* fix(scheduler): remove timed-out requests from queue to prevent memory leak
Fixes#20059
* fix(scheduler): use actual model param instead of hardcoded gpt-3.5-turbo in schedule_acompletion
* trigger CLA recheck
---------
Co-authored-by: Piyush Bhawsar <piyush100x@Piyushs-MacBook-Pro-3.local>
Replace asyncio.wait_for() with anyio.fail_after() in _fetch_tools_with_timeout()
to fix conflict with MCP SDK's anyio TaskGroup that causes 0 tools to be returned
for external StreamableHTTP MCP backends.
Root cause: asyncio.wait_for() wrapping anyio-managed code causes inconsistent
CancelledError propagation, resulting in false cancellations even when the
operation hasn't timed out.
Fixes#20715
Adds a new config option to exclude specific fields from StandardLoggingPayload
before any callback receives it. This provides a general approach to control
what data is logged across ALL integrations (S3, GCS, Datadog, etc.).
## Changes
1. **litellm/__init__.py**: Added new global setting
`standard_logging_payload_excluded_fields: Optional[List[str]] = None`
2. **litellm/integrations/custom_logger.py**: Modified
`redact_standard_logging_payload_from_model_call_details()` to:
- Remove specified fields entirely from the StandardLoggingPayload
- Works alongside existing `turn_off_message_logging` feature
- Excluded fields take precedence (removed rather than redacted)
3. **tests/**: Added comprehensive test suite with 17 tests covering:
- Single/multiple field exclusion
- Interaction with turn_off_message_logging
- Original payload immutability
- Config loading via setattr (proxy pattern)
- Edge cases (empty list, non-existent fields, None standard_logging_object)
## Usage
```yaml
litellm_settings:
success_callback: ["s3"]
standard_logging_payload_excluded_fields: ["response", "messages"]
```
This removes the `response` and `messages` fields from logs before any
callback processes them, reducing log size and improving privacy compliance.
## Available Fields
The fields match StandardLoggingPayload TypedDict keys including:
- messages, response (large payload fields)
- metadata, hidden_params, model_parameters
- error_str, error_information
- And all other StandardLoggingPayload fields
Closes the need for per-integration flags like `s3_log_response`.
When a guardrail (e.g. Zscaler AI Guard) or other exception is raised
during streaming, the error handler in `async_data_generator` includes
the full Python traceback in the SSE response sent to clients:
error_msg = f"{str(e)}\n\n{traceback.format_exc()}"
This leaks internal server details (file paths, line numbers, call
stacks) to end users. The traceback is already logged server-side via
`verbose_proxy_logger.exception()`, so including it in the client
response is unnecessary.
Change to only include the exception message (`str(e)`) in the SSE
error payload, consistent with how `StreamingCallbackError` is already
handled.
Fixes#20610
* [Fix] handle metadata=None in SDK path retry/error logic (utils.py)
Fixes#20871
Same class of bug as #9717 (fixed by #9764 for the proxy path).
The SDK path in utils.py has the same fragile pattern at 7 locations.
Replace `kwargs.get("metadata", {})` with `(kwargs.get("metadata") or {})`
to handle the case where metadata key exists with value None (e.g. from
Azure OpenAI streaming responses).
This is consistent with the existing correct pattern at line 602:
`metadata = kwargs.get("metadata") or {}`
Adds TestMetadataNoneHandling with 6 unit tests in test_utils.py.
* fix: remove duplicate PerplexityResponsesConfig key in lazy imports registry
Removes duplicate dictionary key added in commit be0ebb15 (PR #20860).
The entry at line 1042 is identical to the existing entry at line 906.
This causes ruff F601 lint failure on all PRs targeting main.
Add support for Alibaba Cloud's Qwen3-Max model with:
- 258K input tokens, 65K output tokens
- Tiered pricing based on context window usage (0-32K, 32K-128K, 128K-252K)
- Function calling and tool choice support
- Reasoning capabilities enabled
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
* init schema with TAGS
* ui: add policy test
* resolvePoliciesCall
* add_policy_sources_to_metadata + headers
* types Policy
* preview Impact
* def _describe_match_reason(
* match based on TAGs
* TestTagBasedAttachments
* test fixes
* add policy_resolve_router
* add_guardrails_from_policy_engine
* TestMatchAttribution
* refactor
* fix
* fix: address Greptile review feedback on policy resolve endpoints
- Track unnamed keys/teams as separate counts instead of inflating
affected_keys_count with duplicate "(unnamed key)" placeholders.
Added unnamed_keys_count and unnamed_teams_count to response.
- Push alias pattern matching to DB via _build_alias_where() which
converts exact patterns to Prisma "in" and suffix wildcards to
"startsWith" filters.
- Gate sync_policies_from_db/sync_attachments_from_db behind
force_sync query param (default false) to avoid 2 DB round-trips
on every /policies/resolve request.
- Remove worktree-only conftest.py that cleared sys.modules at import
time — no longer needed since code moved to main repo.
- Rename MAX_ESTIMATE_IMPACT_ROWS → MAX_POLICY_ESTIMATE_IMPACT_ROWS.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: eliminate duplicate DB queries and fix header delimiter ambiguity
- Fetch teams table once in estimate_attachment_impact and reuse for
both tag-based and alias-based lookups (was querying teams twice when
both tag_patterns and team_patterns were provided).
- Convert tag/team filter functions from async DB queries to sync
filters that operate on pre-fetched data (_filter_keys_by_tags,
_filter_teams_by_tags).
- Fix comma ambiguity in x-litellm-policy-sources header: use '; '
as entry delimiter since matched_via values can contain commas.
- Use '+' as the within-value separator in matched_via reason strings
(e.g. "tag:healthcare+team:health-team") to avoid conflict with
header delimiters.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* docs v1 guide with UI imgs
* docs fix
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* init schema with TAGS
* ui: add policy test
* resolvePoliciesCall
* add_policy_sources_to_metadata + headers
* types Policy
* preview Impact
* def _describe_match_reason(
* match based on TAGs
* TestTagBasedAttachments
* test fixes
* add policy_resolve_router
* add_guardrails_from_policy_engine
* TestMatchAttribution
* refactor
* fix
* fix: address Greptile review feedback on policy resolve endpoints
- Track unnamed keys/teams as separate counts instead of inflating
affected_keys_count with duplicate "(unnamed key)" placeholders.
Added unnamed_keys_count and unnamed_teams_count to response.
- Push alias pattern matching to DB via _build_alias_where() which
converts exact patterns to Prisma "in" and suffix wildcards to
"startsWith" filters.
- Gate sync_policies_from_db/sync_attachments_from_db behind
force_sync query param (default false) to avoid 2 DB round-trips
on every /policies/resolve request.
- Remove worktree-only conftest.py that cleared sys.modules at import
time — no longer needed since code moved to main repo.
- Rename MAX_ESTIMATE_IMPACT_ROWS → MAX_POLICY_ESTIMATE_IMPACT_ROWS.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: eliminate duplicate DB queries and fix header delimiter ambiguity
- Fetch teams table once in estimate_attachment_impact and reuse for
both tag-based and alias-based lookups (was querying teams twice when
both tag_patterns and team_patterns were provided).
- Convert tag/team filter functions from async DB queries to sync
filters that operate on pre-fetched data (_filter_keys_by_tags,
_filter_teams_by_tags).
- Fix comma ambiguity in x-litellm-policy-sources header: use '; '
as entry delimiter since matched_via values can contain commas.
- Use '+' as the within-value separator in matched_via reason strings
(e.g. "tag:healthcare+team:health-team") to avoid conflict with
header delimiters.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Update litellm/proxy/policy_engine/policy_resolve_endpoints.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
The template literal in the tags query parameter concatenation included
`${url}` inside a `+=` assignment, causing the full URL to be doubled.
Supersedes #8793 (original fix by @Mte90, now stale).
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* feat(guardrail_hooks/): add guardrail logging to all unified guardrails
ensures unified guardrails use the 'log_guardrail_information' decorator for logging
* fix(custom_guardrail.py): don't log inputs on guardrail response - just emit state
* refactor: don't double log bedrock guardrail information
* feat: add in-product nudges for contributing + trying community custom code guardrails
allows users to contribute / share custom code guardrails
* fix(aiohttp): respect ssl_verify with shared sessions
* fix(aiohttp): resolve mypy error for ssl parameter type
Pass ssl kwarg conditionally to aiohttp request() only when explicitly
configured, since None is not a valid value for the ssl parameter
(expected SSLContext | bool | Fingerprint).
Replace 3 full tree traversals with 1 Pydantic dump + 3 dict lookups
per choice. Deletes _path_matches_pattern, _build_preserved_paths, and
_remove_none_except_preserved helpers (~140 lines removed). Adds 13
regression tests including full output structure snapshots.
When OpenTelemetry is configured via the UI, only OTEL_ENDPOINT and
OTEL_HEADERS are set, but OTEL_EXPORTER is not specified. This caused
the exporter to default to "console", meaning traces were printed to
stdout instead of being sent to the configured endpoint.
This fix adds logic in OpenTelemetryConfig.__post_init__ to automatically
infer "otlp_http" as the exporter when an endpoint is specified but the
exporter is still the default "console".
Fixes issue reported by Elastic team where traces weren't being sent
to their OTEL endpoint when configured through the LiteLLM UI.
