted instead as a live phase span (see
+#                          ``logger.phase_span``) so its DB lookups nest under it,
+#                          not as a flat post-hoc service span.
+_METRICS_ONLY_SERVICES: frozenset[str] = frozenset(
+    {"self", "router", "proxy_pre_call", "auth"}
+)
+
+
+def span_role_for_service(service_name: str) -> SpanRole | None:
+    """The span role for a service call, or ``None`` when it must not be a span.
+
+    ``DB_CALL`` for outbound datastores, ``SERVICE`` for genuine internal work
+    worth a span (background jobs), and ``None`` for framework instrumentation
+    that duplicates a gen-AI span or belongs in metrics only
+    (see ``_METRICS_ONLY_SERVICES``).
+    """
+    if service_name in _METRICS_ONLY_SERVICES:
+        return None
+    return SpanRole.DB_CALL if db_system(service_name) is not None else SpanRole.SERVICE
+
+
+# --- span name builders (the naming convention, per role) ------------------- #
+
+
+# The name the FastAPI instrumentor gives the root server span. V2 never creates
+# this span (the instrumentor owns it), but it anchors request-level spans to it
+# and tests assert against it by name, so the literal lives here with the rest of
+# the span vocabulary rather than being duplicated at each call site.
+LITELLM_PROXY_REQUEST_SPAN_NAME = "Received Proxy Server Request"
+
+
+def llm_call_span_name(data: "LLMCallSpanData") -> str:
+    """``"{operation} {model}"`` e.g. ``"chat gpt-4o"`` (GenAI semconv)."""
+    model = data.request_model or ""
+    return f"{data.operation.value} {model}".strip()
+
+
+def proxy_request_span_name(data: "ProxyRequestSpanData") -> str:
+    """``"{method} {route}"`` (HTTP semconv)."""
+    return f"{data.http_method} {data.route}".strip()
+
+
+def guardrail_span_name(data: "GuardrailSpanData") -> str:
+    return f"execute_guardrail {data.guardrail_name}".strip()
+
+
+def service_span_name(data: "ServiceSpanData") -> str:
+    """``"{service} {call_type}"`` e.g. ``"redis set"`` — service name alone when
+    no call type is known, so identically-named calls stay distinguishable."""
+    return f"{data.service_name} {data.call_type or ''}".strip()
+
+
+def root_roles() -> list[SpanRole]:
+    """Roles that start a new trace (no in-process parent)."""
+    return [role for role, spec in SPAN_REGISTRY.items() if spec.parent is None]
+
+
+def child_roles(parent: SpanRole) -> list[SpanRole]:
+    return [role for role, spec in SPAN_REGISTRY.items() if spec.parent == parent]
+
+
+def validate_registry(
+    registry: dict[SpanRole, SpanSpec] | None = None,
+) -> None:
+    reg = registry if registry is not None else SPAN_REGISTRY
+    for role, spec in reg.items():
+        if spec.role is not role:
+            raise ValueError(f"SPAN_REGISTRY[{role}] has mismatched role {spec.role}")
+        if spec.parent is not None and spec.parent not in reg:
+            raise ValueError(f"span role {role} declares unknown parent {spec.parent}")
+    missing = [role for role in SpanRole if role not in reg]
+    if missing:
+        raise ValueError(f"SPAN_REGISTRY is missing roles: {missing}")

litellm/integrations/otel/model/utils.py

@@ -0,0 +1,103 @@
+"""Shared, OpenTelemetry-free helpers for the otel integration.
+
+Generic value coercion (for reading heterogeneous logging-payload dicts), time
+conversion, and header parsing — pulled out of the individual modules so they
+live in one place. Deliberately free of any ``opentelemetry`` import so the
+OTel-free sources of truth (payloads, semconv, spans, config) can use it too.
+"""
+
+from datetime import datetime
+
+
+def as_str(value: object) -> str | None:
+    if value is None:
+        return None
+    if isinstance(value, str):
+        return value
+    return str(value)
+
+
+def as_int(value: object) -> int | None:
+    if isinstance(value, bool):
+        return int(value)
+    if isinstance(value, int):
+        return value
+    if isinstance(value, float):
+        return int(value)
+    if isinstance(value, str):
+        try:
+            return int(value)
+        except ValueError:
+            return None
+    return None
+
+
+def as_float(value: object) -> float | None:
+    if isinstance(value, bool):
+        return float(value)
+    if isinstance(value, (int, float)):
+        return float(value)
+    if isinstance(value, str):
+        try:
+            return float(value)
+        except ValueError:
+            return None
+    return None
+
+
+def as_bool(value: object) -> bool | None:
+    if value is None:
+        return None
+    if isinstance(value, bool):
+        return value
+    return bool(value)
+
+
+def as_str_tuple(value: object) -> tuple[str, ...] | None:
+    if value is None:
+        return None
+    if isinstance(value, str):
+        return (value,)
+    if isinstance(value, (list, tuple)):
+        return tuple(str(v) for v in value)
+    return None
+
+
+def to_ns(value: datetime | float | int | None) -> int | None:
+    """Coerce a datetime / epoch value to integer nanoseconds."""
+    if value is None:
+        return None
+    if isinstance(value, datetime):
+        return int(value.timestamp() * 1e9)
+    if isinstance(value, (int, float)) and not isinstance(value, bool):
+        return int(float(value) * 1e9)
+    return None
+
+
+def to_seconds(value: datetime | float | int | str | None) -> float | None:
+    """Coerce a datetime / epoch / formatted-string value to epoch seconds."""
+    if value is None:
+        return None
+    if isinstance(value, datetime):
+        return value.timestamp()
+    if isinstance(value, (int, float)) and not isinstance(value, bool):
+        return float(value)
+    if isinstance(value, str):
+        for fmt in ("%Y-%m-%d %H:%M:%S.%f", "%Y-%m-%d %H:%M:%S"):
+            try:
+                return datetime.strptime(value, fmt).timestamp()
+            except ValueError:
+                continue
+    return None
+
+
+def parse_headers(raw: str | None) -> dict[str, str]:
+    """Parse an OTLP ``"k=v,k=v"`` header string into a dict."""
+    headers: dict[str, str] = {}
+    if not raw:
+        return headers
+    for pair in raw.split(","):
+        if "=" in pair:
+            key, _, value = pair.partition("=")
+            headers[key.strip()] = value.strip()
+    return headers

litellm/integrations/otel/mount.py

@@ -0,0 +1,130 @@
+"""FastAPI server-span instrumentation — the proxy mounts this at app creation.
+
+``opentelemetry-instrumentation-fastapi`` creates the SERVER span for each HTTP
+route and extracts inbound ``traceparent`` headers. This module owns the one call
+site that attaches it to the proxy app, plus the passthrough span-naming hook, so
+``proxy_server`` stays free of OTel details.
+
+The ``FastAPIInstrumentor`` import is kept lazy (inside :func:`instrument_fastapi_app`,
+after the gate check) so importing this module never requires the optional
+``opentelemetry-instrumentation-fastapi`` package and pulls in nothing OTel-related
+when the feature gate is off.
+"""
+
+import os
+from typing import Any
+
+from litellm._logging import verbose_logger
+from litellm.integrations.otel.model.config import is_otel_v2_enabled
+
+# Routes excluded from server-span tracing by default: high-frequency pollers and
+# static UI/docs assets, none of which are LLM traffic. Entries are substring-matched
+# against the request path (unanchored, so they survive a ``server_root_path`` prefix
+# and each entry also covers everything beneath it — e.g. ``/health`` covers
+# ``/health/readiness``). Operators override the whole set via the standard
+# ``OTEL_PYTHON_FASTAPI_EXCLUDED_URLS`` env var (set "" to trace everything).
+_DEFAULT_EXCLUDED_ROUTES = (
+    "/health",  # load-balancer liveness/readiness polling
+    "/metrics",  # Prometheus scrape (also drops the /model/metrics admin analytics)
+    "/litellm-asset-prefix",  # hashed UI asset bundles
+    "/_next",  # Next.js static JS/CSS chunks (root-level mount)
+    "/ui",  # admin UI single-page app
+    "/swagger",  # static Swagger UI assets
+    "/docs",  # FastAPI Swagger docs page
+    "/redoc",  # FastAPI ReDoc docs page
+    "/openapi.json",  # OpenAPI schema
+    "favicon",  # /favicon.ico + /get_favicon
+    "/.well-known",  # UI config discovery
+)
+_DEFAULT_EXCLUDED_URLS = ",".join(_DEFAULT_EXCLUDED_ROUTES)
+
+# Passthrough routes are catch-alls (e.g. "/openai/{endpoint:path}"), so the
+# default OTel server-span name "{method} {route}" collapses every upstream
+# endpoint into "POST /openai/{endpoint:path}". The hook below renames those spans
+# to the real request path so each endpoint is distinguishable. Non-catch-all
+# routes keep their low-cardinality template name.
+PASSTHROUGH_PREFIXES = frozenset(
+    {
+        "openai",
+        "openai_passthrough",
+        "anthropic",
+        "azure",
+        "azure_ai",
+        "bedrock",
+        "cohere",
+        "cursor",
+        "gemini",
+        "mistral",
+        "vllm",
+        "vertex_ai",
+        "vertex-ai",
+        "assemblyai",
+        "eu.assemblyai",
+        "milvus",
+    }
+)
+
+
+def _passthrough_span_name_hook(span: Any, scope: dict) -> None:
+    """FastAPI ``server_request_hook``: give passthrough server spans a useful name.
+
+    The instrumentation matches the route at span creation, so both the span name
+    and ``http.route`` are set to the catch-all template (``/openai/{endpoint:path}``)
+    before this hook runs. Rewrite both to the real request path so each upstream
+    endpoint is distinguishable. (The ASGI ``http receive``/``http send`` sub-spans
+    can't be renamed from here — their name is captured at creation — so they are
+    dropped via ``exclude_spans`` at instrumentation time.)
+    """
+    try:
+        if span is None or not span.is_recording():
+            return
+        path = scope.get("path") or ""
+        method = scope.get("method") or ""
+        first_segment = path.lstrip("/").split("/", 1)[0]
+        if first_segment in PASSTHROUGH_PREFIXES:
+            span.update_name(f"{method} {path}".strip())
+            span.set_attribute("http.route", path)
+    except Exception:
+        pass
+
+
+def instrument_fastapi_app(app: Any) -> None:
+    """Attach OTel server-span instrumentation to the proxy FastAPI app.
+
+    Safe no-op when the V2 gate is off or ``opentelemetry-instrumentation-fastapi``
+    is unavailable. This MUST be called at app-creation time — once the lifespan
+    runs, the middleware stack is frozen and ``instrument_app`` raises "Cannot add
+    middleware after an application has started".
+
+    No ``TracerProvider`` is passed, so the instrumentation binds to the OTel global
+    ``ProxyTracerProvider``; the proxy publishes the real provider as the global
+    after config load (see ``proxy_startup_event``), and the proxy delegates to it.
+    That way server spans and gen-ai spans share one provider and the same trace.
+    """
+    try:
+        if not is_otel_v2_enabled():
+            return
+
+        # Lazy: only the V2-enabled path needs the optional
+        # ``opentelemetry-instrumentation-fastapi`` package, which is not part of the
+        # base ``litellm[proxy]`` install. Importing it at module top would make
+        # ``proxy_server``'s unconditional ``import`` of this module crash when the
+        # package is absent, even with the gate off.
+        from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor
+
+        excluded_urls = (
+            os.environ.get("OTEL_PYTHON_FASTAPI_EXCLUDED_URLS")
+            if "OTEL_PYTHON_FASTAPI_EXCLUDED_URLS" in os.environ
+            else _DEFAULT_EXCLUDED_URLS
+        )
+        FastAPIInstrumentor.instrument_app(
+            app,
+            excluded_urls=excluded_urls,
+            server_request_hook=_passthrough_span_name_hook,
+            # Drop the ASGI "http receive"/"http send" lifecycle sub-spans: they
+            # are low-value noise and (for passthrough) carry the catch-all route
+            # template in their name, which can't be rewritten from a hook.
+            exclude_spans=["receive", "send"],
+        )
+    except Exception as e:
+        verbose_logger.debug("Skipping OTel V2 FastAPI instrumentation: %s", e)

litellm/integrations/otel/plumbing/context.py

@@ -0,0 +1,127 @@
+"""Trace-context + Baggage helpers."""
+
+from contextvars import ContextVar
+from typing import Mapping
+
+from opentelemetry import baggage
+from opentelemetry.context import Context, get_current
+from opentelemetry.trace import Span, get_current_span, set_span_in_context
+from opentelemetry.trace.propagation.tracecontext import (
+    TraceContextTextMapPropagator,
+)
+
+_PROPAGATOR = TraceContextTextMapPropagator()
+
+# The request's root span — the FastAPI-owned SERVER span — captured ONCE when the
+# proxy first resolves it, so request-level spans (the LLM call, guardrails) can
+# parent to it EXPLICITLY instead of to whatever span happens to be active at the
+# instant they are emitted. Ambient-only parenting (``get_current_span()``) is
+# wrong at two boundaries:
+#   * inside the ``auth`` phase span the active span is the auth span, so an LLM /
+#     guardrail span emitted there would nest under auth instead of being its
+#     sibling; and
+#   * in a detached success task (pass-through logs success from a fire-and-forget
+#     ``asyncio.create_task``) the server span may not be active at all, orphaning
+#     the span into a brand-new trace.
+# A ``ContextVar`` (not a request attribute) so it rides the request task's context
+# and is inherited by ``asyncio.create_task`` children — i.e. the async logging
+# callbacks that close the span. It is never reset: the contextvar dies with the
+# request task, so there is nothing to leak.
+_request_root_span: "ContextVar[Span | None]" = ContextVar(
+    "litellm_otel_request_root_span", default=None
+)
+
+
+def set_request_root_span(span: Span) -> None:
+    """Anchor the request's root (server) span for explicit child parenting.
+
+    No-ops for a non-recordable span so a bad capture can never replace a good one
+    with a phantom parent. Idempotent — the proxy captures the same server span at
+    more than one entry point.
+    """
+    if is_recordable_span(span):
+        _request_root_span.set(span)
+
+
+def request_root_span() -> "Span | None":
+    """The anchored request root span, or ``None`` outside a proxy request."""
+    span = _request_root_span.get()
+    return span if is_recordable_span(span) else None
+
+
+def set_request_baggage(
+    values: Mapping[str, str], context: Context | None = None
+) -> Context:
+    """Return a context with ``values`` written into Baggage."""
+    ctx = context
+    for key, value in values.items():
+        ctx = baggage.set_baggage(key, value, context=ctx)
+    return ctx if ctx is not None else (context or get_current())
+
+
+def get_baggage_attributes(context: Context | None = None) -> dict[str, str]:
+    """All Baggage entries on ``context`` as strings."""
+    return {key: str(value) for key, value in baggage.get_all(context).items()}
+
+
+def context_from_span(span: Span, context: Context | None = None) -> Context:
+    """A context with ``span`` as the active span (for explicit parenting)."""
+    return set_span_in_context(span, context=context)
+
+
+def resolve_parent_context(threaded: Span | None = None) -> Context:
+    """The context a child span should parent under.
+
+    Ambient-first: parent to the active OTel context (the server span, restored
+    by the logging worker or active in the request task), falling back to a span
+    passed explicitly (``threaded``) only when the ambient context has no
+    recordable span — e.g. a background service call with no request on the
+    stack. When neither is recordable the ambient context is returned unchanged,
+    so the span starts a new root trace.
+
+    Only service/DB spans pass ``threaded`` (the ``parent_otel_span`` handed to
+    the service hook). Request-level spans — the LLM call and guardrails — are
+    created where the server span is genuinely ambient, so they never need it.
+    """
+    ctx = get_current()
+    if is_recordable_span(threaded) and not is_recordable_span(get_current_span(ctx)):
+        ctx = context_from_span(threaded, context=ctx)  # type: ignore[arg-type]
+    return ctx
+
+
+def resolve_request_span_context() -> Context:
+    """The parent context for a request-level span (the LLM call, a guardrail).
+
+    These are direct children of the request's root server span — siblings of the
+    ``auth`` phase span and of each other, never nested under whatever span is
+    momentarily active. So prefer the explicitly anchored root span; fall back to
+    ambient context only when there is no anchor (the SDK / no-proxy path), where
+    the span legitimately starts its own root trace.
+
+    Unlike :func:`resolve_parent_context` (used by DB/service spans, which DO want
+    to nest under the active phase span, e.g. an auth DB lookup under ``auth``),
+    this never returns the active span when an anchor exists.
+    """
+    root = request_root_span()
+    if root is not None:
+        return context_from_span(root)
+    return get_current()
+
+
+def is_recordable_span(obj: object) -> bool:
+    """True if ``obj`` is a live span with a valid context (safe to parent under)."""
+    if not isinstance(obj, Span):
+        return False
+    try:
+        ctx = obj.get_span_context()
+    except Exception:
+        return False
+    return ctx is not None and ctx.is_valid
+
+
+def extract_traceparent(headers: Mapping[str, str]) -> Context | None:
+    """Extract a remote parent context from incoming HTTP headers, if present."""
+    if not any(key.lower() == "traceparent" for key in headers):
+        return None
+    carrier = {str(key).lower(): value for key, value in headers.items()}
+    return _PROPAGATOR.extract(carrier)

litellm/integrations/otel/plumbing/metrics.py

@@ -0,0 +1,28 @@
+"""GenAI client metrics (token usage + operation duration histograms)."""
+
+from dataclasses import dataclass
+
+from opentelemetry.metrics import Histogram, Meter
+
+from litellm.integrations.otel.model.semconv import Metric
+
+
+@dataclass(frozen=True)
+class GenAIMetrics:
+    token_usage: Histogram
+    operation_duration: Histogram
+
+
+def create_genai_metrics(meter: Meter) -> GenAIMetrics:
+    return GenAIMetrics(
+        token_usage=meter.create_histogram(
+            name=Metric.TOKEN_USAGE,
+            unit="{token}",
+            description="Number of tokens used per GenAI request.",
+        ),
+        operation_duration=meter.create_histogram(
+            name=Metric.OPERATION_DURATION,
+            unit="s",
+            description="GenAI operation duration.",
+        ),
+    )

litellm/integrations/otel/plumbing/providers.py

@@ -0,0 +1,220 @@
+"""Provider / exporter factory + the Baggage span processor."""
+
+from typing import Callable, Iterable
+
+from opentelemetry import baggage
+from opentelemetry.context import Context
+from opentelemetry.sdk.resources import Resource
+from opentelemetry.sdk.trace import ReadableSpan, SpanProcessor, TracerProvider
+from opentelemetry.sdk.trace.export import (
+    BatchSpanProcessor,
+    ConsoleSpanExporter,
+    SimpleSpanProcessor,
+    SpanExporter,
+)
+from opentelemetry.sdk.trace.export.in_memory_span_exporter import (
+    InMemorySpanExporter,
+)
+from opentelemetry.trace import Span, SpanKind, Tracer
+
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.model.semconv import LiteLLM
+from litellm.integrations.otel.model.spans import LiteLLMSpanKind
+
+# Re-exported so ``providers.parse_headers`` remains a stable entry point.
+from litellm.integrations.otel.model.utils import parse_headers as parse_headers
+
+_SPAN_KIND_BY_ROLE_KIND: dict[LiteLLMSpanKind, SpanKind] = {
+    LiteLLMSpanKind.SERVER: SpanKind.SERVER,
+    LiteLLMSpanKind.CLIENT: SpanKind.CLIENT,
+    LiteLLMSpanKind.INTERNAL: SpanKind.INTERNAL,
+    LiteLLMSpanKind.PRODUCER: SpanKind.PRODUCER,
+    LiteLLMSpanKind.CONSUMER: SpanKind.CONSUMER,
+}
+
+
+def to_otel_span_kind(kind: LiteLLMSpanKind) -> SpanKind:
+    return _SPAN_KIND_BY_ROLE_KIND[kind]
+
+
+# Custom exporter factories keyed by ``ExporterSpec.kind``. A preset registers
+# one here when its destination needs construction logic the built-in kinds
+# can't express — e.g. an exporter that fetches an auth token lazily on its
+# first export (off the event loop) instead of blocking at config-build time.
+# Keeping the registry here lets this module stay vendor-agnostic: the factory
+# lives with the integration that needs it.
+_EXPORTER_FACTORIES: dict[str, Callable[[ExporterSpec], SpanExporter]] = {}
+
+
+def register_exporter_factory(
+    kind: str, factory: Callable[[ExporterSpec], SpanExporter]
+) -> None:
+    """Register a custom exporter ``factory`` for the exporter ``kind``."""
+    _EXPORTER_FACTORIES[kind.lower()] = factory
+
+
+class LiteLLMBaggageSpanProcessor(SpanProcessor):
+    """Stamps an allowlisted set of Baggage entries onto every span at start."""
+
+    def __init__(
+        self,
+        allowed_keys: Iterable[str],
+        allowed_prefixes: tuple[str, ...] = (LiteLLM.METADATA_PREFIX,),
+    ) -> None:
+        self._allowed_keys = frozenset(allowed_keys)
+        self._allowed_prefixes = tuple(allowed_prefixes)
+
+    def _is_allowed(self, key: str) -> bool:
+        return key in self._allowed_keys or any(
+            key.startswith(prefix) for prefix in self._allowed_prefixes
+        )
+
+    def on_start(self, span: Span, parent_context: Context | None = None) -> None:
+        for key, value in baggage.get_all(parent_context).items():
+            if self._is_allowed(key) and isinstance(value, (str, bool, int, float)):
+                span.set_attribute(key, value)
+
+    def on_end(self, span: ReadableSpan) -> None:  # noqa: D401 - no-op
+        return None
+
+    def shutdown(self) -> None:
+        return None
+
+    def force_flush(self, timeout_millis: int = 30000) -> bool:
+        return True
+
+
+def _otlp_traces_endpoint(endpoint: str | None) -> str | None:
+    """Point an OTLP/HTTP base endpoint at the ``/v1/traces`` signal path.
+
+    ``OTEL_EXPORTER_OTLP_ENDPOINT`` is a base URL (e.g. ``http://host:4318``).
+    The OTLP/HTTP exporter only appends the ``/v1/traces`` path when it reads
+    that env var itself; when an endpoint is passed explicitly it is used
+    verbatim, so a base URL would POST to the root and the collector returns
+    404. Append the signal path here (leaving an already-correct path intact).
+    """
+    if not endpoint:
+        return endpoint
+    endpoint = endpoint.rstrip("/")
+    # Splunk Observability uses ``/v2/trace/otlp``; never rewrite it.
+    if endpoint.endswith("/v1/traces") or "/v2/trace/otlp" in endpoint:
+        return endpoint
+    for other_signal in ("/v1/logs", "/v1/metrics"):
+        if endpoint.endswith(other_signal):
+            return endpoint[: -len(other_signal)] + "/v1/traces"
+    return endpoint + "/v1/traces"
+
+
+def _exporter_from_spec(spec: ExporterSpec) -> SpanExporter:
+    kind = (spec.kind or "console").lower()
+    factory = _EXPORTER_FACTORIES.get(kind)
+    if factory is not None:
+        return factory(spec)
+    if kind in ("in_memory", "inmemory", "memory"):
+        return InMemorySpanExporter()
+    if kind in ("otlp_http", "http", "http/protobuf", "http/json"):
+        from opentelemetry.exporter.otlp.proto.http.trace_exporter import (
+            OTLPSpanExporter as HTTPExporter,
+        )
+
+        return HTTPExporter(
+            endpoint=_otlp_traces_endpoint(spec.endpoint),
+            headers=parse_headers(spec.headers),
+        )
+    if kind in ("otlp_grpc", "grpc"):
+        from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import (
+            OTLPSpanExporter as GRPCExporter,
+        )
+
+        return GRPCExporter(endpoint=spec.endpoint, headers=parse_headers(spec.headers))
+    return ConsoleSpanExporter()
+
+
+def _processor_for(exporter: SpanExporter, use_simple: bool | None) -> SpanProcessor:
+    """Pick a Simple or Batch span processor for ``exporter``.
+
+    When ``use_simple`` is unset, default to Simple for console and in-memory
+    exporters (spans export synchronously, which tests rely on) and Batch for
+    everything else (the right export semantics for production).
+    """
+    if use_simple is None:
+        use_simple = isinstance(exporter, (ConsoleSpanExporter, InMemorySpanExporter))
+    return SimpleSpanProcessor(exporter) if use_simple else BatchSpanProcessor(exporter)
+
+
+def build_span_exporter(config: OpenTelemetryV2Config) -> SpanExporter:
+    """Build a single exporter from the top-level config fields.
+
+    Convenience for the common single-exporter case (and for tests): reads the
+    ``exporter`` / ``endpoint`` / ``headers`` fields. To configure multiple
+    exporters, populate ``config.exporters`` directly.
+    """
+    return _exporter_from_spec(
+        ExporterSpec(
+            kind=config.exporter, endpoint=config.endpoint, headers=config.headers
+        )
+    )
+
+
+def build_resource(config: OpenTelemetryV2Config) -> Resource:
+    attributes: dict[str, str] = {"service.name": config.service_name}
+    if config.deployment_environment:
+        attributes["deployment.environment"] = config.deployment_environment
+    attributes.update(config.resource_attributes)
+    return Resource.create(attributes)
+
+
+def build_tracer_provider(
+    config: OpenTelemetryV2Config,
+    exporter: SpanExporter | None = None,
+    baggage_processor: SpanProcessor | None = None,
+    use_simple_processor: bool | None = None,
+) -> TracerProvider:
+    """Build the shared :class:`TracerProvider`.
+
+    Attach the Baggage processor first (so identity attributes land on each
+    span before any export decision), then add one ``SpanProcessor`` per
+    ``config.exporters`` entry — this is what fans spans out to multiple
+    backends. ``exporter`` and ``use_simple_processor`` are explicit overrides:
+    pass a single exporter to attach exactly that one (used by tests).
+    """
+    provider = TracerProvider(resource=build_resource(config))
+    if baggage_processor is None:
+        baggage_processor = LiteLLMBaggageSpanProcessor(
+            allowed_keys=config.baggage_promoted_keys
+        )
+    provider.add_span_processor(baggage_processor)
+
+    if exporter is not None:
+        provider.add_span_processor(_processor_for(exporter, use_simple_processor))
+        return provider
+
+    # ``config._normalize`` guarantees at least one spec (it folds the top-level
+    # ``exporter``/``endpoint``/``headers`` fields in when ``exporters`` is empty).
+    for spec in config.exporters:
+        exp = _exporter_from_spec(spec)
+        provider.add_span_processor(
+            _processor_for(
+                exp,
+                (
+                    spec.use_simple_processor
+                    if spec.use_simple_processor is not None
+                    else use_simple_processor
+                ),
+            )
+        )
+    return provider
+
+
+def get_tracer(provider: TracerProvider, name: str = "litellm") -> Tracer:
+    return provider.get_tracer(name)
+
+
+def in_memory_provider(
+    config: OpenTelemetryV2Config | None = None,
+) -> tuple[TracerProvider, InMemorySpanExporter]:
+    """Convenience for tests: a provider exporting to an in-memory buffer."""
+    cfg = config or OpenTelemetryV2Config(exporter="in_memory")
+    exporter = InMemorySpanExporter()
+    provider = build_tracer_provider(cfg, exporter=exporter)
+    return provider, exporter

litellm/integrations/otel/plumbing/routing.py

@@ -0,0 +1,101 @@
+"""Per-request multi-tenant tracer routing.
+
+When a request carries team/key vendor credentials in
+``standard_callback_dynamic_params``, its spans must export through a
+``TracerProvider`` whose OTLP headers carry those credentials.
+``TenantTracerCache`` builds and caches one provider per distinct credential
+set, and otherwise hands back the logger's default tracer. This lets a single
+logger fan requests out to many tenants without needing a logger per tenant.
+"""
+
+from collections import OrderedDict
+from typing import Any, Mapping
+
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.trace import Tracer
+
+from litellm._logging import verbose_logger
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.presets import dynamic_otlp_headers
+from litellm.integrations.otel.plumbing.providers import (
+    build_tracer_provider,
+    get_tracer,
+)
+
+# Exporter kinds that ignore headers — never rewritten with dynamic credentials.
+_NON_OTLP_KINDS = ("console", "in_memory", "inmemory", "memory")
+
+# Cap on distinct credential-scoped providers held at once. ``dynamic_params``
+# can be populated from request metadata, so an unbounded cache lets a caller
+# spawn one ``TracerProvider`` (plus its ``BatchSpanProcessor`` background
+# thread) per unique credential set and exhaust the proxy. The LRU bound keeps
+# the working set of active tenants resident while flushing and shutting down
+# evicted providers so their threads are reclaimed.
+_MAX_CACHED_PROVIDERS = 256
+
+
+def _shutdown_provider(provider: TracerProvider) -> None:
+    """Flush + stop an evicted provider's processors (reclaims their threads).
+
+    ``TracerProvider.shutdown`` force-flushes each ``SpanProcessor`` before
+    stopping it, so any spans already handed to a ``BatchSpanProcessor`` are
+    exported rather than dropped. Best-effort: a shutdown failure must not break
+    the request that triggered the eviction.
+    """
+    try:
+        provider.shutdown()
+    except Exception as e:  # pragma: no cover - defensive
+        verbose_logger.debug("OTel V2: error shutting down evicted provider: %s", e)
+
+
+class TenantTracerCache:
+    """Credential-scoped ``TracerProvider`` cache keyed by the dynamic headers."""
+
+    def __init__(
+        self,
+        config: OpenTelemetryV2Config,
+        callback_name: str | None,
+        tracer_name: str,
+    ) -> None:
+        self._config = config
+        self._callback_name = callback_name
+        self._tracer_name = tracer_name
+        self._providers: "OrderedDict[tuple[tuple[str, str], ...], TracerProvider]" = (
+            OrderedDict()
+        )
+
+    def tracer_for(self, default: Tracer, dynamic_params: Any) -> Tracer:
+        """Return the tracer for this request.
+
+        Use ``default`` unless the request's dynamic credentials require a
+        credential-scoped tracer, in which case build (or reuse) one. The cache
+        is a bounded LRU: the least-recently-used provider is flushed and shut
+        down on overflow so its exporter threads don't accumulate.
+        """
+        headers = dynamic_otlp_headers(self._callback_name, dynamic_params)
+        if not headers:
+            return default
+        cache_key = tuple(sorted(headers.items()))
+        provider = self._providers.get(cache_key)
+        if provider is not None:
+            self._providers.move_to_end(cache_key)
+        else:
+            provider = build_tracer_provider(self._config_with_headers(headers))
+            self._providers[cache_key] = provider
+            if len(self._providers) > _MAX_CACHED_PROVIDERS:
+                _, evicted = self._providers.popitem(last=False)
+                _shutdown_provider(evicted)
+        return get_tracer(provider, self._tracer_name)
+
+    def _config_with_headers(self, headers: Mapping[str, str]) -> OpenTelemetryV2Config:
+        """Clone the config, replacing OTLP exporter headers with ``headers``."""
+        header_str = ",".join(f"{key}={value}" for key, value in headers.items())
+        exporters = [
+            (
+                spec
+                if spec.kind.lower() in _NON_OTLP_KINDS
+                else spec.model_copy(update={"headers": header_str})
+            )
+            for spec in self._config.exporters
+        ]
+        return self._config.model_copy(update={"exporters": exporters})

litellm/integrations/otel/presets/__init__.py

@@ -0,0 +1,78 @@
+"""Integration presets — each one returns an :class:`OpenTelemetryV2Config`.
+
+A preset is a callable that reads an integration's env vars and returns an
+``OpenTelemetryV2Config`` describing the exporter destination, the mapper
+vocabularies to apply, and any resource attributes. ``PRESET_BY_CALLBACK``
+maps a callback name (``"arize"``, ``"langfuse_otel"``, ...) to its preset so
+the factory in ``litellm_logging`` can resolve a name and build a single
+``OpenTelemetryV2`` instance from the result.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.presets.agentops import agentops_preset
+from litellm.integrations.otel.presets.arize import arize_dynamic_headers, arize_preset
+from litellm.integrations.otel.presets.base import Preset
+from litellm.integrations.otel.presets.langfuse import (
+    langfuse_dynamic_headers,
+    langfuse_preset,
+)
+from litellm.integrations.otel.presets.langtrace import langtrace_preset
+from litellm.integrations.otel.presets.levo import levo_preset
+from litellm.integrations.otel.presets.phoenix import phoenix_preset
+from litellm.integrations.otel.presets.weave import weave_dynamic_headers, weave_preset
+from litellm.types.utils import StandardCallbackDynamicParams
+
+#: Callback name → preset. The ``Preset`` annotation makes mypy verify every
+#: registered value matches the preset interface.
+PRESET_BY_CALLBACK: dict[str, Preset] = {
+    "agentops": agentops_preset,
+    "arize": arize_preset,
+    "arize_phoenix": phoenix_preset,
+    "langfuse_otel": langfuse_preset,
+    "langtrace": langtrace_preset,
+    "levo": levo_preset,
+    "weave_otel": weave_preset,
+}
+
+#: Callback name → per-request OTLP header builder (team/key multi-tenant
+#: routing). Only integrations that support dynamic credentials appear here —
+#: Arize-Phoenix/Langtrace/Levo/AgentOps don't, so they use the logger's
+#: default tracer.
+DYNAMIC_HEADERS_BY_CALLBACK: dict[
+    str, Callable[[StandardCallbackDynamicParams], dict[str, str]]
+] = {
+    "arize": arize_dynamic_headers,
+    "langfuse_otel": langfuse_dynamic_headers,
+    "weave_otel": weave_dynamic_headers,
+}
+
+
+def dynamic_otlp_headers(
+    callback_name: str | None,
+    dynamic_params: StandardCallbackDynamicParams | None,
+) -> dict[str, str] | None:
+    """Per-request OTLP headers for ``callback_name``, or ``None`` if N/A.
+
+    ``None`` means "no per-request routing" — the caller uses its default tracer.
+    """
+    builder = DYNAMIC_HEADERS_BY_CALLBACK.get(callback_name or "")
+    if builder is None or not dynamic_params:
+        return None
+    headers = builder(dynamic_params)
+    return headers or None
+
+
+__all__ = [
+    "PRESET_BY_CALLBACK",
+    "DYNAMIC_HEADERS_BY_CALLBACK",
+    "Preset",
+    "dynamic_otlp_headers",
+    "agentops_preset",
+    "arize_preset",
+    "langfuse_preset",
+    "langtrace_preset",
+    "levo_preset",
+    "phoenix_preset",
+    "weave_preset",
+]

litellm/integrations/otel/presets/agentops.py

@@ -0,0 +1,139 @@
+"""AgentOps preset — OTLP/HTTP to AgentOps' endpoint with a lazily-fetched JWT.
+
+AgentOps authenticates with a short-lived JWT minted from the API key. Fetching
+it is blocking network I/O, so it must never run on the event loop: callback
+construction (where presets are built) can run inside the proxy's async startup
+or, in the SDK, on the first request. Instead of fetching at config-build time,
+this preset registers a custom exporter (``kind="agentops"``) that mints the JWT
+**on its first export** — which the ``BatchSpanProcessor`` runs in its own
+worker thread, off any event loop — and caches it for the process lifetime.
+"""
+
+from typing import Any
+
+import httpx
+from pydantic import Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from litellm._logging import verbose_logger
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.plumbing.providers import register_exporter_factory
+
+_AGENTOPS_ENDPOINT = "https://otlp.agentops.cloud/v1/traces"
+_AGENTOPS_AUTH_ENDPOINT = "https://api.agentops.ai/v3/auth/token"
+_AGENTOPS_EXPORTER_KIND = "agentops"
+
+
+class _AgentOpsSettings(BaseSettings):
+    model_config = SettingsConfigDict(case_sensitive=False, extra="ignore")
+
+    api_key: str | None = Field(default=None, validation_alias="AGENTOPS_API_KEY")
+    service_name: str = Field(
+        default="agentops", validation_alias="AGENTOPS_SERVICE_NAME"
+    )
+    environment: str | None = Field(
+        default=None, validation_alias="AGENTOPS_ENVIRONMENT"
+    )
+
+
+def agentops_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    """Build the AgentOps config without any network I/O.
+
+    The ``agentops`` exporter mints (and caches) the JWT lazily on its first
+    export, so this stays non-blocking. ``project.id`` is therefore not a
+    resource attribute — it is encoded in the JWT, which AgentOps uses to route
+    the trace to the right project.
+    """
+    settings = _AgentOpsSettings()
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=_AGENTOPS_EXPORTER_KIND,
+                    endpoint=_AGENTOPS_ENDPOINT,
+                    options=(
+                        {"api_key": settings.api_key} if settings.api_key else None
+                    ),
+                ),
+            ],
+            "resource_attributes": {
+                **base.resource_attributes,
+                "service.name": settings.service_name,
+                "telemetry.sdk.name": "agentops",
+                **(
+                    {"deployment.environment": settings.environment}
+                    if settings.environment
+                    else {}
+                ),
+            },
+        }
+    )
+
+
+def _build_agentops_exporter(spec: ExporterSpec) -> Any:
+    """Factory for the ``agentops`` exporter kind: a lazy-auth OTLP/HTTP exporter."""
+    from opentelemetry.exporter.otlp.proto.http.trace_exporter import (
+        OTLPSpanExporter,
+    )
+
+    class _LazyAuthAgentOpsExporter(OTLPSpanExporter):
+        """OTLP/HTTP exporter that mints the AgentOps JWT on its first export.
+
+        ``export`` runs in the ``BatchSpanProcessor`` worker thread, so the
+        blocking token fetch never touches an event loop. The result is cached
+        after the first attempt (success or failure) so it runs at most once.
+        """
+
+        def __init__(self, *, endpoint: str | None, api_key: str | None) -> None:
+            super().__init__(endpoint=endpoint)
+            self._agentops_api_key = api_key
+            self._auth_resolved = False
+
+        def _ensure_authenticated(self) -> None:
+            if self._auth_resolved:
+                return
+            self._auth_resolved = True
+            if not self._agentops_api_key:
+                return
+            try:
+                token = _fetch_agentops_jwt(self._agentops_api_key).get("token")
+                if token:
+                    # ``_session`` is the requests.Session the base exporter
+                    # POSTs through; updating its Authorization header is how the
+                    # minted JWT reaches every subsequent export.
+                    self._session.headers["Authorization"] = f"Bearer {token}"
+            except Exception as e:
+                verbose_logger.debug("AgentOps JWT fetch failed: %s", e)
+
+        def export(self, spans: Any) -> Any:
+            self._ensure_authenticated()
+            return super().export(spans)
+
+    options = spec.options or {}
+    return _LazyAuthAgentOpsExporter(
+        endpoint=spec.endpoint, api_key=options.get("api_key")
+    )
+
+
+def _fetch_agentops_jwt(api_key: str) -> dict[str, Any]:
+    # Own a short-lived client rather than ``_get_httpx_client()``: that returns
+    # a process-wide cached ``HTTPHandler`` whose connection pool is shared by
+    # every caller, so closing it here would break concurrent/subsequent
+    # requests. This one-shot auth call gets its own client to close.
+    with httpx.Client(timeout=10) as client:
+        response = client.post(
+            url=_AGENTOPS_AUTH_ENDPOINT,
+            headers={"Content-Type": "application/json", "Connection": "keep-alive"},
+            json={"api_key": api_key},
+        )
+        if response.status_code != 200:
+            raise RuntimeError(f"Failed to fetch AgentOps token: {response.text}")
+        return response.json()
+
+
+register_exporter_factory(_AGENTOPS_EXPORTER_KIND, _build_agentops_exporter)

litellm/integrations/otel/presets/arize.py

@@ -0,0 +1,75 @@
+"""Arize preset — OTLP exporter to Arize + OpenInference vocabulary."""
+
+from pydantic import Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from litellm.integrations.arize.arize import ArizeLogger as _V1ArizeLogger
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.types.utils import StandardCallbackDynamicParams
+
+
+class _ArizeSettings(BaseSettings):
+    model_config = SettingsConfigDict(case_sensitive=False, extra="ignore")
+
+    # Standard OTLP headers env var, used as the fallback when no Arize
+    # credentials are configured.
+    otlp_traces_headers: str | None = Field(
+        default=None, validation_alias="OTEL_EXPORTER_OTLP_TRACES_HEADERS"
+    )
+
+
+def arize_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    arize_cfg = _V1ArizeLogger.get_arize_config()
+    headers = _arize_headers(arize_cfg)
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=arize_cfg.protocol or "otlp_grpc",
+                    endpoint=arize_cfg.endpoint or "https://otlp.arize.com/v1",
+                    headers=headers,
+                ),
+            ],
+            "mapper_names": ensure_mappers(base.mapper_names, "openinference"),
+            "resource_attributes": {
+                **base.resource_attributes,
+                **(
+                    {"model_id": arize_cfg.project_name}
+                    if arize_cfg.project_name
+                    else {}
+                ),
+            },
+        }
+    )
+
+
+def _arize_headers(arize_cfg) -> str | None:
+    pieces = []
+    if arize_cfg.space_id or arize_cfg.space_key:
+        pieces.append(f"space_id={arize_cfg.space_id or arize_cfg.space_key}")
+    if arize_cfg.api_key:
+        pieces.append(f"api_key={arize_cfg.api_key}")
+    if not pieces:
+        # Fall back to the standard OTLP headers env var when no Arize
+        # credentials are configured.
+        return _ArizeSettings().otlp_traces_headers
+    return ",".join(pieces)
+
+
+def arize_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str, str]:
+    """Per-request Arize OTLP headers from team/key dynamic params."""
+    headers: dict[str, str] = {}
+    # ``arize_space_key`` is the suggested param and wins over ``arize_space_id``.
+    space = params.get("arize_space_key") or params.get("arize_space_id")
+    if space:
+        headers["arize-space-id"] = space
+    api_key = params.get("arize_api_key")
+    if api_key:
+        headers["api_key"] = api_key
+    return headers

litellm/integrations/otel/presets/base.py

@@ -0,0 +1,25 @@
+"""Preset interface.
+
+A preset is a callable that reads its integration's env vars and produces an
+:class:`OpenTelemetryV2Config` (exporter list + mapper-name list + resource
+attributes). This ``Protocol`` pins that contract so ``PRESET_BY_CALLBACK`` and
+the factory in ``litellm_logging`` are type-checked structurally against it,
+matching the ``AttributeMapper`` protocol the mappers use.
+"""
+
+from typing import Protocol, runtime_checkable
+
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+
+
+@runtime_checkable
+class Preset(Protocol):
+    """Reads an integration's env config and returns an ``OpenTelemetryV2Config``.
+
+    ``config_overrides`` lets one preset layer onto another's config (or onto
+    test-supplied defaults); the factory calls presets with no arguments.
+    """
+
+    def __call__(
+        self, *, config_overrides: OpenTelemetryV2Config | None = None
+    ) -> OpenTelemetryV2Config: ...

litellm/integrations/otel/presets/langfuse.py

@@ -0,0 +1,43 @@
+"""Langfuse-OTEL preset."""
+
+from litellm.integrations.langfuse.langfuse_otel import (
+    LangfuseOtelLogger as _V1Langfuse,
+)
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.types.utils import StandardCallbackDynamicParams
+
+
+def langfuse_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    cfg = _V1Langfuse.get_langfuse_otel_config()
+    kind = cfg.exporter if isinstance(cfg.exporter, str) else "otlp_http"
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=kind,
+                    endpoint=cfg.endpoint,
+                    headers=cfg.headers,
+                ),
+            ],
+            "mapper_names": ensure_mappers(base.mapper_names, "langfuse"),
+        }
+    )
+
+
+def langfuse_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str, str]:
+    """Per-request Langfuse OTLP headers from team/key dynamic params."""
+    public_key = params.get("langfuse_public_key")
+    secret_key = params.get("langfuse_secret_key")
+    if public_key and secret_key:
+        return {
+            "Authorization": _V1Langfuse._get_langfuse_authorization_header(
+                public_key=public_key, secret_key=secret_key
+            )
+        }
+    return {}

litellm/integrations/otel/presets/langtrace.py

@@ -0,0 +1,22 @@
+"""Langtrace preset — Langtrace consumes generic OTLP + a vendor mapper."""
+
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+
+
+def langtrace_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    """Compose the Langtrace mapper on top of the customer's OTLP destination.
+
+    Unlike Arize / Phoenix / Langfuse, Langtrace doesn't ship its own endpoint
+    — users point their existing OTLP collector at Langtrace and just
+    need the vendor attribute schema applied to outgoing spans.
+    """
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "mapper_names": ensure_mappers(base.mapper_names, "langtrace"),
+        }
+    )

litellm/integrations/otel/presets/levo.py

@@ -0,0 +1,24 @@
+"""Levo preset — OTLP/HTTP to a Levo collector with org+workspace headers."""
+
+from litellm.integrations.levo.levo import LevoLogger as _V1Levo
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+
+
+def levo_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    cfg = _V1Levo.get_levo_config()
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind="otlp_http",
+                    endpoint=cfg.endpoint,
+                    headers=cfg.otlp_auth_headers,
+                ),
+            ],
+        }
+    )

litellm/integrations/otel/presets/phoenix.py

@@ -0,0 +1,48 @@
+"""Arize-Phoenix preset."""
+
+from pydantic import AliasChoices, Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from litellm.integrations.arize.arize_phoenix import (
+    ArizePhoenixLogger as _V1Phoenix,
+)
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+
+
+class _PhoenixSettings(BaseSettings):
+    model_config = SettingsConfigDict(case_sensitive=False, extra="ignore")
+
+    project_name: str = Field(
+        default="default",
+        validation_alias=AliasChoices(
+            "PHOENIX_PROJECT_NAME", "PHOENIX_COLLECTOR_PROJECT_NAME"
+        ),
+    )
+
+
+def phoenix_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    cfg = _V1Phoenix.get_arize_phoenix_config()
+    headers = cfg.otlp_auth_headers if hasattr(cfg, "otlp_auth_headers") else None
+    project_name = _PhoenixSettings().project_name
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=cfg.protocol if hasattr(cfg, "protocol") else "otlp_http",
+                    endpoint=cfg.endpoint,
+                    headers=headers,
+                ),
+            ],
+            "mapper_names": ensure_mappers(base.mapper_names, "openinference"),
+            "resource_attributes": {
+                **base.resource_attributes,
+                "openinference.project.name": project_name,
+            },
+        }
+    )

litellm/integrations/otel/presets/utils.py

@@ -0,0 +1,16 @@
+"""Shared helpers for the integration presets."""
+
+from typing import Iterable
+
+
+def ensure_mappers(mapper_names: Iterable[str], *names: str) -> list[str]:
+    """Return ``mapper_names`` with each of ``names`` appended if not already present.
+
+    Order is preserved and duplicates are skipped, so composing several presets
+    (or re-applying one) never double-adds a vocabulary.
+    """
+    result = list(mapper_names)
+    for name in names:
+        if name not in result:
+            result.append(name)
+    return result

litellm/integrations/otel/presets/weave.py

@@ -0,0 +1,43 @@
+"""Weave (W&B) preset."""
+
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.integrations.weave.weave_otel import (
+    _get_weave_authorization_header,
+    get_weave_otel_config,
+)
+from litellm.types.utils import StandardCallbackDynamicParams
+
+
+def weave_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    weave_cfg = get_weave_otel_config()
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=weave_cfg.protocol or "otlp_http",
+                    endpoint=weave_cfg.endpoint,
+                    headers=weave_cfg.otlp_auth_headers,
+                ),
+            ],
+            # Weave consumes OpenInference + a small Weave-specific overlay.
+            "mapper_names": ensure_mappers(base.mapper_names, "openinference", "weave"),
+        }
+    )
+
+
+def weave_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str, str]:
+    """Per-request Weave OTLP headers from team/key dynamic params."""
+    headers: dict[str, str] = {}
+    api_key = params.get("wandb_api_key")
+    if api_key:
+        headers["Authorization"] = _get_weave_authorization_header(api_key=api_key)
+    project_id = params.get("weave_project_id")
+    if project_id:
+        headers["project_id"] = project_id
+    return headers

litellm/integrations/otel/runtime.py

@@ -0,0 +1,38 @@
+"""SDK-free entrypoints for proxy-core call sites (auth, …).
+
+Proxy code may run without the OpenTelemetry SDK installed, so it must not import
+``litellm.integrations.otel.logger`` (which imports the SDK at module scope) at
+module load. These wrappers import it lazily and no-op when the SDK is absent or
+V2 is not the active logger — so a call site can wrap a request phase or seed
+identity unconditionally.
+"""
+
+from contextlib import contextmanager
+from typing import Any, Iterator
+
+
+@contextmanager
+def phase_span(name: str) -> "Iterator[Any]":
+    """Run a request phase inside a live active span so its DB/service calls nest.
+
+    Yields ``None`` (a plain no-op) when the OTel SDK is unavailable or V2 is not
+    the active logger.
+    """
+    try:
+        from litellm.integrations.otel.logger import phase_span as _phase_span
+    except Exception:
+        yield None
+        return
+    with _phase_span(name) as span:
+        yield span
+
+
+def seed_request_identity(user_api_key_dict: Any, model: Any = None) -> None:
+    """Seed request-identity Baggage at the auth boundary (no-op without V2)."""
+    try:
+        from litellm.integrations.otel.logger import (
+            seed_request_identity as _seed_request_identity,
+        )
+    except Exception:
+        return
+    _seed_request_identity(user_api_key_dict, model=model)

litellm/integrations/prometheus.py

@@ -166,6 +166,53 @@ class PrometheusLogger(CustomLogger):
                 labelnames=self.get_labels_for_metric("litellm_output_tokens_metric"),
             )
 
+            # Token-type detail metrics. These break out cached, cache-creation,
+            # audio and reasoning tokens that providers report inside
+            # prompt_tokens_details / completion_tokens_details on the usage
+            # object. They are sparse (only incremented when the provider
+            # reports a non-zero value) and are additive to the existing
+            # input/output token totals — no breaking change for existing
+            # dashboards built on the totals.
+            self.litellm_input_cached_tokens_metric = self._counter_factory(
+                "litellm_input_cached_tokens_metric",
+                "Provider-side cached input tokens (e.g. OpenAI prompt_tokens_details.cached_tokens, Anthropic cache_read_input_tokens)",
+                labelnames=self.get_labels_for_metric(
+                    "litellm_input_cached_tokens_metric"
+                ),
+            )
+
+            self.litellm_input_cache_creation_tokens_metric = self._counter_factory(
+                "litellm_input_cache_creation_tokens_metric",
+                "Provider-side input tokens written to prompt cache (e.g. Anthropic cache_creation_input_tokens)",
+                labelnames=self.get_labels_for_metric(
+                    "litellm_input_cache_creation_tokens_metric"
+                ),
+            )
+
+            self.litellm_input_audio_tokens_metric = self._counter_factory(
+                "litellm_input_audio_tokens_metric",
+                "Audio input tokens reported in prompt_tokens_details.audio_tokens",
+                labelnames=self.get_labels_for_metric(
+                    "litellm_input_audio_tokens_metric"
+                ),
+            )
+
+            self.litellm_output_reasoning_tokens_metric = self._counter_factory(
+                "litellm_output_reasoning_tokens_metric",
+                "Reasoning tokens reported in completion_tokens_details.reasoning_tokens",
+                labelnames=self.get_labels_for_metric(
+                    "litellm_output_reasoning_tokens_metric"
+                ),
+            )
+
+            self.litellm_output_audio_tokens_metric = self._counter_factory(
+                "litellm_output_audio_tokens_metric",
+                "Audio output tokens reported in completion_tokens_details.audio_tokens",
+                labelnames=self.get_labels_for_metric(
+                    "litellm_output_audio_tokens_metric"
+                ),
+            )
+
             # Remaining Budget for Team
             self.litellm_remaining_team_budget_metric = self._gauge_factory(
                 "litellm_remaining_team_budget_metric",
@@ -1301,6 +1348,101 @@ class PrometheusLogger(CustomLogger):
             amount=float(standard_logging_payload["completion_tokens"]),
         )
 
+        # Token-type detail metrics — sparse, only emitted when the provider
+        # reports a non-zero value in usage.prompt_tokens_details /
+        # usage.completion_tokens_details.
+        self._increment_token_detail_metrics(
+            standard_logging_payload=standard_logging_payload,
+            enum_values=enum_values,
+            label_context=label_context,
+        )
+
+    def _increment_token_detail_metrics(
+        self,
+        standard_logging_payload: StandardLoggingPayload,
+        enum_values: UserAPIKeyLabelValues,
+        label_context: Optional[PrometheusLabelFactoryContext] = None,
+    ) -> None:
+        """
+        Increment per-token-type counters from the Usage object that providers
+        attach to the request. The Usage dict is plumbed onto
+        ``standard_logging_payload["metadata"]["usage_object"]`` by
+        ``get_standard_logging_object_payload``.
+
+        Each counter is only incremented when the underlying value is > 0, so
+        scrape output stays sparse for providers that don't report these
+        details (most non-OpenAI/Anthropic models).
+        """
+        metadata = standard_logging_payload.get("metadata") or {}
+        usage_object = (
+            metadata.get("usage_object") if isinstance(metadata, dict) else None
+        )
+        if not isinstance(usage_object, dict):
+            return
+
+        prompt_details = usage_object.get("prompt_tokens_details") or {}
+        completion_details = usage_object.get("completion_tokens_details") or {}
+
+        detail_metrics: List[Tuple[Any, DEFINED_PROMETHEUS_METRICS, Any]] = [
+            (
+                self.litellm_input_cached_tokens_metric,
+                "litellm_input_cached_tokens_metric",
+                (
+                    prompt_details.get("cached_tokens")
+                    if isinstance(prompt_details, dict)
+                    else None
+                ),
+            ),
+            (
+                self.litellm_input_cache_creation_tokens_metric,
+                "litellm_input_cache_creation_tokens_metric",
+                (
+                    prompt_details.get("cache_creation_tokens")
+                    if isinstance(prompt_details, dict)
+                    else None
+                ),
+            ),
+            (
+                self.litellm_input_audio_tokens_metric,
+                "litellm_input_audio_tokens_metric",
+                (
+                    prompt_details.get("audio_tokens")
+                    if isinstance(prompt_details, dict)
+                    else None
+                ),
+            ),
+            (
+                self.litellm_output_reasoning_tokens_metric,
+                "litellm_output_reasoning_tokens_metric",
+                (
+                    completion_details.get("reasoning_tokens")
+                    if isinstance(completion_details, dict)
+                    else None
+                ),
+            ),
+            (
+                self.litellm_output_audio_tokens_metric,
+                "litellm_output_audio_tokens_metric",
+                (
+                    completion_details.get("audio_tokens")
+                    if isinstance(completion_details, dict)
+                    else None
+                ),
+            ),
+        ]
+
+        for counter, metric_name, value in detail_metrics:
+            if not isinstance(value, (int, float)) or value <= 0:
+                continue
+            PrometheusLogger._inc_labeled_counter(
+                self,
+                counter,
+                metric_name,
+                enum_values,
+                label_context=label_context,
+                amount=float(value),
+            )
+
     def _increment_cache_metrics(
         self,
         standard_logging_payload: StandardLoggingPayload,

litellm/litellm_core_utils/get_litellm_params.py

@@ -1,5 +1,7 @@
 from typing import Optional
 
+from litellm.llms.openai.data_residency import infer_openai_data_residency
+
 # Pre-define optional kwargs keys as frozenset for O(1) lookups
 # These are extracted from kwargs only if present, avoiding unnecessary .get() calls
 _OPTIONAL_KWARGS_KEYS = frozenset(
@@ -103,6 +105,10 @@ def get_litellm_params(
     if litellm_trace_id is None:
         litellm_trace_id = _meta.get("trace_id") or _meta.get("session_id")
 
+    data_residency: Optional[str] = infer_openai_data_residency(
+        custom_llm_provider, api_base
+    )
+
     # Build base dict with explicit parameters (always included)
     litellm_params = {
         "acompletion": acompletion,
@@ -112,6 +118,7 @@ def get_litellm_params(
         "verbose": verbose,
         "custom_llm_provider": custom_llm_provider,
         "api_base": api_base,
+        "data_residency": data_residency,
         "litellm_call_id": litellm_call_id,
         "model_alias_map": model_alias_map,
         "completion_call_id": completion_call_id,

litellm/litellm_core_utils/get_supported_openai_params.py

@@ -11,6 +11,7 @@ def get_supported_openai_params(  # noqa: PLR0915
     request_type: Literal[
         "chat_completion", "embeddings", "transcription"
     ] = "chat_completion",
+    base_model: Optional[str] = None,
 ) -> Optional[list]:
     """
     Returns the supported openai params for a given model + provider
@@ -20,6 +21,11 @@ def get_supported_openai_params(  # noqa: PLR0915
     get_supported_openai_params(model="anthropic.claude-3", custom_llm_provider="bedrock")
     ```
 
+    Args:
+        base_model: For Azure, the true underlying model (e.g. ``"azure/gpt-5.2"``)
+            when the deployment name differs. Used for model-type detection so that
+            non-standard deployment names route to the correct config.
+
     Returns:
     - List if custom_llm_provider is mapped
     - None if unmapped
@@ -32,17 +38,21 @@ def get_supported_openai_params(  # noqa: PLR0915
 
     if custom_llm_provider in LlmProvidersSet:
         provider_config = litellm.ProviderConfigManager.get_provider_chat_config(
-            model=model, provider=LlmProviders(custom_llm_provider)
+            model=model,
+            provider=LlmProviders(custom_llm_provider),
+            base_model=base_model,
         )
     elif custom_llm_provider.split("/")[0] in LlmProvidersSet:
         provider_config = litellm.ProviderConfigManager.get_provider_chat_config(
-            model=model, provider=LlmProviders(custom_llm_provider.split("/")[0])
+            model=model,
+            provider=LlmProviders(custom_llm_provider.split("/")[0]),
+            base_model=base_model,
         )
     else:
         provider_config = None
 
     if provider_config and request_type == "chat_completion":
-        return provider_config.get_supported_openai_params(model=model)
+        return provider_config.get_supported_openai_params(model=base_model or model)
 
     if custom_llm_provider == "bedrock":
         return litellm.AmazonConverseConfig().get_supported_openai_params(model=model)
@@ -130,16 +140,23 @@ def get_supported_openai_params(  # noqa: PLR0915
                 model=model
             )
     elif custom_llm_provider == "azure":
-        if litellm.AzureOpenAIO1Config().is_o_series_model(model=model):
+        _azure_detection_model = base_model or model
+        if litellm.AzureOpenAIO1Config().is_o_series_model(
+            model=_azure_detection_model
+        ):
             return litellm.AzureOpenAIO1Config().get_supported_openai_params(
-                model=model
+                model=_azure_detection_model
             )
-        elif litellm.AzureOpenAIGPT5Config.is_model_gpt_5_model(model=model):
+        elif litellm.AzureOpenAIGPT5Config.is_model_gpt_5_model(
+            model=_azure_detection_model
+        ):
             return litellm.AzureOpenAIGPT5Config().get_supported_openai_params(
-                model=model
+                model=_azure_detection_model
             )
         else:
-            return litellm.AzureOpenAIConfig().get_supported_openai_params(model=model)
+            return litellm.AzureOpenAIConfig().get_supported_openai_params(
+                model=_azure_detection_model
+            )
     elif custom_llm_provider == "openrouter":
         return litellm.OpenrouterConfig().get_supported_openai_params(model=model)
     elif custom_llm_provider == "vercel_ai_gateway":

litellm/litellm_core_utils/litellm_logging.py

@@ -1546,6 +1546,11 @@ class Logging(LiteLLMLoggingBaseClass):
                     if self.optional_params
                     else None
                 ),
+                "data_residency": (
+                    self.litellm_params.get("data_residency")
+                    if hasattr(self, "litellm_params") and self.litellm_params
+                    else None
+                ),
             }
         except Exception as e:  # error creating kwargs for cost calculation
             debug_info = StandardLoggingModelCostFailureDebugInformation(
@@ -1607,6 +1612,90 @@ class Logging(LiteLLMLoggingBaseClass):
     ) -> Optional[float]:
         return self._response_cost_calculator(result=result, cache_hit=cache_hit)
 
+    @staticmethod
+    def _is_sync_litellm_request(litellm_params: dict) -> bool:
+        """True for sync SDK entrypoints (``completion``), false for async (``acompletion``, etc.)."""
+        return (
+            litellm_params.get(CallTypes.acompletion.value, False) is not True
+            and litellm_params.get(CallTypes.aresponses.value, False) is not True
+            and litellm_params.get(CallTypes.aembedding.value, False) is not True
+            and litellm_params.get(CallTypes.aimage_generation.value, False) is not True
+            and litellm_params.get(CallTypes.atranscription.value, False) is not True
+        )
+
+    def _is_assembled_stream_success(self, result=None) -> bool:
+        """Final assembled stream export (not a per-chunk success call).
+
+        Per-chunk callers pass a ``ModelResponseStream`` (or ``None``); the
+        final assembled response is any other non-``None`` value (typically a
+        ``ModelResponse``). Treating a chunk as the assembled response would
+        prematurely set the ``has_dispatched_final_stream_success`` dedup
+        guard and silently suppress the real final stream log.
+        """
+        if self.stream is not True:
+            return False
+        if result is not None and not isinstance(result, ModelResponseStream):
+            return True
+        return (
+            "async_complete_streaming_response" in self.model_call_details
+            or self.model_call_details.get("complete_streaming_response") is not None
+        )
+
+    async def dispatch_success_handlers(
+        self,
+        result=None,
+        start_time=None,
+        end_time=None,
+        cache_hit=None,
+        prefer_async_handlers: bool = False,
+        **kwargs,
+    ) -> None:
+        """Route success logging to async and/or sync handlers for this request.
+
+        ``prefer_async_handlers`` only bypasses the sync-SDK-only shortcut (e.g.
+        ``async for`` on a stream from ``completion()``). Legacy string callbacks
+        still run via ``executor.submit(success_handler)`` when configured.
+        """
+        from litellm.litellm_core_utils.thread_pool_executor import executor
+
+        if self._is_assembled_stream_success(result):
+            if self.model_call_details.get("has_dispatched_final_stream_success"):
+                return
+            self.model_call_details["has_dispatched_final_stream_success"] = True
+
+        litellm_params = self.model_call_details.get("litellm_params", {}) or {}
+        sync_sdk = self._is_sync_litellm_request(litellm_params)
+        passthrough = self.call_type == CallTypes.pass_through.value
+        if sync_sdk and not prefer_async_handlers and not passthrough:
+            self.success_handler(
+                result,
+                start_time=start_time,
+                end_time=end_time,
+                cache_hit=cache_hit,
+                **kwargs,
+            )
+            return
+
+        await self.async_success_handler(
+            result,
+            start_time=start_time,
+            end_time=end_time,
+            cache_hit=cache_hit,
+            **kwargs,
+        )
+
+        if not self._should_run_sync_callbacks_for_async_calls():
+            return
+
+        executor.submit(
+            self.success_handler,
+            result,
+            start_time=start_time,
+            end_time=end_time,
+            cache_hit=cache_hit,
+            **kwargs,
+        )
+
     def should_run_logging(
         self,
         event_type: Literal[
@@ -2029,13 +2118,7 @@ class Logging(LiteLLMLoggingBaseClass):
             standard_logging_object=kwargs.get("standard_logging_object", None),
         )
         litellm_params = self.model_call_details.get("litellm_params", {})
-        is_sync_request = (
-            litellm_params.get(CallTypes.acompletion.value, False) is not True
-            and litellm_params.get(CallTypes.aresponses.value, False) is not True
-            and litellm_params.get(CallTypes.aembedding.value, False) is not True
-            and litellm_params.get(CallTypes.aimage_generation.value, False) is not True
-            and litellm_params.get(CallTypes.atranscription.value, False) is not True
-        )
+        is_sync_request = self._is_sync_litellm_request(litellm_params)
         try:
             ## BUILD COMPLETE STREAMED RESPONSE
             complete_streaming_response: Optional[
@@ -2491,9 +2574,11 @@ class Logging(LiteLLMLoggingBaseClass):
         print_verbose(
             "Logging Details LiteLLM-Async Success Call, cache_hit={}".format(cache_hit)
         )
-        if not self.should_run_logging(
+        if not self._is_assembled_stream_success(
+            result
+        ) and not self.should_run_logging(
             event_type="async_success"
-        ):  # prevent double logging
+        ):  # prevent double logging (non-streaming)
             return
 
         ## CALCULATE COST FOR BATCH JOBS
@@ -2943,13 +3028,7 @@ class Logging(LiteLLMLoggingBaseClass):
         ):  # prevent double logging
             return
         litellm_params = self.model_call_details.get("litellm_params", {})
-        is_sync_request = (
-            litellm_params.get(CallTypes.acompletion.value, False) is not True
-            and litellm_params.get(CallTypes.aresponses.value, False) is not True
-            and litellm_params.get(CallTypes.aembedding.value, False) is not True
-            and litellm_params.get(CallTypes.aimage_generation.value, False) is not True
-            and litellm_params.get(CallTypes.atranscription.value, False) is not True
-        )
+        is_sync_request = self._is_sync_litellm_request(litellm_params)
 
         try:
             start_time, end_time = self._failure_handler_helper_fn(
@@ -3713,6 +3792,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
     try:
         custom_logger_init_args = custom_logger_init_args or {}
         if logging_integration == "agentops":  # Add AgentOps initialization
+            _v2 = _maybe_construct_otel_v2("agentops", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             for callback in _in_memory_loggers:
                 if isinstance(callback, AgentOps):
                     return callback  # type: ignore
@@ -3865,6 +3947,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_opik_logger)
             return _opik_logger  # type: ignore
         elif logging_integration == "arize":
+            _v2 = _maybe_construct_otel_v2("arize", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
                 OpenTelemetryConfig,
@@ -3894,6 +3979,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_arize_otel_logger)
             return _arize_otel_logger  # type: ignore
         elif logging_integration == "arize_phoenix":
+            _v2 = _maybe_construct_otel_v2("arize_phoenix", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
                 OpenTelemetryConfig,
@@ -3905,31 +3993,6 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
                 endpoint=arize_phoenix_config.endpoint,
                 headers=arize_phoenix_config.otlp_auth_headers,
             )
-            if arize_phoenix_config.project_name:
-                existing_attrs = os.environ.get("OTEL_RESOURCE_ATTRIBUTES", "")
-                # Add openinference.project.name attribute
-                if existing_attrs:
-                    os.environ["OTEL_RESOURCE_ATTRIBUTES"] = (
-                        f"{existing_attrs},openinference.project.name={arize_phoenix_config.project_name}"
-                    )
-                else:
-                    os.environ["OTEL_RESOURCE_ATTRIBUTES"] = (
-                        f"openinference.project.name={arize_phoenix_config.project_name}"
-                    )
-
-            # Set Phoenix project name from environment variable
-            phoenix_project_name = os.environ.get("PHOENIX_PROJECT_NAME", None)
-            if phoenix_project_name:
-                existing_attrs = os.environ.get("OTEL_RESOURCE_ATTRIBUTES", "")
-                # Add openinference.project.name attribute
-                if existing_attrs:
-                    os.environ["OTEL_RESOURCE_ATTRIBUTES"] = (
-                        f"{existing_attrs},openinference.project.name={phoenix_project_name}"
-                    )
-                else:
-                    os.environ["OTEL_RESOURCE_ATTRIBUTES"] = (
-                        f"openinference.project.name={phoenix_project_name}"
-                    )
 
             # auth can be disabled on local deployments of arize phoenix
             if arize_phoenix_config.otlp_auth_headers is not None:
@@ -3949,6 +4012,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_arize_phoenix_otel_logger)
             return _arize_phoenix_otel_logger  # type: ignore
         elif logging_integration == "levo":
+            _v2 = _maybe_construct_otel_v2("levo", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.levo.levo import LevoLogger
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
@@ -3974,6 +4040,28 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_levo_otel_logger)
             return _levo_otel_logger  # type: ignore
         elif logging_integration == "otel":
+            # Gate the new typed V2 adapter behind LITELLM_OTEL_V2. When off,
+            # the legacy 3,227-line god-class is used unchanged. The two are
+            # never registered simultaneously — the dedup loop below treats
+            # any module under ``litellm.integrations.otel`` or
+            # ``litellm.integrations.opentelemetry`` as "the OTel callback".
+            from litellm.integrations.otel.model.config import is_otel_v2_enabled
+
+            if is_otel_v2_enabled():
+                from litellm.integrations.otel.logger import OpenTelemetryV2
+
+                for callback in _in_memory_loggers:
+                    if type(callback) is OpenTelemetryV2:
+                        return callback  # type: ignore
+                otel_logger_v2 = OpenTelemetryV2(
+                    **_get_custom_logger_settings_from_proxy_server(
+                        callback_name=logging_integration
+                    )
+                )
+                _in_memory_loggers.append(otel_logger_v2)
+                _maybe_auto_initialize_arize_phoenix(_in_memory_loggers)
+                return otel_logger_v2  # type: ignore
+
             from litellm.integrations.opentelemetry import OpenTelemetry
 
             for callback in _in_memory_loggers:
@@ -4112,6 +4200,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
         elif logging_integration == "langtrace":
             if "LANGTRACE_API_KEY" not in os.environ:
                 raise ValueError("LANGTRACE_API_KEY not found in environment variables")
+            _v2 = _maybe_construct_otel_v2("langtrace", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
 
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
@@ -4152,6 +4243,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(langfuse_logger)
             return langfuse_logger  # type: ignore
         elif logging_integration == "langfuse_otel":
+            _v2 = _maybe_construct_otel_v2("langfuse_otel", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.langfuse.langfuse_otel import LangfuseOtelLogger
 
             for callback in _in_memory_loggers:
@@ -4168,6 +4262,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_otel_logger)
             return _otel_logger  # type: ignore
         elif logging_integration == "weave_otel":
+            _v2 = _maybe_construct_otel_v2("weave_otel", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.opentelemetry import OpenTelemetryConfig
             from litellm.integrations.weave.weave_otel import (
                 WeaveOtelLogger,
@@ -4316,6 +4413,42 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
     return None
 
 
+def _maybe_construct_otel_v2(
+    callback_name: str, _in_memory_loggers: list
+) -> Optional[Any]:
+    """If ``LITELLM_OTEL_V2`` is on, build (or reuse) a single ``OpenTelemetryV2``
+    instance configured via the preset for ``callback_name``.
+
+    Returns ``None`` when V2 is off OR when there's no preset registered for
+    ``callback_name`` — callers should then fall through to the legacy path.
+    """
+    from litellm.integrations.otel.model.config import is_otel_v2_enabled
+
+    if not is_otel_v2_enabled():
+        return None
+    from litellm.integrations.otel.logger import OpenTelemetryV2
+    from litellm.integrations.otel.presets import PRESET_BY_CALLBACK
+
+    preset_fn = PRESET_BY_CALLBACK.get(callback_name)
+    if preset_fn is None:
+        return None
+    for callback in _in_memory_loggers:
+        if (
+            isinstance(callback, OpenTelemetryV2)
+            and getattr(callback, "callback_name", None) == callback_name
+        ):
+            return callback
+    try:
+        config = preset_fn()
+    except Exception:
+        # If env vars are missing or the preset raises, defer to the legacy path
+        # so customers get the same error story they had before V2 landed.
+        return None
+    v2_logger = OpenTelemetryV2(config=config, callback_name=callback_name)
+    _in_memory_loggers.append(v2_logger)
+    return v2_logger
+
+
 def _maybe_auto_initialize_arize_phoenix(_in_memory_loggers: list) -> None:
     """
     Auto-initialize ArizePhoenixLogger when Phoenix env vars are detected.
@@ -5140,13 +5273,17 @@ class StandardLoggingPayloadSetup:
     ) -> StandardLoggingPayloadErrorInformation:
         from litellm.constants import MAXIMUM_TRACEBACK_LINES_TO_LOG
 
-        # Check for 'code' first (used by ProxyException), then fall back to 'status_code' (used by LiteLLM exceptions)
-        # Ensure error_code is always a string for Prisma Python JSON field compatibility
+        # ProxyException uses .code, LiteLLM exceptions use .status_code,
+        # httpx.HTTPStatusError exposes status only as .response.status_code.
+        # Stringified for Prisma JSON compatibility.
         error_code_attr = getattr(original_exception, "code", None)
         if error_code_attr is not None and str(error_code_attr) not in ("", "None"):
             error_status: str = str(error_code_attr)
         else:
             status_code_attr = getattr(original_exception, "status_code", None)
+            if status_code_attr is None:
+                response_attr = getattr(original_exception, "response", None)
+                status_code_attr = getattr(response_attr, "status_code", None)
             error_status = str(status_code_attr) if status_code_attr is not None else ""
         error_class: str = (
             str(original_exception.__class__.__name__) if original_exception else ""

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -9,6 +9,7 @@ from litellm.types.utils import (
     CacheCreationTokenDetails,
     CallTypes,
     CompletionTokensDetailsWrapper,
+    DataResidency,
     ImageResponse,
     ModelInfo,
     PassthroughCallTypes,
@@ -29,6 +30,9 @@ _IMAGE_RESPONSE_CALL_TYPES = frozenset(
     }
 )
 
+# Pre-resolved DataResidency enum values for fast membership checks
+_VALID_DATA_RESIDENCIES = frozenset(r.value for r in DataResidency)
+
 
 def _is_above_128k(tokens: float) -> bool:
     if tokens > 128000:
@@ -617,11 +621,46 @@ def _calculate_input_cost(
     return prompt_cost
 
 
+def _get_regional_uplift_multiplier(
+    model_info: ModelInfo, data_residency: Optional[str]
+) -> float:
+    """
+    Resolve the per-model regional-processing uplift multiplier for a given
+    data-residency region.
+
+    OpenAI applies a flat percentage uplift (e.g. +10%) on all token costs for
+    requests served from a regionalized hostname (eu./us.api.openai.com). The
+    multiplier is stored on the model entry as
+    ``regional_processing_uplift_multiplier_<region>`` (e.g. 1.10).
+
+    Returns 1.0 (no uplift) when ``data_residency`` is ``None`` or when the
+    model has no multiplier configured for the given region.
+    """
+    if data_residency is None:
+        return 1.0
+    residency = data_residency.lower()
+    if residency not in _VALID_DATA_RESIDENCIES:
+        return 1.0
+    multiplier = model_info.get(f"regional_processing_uplift_multiplier_{residency}")
+    if multiplier is None:
+        return 1.0
+    try:
+        return float(cast(float, multiplier))
+    except (TypeError, ValueError):
+        verbose_logger.exception(
+            "Invalid regional_processing_uplift_multiplier_%s for model; "
+            "defaulting to 1.0",
+            residency,
+        )
+        return 1.0
+
+
 def generic_cost_per_token(  # noqa: PLR0915
     model: str,
     usage: Usage,
     custom_llm_provider: str,
     service_tier: Optional[str] = None,
+    data_residency: Optional[str] = None,
 ) -> Tuple[float, float]:
     """
     Calculates the cost per token for a given model, prompt tokens, and completion tokens.
@@ -631,6 +670,8 @@ def generic_cost_per_token(  # noqa: PLR0915
     Input:
         - model: str, the model name without provider prefix
         - usage: LiteLLM Usage block, containing anthropic caching information
+        - data_residency: optional OpenAI data-residency region (e.g. "eu", "us"),
+          used to apply the per-model regional-processing uplift multiplier.
 
     Returns:
         Tuple[float, float] - prompt_cost_in_usd, completion_cost_in_usd
@@ -781,6 +822,14 @@ def generic_cost_per_token(  # noqa: PLR0915
         )
         completion_cost += float(image_tokens) * _output_cost_per_image_token
 
+    ## REGIONAL DATA-RESIDENCY UPLIFT
+    # Applied as a flat multiplier across all token costs for the request
+    # when the upstream is a regionalized OpenAI host (eu./us.api.openai.com).
+    uplift = _get_regional_uplift_multiplier(model_info, data_residency)
+    if uplift != 1.0:
+        prompt_cost *= uplift
+        completion_cost *= uplift
+
     return prompt_cost, completion_cost
 
 

litellm/litellm_core_utils/prompt_templates/common_utils.py

@@ -3,6 +3,7 @@ Common utility functions used for translating messages across providers
 """
 
 import io
+import json
 import mimetypes
 import re
 from os import PathLike
@@ -132,6 +133,39 @@ def strip_none_values_from_message(message: AllMessageValues) -> AllMessageValue
     return cast(AllMessageValues, {k: v for k, v in message.items() if v is not None})
 
 
+def extract_search_results_text(search_results: object) -> str:
+    """
+    Extract model-visible text from OpenAI tool-message ``search_results``.
+
+    Used by token estimators and TPM limiters so large search result payloads
+    cannot bypass preflight checks via a small ``content`` field.
+
+    Counts every string field forwarded on Bedrock ``SearchResultBlock``:
+    ``source``, ``title``, ``content[].text``, and ``citations``.
+    """
+    if not isinstance(search_results, list):
+        return ""
+    texts = ""
+    for result in search_results:
+        if not isinstance(result, dict):
+            continue
+        for key in ("source", "title"):
+            value = result.get(key)
+            if isinstance(value, str):
+                texts += value
+        content = result.get("content")
+        if isinstance(content, list):
+            for block in content:
+                if isinstance(block, dict):
+                    text = block.get("text")
+                    if isinstance(text, str):
+                        texts += text
+        citations = result.get("citations")
+        if citations is not None:
+            texts += json.dumps(citations, separators=(",", ":"))
+    return texts
+
+
 def convert_content_list_to_str(
     message: Union[AllMessageValues, ChatCompletionResponseMessage],
 ) -> str:
@@ -152,6 +186,7 @@ def convert_content_list_to_str(
         elif message_content is not None and isinstance(message_content, str):
             texts = message_content
 
+    texts += extract_search_results_text(message.get("search_results"))
     return texts
 
 
@@ -1204,12 +1239,8 @@ def get_last_user_message(messages: List[AllMessageValues]) -> Optional[str]:
         {"role": "assistant", "content": "I'm good, thank you!"},
         {"role": "user", "content": "What is the weather in Tokyo?"},
     ]
-    get_user_prompt(messages) -> "What is the weather in Tokyo?"
+    get_last_user_message(messages) -> "What is the weather in Tokyo?"
     """
-    from litellm.litellm_core_utils.prompt_templates.common_utils import (
-        convert_content_list_to_str,
-    )
-
     if not messages:
         return None
 

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -3658,6 +3658,7 @@ from litellm.types.llms.bedrock import (
     ToolInputSchemaBlock as BedrockToolInputSchemaBlock,
 )
 from litellm.types.llms.bedrock import ToolJsonSchemaBlock as BedrockToolJsonSchemaBlock
+from litellm.types.llms.bedrock import SearchResultBlock
 from litellm.types.llms.bedrock import ToolResultBlock as BedrockToolResultBlock
 from litellm.types.llms.bedrock import (
     ToolResultContentBlock as BedrockToolResultContentBlock,
@@ -3997,7 +3998,7 @@ def _convert_to_bedrock_tool_call_invoke(
         for tool in tool_calls:
             if "function" in tool:
                 tool_id = tool["id"]
-                name = tool["function"].get("name", "")
+                name = make_valid_bedrock_tool_name(tool["function"].get("name", ""))
                 arguments = tool["function"].get("arguments", "")
 
                 if not arguments or not arguments.strip():
@@ -4063,6 +4064,122 @@ def _convert_to_bedrock_tool_call_invoke(
         )
 
 
+def _append_bedrock_tool_result_media_block(
+    tool_result_content_blocks: List[BedrockToolResultContentBlock],
+    processed_block: BedrockContentBlock,
+    content: dict,
+    content_type: str,
+) -> None:
+    if "image" in processed_block:
+        tool_result_content_blocks.append(
+            BedrockToolResultContentBlock(image=processed_block["image"])
+        )
+    elif "document" in processed_block:
+        tool_result_content_blocks.append(
+            BedrockToolResultContentBlock(document=processed_block["document"])
+        )
+    else:
+        verbose_logger.warning(
+            "Bedrock Converse: unrecognized BedrockContentBlock keys "
+            "%s for %s tool-result block %s; dropping.",
+            list(processed_block.keys()),
+            content_type,
+            content,
+        )
+
+
+def _append_bedrock_tool_result_image_url_block(
+    tool_result_content_blocks: List[BedrockToolResultContentBlock],
+    content: dict,
+) -> None:
+    format: Optional[str] = None
+    if isinstance(content["image_url"], dict):
+        image_url = content["image_url"]["url"]
+        format = content["image_url"].get("format")
+    else:
+        image_url = content["image_url"]
+    processed_block = BedrockImageProcessor.process_image_sync(
+        image_url=image_url,
+        format=format,
+    )
+    _append_bedrock_tool_result_media_block(
+        tool_result_content_blocks, processed_block, content, "image_url"
+    )
+
+
+def _append_bedrock_tool_result_file_block(
+    tool_result_content_blocks: List[BedrockToolResultContentBlock],
+    content: dict,
+) -> None:
+    # Match the user-message path (_process_file_message): accept either
+    # file_data (base64 data URI) or file_id (server-side reference / URL).
+    file_obj = content.get("file") or {}
+    file_data = file_obj.get("file_data")
+    file_id = file_obj.get("file_id")
+    if file_data is None and file_id is None:
+        raise litellm.BadRequestError(
+            message="file_data and file_id cannot both be None. Got={}".format(content),
+            model="",
+            llm_provider="bedrock",
+        )
+    processed_block = BedrockImageProcessor.process_image_sync(
+        image_url=cast(str, file_id or file_data),
+        format=file_obj.get("format"),
+    )
+    _append_bedrock_tool_result_media_block(
+        tool_result_content_blocks, processed_block, content, "file"
+    )
+
+
+def _parse_bedrock_tool_result_content_list(
+    content_list: List,
+) -> List[BedrockToolResultContentBlock]:
+    tool_result_content_blocks: List[BedrockToolResultContentBlock] = []
+    for content in content_list:
+        if content["type"] == "text":
+            tool_result_content_blocks.append(
+                BedrockToolResultContentBlock(text=content["text"])
+            )
+        elif content["type"] == "image_url":
+            _append_bedrock_tool_result_image_url_block(
+                tool_result_content_blocks, content
+            )
+        elif content["type"] == "file":
+            _append_bedrock_tool_result_file_block(tool_result_content_blocks, content)
+    return tool_result_content_blocks
+
+
+def _build_bedrock_tool_result_content_blocks(
+    message: Union[ChatCompletionToolMessage, ChatCompletionFunctionMessage],
+) -> tuple[List[BedrockToolResultContentBlock], bool]:
+    # Optional OpenAI tool-message extension:
+    # allow structured Bedrock search results on tool messages and map them
+    # directly to toolResult.content[].searchResult for Converse API.
+    #
+    # If `search_results` is present, we intentionally prefer it over `content`
+    # to avoid generating mixed text + searchResult blocks.
+    search_results = message.get("search_results")
+    if isinstance(search_results, list):
+        tool_result_content_blocks: List[BedrockToolResultContentBlock] = []
+        for result in search_results:
+            if not isinstance(result, dict):
+                continue
+            tool_result_content_blocks.append(
+                BedrockToolResultContentBlock(
+                    searchResult=cast(SearchResultBlock, result)
+                )
+            )
+        if tool_result_content_blocks:
+            return tool_result_content_blocks, True
+
+    message_content = message["content"]
+    if isinstance(message_content, str):
+        return [BedrockToolResultContentBlock(text=message_content)], False
+    if isinstance(message_content, List):
+        return _parse_bedrock_tool_result_content_list(message_content), False
+    return [], False
+
+
 def _convert_to_bedrock_tool_call_result(
     message: Union[ChatCompletionToolMessage, ChatCompletionFunctionMessage],
 ) -> BedrockContentBlock:
@@ -4106,90 +4223,18 @@ def _convert_to_bedrock_tool_call_result(
     """
     - 
     """
-    tool_result_content_blocks: List[BedrockToolResultContentBlock] = []
-    if isinstance(message["content"], str):
-        tool_result_content_blocks.append(
-            BedrockToolResultContentBlock(text=message["content"])
-        )
-    elif isinstance(message["content"], List):
-        content_list = message["content"]
-        for content in content_list:
-            if content["type"] == "text":
-                tool_result_content_blocks.append(
-                    BedrockToolResultContentBlock(text=content["text"])
-                )
-            elif content["type"] == "image_url":
-                format: Optional[str] = None
-                if isinstance(content["image_url"], dict):
-                    image_url = content["image_url"]["url"]
-                    format = content["image_url"].get("format")
-                else:
-                    image_url = content["image_url"]
-                _block: BedrockContentBlock = BedrockImageProcessor.process_image_sync(
-                    image_url=image_url,
-                    format=format,
-                )
-                if "image" in _block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(image=_block["image"])
-                    )
-                elif "document" in _block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(document=_block["document"])
-                    )
-                else:
-                    verbose_logger.warning(
-                        "Bedrock Converse: unrecognized BedrockContentBlock keys "
-                        "%s for image_url tool-result block %s; dropping.",
-                        list(_block.keys()),
-                        content,
-                    )
-            elif content["type"] == "file":
-                # Match the user-message path (_process_file_message): accept
-                # either file_data (base64 data URI) or file_id (server-side
-                # reference / URL) and hand off to BedrockImageProcessor. Raise
-                # BadRequestError on both-None rather than silently dropping.
-                file_obj = content.get("file") or {}
-                file_data = file_obj.get("file_data")
-                file_id = file_obj.get("file_id")
-                if file_data is None and file_id is None:
-                    raise litellm.BadRequestError(
-                        message="file_data and file_id cannot both be None. Got={}".format(
-                            content
-                        ),
-                        model="",
-                        llm_provider="bedrock",
-                    )
-                file_format = file_obj.get("format")
-                _file_block: BedrockContentBlock = (
-                    BedrockImageProcessor.process_image_sync(
-                        image_url=cast(str, file_id or file_data),
-                        format=file_format,
-                    )
-                )
-                if "document" in _file_block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(document=_file_block["document"])
-                    )
-                elif "image" in _file_block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(image=_file_block["image"])
-                    )
-                else:
-                    verbose_logger.warning(
-                        "Bedrock Converse: unrecognized BedrockContentBlock keys "
-                        "%s for file tool-result block %s; dropping.",
-                        list(_file_block.keys()),
-                        content,
-                    )
+    tool_result_content_blocks, used_search_results = (
+        _build_bedrock_tool_result_content_blocks(message)
+    )
 
     message.get("name", "")
     id = str(message.get("tool_call_id", str(uuid.uuid4())))
 
     tool_result = BedrockToolResultBlock(
-        content=tool_result_content_blocks,
-        toolUseId=id,
+        content=tool_result_content_blocks, toolUseId=id
     )
+    if used_search_results:
+        tool_result["status"] = cast(Literal["success"], "success")
 
     content_block = BedrockContentBlock(toolResult=tool_result)
 
@@ -5323,16 +5368,10 @@ def _bedrock_converse_messages_pt(  # noqa: PLR0915
 
 
 def make_valid_bedrock_tool_name(input_tool_name: str) -> str:
-    """
-    Replaces any invalid characters in the input tool name with underscores
-    and ensures the resulting string is a valid identifier for Bedrock tools
-    """
+    """Normalize tool names to Bedrock pattern [a-zA-Z][a-zA-Z0-9_-]*."""
 
     def replace_invalid(char):
-        """
-        Bedrock tool names only supports alpha-numeric characters and underscores
-        """
-        if char.isalnum() or char == "_":
+        if char.isalnum() or char in ("_", "-"):
             return char
         return "_"
 
@@ -5492,7 +5531,7 @@ def _bedrock_tools_pt(
             raw_name = f"litellm_unnamed_tool_{tool_idx}"
 
         # related issue: https://github.com/BerriAI/litellm/issues/5007
-        # Bedrock tool names must satisfy regular expression pattern: [a-zA-Z][a-zA-Z0-9_]* ensure this is true
+        # Bedrock tool names must satisfy pattern: [a-zA-Z][a-zA-Z0-9_-]*
         name = make_valid_bedrock_tool_name(input_tool_name=raw_name)
         if _tool_description:  # bedrock doesn't accept empty "" or None descriptions
             description = _tool_description

litellm/litellm_core_utils/realtime_streaming.py

@@ -86,6 +86,12 @@ class RealTimeStreaming:
         # When a text message is blocked, hold the guardrail reason so the next
         # response.create can be rewritten to include the failure context.
         self._pending_guardrail_message: Optional[str] = None
+        # Track whether session.created has already been sent to the client
+        # (e.g. synthetic event in deferred setup mode).
+        self._session_created_sent_to_client: bool = False
+        # Track whether we have already sent the guardrail turn-detection update
+        # that disables provider auto-response for transcription guardrails.
+        self._guardrail_turn_detection_update_sent: bool = False
 
     _SESSION_EVENT_TYPES = frozenset(["session.created", "session.updated"])
     _AUDIO_FORMAT_MAP: Dict[str, Dict[str, Any]] = {
@@ -248,40 +254,82 @@ class RealTimeStreaming:
             ## SYNC LOGGING
             executor.submit(self.logging_obj.success_handler(self.messages))
 
-    async def _send_to_backend(self, message: str) -> None:
+    async def _send_to_backend(self, message: str) -> bool:
         """Send a message to the backend WebSocket.
 
         If a provider_config is set the message is first passed through
         transform_realtime_request so that provider-specific translation
         (e.g. dropping session.update for Vertex AI) is applied even for
         guardrail-injected messages.
+
+        Returns True if at least one message was actually delivered to the
+        backend, False if the provider transformation produced no output and
+        the message was effectively dropped.
         """
         if self.provider_config:
             transformed = self.provider_config.transform_realtime_request(
                 message, self.model, self.session_configuration_request
             )
+            sent = False
             for msg in transformed:
+                # Send first; only cache the setup payload once the backend
+                # has actually accepted it. Caching before send would leave
+                # ``session_configuration_request`` populated after a failed
+                # send, causing subsequent client session.update messages to
+                # be treated as "subsequent" and dropped even though the
+                # backend never received the original setup.
                 await self.backend_ws.send(msg)  # type: ignore[union-attr, attr-defined]
-        else:
-            await self.backend_ws.send(message)  # type: ignore[union-attr, attr-defined]
+                self._cache_session_configuration_request(msg)
+                sent = True
+            return sent
+        await self.backend_ws.send(message)  # type: ignore[union-attr, attr-defined]
+        return True
+
+    def _cache_session_configuration_request(self, transformed_message: str) -> None:
+        """Store setup payload once sent to backend.
+
+        Updates the cached setup on every successful setup send so follow-up
+        ``session.update`` messages (which produce a merged setup with new
+        ``generationConfig`` / ``systemInstruction`` / etc.) are reflected in
+        the cache used by downstream readers (``transform_session_created_event``,
+        ``return_new_content_delta_events`` modality lookup, ...).
+        """
+        try:
+            message_obj = json.loads(transformed_message)
+            if "setup" in message_obj:
+                self.session_configuration_request = transformed_message
+        except (json.JSONDecodeError, TypeError):
+            return
 
     def _make_disable_auto_response_message(self) -> str:
         """Return a session.update that disables VAD auto-response."""
+        turn_detection: Dict[str, Any] = {
+            "type": "server_vad",
+            "create_response": False,
+        }
         if self._backend_uses_beta_protocol:
-            session: Dict[str, Any] = {
-                "turn_detection": {"create_response": False},
-            }
+            session: Dict[str, Any] = {"turn_detection": turn_detection}
         else:
             session = {
                 "type": "realtime",
-                "audio": {
-                    "input": {
-                        "turn_detection": {"create_response": False},
-                    }
-                },
+                "audio": {"input": {"turn_detection": turn_detection}},
             }
         return json.dumps({"type": "session.update", "session": session})
 
+    async def _maybe_send_guardrail_turn_detection_update(self) -> None:
+        """Disable provider auto-response once when transcription guardrails are enabled."""
+        if self._guardrail_turn_detection_update_sent:
+            return
+        if not self._has_audio_transcription_guardrails():
+            return
+        sent = await self._send_to_backend(self._make_disable_auto_response_message())
+        # Only mark as sent when the provider transformation actually delivered
+        # the update to the backend. Otherwise (e.g. Gemini drops session.update
+        # after the initial setup), leave the flag unset so future opportunities
+        # — such as a duplicate session.created — can retry.
+        if sent:
+            self._guardrail_turn_detection_update_sent = True
+
     def _has_realtime_guardrails(self) -> bool:
         """Return True if any callback is registered for realtime guardrail event types."""
         from litellm.integrations.custom_guardrail import CustomGuardrail
@@ -320,12 +368,20 @@ class RealTimeStreaming:
         self,
         transcript: str,
         item_id: Optional[str] = None,
+        pre_block_backend_message: Optional[str] = None,
     ) -> bool:
         """
         Run registered guardrails on a completed speech transcription.
 
         Returns True if blocked (synthetic warning already sent to client).
         Returns False if clean (caller should send response.create to the backend).
+
+        ``pre_block_backend_message`` (if provided) is sent to the backend
+        BEFORE any of the guardrail's own backend messages when a block is
+        triggered. This is needed for protocol contracts that require a
+        specific message to be sent first — e.g. Gemini Live requires a
+        matching ``toolResponse`` immediately after a ``toolCall`` before any
+        other client messages can be accepted.
         """
         from litellm.integrations.custom_guardrail import CustomGuardrail
         from litellm.types.guardrails import GuardrailEventHooks
@@ -385,6 +441,13 @@ class RealTimeStreaming:
                     getattr(callback, "realtime_violation_message", None) or safe_msg
                 )
 
+                # Deliver any caller-supplied backend message FIRST so that
+                # protocol contracts requiring a specific ordering (e.g.
+                # Gemini Live's mandatory ``toolResponse`` after a
+                # ``toolCall``) are honored before the guardrail's own
+                # clientContent / cancel messages are sent.
+                if pre_block_backend_message is not None:
+                    await self._send_to_backend(pre_block_backend_message)
                 # Cancel any in-progress LLM response (e.g. VAD auto-response).
                 await self._send_to_backend(json.dumps({"type": "response.cancel"}))
                 # Send the policy violation hint (shows as small gray status text in UI).
@@ -480,16 +543,34 @@ class RealTimeStreaming:
             else [transformed_response]
         )
         for event in events:
+            is_session_created_event = (
+                isinstance(event, dict) and event.get("type") == "session.created"
+            )
+            if is_session_created_event:
+                if self._session_created_sent_to_client:
+                    # A synthetic session.created (with placeholder defaults) was
+                    # already forwarded to the client when we connected.  The
+                    # provider's real session.created (e.g. emitted from Gemini
+                    # `setupComplete`) carries the authoritative modalities/model
+                    # from the client's session.update.  Re-emit it as
+                    # `session.updated` so the client learns the corrected
+                    # configuration without seeing two `session.created` events.
+                    event = {**event, "type": "session.updated"}
+                else:
+                    self._session_created_sent_to_client = True
             event_str = json.dumps(event)
-            ## For audio/VAD guardrail path: forward session.created first, then inject.
-            if (
-                isinstance(event, dict)
-                and event.get("type") == "session.created"
-                and self._has_audio_transcription_guardrails()
-            ):
+            ## For audio/VAD guardrail path: forward the (possibly retyped)
+            ## session.created first, then invoke the one-time guardrail
+            ## turn-detection update.  ``_maybe_send_guardrail_turn_detection_update``
+            ## is idempotent (gated by ``_guardrail_turn_detection_update_sent``),
+            ## so duplicate session.created events — including those emitted
+            ## after a synthetic session.created from ``llm_http_handler`` in
+            ## deferred-setup mode — still get a single chance to inject the
+            ## update if a prior attempt was dropped by the provider transform.
+            if is_session_created_event and self._has_audio_transcription_guardrails():
                 self.store_message(event_str)
                 await self.websocket.send_text(event_str)
-                await self._send_to_backend(self._make_disable_auto_response_message())
+                await self._maybe_send_guardrail_turn_detection_update()
                 continue
             ## GUARDRAIL: run on transcription events in provider_config path too
             if (
@@ -564,10 +645,19 @@ class RealTimeStreaming:
                 try:
                     raw_response = await self.backend_ws.recv(  # type: ignore[union-attr]
                         decode=False
-                    )  # improves performance
+                    )
                 except TypeError:
                     raw_response = await self.backend_ws.recv()  # type: ignore[union-attr, assignment]
 
+                if isinstance(raw_response, bytes):
+                    try:
+                        raw_response = raw_response.decode("utf-8")
+                    except UnicodeDecodeError:
+                        verbose_logger.warning(
+                            "Received non-UTF-8 binary frame from backend, skipping."
+                        )
+                        continue
+
                 if self.provider_config:
                     try:
                         await self._handle_provider_config_message(raw_response)
@@ -783,12 +873,13 @@ class RealTimeStreaming:
         item["content"] = new_content
         return item
 
-    async def client_ack_messages(self):
+    async def client_ack_messages(self):  # noqa: PLR0915
         try:
             while True:
                 message = await self.websocket.receive_text()
 
                 ## GUARDRAIL: intercept conversation.item.create for text-based injection.
+                guardrail_turn_detection_injected = False
                 try:
                     msg_obj = json.loads(message)
                     msg_type = msg_obj.get("type")
@@ -796,7 +887,68 @@ class RealTimeStreaming:
                     if msg_type == "conversation.item.create":
                         # Check user text messages for prompt injection
                         item = msg_obj.get("item", {})
-                        if item.get("role") == "user":
+                        # Check function_call_output first so a client cannot
+                        # bypass the tool-result guardrail by also setting
+                        # role="user" on a function_call_output item.
+                        if item.get("type") == "function_call_output":
+                            # Tool results are client-controlled and fed to the
+                            # model; check them with the same guardrail used for
+                            # user text so an attacker cannot smuggle blocked
+                            # content into a function_call_output.
+                            output = item.get("output", "")
+                            output_text = (
+                                output
+                                if isinstance(output, str)
+                                else json.dumps(output)
+                            )
+                            if output_text:
+                                # Build the sanitized function_call_output up
+                                # front so we can hand it to the guardrail
+                                # runner as the pre-block message. Providers
+                                # that pair every toolCall with a toolResponse
+                                # (e.g. Gemini/Vertex Live) require the
+                                # toolResponse to arrive BEFORE any other
+                                # client message — otherwise the guardrail's
+                                # own clientContent would violate the
+                                # pending-tool-call protocol contract and the
+                                # backend could close the connection before
+                                # the sanitized response ever lands. Dropping
+                                # the blocked item outright would similarly
+                                # leave such providers waiting indefinitely.
+                                # The sanitized payload carries no blocked
+                                # content — only a generic policy marker.
+                                sanitized_msg = json.dumps(
+                                    {
+                                        **msg_obj,
+                                        "item": {
+                                            **item,
+                                            "output": json.dumps(
+                                                {
+                                                    "error": "Tool output blocked by content policy",
+                                                }
+                                            ),
+                                        },
+                                    }
+                                )
+                                blocked = await self.run_realtime_guardrails(
+                                    output_text,
+                                    pre_block_backend_message=sanitized_msg,
+                                )
+                                if blocked:
+                                    # ``_pending_guardrail_message`` is
+                                    # intentionally NOT set here. That flag
+                                    # exists to swallow the reflexive
+                                    # ``response.create`` an OpenAI client
+                                    # sends immediately after a user text
+                                    # message. In a tool-calling flow the
+                                    # client may not send a ``response.create``
+                                    # at all (e.g. Gemini SDKs auto-respond),
+                                    # so leaving the flag set would
+                                    # incorrectly drop an unrelated
+                                    # ``response.create`` from a later
+                                    # interaction turn.
+                                    continue
+                        elif item.get("role") == "user":
                             content_list = item.get("content", [])
                             texts = [
                                 c.get("text", "")
@@ -824,6 +976,89 @@ class RealTimeStreaming:
                         self._pending_guardrail_message = None
                         continue
 
+                    ## GUARDRAIL: Inject turn_detection into first session.update
+                    # if needed. Done BEFORE the GA remap so the injected
+                    # ``create_response`` rides along with any client-provided
+                    # turn_detection fields (e.g. silence_duration_ms) into the
+                    # nested ``audio.input.turn_detection`` path produced by the
+                    # remap. Doing this after the remap would create a separate
+                    # minimal root-level ``turn_detection`` and silently drop
+                    # the client's nested settings.
+                    if (
+                        msg_type == "session.update"
+                        and self.session_configuration_request is None
+                        and not self._guardrail_turn_detection_update_sent
+                        and self._has_audio_transcription_guardrails()
+                    ):
+                        session = msg_obj.setdefault("session", {})
+                        if isinstance(session, dict):
+                            existing_td = session.get("turn_detection")
+                            if not isinstance(existing_td, dict):
+                                existing_td = {}
+                            existing_td["create_response"] = False
+                            session["turn_detection"] = existing_td
+                            message = json.dumps(msg_obj)
+                            guardrail_turn_detection_injected = True
+                            verbose_logger.debug(
+                                "Injected turn_detection into first session.update for audio transcription guardrails"
+                            )
+
+                    ## GUARDRAIL: Force ``create_response`` to False in any
+                    # client-provided ``turn_detection`` so a later
+                    # ``session.update`` cannot re-enable VAD auto-response
+                    # and bypass the transcription guardrail after the
+                    # initial disable. Covers both the flat beta key and the
+                    # nested GA ``audio.input.turn_detection`` shape, since
+                    # the GA remap below also accepts either form. Skipped
+                    # when the injection block above already ran for this
+                    # message, to avoid redundant double-serialization.
+                    if (
+                        msg_type == "session.update"
+                        and not guardrail_turn_detection_injected
+                        and self._has_audio_transcription_guardrails()
+                    ):
+                        session = msg_obj.get("session")
+                        if isinstance(session, dict):
+                            td_overridden = False
+                            flat_td = session.get("turn_detection")
+                            flat_td_present = flat_td is not None
+                            if flat_td_present:
+                                if not isinstance(flat_td, dict):
+                                    flat_td = {}
+                                if flat_td.get("create_response") is not False:
+                                    flat_td["create_response"] = False
+                                    session["turn_detection"] = flat_td
+                                    td_overridden = True
+                            nested_td_present = False
+                            audio = session.get("audio")
+                            if isinstance(audio, dict):
+                                audio_input = audio.get("input")
+                                if isinstance(audio_input, dict):
+                                    nested_td = audio_input.get("turn_detection")
+                                    if nested_td is not None:
+                                        nested_td_present = True
+                                        if not isinstance(nested_td, dict):
+                                            nested_td = {}
+                                        if (
+                                            nested_td.get("create_response")
+                                            is not False
+                                        ):
+                                            nested_td["create_response"] = False
+                                            audio_input["turn_detection"] = nested_td
+                                            td_overridden = True
+                            # Symmetric with the first-update injection block:
+                            # if the client omitted turn_detection entirely on
+                            # a subsequent session.update, still inject the
+                            # ``create_response: False`` override so the
+                            # transcription guardrail cannot be re-enabled by
+                            # any downstream merge that drops the original
+                            # disable.
+                            if not flat_td_present and not nested_td_present:
+                                session["turn_detection"] = {"create_response": False}
+                                td_overridden = True
+                            if td_overridden:
+                                message = json.dumps(msg_obj)
+
                     # GA compatibility: remap beta-style session fields only when
                     # the upstream is in GA mode. Beta upstreams expect the flat
                     # session shape unchanged.
@@ -841,17 +1076,20 @@ class RealTimeStreaming:
                     pass
 
                 ## LOGGING
+                # Log after any in-place modifications (GA remap, guardrail
+                # turn_detection injection) so audit logs reflect what we
+                # actually forward to the backend.
                 self.store_input(message=message)
-                ## FORWARD TO BACKEND
-                if self.provider_config:
-                    message = self.provider_config.transform_realtime_request(
-                        message, self.model
-                    )
 
-                    for msg in message:
-                        await self.backend_ws.send(msg)  # type: ignore[union-attr]
-                else:
-                    await self.backend_ws.send(message)  # type: ignore[union-attr]
+                ## FORWARD TO BACKEND
+                # Only mark the guardrail turn_detection update as sent after the
+                # backend actually accepted the message. Setting the flag earlier
+                # would permanently disable the injection if ``_send_to_backend``
+                # raised — neither this loop nor
+                # ``_maybe_send_guardrail_turn_detection_update`` would retry.
+                sent = await self._send_to_backend(message)
+                if guardrail_turn_detection_injected and sent:
+                    self._guardrail_turn_detection_update_sent = True
 
         except Exception as e:
             verbose_logger.debug(f"Error in client ack messages: {e}")

litellm/litellm_core_utils/sensitive_data_masker.py

@@ -146,6 +146,37 @@ class SensitiveDataMasker:
         return masked_data
 
 
+_default_masker = SensitiveDataMasker()
+
+
+def mask_sensitive_keys(
+    data: Dict[str, Any], sensitive_fields: Set[str]
+) -> Dict[str, Any]:
+    """Return a new dict with values masked for keys listed in ``sensitive_fields``.
+
+    Unlike :meth:`SensitiveDataMasker.mask_dict`, this does exact key-name
+    matching (not segment matching), so callers explicitly enumerate which
+    fields to mask. Non-string and None values are passed through unchanged.
+
+    Values shorter than ``visible_prefix + visible_suffix`` (8 by default)
+    fall outside :meth:`SensitiveDataMasker._mask_value`'s partial-reveal
+    range and are replaced with a fixed-length all-mask string, so a short
+    credential is never returned verbatim.
+    """
+    masked: Dict[str, Any] = {}
+    mask_char = _default_masker.mask_char
+    min_visible = _default_masker.visible_prefix + _default_masker.visible_suffix
+    for key, value in data.items():
+        if value is not None and key in sensitive_fields and isinstance(value, str):
+            if len(value) < min_visible:
+                masked[key] = mask_char * len(value) if value else value
+            else:
+                masked[key] = _default_masker._mask_value(value)
+        else:
+            masked[key] = value
+    return masked
+
+
 # Usage example:
 """
 masker = SensitiveDataMasker()

litellm/litellm_core_utils/streaming_handler.py

@@ -59,6 +59,8 @@ FUNCTION_CALL_ATTRIBUTE = "function_call"
 
 _SYNC_ITER_EXHAUSTED = object()
 
+_GCHUNK_FIELDS: frozenset = frozenset(GChunk.__annotations__)
+
 
 def _next_sync_or_exhausted(it: Any) -> Any:
     """
@@ -181,6 +183,30 @@ class CustomStreamWrapper:
         self.created: Optional[int] = None
         self._last_returned_hidden_params: Optional[dict] = None
 
+        _cached_logging_provider = self.logging_obj.model_call_details.get(
+            "custom_llm_provider", None
+        )
+        self._cached_logging_llm_provider: Optional[str] = _cached_logging_provider
+        _effective_model = model or ""
+        if (
+            custom_llm_provider == "openai"
+            and custom_llm_provider != _cached_logging_provider
+        ):
+            _effective_model = "{}/{}".format(
+                _cached_logging_provider, _effective_model
+            )
+        self._cached_model_name: str = _effective_model
+
+        # Snapshot assumes self._hidden_params is populated from litellm_params
+        # at init and never mutated during the stream. If that ever changes,
+        # this cache must be removed.
+        self._base_hidden_params: Dict[str, Any] = {
+            **self._hidden_params,
+            "response_cost": None,
+        }
+
+        self._post_streaming_hooks: Optional[List] = None
+
     def _check_max_streaming_duration(self) -> None:
         """Raise litellm.Timeout if the stream has exceeded LITELLM_MAX_STREAMING_DURATION_SECONDS."""
         from litellm.constants import LITELLM_MAX_STREAMING_DURATION_SECONDS
@@ -681,29 +707,16 @@ class CustomStreamWrapper:
     def model_response_creator(
         self, chunk: Optional[dict] = None, hidden_params: Optional[dict] = None
     ):
-        _model = self.model
-        _received_llm_provider = self.custom_llm_provider
-        _logging_obj_llm_provider = self.logging_obj.model_call_details.get("custom_llm_provider", None)  # type: ignore
-        if (
-            _received_llm_provider == "openai"
-            and _received_llm_provider != _logging_obj_llm_provider
-        ):
-            _model = "{}/{}".format(_logging_obj_llm_provider, _model)
+        _model = self._cached_model_name
+        _logging_obj_llm_provider = self._cached_logging_llm_provider
+
         if chunk is None:
-            chunk = {}
+            args: Dict[str, Any] = {"model": _model}
         else:
-            # pop model keyword
             chunk.pop("model", None)
-
-        chunk_dict = {}
-        for key, value in chunk.items():
-            if key != "stream":
-                chunk_dict[key] = value
-
-        args = {
-            "model": _model,
-            **chunk_dict,
-        }
+            args = {"model": _model}
+            if chunk:
+                args.update({k: v for k, v in chunk.items() if k != "stream"})
 
         model_response = ModelResponseStream(**args)
         if self.response_id is not None:
@@ -717,15 +730,23 @@ class CustomStreamWrapper:
             model_response.created = self.created
         else:
             self.created = model_response.created
+
+        # Spread order is load-bearing: _base_hidden_params (model_id, api_base, ...)
+        # must win over both caller-supplied hidden_params and the computed
+        # custom_llm_provider/created_at values, so it comes last.
         if hidden_params is not None:
-            model_response._hidden_params = hidden_params
-        model_response._hidden_params["custom_llm_provider"] = _logging_obj_llm_provider
-        model_response._hidden_params["created_at"] = time.time()
-        model_response._hidden_params = {
-            **model_response._hidden_params,
-            **self._hidden_params,
-            "response_cost": None,
-        }
+            model_response._hidden_params = {
+                **hidden_params,
+                "custom_llm_provider": _logging_obj_llm_provider,
+                "created_at": time.time(),
+                **self._base_hidden_params,
+            }
+        else:
+            model_response._hidden_params = {
+                "custom_llm_provider": _logging_obj_llm_provider,
+                "created_at": time.time(),
+                **self._base_hidden_params,
+            }
 
         if (
             len(model_response.choices) > 0
@@ -1627,7 +1648,17 @@ class CustomStreamWrapper:
             from litellm.integrations.custom_logger import CustomLogger
             from litellm.types.utils import CallTypes
 
-            # Get request kwargs from logging object
+            if self._post_streaming_hooks is None:
+                self._post_streaming_hooks = [
+                    cb
+                    for cb in litellm.callbacks
+                    if isinstance(cb, CustomLogger)
+                    and hasattr(cb, "async_post_call_streaming_deployment_hook")
+                ]
+
+            if not self._post_streaming_hooks:
+                return chunk
+
             request_data = self.logging_obj.model_call_details
             call_type_str = self.logging_obj.call_type
 
@@ -1636,18 +1667,14 @@ class CustomStreamWrapper:
             except ValueError:
                 typed_call_type = None
 
-            # Call hooks for all callbacks
-            for callback in litellm.callbacks:
-                if isinstance(callback, CustomLogger) and hasattr(
-                    callback, "async_post_call_streaming_deployment_hook"
-                ):
-                    result = await callback.async_post_call_streaming_deployment_hook(
-                        request_data=request_data,
-                        response_chunk=chunk,
-                        call_type=typed_call_type,
-                    )
-                    if result is not None:
-                        chunk = result
+            for callback in self._post_streaming_hooks:
+                result = await callback.async_post_call_streaming_deployment_hook(
+                    request_data=request_data,
+                    response_chunk=chunk,
+                    call_type=typed_call_type,
+                )
+                if result is not None:
+                    chunk = result
 
             return chunk
         except Exception as e:
@@ -1808,8 +1835,10 @@ class CustomStreamWrapper:
                     processed_chunk, None, None, cache_hit
                 )
             )
-        ## SYNC LOGGING
-        self.logging_obj.success_handler(processed_chunk, None, None, cache_hit)
+        ## SYNC LOGGING — only for sync SDK entrypoints; async proxy paths export via async_success_handler
+        litellm_params = self.logging_obj.model_call_details.get("litellm_params", {})
+        if self.logging_obj._is_sync_litellm_request(litellm_params):
+            self.logging_obj.success_handler(processed_chunk, None, None, cache_hit)
 
     def finish_reason_handler(self):
         model_response = self.model_response_creator()
@@ -1888,17 +1917,15 @@ class CustomStreamWrapper:
                         response = self._add_mcp_list_tools_to_first_chunk(response)
                         self.sent_first_chunk = True
 
-                    if hasattr(
-                        response, "usage"
-                    ):  # remove usage from chunk, only send on final chunk
-                        # Convert the object to a dictionary
+                    # ModelResponseStream declares `usage` as a field, so
+                    # hasattr(response, "usage") is always True — must check
+                    # `is not None` to avoid running this path on every chunk.
+                    if getattr(response, "usage", None) is not None:
                         obj_dict = response.model_dump()
 
-                        # Remove an attribute (e.g., 'attr2')
                         if "usage" in obj_dict:
                             del obj_dict["usage"]
 
-                        # Create a new object without the removed attribute
                         response = self.model_response_creator(
                             chunk=obj_dict, hidden_params=response._hidden_params
                         )
@@ -2206,23 +2233,19 @@ class CustomStreamWrapper:
                         cache_hit,
                     )
                 else:
+                    # prefer_async_handlers routes CustomLogger to async_success_handler
+                    # when consumers use ``async for`` on sync-SDK streams. Legacy string
+                    # callbacks still run via executor.submit inside dispatch_success_handlers.
                     asyncio.create_task(
-                        self.logging_obj.async_success_handler(
+                        self.logging_obj.dispatch_success_handlers(
                             complete_streaming_response,
                             cache_hit=cache_hit,
                             start_time=None,
                             end_time=None,
+                            prefer_async_handlers=True,
                         )
                     )
 
-                    executor.submit(
-                        self.logging_obj.success_handler,
-                        complete_streaming_response,
-                        cache_hit=cache_hit,
-                        start_time=None,
-                        end_time=None,
-                    )
-
                 raise StopAsyncIteration  # Re-raise StopIteration
             else:
                 self.sent_last_chunk = True
@@ -2398,10 +2421,7 @@ def generic_chunk_has_all_required_fields(chunk: dict) -> bool:
     :param chunk: The dictionary to check.
     :return: True if all required fields are present, False otherwise.
     """
-    _all_fields = GChunk.__annotations__
-
-    decision = all(key in _all_fields for key in chunk)
-    return decision
+    return all(key in _GCHUNK_FIELDS for key in chunk)
 
 
 def convert_generic_chunk_to_model_response_stream(

litellm/litellm_core_utils/token_counter.py

@@ -486,6 +486,14 @@ def _count_messages(
                     use_default_image_token_count,
                     default_token_count,
                 )
+            elif key == "search_results" and isinstance(value, list):
+                from litellm.litellm_core_utils.prompt_templates.common_utils import (
+                    extract_search_results_text,
+                )
+
+                search_results_text = extract_search_results_text(value)
+                if search_results_text:
+                    num_tokens += params.count_function(search_results_text)
             else:
                 # Skip unsupported keys instead of raising an error
                 continue
@@ -764,11 +772,29 @@ def _format_function_definitions(tools):
     lines.append("namespace functions {")
     lines.append("")
     for tool in tools:
+        if not isinstance(tool, dict):
+            continue
         function = tool.get("function")
+        if not isinstance(function, dict):
+            # Anthropic tool shape → OpenAI function dict for token counting.
+            params = tool.get("input_schema") or tool.get("parameters") or {}
+            if not isinstance(params, dict):
+                params = {}
+            function = {
+                "name": tool.get("name"),
+                "description": tool.get("description"),
+                "parameters": params,
+            }
+        function_name = function.get("name")
+        if not function_name:
+            # Skip malformed tools missing a name to avoid emitting
+            # ``type None = ...`` which would produce inaccurate token counts.
+            continue
         if function_description := function.get("description"):
             lines.append(f"// {function_description}")
-        function_name = function.get("name")
-        parameters = function.get("parameters", {})
+        parameters = function.get("parameters") or {}
+        if not isinstance(parameters, dict):
+            parameters = {}
         properties = parameters.get("properties")
         if properties and properties.keys():
             lines.append(f"type {function_name} = (_: {{")

litellm/llms/anthropic/chat/transformation.py

@@ -337,13 +337,12 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
         )
 
     @staticmethod
-    def _supports_effort_level(model: str, level: str) -> bool:
-        """Check ``supports_{level}_reasoning_effort`` in the model map.
+    def _supports_model_capability(model: str, key: str) -> bool:
+        """Check a boolean capability ``key`` in the model map.
 
         Strips bedrock/vertex prefixes so a provider-routed Claude still
         resolves to the Anthropic model-map entry.
         """
-        key = f"supports_{level}_reasoning_effort"
         try:
             if _supports_factory(
                 model=model,
@@ -372,8 +371,6 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
         except Exception:
             pass
         try:
-            import litellm
-
             for cand in candidates:
                 if cand in litellm.model_cost and (
                     litellm.model_cost[cand].get(key) is True
@@ -383,6 +380,13 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
             pass
         return False
 
+    @staticmethod
+    def _supports_effort_level(model: str, level: str) -> bool:
+        """Check ``supports_{level}_reasoning_effort`` in the model map."""
+        return AnthropicConfig._supports_model_capability(
+            model, f"supports_{level}_reasoning_effort"
+        )
+
     @staticmethod
     def _validate_effort_for_model(model: str, effort: Optional[str]) -> Optional[str]:
         """Return ``None`` if ``effort`` is allowed on ``model``, else an error message."""
@@ -400,7 +404,15 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
 
     @staticmethod
     def _model_supports_effort_param(model: str) -> bool:
-        """Whether the model accepts ``output_config.effort`` at all."""
+        """Whether the model accepts ``output_config.effort`` at all.
+
+        A model qualifies if its map entry advertises ``supports_output_config``
+        or any ``supports_*_reasoning_effort`` flag. The two are independent
+        signals: e.g. Claude Opus 4.5 supports ``output_config`` without
+        advertising a non-default (max/xhigh) effort level.
+        """
+        if AnthropicConfig._supports_model_capability(model, "supports_output_config"):
+            return True
         return any(
             AnthropicConfig._supports_effort_level(model, level)
             for level in ("low", "minimal", "medium", "high", "xhigh", "max")
@@ -1793,7 +1805,10 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
             self._ensure_context_management_beta_header(
                 headers, optional_params["context_management"]
             )
-        if optional_params.get("output_format") is not None:
+        output_config = optional_params.get("output_config")
+        if optional_params.get("output_format") is not None or (
+            isinstance(output_config, dict) and output_config.get("format") is not None
+        ):
             self._ensure_beta_header(
                 headers, ANTHROPIC_BETA_HEADER_VALUES.STRUCTURED_OUTPUT_2025_09_25.value
             )

litellm/llms/anthropic/experimental_pass_through/adapters/handler.py

@@ -4,6 +4,7 @@ from typing import (
     AsyncIterator,
     Coroutine,
     Dict,
+    Iterator,
     List,
     Optional,
     Tuple,
@@ -12,9 +13,16 @@ from typing import (
 )
 
 import litellm
+from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.asyncify import run_async_function
 from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
     AnthropicAdapter,
 )
+from litellm.llms.anthropic.experimental_pass_through.context_management import (
+    AnthropicContextManagementError,
+    PolyfillResult,
+    apply_context_management,
+)
 from litellm.llms.anthropic.experimental_pass_through.utils import (
     is_reasoning_auto_summary_enabled,
 )
@@ -28,15 +36,266 @@ if TYPE_CHECKING:
     pass
 
 
-# Anthropic-only fields that the translator above already maps into the
-# OpenAI-format completion_kwargs (output_config → reasoning_effort /
-# response_format, etc.). They must be filtered out of the raw
-# extra_kwargs re-merge below or non-Anthropic backends reject the call
-# with 400 "Extra inputs are not permitted". Add new entries here when
-# extending AnthropicMessagesRequestOptionalParams with another Anthropic-
-# specific key.
+# Anthropic-only keys already mapped by the translator; strip on extra_kwargs re-merge.
 ANTHROPIC_ONLY_REQUEST_KEYS: frozenset[str] = frozenset({"output_config"})
 
+
+def _messages_have_compaction_block(messages: List[Dict]) -> bool:
+    """Return True when any message carries a ``compaction`` content block."""
+    for msg in messages:
+        content = msg.get("content")
+        if not isinstance(content, list):
+            continue
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "compaction":
+                return True
+    return False
+
+
+def _extract_proxy_litellm_metadata(kwargs: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    """Return ``kwargs["litellm_metadata"]`` when it's a dict; ``None`` otherwise.
+
+    The proxy attaches its auth/spend-attribution fields (``user_api_key``,
+    ``user_api_key_team_id``, ``litellm_call_id``, the full ``UserAPIKeyAuth``
+    object under ``user_api_key_auth``, ...) to ``data["litellm_metadata"]``
+    for ``/v1/messages`` (see
+    ``LiteLLMProxyRequestSetup.add_user_api_key_auth_to_request_metadata`` and
+    ``LITELLM_METADATA_ROUTES``). The Anthropic-shape ``metadata`` arg only
+    carries ``user_id`` and must not be conflated. Returns ``None`` for SDK
+    callers that bypass the proxy entirely.
+    """
+    litellm_metadata = kwargs.get("litellm_metadata")
+    if not isinstance(litellm_metadata, dict):
+        return None
+    return litellm_metadata
+
+
+async def _prepare_context_managed_request(
+    *,
+    model: str,
+    messages: List[Dict],
+    tools: Optional[List[Dict]],
+    system: Optional[Any],
+    context_management_spec: Any,
+    litellm_metadata: Optional[Dict],
+    drop_params: Optional[bool],
+    llm_router: Any,
+    user_api_key_auth: Any = None,
+) -> Optional[PolyfillResult]:
+    """Apply client compaction history, then optional context_management polyfill."""
+    from litellm.llms.anthropic.experimental_pass_through.context_management.editors.compact import (
+        apply_client_compaction_block_history,
+    )
+
+    # Skip the client-history pre-processing when a ``compact_20260112``
+    # polyfill spec will run: that editor already slices around any client-sent
+    # compaction block in its Phase A (and uses the full post-compaction tail
+    # for its token-threshold check). Pre-collapsing to just the latest user
+    # question here would starve the polyfill of conversation context and
+    # silently drop intermediate turns.
+    polyfill_will_run = _polyfill_will_run(
+        context_management_spec=context_management_spec,
+        drop_params=drop_params,
+    )
+
+    if polyfill_will_run:
+        history_result: Optional[PolyfillResult] = None
+        working_messages: List[Dict] = messages
+        working_system: Optional[Any] = system
+    else:
+        history_result = apply_client_compaction_block_history(
+            messages=cast(List[Dict[str, Any]], messages),
+            system=system,
+        )
+        working_messages = (
+            history_result.messages if history_result is not None else messages
+        )
+        working_system = history_result.system if history_result is not None else system
+
+    polyfill_result = await _run_polyfill_if_enabled(
+        model=model,
+        messages=working_messages,
+        tools=tools,
+        system=working_system,
+        context_management_spec=context_management_spec,
+        litellm_metadata=litellm_metadata,
+        drop_params=drop_params,
+        llm_router=llm_router,
+        user_api_key_auth=user_api_key_auth,
+    )
+
+    if polyfill_result is not None:
+        return polyfill_result
+
+    # Safety net: if we skipped client-history pre-processing because a
+    # ``compact_20260112`` polyfill was expected to handle the compaction
+    # block itself but the polyfill ultimately did not produce a result
+    # (e.g. it crashed and was best-effort swallowed in
+    # ``_run_polyfill_if_enabled``), apply the slice-only fallback now so
+    # Anthropic-specific ``compaction`` content blocks don't leak through
+    # to non-Anthropic backends that would reject them.
+    if polyfill_will_run and history_result is None:
+        history_result = apply_client_compaction_block_history(
+            messages=cast(List[Dict[str, Any]], messages),
+            system=system,
+        )
+    return history_result
+
+
+def _polyfill_will_run(
+    *,
+    context_management_spec: Any,
+    drop_params: Optional[bool],
+) -> bool:
+    """Return True when ``compact_20260112`` will run via the polyfill dispatcher.
+
+    Mirrors the gating in ``_run_polyfill_if_enabled``: an empty spec or
+    effective ``drop_params`` short-circuits the polyfill. The pre-processing
+    skip only applies when the dispatcher will actually invoke
+    ``apply_compact_20260112`` (which has its own compaction-block slicing).
+    """
+    edits = _normalize_spec_edits(
+        context_management_spec=context_management_spec,
+        drop_params=drop_params,
+    )
+    if edits is None:
+        return False
+
+    from litellm.llms.anthropic.experimental_pass_through.context_management.constants import (
+        COMPACT_EDIT_TYPE,
+    )
+
+    return any(
+        isinstance(edit, dict) and edit.get("type") == COMPACT_EDIT_TYPE
+        for edit in edits
+    )
+
+
+def _spec_has_non_compact_edits(
+    *,
+    context_management_spec: Any,
+    drop_params: Optional[bool],
+) -> bool:
+    """Return True when the spec includes edits other than ``compact_20260112``.
+
+    Used to decide whether a polyfill failure can be silently swallowed
+    (compact-only specs have a safe compaction-block slicing fallback) or
+    must be surfaced (other editors like ``clear_tool_uses_20250919`` have
+    no slice-only fallback and would otherwise be dropped without notice).
+    """
+    edits = _normalize_spec_edits(
+        context_management_spec=context_management_spec,
+        drop_params=drop_params,
+    )
+    if edits is None:
+        return False
+
+    from litellm.llms.anthropic.experimental_pass_through.context_management.constants import (
+        COMPACT_EDIT_TYPE,
+    )
+
+    return any(
+        isinstance(edit, dict)
+        and isinstance(edit.get("type"), str)
+        and edit.get("type") != COMPACT_EDIT_TYPE
+        for edit in edits
+    )
+
+
+def _normalize_spec_edits(
+    *,
+    context_management_spec: Any,
+    drop_params: Optional[bool],
+) -> Optional[List[Dict[str, Any]]]:
+    """Return the normalized ``edits`` list, or ``None`` if the polyfill won't run.
+
+    Delegates spec-shape normalization to the dispatcher's ``_normalize_spec``
+    so the prediction here can't drift from what the dispatcher actually does.
+    """
+    if not context_management_spec:
+        return None
+
+    effective_drop_params = (
+        drop_params if drop_params is not None else litellm.drop_params
+    )
+    if effective_drop_params:
+        return None
+
+    from litellm.llms.anthropic.experimental_pass_through.context_management.dispatcher import (
+        _normalize_spec,
+    )
+
+    try:
+        return _normalize_spec(context_management_spec)
+    except Exception:
+        return None
+
+
+async def _run_polyfill_if_enabled(
+    *,
+    model: str,
+    messages: List[Dict],
+    tools: Optional[List[Dict]],
+    system: Optional[Any],
+    context_management_spec: Any,
+    litellm_metadata: Optional[Dict],
+    drop_params: Optional[bool],
+    llm_router: Any,
+    user_api_key_auth: Any = None,
+) -> Optional[PolyfillResult]:
+    """Run the async context_management polyfill if a spec is present.
+
+    Returns ``None`` when the spec is empty or drop_params is on. Raises
+    ``AnthropicContextManagementError`` so the /v1/messages endpoint can
+    emit an Anthropic-format 400. All other exceptions are best-effort
+    swallowed (matches v0 behavior).
+    """
+    if not context_management_spec:
+        return None
+
+    effective_drop_params = (
+        drop_params if drop_params is not None else litellm.drop_params
+    )
+    if effective_drop_params:
+        return None
+
+    try:
+        return await apply_context_management(
+            model=model,
+            messages=messages,
+            tools=tools,
+            system=system,
+            context_management_spec=context_management_spec,
+            litellm_metadata=litellm_metadata,
+            llm_router=llm_router,
+            user_api_key_auth=user_api_key_auth,
+        )
+    except AnthropicContextManagementError:
+        # Surface validation errors so the endpoint can emit an Anthropic-format
+        # 400. Other exception types fall into the best-effort branch below.
+        raise
+    except Exception as e:
+        verbose_logger.exception(
+            "context_management polyfill: skipping edits due to error: %s", e
+        )
+        # Best-effort swallow is only safe for compact-only specs, where the
+        # caller's compaction-block-slicing safety net produces a correct
+        # (if degraded) result. When the spec also requested non-compact
+        # edits (e.g. ``clear_tool_uses_20250919``), the safety net does
+        # NOT re-run those editors, so silently returning ``None`` would
+        # drop them with no error surface. Raise instead so the endpoint
+        # emits an Anthropic-format error.
+        if _spec_has_non_compact_edits(
+            context_management_spec=context_management_spec,
+            drop_params=drop_params,
+        ):
+            raise AnthropicContextManagementError(
+                status_code=500,
+                message=f"context_management polyfill failed: {e}",
+            ) from e
+        return None
+
+
 ########################################################
 # init adapter
 ANTHROPIC_ADAPTER = AnthropicAdapter()
@@ -163,7 +422,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
         metadata: Optional[Dict] = None,
         stop_sequences: Optional[List[str]] = None,
         stream: Optional[bool] = False,
-        system: Optional[str] = None,
+        system: Optional[Union[str, List[Dict[str, Any]]]] = None,
         temperature: Optional[float] = None,
         thinking: Optional[Dict] = None,
         tool_choice: Optional[Dict] = None,
@@ -307,19 +566,56 @@ class LiteLLMMessagesToCompletionTransformationHandler:
         top_p: Optional[float] = None,
         output_format: Optional[Dict] = None,
         **kwargs,
-    ) -> Union[AnthropicMessagesResponse, AsyncIterator]:
+    ) -> Union[AnthropicMessagesResponse, AsyncIterator[Any], Iterator[bytes]]:
         """Handle non-Anthropic models asynchronously using the adapter"""
+        context_management = kwargs.pop("context_management", None)
+        drop_params: Optional[bool] = kwargs.get("drop_params", None)
+        litellm_router = kwargs.pop("litellm_router", None)
+        if litellm_router is None:
+            try:
+                from litellm.proxy.proxy_server import llm_router as _proxy_router
+
+                litellm_router = _proxy_router
+            except Exception:
+                pass
+
+        proxy_litellm_metadata = _extract_proxy_litellm_metadata(kwargs)
+        user_api_key_auth = (
+            proxy_litellm_metadata.get("user_api_key_auth")
+            if proxy_litellm_metadata is not None
+            else None
+        )
+
+        polyfill_result = await _prepare_context_managed_request(
+            model=model,
+            messages=messages,
+            tools=tools,
+            system=system,
+            context_management_spec=context_management,
+            litellm_metadata=proxy_litellm_metadata,
+            drop_params=drop_params,
+            llm_router=litellm_router,
+            user_api_key_auth=user_api_key_auth,
+        )
+
+        effective_messages = (
+            polyfill_result.messages if polyfill_result is not None else messages
+        )
+        effective_system = (
+            polyfill_result.system if polyfill_result is not None else system
+        )
+
         (
             completion_kwargs,
             tool_name_mapping,
         ) = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs(
             max_tokens=max_tokens,
-            messages=messages,
+            messages=effective_messages,
             model=model,
             metadata=metadata,
             stop_sequences=stop_sequences,
             stream=stream,
-            system=system,
+            system=effective_system,
             temperature=temperature,
             thinking=thinking,
             tool_choice=tool_choice,
@@ -338,6 +634,8 @@ class LiteLLMMessagesToCompletionTransformationHandler:
                     completion_response,
                     model=model,
                     tool_name_mapping=tool_name_mapping,
+                    polyfill_result=polyfill_result,
+                    is_async=True,
                 )
             )
             if transformed_stream is not None:
@@ -347,6 +645,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             anthropic_response = ANTHROPIC_ADAPTER.translate_completion_output_params(
                 cast(ModelResponse, completion_response),
                 tool_name_mapping=tool_name_mapping,
+                polyfill_result=polyfill_result,
             )
             if anthropic_response is not None:
                 return anthropic_response
@@ -372,8 +671,13 @@ class LiteLLMMessagesToCompletionTransformationHandler:
         **kwargs,
     ) -> Union[
         AnthropicMessagesResponse,
+        Iterator[bytes],
         AsyncIterator[Any],
-        Coroutine[Any, Any, Union[AnthropicMessagesResponse, AsyncIterator[Any]]],
+        Coroutine[
+            Any,
+            Any,
+            Union[AnthropicMessagesResponse, AsyncIterator[Any], Iterator[bytes]],
+        ],
     ]:
         """Handle non-Anthropic models using the adapter."""
         if _is_async is True:
@@ -395,17 +699,72 @@ class LiteLLMMessagesToCompletionTransformationHandler:
                 **kwargs,
             )
 
+        # Run the context_management polyfill on the sync path too so that
+        # ``litellm.messages.create()`` callers don't silently lose edits like
+        # ``clear_tool_uses_20250919``. The dispatcher is async (so the
+        # ``compact_20260112`` editor can ``await`` the summarization model);
+        # bridge to it via ``run_async_function``.
+        context_management = kwargs.pop("context_management", None)
+        drop_params: Optional[bool] = kwargs.get("drop_params", None)
+        # Deliberately do NOT auto-attach the proxy ``llm_router`` here:
+        # ``run_async_function`` spawns a new event loop in a worker thread
+        # to bridge to the async dispatcher, but the proxy router's httpx
+        # ``AsyncClient`` instances are bound to the proxy's main event loop.
+        # Reusing them from the new thread's loop violates httpx's single-loop
+        # invariant and can raise ``RuntimeError: Event loop is closed`` or
+        # produce stalled connections. The summary editor falls back to
+        # ``litellm.acompletion`` (which creates a fresh client per call) when
+        # ``llm_router`` is ``None``, which is safe to call from the bridged
+        # loop. The async ``async_anthropic_messages_handler`` path is
+        # unaffected because it ``await``s within the original event loop.
+        litellm_router = kwargs.pop("litellm_router", None)
+
+        # Skip the async bridge entirely when there is nothing for either the
+        # polyfill or the client-history slice-only fallback to do. The vast
+        # majority of sync ``litellm.messages.create()`` requests carry no
+        # ``context_management`` spec and no client-sent ``compaction`` block,
+        # and bridging through a worker-thread event loop just to discover
+        # there is no work is pure overhead.
+        if context_management is None and not _messages_have_compaction_block(messages):
+            polyfill_result: Optional[PolyfillResult] = None
+        else:
+            proxy_litellm_metadata = _extract_proxy_litellm_metadata(kwargs)
+            user_api_key_auth = (
+                proxy_litellm_metadata.get("user_api_key_auth")
+                if proxy_litellm_metadata is not None
+                else None
+            )
+            polyfill_result = run_async_function(
+                _prepare_context_managed_request,
+                model=model,
+                messages=messages,
+                tools=tools,
+                system=system,
+                context_management_spec=context_management,
+                litellm_metadata=proxy_litellm_metadata,
+                drop_params=drop_params,
+                llm_router=litellm_router,
+                user_api_key_auth=user_api_key_auth,
+            )
+
+        effective_messages = (
+            polyfill_result.messages if polyfill_result is not None else messages
+        )
+        effective_system = (
+            polyfill_result.system if polyfill_result is not None else system
+        )
+
         (
             completion_kwargs,
             tool_name_mapping,
         ) = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs(
             max_tokens=max_tokens,
-            messages=messages,
+            messages=effective_messages,
             model=model,
             metadata=metadata,
             stop_sequences=stop_sequences,
             stream=stream,
-            system=system,
+            system=effective_system,
             temperature=temperature,
             thinking=thinking,
             tool_choice=tool_choice,
@@ -424,6 +783,8 @@ class LiteLLMMessagesToCompletionTransformationHandler:
                     completion_response,
                     model=model,
                     tool_name_mapping=tool_name_mapping,
+                    polyfill_result=polyfill_result,
+                    is_async=False,
                 )
             )
             if transformed_stream is not None:
@@ -433,6 +794,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             anthropic_response = ANTHROPIC_ADAPTER.translate_completion_output_params(
                 cast(ModelResponse, completion_response),
                 tool_name_mapping=tool_name_mapping,
+                polyfill_result=polyfill_result,
             )
             if anthropic_response is not None:
                 return anthropic_response

litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py

@@ -3,11 +3,26 @@
 import json
 import traceback
 from collections import deque
-from typing import TYPE_CHECKING, Any, AsyncIterator, Dict, Iterator, Literal, Optional
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    AsyncIterator,
+    Dict,
+    Iterator,
+    List,
+    Literal,
+    Optional,
+)
 
-from litellm import verbose_logger
+from litellm._logging import verbose_logger
 from litellm._uuid import uuid
-from litellm.types.llms.anthropic import UsageDelta
+from litellm.types.llms.anthropic import (
+    AppliedEdit,
+    CompactionBlock,
+    ContextManagementResponse,
+    UsageDelta,
+    UsageIteration,
+)
 from litellm.types.utils import AdapterCompletionStreamWrapper
 
 if TYPE_CHECKING:
@@ -37,22 +52,208 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
     holding_stop_reason_chunk: Optional[Any] = None
     queued_usage_chunk: bool = False
     current_content_block_index: int = 0
-    current_content_block_start: ContentBlockContentBlockDict = TextBlock(
-        type="text",
-        text="",
-    )
-    chunk_queue: deque = deque()  # Queue for buffering multiple chunks
 
     def __init__(
         self,
         completion_stream: Any,
         model: str,
         tool_name_mapping: Optional[Dict[str, str]] = None,
+        applied_edits: Optional[List[AppliedEdit]] = None,
+        compaction_block: Optional[CompactionBlock] = None,
+        iterations_usage: Optional[List[UsageIteration]] = None,
     ):
         super().__init__(completion_stream)
         self.model = model
         # Mapping of truncated tool names to original names (for OpenAI's 64-char limit)
         self.tool_name_mapping = tool_name_mapping or {}
+        # Polyfill applied_edits on final message_delta.
+        self.applied_edits: List[AppliedEdit] = list(applied_edits or [])
+        # Synthesized compaction block from compact_20260112 polyfill (streaming).
+        self.compaction_block = compaction_block
+        self.iterations_usage = iterations_usage
+        self.sent_compaction_block: bool = False
+        # Per-phase flags so the compaction block's start/delta/stop events
+        # are emitted (and the public state machine is advanced) in
+        # lock-step with the caller actually consuming each event. Pre-
+        # queuing all three would set ``sent_content_block_finish=True``
+        # before the client received ``content_block_stop``, leaving the
+        # observable state inconsistent during the drain window.
+        self.sent_compaction_block_start: bool = False
+        self.sent_compaction_block_delta: bool = False
+        # Per-instance queue for buffering multiple chunks. Must be initialized
+        # here (not at class level) so concurrent streams don't share the same
+        # deque and corrupt each other's SSE event order.
+        self.chunk_queue: deque = deque()
+        # Per-instance default content block. Must be initialized here (not at
+        # class level) so concurrent streams don't share the same mutable dict
+        # — `_should_start_new_content_block` mutates `tool_block["name"]` in
+        # place, which would otherwise leak across streams.
+        self.current_content_block_start: (
+            "AnthropicStreamWrapper.ContentBlockContentBlockDict"
+        ) = self.TextBlock(
+            type="text",
+            text="",
+        )
+
+    def _merge_usage_into_held_stop_reason_chunk(self, chunk: Any) -> Dict[str, Any]:
+        """Merge usage data from ``chunk`` into the held ``message_delta`` chunk.
+
+        Shared by both the sync ``__next__`` and async ``__anext__`` paths so
+        the subtle hold-and-merge logic (cache tokens, ``context_management``
+        attachment, ``UsageDelta`` shape) lives in exactly one place.
+
+        Caller is responsible for managing ``self.holding_stop_reason_chunk``
+        and ``self.queued_usage_chunk`` state and for queuing the returned
+        merged chunk.
+        """
+        assert self.holding_stop_reason_chunk is not None
+        merged_chunk = self.holding_stop_reason_chunk.copy()
+        if "delta" not in merged_chunk:
+            merged_chunk["delta"] = {}
+
+        uncached_input_tokens = chunk.usage.prompt_tokens or 0
+        if (
+            hasattr(chunk.usage, "prompt_tokens_details")
+            and chunk.usage.prompt_tokens_details
+        ):
+            cached_tokens = (
+                getattr(chunk.usage.prompt_tokens_details, "cached_tokens", 0) or 0
+            )
+            uncached_input_tokens -= cached_tokens
+
+        usage_dict: UsageDelta = {
+            "input_tokens": uncached_input_tokens,
+            "output_tokens": chunk.usage.completion_tokens or 0,
+        }
+        if (
+            hasattr(chunk.usage, "_cache_creation_input_tokens")
+            and chunk.usage._cache_creation_input_tokens > 0
+        ):
+            usage_dict["cache_creation_input_tokens"] = (
+                chunk.usage._cache_creation_input_tokens
+            )
+        if (
+            hasattr(chunk.usage, "_cache_read_input_tokens")
+            and chunk.usage._cache_read_input_tokens > 0
+        ):
+            usage_dict["cache_read_input_tokens"] = chunk.usage._cache_read_input_tokens
+        merged_chunk["usage"] = usage_dict
+        if self.applied_edits and "context_management" not in merged_chunk:
+            merged_chunk["context_management"] = ContextManagementResponse(
+                applied_edits=list(self.applied_edits)
+            )
+        return self._augment_message_delta_usage(merged_chunk)
+
+    def _ensure_context_management_attached(
+        self, message_delta_chunk: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """Attach ``context_management`` to a ``message_delta`` chunk if
+        ``self.applied_edits`` is non-empty and the chunk does not already
+        carry it. Returns the (possibly new) chunk dict.
+
+        Centralizing this guard ensures every ``message_delta`` emission
+        path (merge-with-usage and direct-flush-of-held) consistently
+        surfaces ``applied_edits`` to the client.
+        """
+        if not self.applied_edits or "context_management" in message_delta_chunk:
+            return message_delta_chunk
+        augmented = message_delta_chunk.copy()
+        augmented["context_management"] = ContextManagementResponse(
+            applied_edits=list(self.applied_edits)
+        )
+        return augmented
+
+    def _augment_message_delta_usage(
+        self, message_delta_chunk: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """Attach polyfill compaction iteration usage to the final message_delta.
+
+        Also defensively re-attaches ``context_management`` so the direct
+        held-chunk flush path stays in sync with the merge path's guarantee
+        when ``self.applied_edits`` is non-empty.
+        """
+        message_delta_chunk = self._ensure_context_management_attached(
+            message_delta_chunk
+        )
+        if self.iterations_usage is None:
+            return message_delta_chunk
+        usage = message_delta_chunk.get("usage")
+        if not isinstance(usage, dict) or "iterations" in usage:
+            return message_delta_chunk
+
+        input_tokens = usage.get("input_tokens", 0) or 0
+        output_tokens = usage.get("output_tokens", 0) or 0
+        augmented = message_delta_chunk.copy()
+        augmented_usage = dict(usage)
+        iterations: List[UsageIteration] = list(self.iterations_usage)
+        # Only emit a ``message`` iteration when we have real token data.
+        # Without a separate usage chunk (e.g. provider sent finish_reason
+        # alone), the held ``message_delta`` carries placeholder zeros from
+        # the translate step; reporting a zero-token iteration would be
+        # misleading and inconsistent with the non-streaming path.
+        if input_tokens > 0 or output_tokens > 0:
+            message_iteration: UsageIteration = {
+                "type": "message",
+                "input_tokens": input_tokens,
+                "output_tokens": output_tokens,
+            }
+            iterations.append(message_iteration)
+        augmented_usage["iterations"] = iterations  # type: ignore[typeddict-unknown-key]
+        augmented["usage"] = augmented_usage
+        return augmented
+
+    def _next_compaction_event(self) -> Optional[Dict[str, Any]]:
+        """Return the next compaction content-block SSE event, or ``None``.
+
+        Anthropic delivers compaction as a single delta (no token-by-token
+        streaming), but we still surface it as a proper
+        start → delta → stop trio. Each call returns exactly one event so
+        the state machine (``sent_content_block_finish``,
+        ``current_content_block_index``) is advanced *only* when the
+        terminal stop event is actually handed back to the caller. This
+        prevents an observable window where the flags claim the block is
+        finished while the stop event is still buffered.
+        """
+        if self.compaction_block is None or self.sent_compaction_block:
+            return None
+
+        compaction_index = self.current_content_block_index
+
+        if not self.sent_compaction_block_start:
+            self.sent_compaction_block_start = True
+            return {
+                "type": "content_block_start",
+                "index": compaction_index,
+                # Mirror the text-block shape ({"type": "text", "text": ""}):
+                # send an empty ``content`` field so clients that introspect
+                # ``content_block_start`` see the full block schema. The
+                # actual summary text arrives via the ``content_block_delta``
+                # below.
+                "content_block": {"type": "compaction", "content": ""},
+            }
+
+        if not self.sent_compaction_block_delta:
+            self.sent_compaction_block_delta = True
+            summary_content = self.compaction_block.get("content") or ""
+            return {
+                "type": "content_block_delta",
+                "index": compaction_index,
+                "delta": {"type": "compaction_delta", "content": summary_content},
+            }
+
+        stop_event = {
+            "type": "content_block_stop",
+            "index": compaction_index,
+        }
+        # Don't touch ``sent_content_block_finish`` here: that flag is the
+        # state machine for the regular text/tool_use/thinking block and is
+        # independent of the synthetic compaction block lifecycle. Conflating
+        # them would let outside observers (subclass overrides, introspection
+        # hooks, exception paths) see ``sent_content_block_finish=True``
+        # without any regular content block ever having started.
+        self._increment_content_block_index()
+        self.sent_compaction_block = True
+        return stop_event
 
     def _create_initial_usage_delta(self) -> UsageDelta:
         """
@@ -75,7 +276,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             cache_read_input_tokens=0,
         )
 
-    def __next__(self):
+    def __next__(self):  # noqa: PLR0915
         from .transformation import LiteLLMAnthropicMessagesAdapter
 
         try:
@@ -103,21 +304,55 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                 )
                 return self.chunk_queue.popleft()
 
+            if (
+                self.sent_compaction_block is False
+                and self.compaction_block is not None
+            ):
+                compaction_event = self._next_compaction_event()
+                if compaction_event is not None:
+                    return compaction_event
+
+            if (
+                self.sent_compaction_block is False
+                and self.compaction_block is not None
+            ):
+                compaction_event = self._next_compaction_event()
+                if compaction_event is not None:
+                    return compaction_event
+
             for chunk in self.completion_stream:
                 if chunk == "None" or chunk is None:
                     raise Exception
 
                 should_start_new_block = self._should_start_new_content_block(chunk)
-                if should_start_new_block and self.sent_content_block_start:
+                if should_start_new_block:
                     self._increment_content_block_index()
 
+                # applied_edits only needs to flow to the final message_delta
+                # (when finish_reason is set); skip threading it through every
+                # intermediate chunk. For the hold-and-merge path below,
+                # context_management is attached directly to the merged chunk,
+                # so the translated ``processed_chunk`` would be discarded —
+                # skip the applied_edits attachment in that case to avoid
+                # allocating a throwaway ``MessageBlockDelta``.
+                will_merge_into_held = (
+                    self.holding_stop_reason_chunk is not None
+                    and getattr(chunk, "usage", None) is not None
+                )
+                is_final_chunk = chunk.choices[0].finish_reason is not None
                 processed_chunk = LiteLLMAnthropicMessagesAdapter().translate_streaming_openai_response_to_anthropic(
                     response=chunk,
                     current_content_block_index=self.current_content_block_index,
+                    applied_edits=(
+                        self.applied_edits
+                        if is_final_chunk and not will_merge_into_held
+                        else None
+                    ),
                 )
 
                 if not self.sent_content_block_start:
                     self.sent_content_block_start = True
+                    self.sent_content_block_finish = False
                     self.chunk_queue.append(
                         {
                             "type": "content_block_start",
@@ -137,6 +372,45 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                         self.chunk_queue.append(processed_chunk)
                     return self.chunk_queue.popleft()
 
+                if not self.sent_content_block_start:
+                    self.sent_content_block_start = True
+                    self.sent_content_block_finish = False
+                    self.chunk_queue.append(
+                        {
+                            "type": "content_block_start",
+                            "index": self.current_content_block_index,
+                            "content_block": self.current_content_block_start,
+                        }
+                    )
+                    if (
+                        processed_chunk.get("type") == "content_block_delta"
+                        and isinstance(processed_chunk.get("delta"), dict)
+                        and processed_chunk["delta"].get("type")
+                        in ("text_delta", "input_json_delta")
+                        and processed_chunk["delta"].get(
+                            "text", processed_chunk["delta"].get("partial_json", "")
+                        )
+                    ):
+                        self.chunk_queue.append(processed_chunk)
+                    return self.chunk_queue.popleft()
+
+                # Check if this is a usage chunk and we have a held stop_reason chunk
+                if will_merge_into_held:
+                    merged_chunk = self._merge_usage_into_held_stop_reason_chunk(chunk)
+                    self.chunk_queue.append(merged_chunk)
+                    self.queued_usage_chunk = True
+                    self.holding_stop_reason_chunk = None
+                    return self.chunk_queue.popleft()
+
+                if self.queued_usage_chunk:
+                    # Usage has already been merged + emitted. Any trailing
+                    # provider events would violate Anthropic SSE ordering
+                    # (no chunks may follow the final ``message_delta``), so
+                    # silently drop them — matches the async ``__anext__``
+                    # behavior where the block-handling logic is gated on
+                    # ``not self.queued_usage_chunk``.
+                    continue
+
                 if should_start_new_block and not self.sent_content_block_finish:
                     # Queue the sequence: content_block_stop -> content_block_start
                     # For text blocks the trigger chunk is not emitted as a separate
@@ -188,20 +462,64 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                         }
                     )
                     self.sent_content_block_finish = True
-                    self.chunk_queue.append(processed_chunk)
+                    if processed_chunk.get("delta", {}).get("stop_reason") is not None:
+                        self.holding_stop_reason_chunk = processed_chunk
+                    else:
+                        processed_chunk = self._augment_message_delta_usage(
+                            processed_chunk
+                        )
+                        self.chunk_queue.append(processed_chunk)
                     return self.chunk_queue.popleft()
                 elif self.holding_chunk is not None:
                     self.chunk_queue.append(self.holding_chunk)
+                    if processed_chunk.get("type") == "message_delta":
+                        processed_chunk = self._augment_message_delta_usage(
+                            processed_chunk
+                        )
                     self.chunk_queue.append(processed_chunk)
                     self.holding_chunk = None
                     return self.chunk_queue.popleft()
                 else:
+                    if processed_chunk.get("type") == "message_delta":
+                        processed_chunk = self._augment_message_delta_usage(
+                            processed_chunk
+                        )
                     self.chunk_queue.append(processed_chunk)
                     return self.chunk_queue.popleft()
 
-            # Handle any remaining held chunks after stream ends
-            if self.holding_chunk is not None:
-                self.chunk_queue.append(self.holding_chunk)
+            # Handle any remaining held chunks after stream ends. The
+            # buffered ``holding_chunk`` (a ``content_block_delta``) must
+            # precede the final ``message_delta`` so Anthropic SSE event
+            # ordering is preserved. When ``queued_usage_chunk`` is True,
+            # the final ``message_delta`` has already been emitted; any
+            # buffered content delta is dropped rather than emitted after
+            # ``message_delta`` (which would violate SSE ordering and may
+            # confuse strict Anthropic SDK clients).
+            if not self.queued_usage_chunk:
+                if self.holding_chunk is not None:
+                    self.chunk_queue.append(self.holding_chunk)
+                    self.holding_chunk = None
+                if self.holding_stop_reason_chunk is not None:
+                    # A final ``message_delta`` must be preceded by
+                    # ``content_block_stop`` so the emitted SSE stays in
+                    # valid Anthropic order (... -> content_block_stop ->
+                    # message_delta). Emit ``content_block_stop`` here if
+                    # the active content block was not already closed.
+                    if not self.sent_content_block_finish:
+                        self.chunk_queue.append(
+                            {
+                                "type": "content_block_stop",
+                                "index": self.current_content_block_index,
+                            }
+                        )
+                        self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+            else:
                 self.holding_chunk = None
 
             if not self.sent_last_message:
@@ -215,6 +533,26 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
         except StopIteration:
             if self.chunk_queue:
                 return self.chunk_queue.popleft()
+            # Handle any held stop_reason chunk. Emit ``content_block_stop``
+            # first if the active content block was not already closed, so
+            # Anthropic SSE ordering is preserved (content_block_stop ->
+            # message_delta).
+            if self.holding_stop_reason_chunk is not None:
+                if not self.sent_content_block_finish:
+                    self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+                    return {
+                        "type": "content_block_stop",
+                        "index": self.current_content_block_index,
+                    }
+                held = self._augment_message_delta_usage(self.holding_stop_reason_chunk)
+                self.holding_stop_reason_chunk = None
+                return held
             if self.sent_last_message is False:
                 self.sent_last_message = True
                 return {"type": "message_stop"}
@@ -223,7 +561,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             verbose_logger.error(
                 "Anthropic Adapter - {}\n{}".format(e, traceback.format_exc())
             )
-            raise StopAsyncIteration
+            raise StopIteration
 
     async def __anext__(self):  # noqa: PLR0915
         from .transformation import LiteLLMAnthropicMessagesAdapter
@@ -253,87 +591,56 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                 )
                 return self.chunk_queue.popleft()
 
+            if (
+                self.sent_compaction_block is False
+                and self.compaction_block is not None
+            ):
+                compaction_event = self._next_compaction_event()
+                if compaction_event is not None:
+                    return compaction_event
+
+            if (
+                self.sent_compaction_block is False
+                and self.compaction_block is not None
+            ):
+                compaction_event = self._next_compaction_event()
+                if compaction_event is not None:
+                    return compaction_event
+
             async for chunk in self.completion_stream:
                 if chunk == "None" or chunk is None:
                     raise Exception
 
                 # Check if we need to start a new content block
                 should_start_new_block = self._should_start_new_content_block(chunk)
-                if should_start_new_block and self.sent_content_block_start:
+                if should_start_new_block:
                     self._increment_content_block_index()
 
+                # applied_edits only needs to flow to the final message_delta
+                # (when finish_reason is set); skip threading it through every
+                # intermediate chunk. For the hold-and-merge path below,
+                # context_management is attached directly to the merged chunk,
+                # so the translated ``processed_chunk`` would be discarded —
+                # skip the applied_edits attachment in that case to avoid
+                # allocating a throwaway ``MessageBlockDelta``.
+                will_merge_into_held = (
+                    self.holding_stop_reason_chunk is not None
+                    and getattr(chunk, "usage", None) is not None
+                )
+                is_final_chunk = chunk.choices[0].finish_reason is not None
                 processed_chunk = LiteLLMAnthropicMessagesAdapter().translate_streaming_openai_response_to_anthropic(
                     response=chunk,
                     current_content_block_index=self.current_content_block_index,
+                    applied_edits=(
+                        self.applied_edits
+                        if is_final_chunk and not will_merge_into_held
+                        else None
+                    ),
                 )
 
-                if not self.sent_content_block_start:
-                    self.sent_content_block_start = True
-                    self.chunk_queue.append(
-                        {
-                            "type": "content_block_start",
-                            "index": self.current_content_block_index,
-                            "content_block": self.current_content_block_start,
-                        }
-                    )
-                    if (
-                        processed_chunk.get("type") == "content_block_delta"
-                        and isinstance(processed_chunk.get("delta"), dict)
-                        and processed_chunk["delta"].get("type")
-                        in ("text_delta", "input_json_delta")
-                        and processed_chunk["delta"].get(
-                            "text", processed_chunk["delta"].get("partial_json", "")
-                        )
-                    ):
-                        self.chunk_queue.append(processed_chunk)
-                    return self.chunk_queue.popleft()
-
                 # Check if this is a usage chunk and we have a held stop_reason chunk
-                if (
-                    self.holding_stop_reason_chunk is not None
-                    and getattr(chunk, "usage", None) is not None
-                ):
-                    # Merge usage into the held stop_reason chunk
-                    merged_chunk = self.holding_stop_reason_chunk.copy()
-                    if "delta" not in merged_chunk:
-                        merged_chunk["delta"] = {}
-
-                    # Add usage to the held chunk
-                    uncached_input_tokens = chunk.usage.prompt_tokens or 0
-                    if (
-                        hasattr(chunk.usage, "prompt_tokens_details")
-                        and chunk.usage.prompt_tokens_details
-                    ):
-                        cached_tokens = (
-                            getattr(
-                                chunk.usage.prompt_tokens_details, "cached_tokens", 0
-                            )
-                            or 0
-                        )
-                        uncached_input_tokens -= cached_tokens
-
-                    usage_dict: UsageDelta = {
-                        "input_tokens": uncached_input_tokens,
-                        "output_tokens": chunk.usage.completion_tokens or 0,
-                    }
-                    # Add cache tokens if available (for prompt caching support)
-                    if (
-                        hasattr(chunk.usage, "_cache_creation_input_tokens")
-                        and chunk.usage._cache_creation_input_tokens > 0
-                    ):
-                        usage_dict["cache_creation_input_tokens"] = (
-                            chunk.usage._cache_creation_input_tokens
-                        )
-                    if (
-                        hasattr(chunk.usage, "_cache_read_input_tokens")
-                        and chunk.usage._cache_read_input_tokens > 0
-                    ):
-                        usage_dict["cache_read_input_tokens"] = (
-                            chunk.usage._cache_read_input_tokens
-                        )
-                    merged_chunk["usage"] = usage_dict
-
-                    # Queue the merged chunk and reset
+                if will_merge_into_held:
+                    merged_chunk = self._merge_usage_into_held_stop_reason_chunk(chunk)
                     self.chunk_queue.append(merged_chunk)
                     self.queued_usage_chunk = True
                     self.holding_stop_reason_chunk = None
@@ -399,28 +706,63 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                         ):
                             self.holding_stop_reason_chunk = processed_chunk
                         else:
+                            processed_chunk = self._augment_message_delta_usage(
+                                processed_chunk
+                            )
                             self.chunk_queue.append(processed_chunk)
                         return self.chunk_queue.popleft()
                     elif self.holding_chunk is not None:
                         # Queue both chunks
                         self.chunk_queue.append(self.holding_chunk)
+                        if processed_chunk.get("type") == "message_delta":
+                            processed_chunk = self._augment_message_delta_usage(
+                                processed_chunk
+                            )
                         self.chunk_queue.append(processed_chunk)
                         self.holding_chunk = None
                         return self.chunk_queue.popleft()
                     else:
-                        # Queue the current chunk
+                        if processed_chunk.get("type") == "message_delta":
+                            processed_chunk = self._augment_message_delta_usage(
+                                processed_chunk
+                            )
                         self.chunk_queue.append(processed_chunk)
                         return self.chunk_queue.popleft()
 
-            # Handle any remaining held chunks after stream ends
+            # Handle any remaining held chunks after stream ends. The
+            # buffered ``holding_chunk`` (a ``content_block_delta``) must
+            # precede the final ``message_delta`` so Anthropic SSE event
+            # ordering is preserved. When ``queued_usage_chunk`` is True,
+            # the final ``message_delta`` has already been emitted; any
+            # buffered content delta is dropped rather than emitted after
+            # ``message_delta`` (which would violate SSE ordering and may
+            # confuse strict Anthropic SDK clients).
             if not self.queued_usage_chunk:
-                if self.holding_stop_reason_chunk is not None:
-                    self.chunk_queue.append(self.holding_stop_reason_chunk)
-                    self.holding_stop_reason_chunk = None
-
                 if self.holding_chunk is not None:
                     self.chunk_queue.append(self.holding_chunk)
                     self.holding_chunk = None
+                if self.holding_stop_reason_chunk is not None:
+                    # A final ``message_delta`` must be preceded by
+                    # ``content_block_stop`` so the emitted SSE stays in
+                    # valid Anthropic order (... -> content_block_stop ->
+                    # message_delta). Emit ``content_block_stop`` here if
+                    # the active content block was not already closed.
+                    if not self.sent_content_block_finish:
+                        self.chunk_queue.append(
+                            {
+                                "type": "content_block_stop",
+                                "index": self.current_content_block_index,
+                            }
+                        )
+                        self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+            else:
+                self.holding_chunk = None
 
             if not self.sent_last_message:
                 self.sent_last_message = True
@@ -436,9 +778,28 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             # Handle any remaining queued chunks before stopping
             if self.chunk_queue:
                 return self.chunk_queue.popleft()
-            # Handle any held stop_reason chunk
+            # Handle any held stop_reason chunk — clear after capturing so a
+            # subsequent ``__anext__`` call doesn't re-emit the same chunk
+            # (matches the sync ``__next__`` path). Emit ``content_block_stop``
+            # first if the active content block was not already closed, so
+            # Anthropic SSE ordering is preserved (content_block_stop ->
+            # message_delta).
             if self.holding_stop_reason_chunk is not None:
-                return self.holding_stop_reason_chunk
+                if not self.sent_content_block_finish:
+                    self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+                    return {
+                        "type": "content_block_stop",
+                        "index": self.current_content_block_index,
+                    }
+                held = self._augment_message_delta_usage(self.holding_stop_reason_chunk)
+                self.holding_stop_reason_chunk = None
+                return held
             if not self.sent_last_message:
                 self.sent_last_message = True
                 return {"type": "message_stop"}