- Hoist _resolve_ref and _resolve_param_list to module level in
openapi_to_mcp_generator.py (were being redefined on every loop iteration)
- _resolve_ref now returns None for unresolvable $refs instead of
the stub dict, preventing (None, None) from poisoning deduplication
- Add resolve_operation_params() as a shared helper that handles both
$ref resolution and path-level param merging
- Replace duplicated inline logic in mcp_server_manager.py and
rest_endpoints.py with calls to resolve_operation_params()
- Add TestResolveRef, TestResolveParamList, TestResolveOperationParams
test classes covering $ref resolution, path-level merging, collision
semantics, unresolvable ref filtering, and a GitHub-style spec fixture
When the messages or response JSON fields in spend logs are truncated
before being written to the database, the truncation marker now includes
a note explaining:
- This is a DB storage safeguard
- Full, untruncated data is still sent to logging callbacks (OTEL, Datadog, etc.)
- The MAX_STRING_LENGTH_PROMPT_IN_DB env var can be used to increase the limit
Also emits a verbose_proxy_logger.info message when truncation occurs in
the request body or response spend log paths.
Adds 3 new tests:
- test_truncation_includes_db_safeguard_note
- test_response_truncation_logs_info_message
- test_request_body_truncation_logs_info_message
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Previously, model_dump(exclude_none=True) included all bool fields (since
False != None), causing a partial PATCH to overwrite every other setting to
its default. Fix uses exclude_unset=True and reads the existing DB record
before merging, giving proper PATCH semantics.
This was a pre-existing bug but is fixed here since we're touching this code.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Claude Code v2.1.69+ sends `custom: {defer_loading: true}` on tool
definitions. Anthropic's API accepts this field, but Bedrock rejects it
with "Extra inputs are not permitted", causing ~90% of requests to fail.
Strip the `custom` field from each tool in the request body before
sending to Bedrock, in both the Messages API and Chat API invoke paths.
Fixes#22847
Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com>
* fix(proxy): readiness check returns 200 when database is unreachable
_db_health_readiness_check() catches health_check() exceptions but
never updates db_health_cache to "disconnected" and never re-raises.
The caller health_readiness() always returns 200 with "db": "connected"
hardcoded, regardless of actual DB state.
In Kubernetes, this means pods with dead database connections stay in
the Service endpoints and continue receiving traffic they cannot serve.
Changes:
- Set db_health_cache to "disconnected" and re-raise the exception on
health_check failure so health_readiness() returns 503
- Use actual db_health_status["status"] in the response instead of
hardcoding "db": "connected"
- Reduce cache TTL from 2 minutes to 15 seconds. The 2-minute window
is too wide for readiness probes (typically 10-15s intervals) and
means a pod can report healthy for up to 2 minutes after the DB dies
- Only serve cached results when status is "connected". The previous
condition (status != "unknown") would also cache "disconnected" for
2 minutes, delaying recovery detection after a DB comes back
* fix(proxy): add DB connection self-healing to readiness check
When the Prisma query engine's internal TCP connection pool holds dead
connections (caused by network blips, Cloud SQL proxy restarts, or
node-level issues), health_check() fails with httpx.ConnectError.
The engine never recovers on its own because nothing triggers a
disconnect/connect cycle to restart the subprocess with fresh
connections.
This leaves pods permanently failing readiness checks until they are
manually restarted, even after the underlying DB becomes reachable
again.
Add a reconnect attempt to _db_health_readiness_check() when
health_check() fails:
1. disconnect() - kills the query engine subprocess and closes all
connections (has built-in backoff retry: 3 tries, 10s max)
2. connect() - starts a new engine with fresh TCP connections (has
built-in backoff retry: 3 tries, 10s max)
3. health_check() - verifies the new connection works (has built-in
backoff retry: 3 tries, 10s max)
If reconnect succeeds, the pod immediately returns to service (200).
If it fails, the original exception is re-raised (503). Reconnect
attempts are rate-limited by probe frequency (~10-15s), so a
permanently unreachable DB gets one attempt per cycle with no retry
loops.
This uses the same disconnect/connect mechanism that
PrismaWrapper.recreate_prisma_client() uses for IAM token refresh,
and aligns with the community-documented pattern for Prisma connection
recovery in long-running processes (prisma/prisma#24718, #27024).
* Add poetry lock and modify test_health_endpoints
* Address allow_requests_on_db_unavailable regression
* Address comments
* resolve greptile issue
* Restore accidentally deleted UI HTML files
These were removed in an earlier commit but still exist on main.
Restoring to keep the PR diff clean.
* Guard reconnect with is_database_transport_error
Only attempt disconnect/connect/health_check cycle for transport-level
failures (unreachable DB, dropped connection). Data-layer errors like
UniqueViolationError indicate the DB is reachable, so reconnecting
would be pointless churn.
* Address greptile's comments
* Fix module alias after rebase and add adversarial test coverage
- Unify module alias to _health_endpoints_module after rebase conflict
- Add test for non-transport error with flag on (exercises is_database_transport_error guard)
- Add test for disconnect() failure during reconnect cycle
- Split non-transport error test into flag-off (re-raises) and flag-on (skips reconnect) variants
* Remove stale UI HTML files reintroduced during rebase
* fix: don't close HTTP/SDK clients on LLMClientCache eviction
Removing the _remove_key override that eagerly called aclose()/close()
on evicted clients. Evicted clients may still be held by in-flight
streaming requests; closing them causes:
RuntimeError: Cannot send a request, as the client has been closed.
This is a regression from commit fb72979432. Clients that are no longer
referenced will be garbage-collected naturally. Explicit shutdown cleanup
happens via close_litellm_async_clients().
Fixes production crashes after the 1-hour cache TTL expires.
* test: update LLMClientCache unit tests for no-close-on-eviction behavior
Flip the assertions: evicted clients must NOT be closed. Replace
test_remove_key_closes_async_client → test_remove_key_does_not_close_async_client
and equivalents for sync/eviction paths.
Add test_remove_key_removes_plain_values for non-client cache entries.
Remove test_background_tasks_cleaned_up_after_completion (no more _background_tasks).
Remove test_remove_key_no_event_loop variant that depended on old behavior.
* test: add e2e tests for OpenAI SDK client surviving cache eviction
Add two new e2e tests using real AsyncOpenAI clients:
- test_evicted_openai_sdk_client_stays_usable: verifies size-based eviction
doesn't close the client
- test_ttl_expired_openai_sdk_client_stays_usable: verifies TTL expiry
eviction doesn't close the client
Both tests sleep after eviction so any create_task()-based close would
have time to run, making the regression detectable.
Also expand the module docstring to explain why the sleep is required.
* docs(AGENTS.md): add rule — never close HTTP/SDK clients on cache eviction
* docs(CLAUDE.md): add HTTP client cache safety guideline
* Include user_email in new user creation within get_user_object
Enhance the get_user_object function to include user_email in the parameters when creating a new user. This change is accompanied by a new test to verify that user_email is correctly included during the upsert process.
* Improve error handling in test_get_user_object by logging exceptions
Updated the test_get_user_object_upsert_includes_user_email function to log exceptions when they occur, enhancing the visibility of potential issues during testing. This change helps in diagnosing failures related to the mock LiteLLM_UserTable.
* fix(passthrough): raise_for_status in _async_streaming to propagate Azure 429s
* address greptile review feedback (greploop iteration 1)
Guard data/json args when content is provided to avoid httpx ValueError
* address greptile review feedback (greploop iteration 2)
Use bare raise to preserve original traceback in _async_streaming exception handler
* address greptile review feedback (greploop iteration 3)
Close httpx streaming response on error to prevent connection pool exhaustion
* address greptile review feedback (greploop iteration 4)
Guard aclose() call to prevent masking original exception; add explicit test for content param forwarding
* address greptile review feedback (greploop iteration 5)
Pass content to sign_request so AWS body-hash signing is correct when content is the sole body source
* revert sign_request content change - request_data expects dict, not bytes
Bedrock's sign_request calls json.dumps(request_data) — passing content bytes
would TypeError. sign_request should only receive data/json (dict), not raw bytes.
gemini/gemini-live-2.5-flash-preview-native-audio-09-2025 uses mode='realtime'
but the schema in test_aaamodel_prices_and_context_window_json_is_valid did
not include 'realtime' as a valid enum value, causing a ValidationError.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Vertex AI does not support the output_config parameter in its API.
This parameter is being added by Anthropic/Gemini transformations but needs
to be removed before sending requests to Vertex AI endpoints.
This fix addresses the "Extra inputs are not permitted" error (issue #22312)
when using Claude models with structured outputs on Vertex AI.
Changes:
- Drop output_config in Gemini model transformation
- Drop output_config in Anthropic partner model transformation
- Drop output_config in Anthropic experimental pass-through transformation
- Add comprehensive tests to verify output_config is dropped
Fixes: #22312
Made-with: Cursor
Azure AI Foundry's Anthropic endpoint does not support the scope field in
cache_control. Strip it from both system and messages before sending.
Made-with: Cursor
Bedrock does not support the scope field in cache_control (e.g. 'global' for
cross-request caching). Only type and ttl are supported per AWS docs.
- Remove scope from cache_control in both system and messages
- Extend _remove_ttl_from_cache_control to process system blocks
- Add test for scope removal
Made-with: Cursor
* feat(mcp): BYOK (Bring Your Own Key) for OpenAPI MCP servers with OAuth 2.1 flow
Adds per-user credential storage for BYOK MCP servers so external clients
can authenticate via standard OAuth 2.1 PKCE without needing a full identity
provider.
Backend:
- New DB table LiteLLM_MCPUserCredentials (user_id, server_id, credential_b64)
- is_byok, byok_description, byok_api_key_help_url fields on MCPServerTable
- OAuth 2.1 authorization server endpoints (/.well-known/oauth-authorization-server,
/.well-known/oauth-protected-resource, /v1/mcp/oauth/authorize, /v1/mcp/oauth/token)
- 401 challenge with WWW-Authenticate header when BYOK server has no credential
- CRUD endpoints: POST/DELETE /v1/mcp/server/{id}/user-credential
- has_user_credential annotated on GET /v1/mcp/server response
UI:
- ByokCredentialModal: 2-step Connect flow (access description + API key entry)
- BYOK toggle + description fields on admin MCP server create form
- Connect/Connected state in MCP server table
- BYOK Demo page (/tools/byok-demo) showing full OAuth 2.1 PKCE flow
* feat(mcp/byok): redesign OAuth authorize page to match 2-step Connect mockup
- Step 1: L→S logos, requested access checklist, How it works box, Continue button
- Step 2: API key input, Save toggle, Duration pills (1h/24h/7d/30d/until_revoked), security note
- Matches screenshots: white modal on dark bg, progress dots, dark CTA buttons
- Authorize handler now fetches byok_description and byok_api_key_help_url from server registry
- CLAUDE.md: replace SQL snippet with proper DB migration troubleshooting guidance
* fix: address greptile review feedback (greploop iteration 1)
- XSS: escape all user-supplied values in _build_authorize_html() with html.escape()
- Open redirect: validate redirect_uri scheme and URL-encode code/state in redirect
- N+1 query: batch BYOK credential lookup into single find_many() call
- Critical path DB: add 60s TTL in-memory cache to _check_byok_credential()
- Encrypt BYOK credentials at rest using encrypt_value_helper/decrypt_value_helper
* fix(byok): update OAuth popup with LiteLLM logo, MCP title suffix, remove emojis
* fix(byok-demo): fix token endpoint URL (/v1/mcp/oauth/token not /v1/mcp/token)
* feat(byok): inject stored BYOK credential as mcp_auth_header on tool execution
* feat(byok): use contextvars to inject per-user credential into OpenAPI tool closures; remove byok-demo from LiteLLM UI
OpenAPI tools have auth headers baked into their closures at registration time. BYOK servers have
no static auth token, so per-user credentials were never reaching the HTTP calls.
Fix: add _request_auth_header ContextVar in openapi_to_mcp_generator.py. create_tool_function now
reads this var at call time and overrides the Authorization header if set. execute_mcp_tool resolves
the MCP server and performs BYOK checks before the local-tool dispatch branch, then sets the
ContextVar around _handle_local_mcp_tool so the credential flows into the HTTP request.
Also remove the /tools/byok-demo page from the LiteLLM UI dashboard — the demo lives at
~/Downloads/litellm-byok-demo/index.html (served separately on port 8080).
* fix: address greptile review feedback (greploop iteration 2)
- Cache invalidation: add _invalidate_byok_cred_cache() and call it after
store_user_credential() in both token endpoint and management endpoint
- Unbounded cache: add _BYOK_CRED_CACHE_MAX_SIZE=4096 with clear-on-overflow
- Unbounded auth codes: add _AUTH_CODES_MAX_SIZE=1000 with 503 on overflow
- Double DB query: merge _check_byok_credential + _get_byok_credential into
single _get_byok_credential call; raise 401 inline if None returned
- Sidebar: remove byok-demo entry (page was deleted in prior commit)
- JWT comment: document why byok_session HS256 token can't be used as proxy auth
* fix: address greptile review feedback (greploop iteration 3)
- auth_type: pre-format Authorization header (Bearer/ApiKey/Basic) in server.py
before setting ContextVar so openapi_to_mcp_generator respects server auth_type
- cache invalidation on delete: call _invalidate_byok_cred_cache after
delete_user_credential so stale True entries don't persist for 60s
- ContextVar guard: only set _request_auth_header when mcp_auth_header is set,
avoiding unnecessary ContextVar overhead on non-BYOK tool calls
* fix: address greptile review feedback (greploop iteration 4)
- Unified credential cache: store actual credential value (Optional[str])
instead of just bool so _get_byok_credential also benefits from caching —
eliminates the DB hit on every BYOK tool call within the 60s TTL window
- Extracted _write_byok_cred_cache() helper for consistent cache writes
- Replaced has_user_credential with get_user_credential in _check_byok_credential
so one DB call satisfies both existence check and value retrieval
- Remove false 'encrypted at rest' claim from OAuth HTML and ByokCredentialModal
* Update tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* Update tests/test_litellm/proxy/_experimental/mcp_server/test_byok_oauth_endpoints.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
---------
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix: complexity_router fails on list-format message content (OpenAI multi-part messages)
When a client sends messages with list-format content
(e.g. [{"type": "text", "text": "..."}] as used by the OpenAI JS SDK
and other clients), the complexity_router's async_pre_routing_hook
skipped those messages because it only handled str content. This caused
user_message to be None, the hook returned None, and the router fell
through to selecting the complexity_router deployment itself
(model="auto_router/complexity_router") which litellm cannot dispatch,
resulting in LiteLLMUnknownProvider.
Fixes:
- Extract text from list-format content parts (type=text) before
classifying
- Return default_model instead of None when no user message can be
extracted, preventing the crash fallthrough
- Loosen PreRoutingHookResponse.messages type from Dict[str, str] to
Dict[str, Any] to accommodate list-format content values
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: update messages type annotation in async_pre_routing_hook to Dict[str, Any]
Consistent with PreRoutingHookResponse.messages type change and the
list-format content support added in the previous commit.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: normalize None content to empty string in complexity_router message parsing
msg.get("content", "") returns None when the key exists with value None
(e.g. assistant messages with tool calls). Use `or ""` to normalize
None to an empty string explicitly.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: strip whitespace from joined list content parts in complexity_router
Prevents leading/trailing spaces when some content parts have empty
text values (e.g. " ".join(["", "hello"]) → " hello").
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* (sap) ensure tool parameters have type='object' for SAP compatibility
Fix SAP GenAI Hub Orchestration Service rejecting tool calls with error:
"400 - LLM Module: tools.0.custom.input_schema.type: Input should be 'object'"
Root cause: When Claude Code uses tools (like web_search) with the SAP provider
through LiteLLM's Anthropic experimental pass-through adapter, Anthropic's
input_schema format doesn't always include the required type="object" field.
The adapter's translate_anthropic_tools_to_openai() function was directly
copying input_schema to OpenAI's parameters field without ensuring the
type="object" requirement that SAP's API strictly enforces.
Changes:
- Modified translate_anthropic_tools_to_openai() to check if input_schema
is missing the type field and add type="object" if absent
- Preserves existing type field if already present
- Added comprehensive test suite (6 tests) covering:
- Missing type field scenario (now adds type="object")
- Existing type preservation
- Empty input_schema handling
- Multiple tools transformation
- Additional schema properties preservation
- SAP-specific compatibility regression test
Testing:
- All new tests pass (6/6 in test_anthropic_tool_schema_fix.py)
- All existing Anthropic tool tests pass (57/57 tool-related tests)
- SAP tool parameter validation tests pass (9/9 in test_sap_tool_parameters.py)
* (sap) enable native response_format for anthropic models
* (sap) filter strict param from model_params for GPT models only
* (sap) revert Anthropic adapter type='object' fix
The SAP FunctionTool Pydantic validator in litellm/llms/sap/chat/models.py
already ensures type='object' is added to all tool parameters for SAP
API compatibility.
The Anthropic adapter change affected ALL consumers, not just SAP, which
was broader scope than intended for this PR.
- Revert input_schema modification in Anthropic adapter
- Remove Anthropic-specific test file (SAP tests still cover this case)
* (sap) gate markdown stripping to Anthropic models only
SAP GenAI Hub with Anthropic models sometimes returns JSON wrapped in
markdown code blocks. GPT/Gemini/Mistral models don't exhibit this
behavior, so stripping is now gated to avoid accidentally modifying
valid responses that may contain markdown in JSON string values.
* azure content enhancement...
* rafactored to increase confidence score
* improvements based on additional feedback
* removed unused import
* Force-split any word longer than max length allowed
* preserve whitespace in text splitting
* moving common initialization to base class
* consolidate enforcement into async_make_request as single point, remove redundant caller-side checks, extract shared init/HTTP logic into base, and fix stale log messages
* clean up
* clean up tests
Add cache_read_input_token_cost_per_audio_token, supports_code_execution,
and supports_file_search to the JSON schema used by the model prices
validation test.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The _encrypt_response_id method now receives request_cache=None as a
keyword argument from async_post_call_success_hook. Updated the mock
assertion to expect this parameter.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Tests used call_policy throughout but the actual API model uses
input_policy and output_policy. Updated _make_tool_row helper,
list filter query param, and policy update request/response assertions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The aresponses_websocket CallType was recently added but not included
in the test exclusion list. It uses WebSocket passthrough (not Azure SDK
client initialization), so it correctly doesn't call
initialize_azure_sdk_client.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add test_parallel_tool_calls_comprehensive_streaming_integration which
synthesizes the full 10-event Responses API SSE sequence with split
argument deltas and asserts all fix invariants together:
1. output_item.done emits no finish_reason (no premature stream end)
2. Each call_id appears exactly once (no duplicate tool_call chunks)
3. Split argument deltas assemble to correct final JSON
4. Exactly one finish event, at the terminal response.completed chunk
5. Parallel tool calls have distinct indices (output_index 0 and 1)
All 24 unit tests pass.
* feat(proxy): add key_alias, key_hash, requested_model tags to DD APM spans
* refactor(proxy): consolidate DD APM tag helpers into DDSpanTagger class
* refactor(proxy): move DDSpanTagger to its own file litellm/proxy/dd_span_tagger.py
* style(ui/): distinguish agent calls from llm calls on ui
* feat: initial grouping working
* feat: set stable contextid for a2a calls - allows for easily passing to downstream llm/mcp calls
* feat(a2a_endpoints.py): fix tracing to avoid recreating logging objects for the same call
allows stable trace id usage
* fix(guardrail_endpoints): handle string ui_type values in _build_field_dict
_build_field_dict unconditionally called .value on ui_type, which crashes
for guardrail configs that use plain strings (e.g. BlockCodeExecutionGuardrailConfigModel
uses "multiselect" and "percentage"). Now checks with hasattr before calling .value.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: propagate trace/session id from headers in MCP server calls
Cherry-picked mcp_server/server.py fixes from 6feb9bab: adds
get_chain_id_from_headers to extract x-litellm-trace-id /
x-litellm-session-id from raw headers, and uses it in call_tool
and list_tools to keep spend logs and tracing consistent with A2A.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
