Split the single try/except in the _exception_raised cleanup path into
separate try blocks for asyncio.create_task and executor.submit, matching
the pattern used in _run_deferred_stream_guardrails. If create_task
raises, sync logging via executor.submit still fires.
Replace sys.exc_info()[1] check with an explicit _exception_raised boolean
sentinel. The flag is function-scoped, immune to outer exception context,
and only set when an exception actually occurs in base_process_llm_request.
This prevents false positives when called from a caller's except block.
- Replace hand-rolled _extract_search_query with existing
get_last_user_message from common_utils
- Use full UUID (str(uuid.uuid4())) to match codebase convention
- Move uuid import to module level per CLAUDE.md
Addresses Greptile review feedback:
- Save original stream flag before pre-request hooks convert it, so
streaming callers get SSE events instead of a plain dict
- Propagate custom_llm_provider derived inside _execute_pre_request_hooks
when it was not explicitly passed by the caller
- Add tests covering both scenarios
Move non-essential lazy imports (llm_router, _check_and_merge,
unified_guardrail) inside the try block of _run_deferred_stream_guardrails
so that import failures are caught and the finally block still fires
logging. Only executor stays outside since the finally block needs it.
Add _on_deferred_stream_complete orphan cleanup in the finally block of
base_process_llm_request. If an exception propagates after the streaming
closure is stored but before a StreamingResponse is returned, the closure
is orphaned (CSW never consumes the stream). Detect this via
sys.exc_info() and fire logging directly to prevent silent loss.
For providers like github_copilot that don't natively support web search,
Claude Code's search sub-conversations were falling through to the adapter
path which strips the web_search tool and has no stream reconversion.
Instead of routing search requests through the full LLM pipeline, detect
web-search-only requests early (all tools are web_search, simple prompt)
and execute the search directly via Tavily/Perplexity, returning a
synthetic Anthropic response. No adapter, no backend LLM call needed.
Fixes#21733
Reuse the module-level unified_guardrail singleton from proxy/utils.py
in _run_deferred_stream_guardrails instead of creating a new instance
per call, matching the pattern used by post_call_success_hook.
Rename local variable _has_post_call_guardrails to
_post_call_guardrails_active to avoid shadowing the static method name.
* add DD Tracing (#24033)
* feat(models): add Azure GPT-5.4 mini and nano variants (#24045)
Add `azure/gpt-5.4-mini` and `azure/gpt-5.4-nano` to the model
database with official pricing from Azure OpenAI:
- GPT-5.4 mini: $0.75/M input, $0.075/M cached, $4.5/M output
- GPT-5.4 nano: $0.20/M input, $0.02/M cached, $1.25/M output
Both models support:
- 1.05M input / 128K output context window
- Chat, batch, and responses endpoints
- Function calling, tools, vision, reasoning
- Prompt caching with automatic tiered pricing
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* Add new model pricing details for volcengine Doubao-Seed-2.0 series (#23871)
Add entries for volcengine Doubao-Seed-2.0 series
* fix(mcp): support refresh_token grant type in OAuth token endpoint (#23701)
* fix(mcp): support refresh_token grant type in OAuth token endpoint (#23700)
The .well-known/oauth-authorization-server metadata advertises
refresh_token as a supported grant type, but the token endpoint
rejected it with HTTP 400. This adds refresh_token grant support
so MCP clients can refresh expired tokens without re-authenticating.
* test(mcp): add tests for refresh_token grant type in OAuth token endpoint
* fix(mcp): move code_verifier guard into authorization_code branch
code_verifier is only relevant for authorization_code grants (PKCE).
Move it inside the else branch so it doesn't apply to refresh_token.
* fix(mcp): guard None client_secret and forward scope in token exchange
- Conditionally include client_secret in form data to prevent httpx
from sending the literal string "None" (applies to both
authorization_code and refresh_token branches)
- Forward optional scope parameter per RFC 6749 §6, allowing clients
to request a subset of originally-granted scopes on refresh
* fix(mcp): validate code param in authorization_code grant
Guard against None code being form-encoded as literal string "None"
by httpx, symmetric with the existing refresh_token guard.
* docs: add incident report for guardrail logging secret exposure (#24059)
Add blog post documenting the guardrail logging path exposing internal
request data (e.g. Authorization headers) in spend logs and OTEL traces.
Fix available in LiteLLM 1.82.3+.
Made-with: Cursor
* [Fix] Datadog LLM Observability tags format (env, service, version missing) (#23673)
* tag fix
* greptile comment
* fix(ci): stabilize 6 failing CI jobs
1. mypy: remove duplicate type annotation for token_data in discoverable_endpoints.py
2. integrations tests: add parameterized to CI test deps
3. doc quality: document OTEL_IGNORE_CONTEXT_PROPAGATION env key
4. security: allowlist CVE-2026-2673, CVE-2026-3644, CVE-2026-4224 (no fix available)
5. proxy_store_model_in_db: fix missing x-litellm-call-id header on error responses
6. google tests: add --retries 3 for transient Vertex AI rate limits
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
* fix(streaming): handle RuntimeError during model_copy in streaming handler
The race condition occurs when model_copy(deep=True) tries to deepcopy
_hidden_params dict while it's being concurrently modified by logging
callbacks. Fall back to shallow copy if the deep copy fails.
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
* fix(cost): handle non-string traffic_type in cost calculator + add retries
1. Fix AttributeError in _map_traffic_type_to_service_tier when traffic_type
is an integer (cast to str before calling .upper()). This was causing
pass-through vertex spend logging to fail silently.
2. Add --retries to llm_translation_testing for flaky external API calls.
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
---------
Co-authored-by: Emerson Gomes <emerson.gomes@thalesgroup.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: ExMatics HydrogenC <33123710+HydrogenC@users.noreply.github.com>
Co-authored-by: Jack Venberg <jack.venberg@rover.com>
Co-authored-by: milan-berri <milan@berri.ai>
Co-authored-by: Shivam Rawat <161387515+shivamrawat1@users.noreply.github.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
Wrap _run_deferred_stream_guardrails initialization (UnifiedLLMGuardrails
constructor and _check_and_merge_model_level_guardrails) in try/finally
so logging always fires even if init throws. Prevents silent logging loss
on transient errors.
Move fastapi.HTTPException import from module-level to local test-function
scope. Add test_logging_fires_even_if_guardrail_init_raises to verify the
try/finally guard.
Use the merged guardrail_data dict (from _check_and_merge_model_level_guardrails)
for hook invocations in _run_deferred_stream_guardrails, instead of the original
captured_data. This ensures model-level non-default guardrails are visible to
inner should_run_guardrail re-checks inside UnifiedLLMGuardrails.
Rewrite three hand-crafted closure tests to exercise the production
_run_deferred_stream_guardrails exception-handling path. Add three new tests
that use deep-copy mocks to prove hooks receive the merged dict.
Add a CopyOutlined icon next to the truncated User ID that copies
the full UUID to clipboard on click. Follows the existing pattern
used in model_hub_table_columns.tsx.
guardrail_information is None in StandardLoggingPayload because logging
fires before post-call guardrails write to metadata.
Non-streaming: wrapper_async stores a closure instead of calling
create_task immediately. The proxy fires it in a try/finally after
post_call_success_hook so the SLP is built with guardrail info.
Streaming: a closure on logging_obj is called by CSW.__anext__ at
stream end. The closure runs only guardrail hooks (not all callbacks)
on the assembled response, then fires both logging handlers. This
avoids behavioral changes for non-guardrail callbacks on streaming.
- Change `if team_limit:` to `if team_limit is not None:` in both
get_key_model_rpm_limit and get_key_model_tpm_limit so that an
explicitly-empty team rate-limit map ({}) is returned as-is instead
of silently falling through to deployment defaults (P1 fix).
- Replace the bare `int()` list comprehension in _get_deployment_default_limit
with a loop that catches ValueError/TypeError so malformed config strings
do not raise an unhandled exception during request handling (P2 fix).
- Add corresponding unit tests for both edge cases.
Co-Authored-By: Claude (claude-sonnet-4-6) <noreply@anthropic.com>
- Merge _get_deployment_default_rpm_limit and _get_deployment_default_tpm_limit
into a single _get_deployment_default_limit(model_name, field) helper; the two
thin wrappers are preserved for callers but share one implementation
- Compute _success_tpm_limit / _success_rpm_limit once before the guard condition
in async_log_success_event, eliminating the previous two unconditional
get_key_model_* calls (each of which could hit llm_router.get_model_list)
- Replace fragile llm_model_list=[{}] sentinel in test with []
Co-Authored-By: Claude (claude-sonnet-4-6) <noreply@anthropic.com>
- Extract helper methods in langsmith._prepare_log_data to reduce from 51 to <50 statements
- Extract helper methods in anthropic.transform_parsed_response to reduce from 57 to <50 statements
- Fixes PLR0915 linter errors
- All existing tests pass (10 langsmith tests, 126 anthropic tests)
Made-with: Cursor
Aligns proxy default with litellm.AZURE_DEFAULT_API_VERSION (2025-02-01-preview)
so Azure response_format + json_schema works without tools fallback.
Made-with: Cursor
Enable deployment_affinity, responses_api_deployment_check, and session_affinity to be configured per model group via router_settings.model_group_affinity_config, falling back to global settings for unconfigured groups.
- Add model_group_affinity_config parameter to Router and DeploymentAffinityCheck
- Add _get_effective_flags helper to resolve flags per model group
- Update async_filter_deployments and async_pre_call_deployment_hook to use per-group config
- Add 4 comprehensive tests covering per-group config, fallback, and override scenarios
This allows fine-grained control of affinity behavior across model groups, e.g., enabling stickiness only for cross-provider deployments while leaving other groups free to load-balance.
Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
Previously the test called common_processing_pre_call_logic in isolation,
making generate_polling_id.assert_not_called() vacuously true. Now the test
calls responses_api() end-to-end so it actually verifies that a rate-limited
request never receives a polling ID.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Guard logging_obj for None when skip_pre_call_logic=True: raise ValueError
if litellm_logging_obj not in data, preventing AttributeError downstream
- Add model=None to common_processing_pre_call_logic call in endpoints.py
to match style of other call sites
- Add test verifying rate-limited request never receives polling ID
Move pre-call checks (rate limits, guardrails, budget) to run BEFORE
polling ID creation in the background streaming flow. This prevents the
edge case where a rate-limited request receives a polling ID that
immediately fails.
Changes:
- Add skip_pre_call_logic parameter to base_process_llm_request to allow
skipping pre-call checks (avoiding double-counting of RPM/parallel requests)
- Run common_processing_pre_call_logic before generating polling ID in the
responses API endpoint. If rate limits/guardrails fail, return error
immediately without creating a polling ID
- Background streaming task passes skip_pre_call_logic=True to avoid re-running
pre-call checks that were already done before polling ID creation
- Add tests verifying skip_pre_call_logic parameter works correctly
Fixes the edge case where polling_via_cache would return a polling ID
for a request that immediately fails due to rate limiting.
- Change chunk["id"] to chunk.get("id") for compatibility with MiniMax
- ModelResponseStream auto-generates id when None is passed
- Add regression test test_chunk_parser_without_id_field
Replace bare _get_deployment_default_tpm/rpm_limit calls in the
async_log_success_event condition with get_key_model_tpm/rpm_limit
(model_name=model_group). The higher-level getters short-circuit on
key/team metadata hits before ever reaching the router, so requests
that don't use deployment defaults incur no extra router lookup. Remove
the now-unused bare helper imports.
Also fix invalid `int = None` type hints in test helper signatures
to `Optional[int] = None`.
Co-Authored-By: Claude (claude-sonnet-4-6) <noreply@anthropic.com>
async_log_success_event only updated the per-model cache counter when
model_rpm_limit / model_tpm_limit were present in key metadata or
model_max_budget was set. For the new deployment-default path
(default_api_key_tpm_limit / default_api_key_rpm_limit), none of those
conditions held, so current_tpm stayed at zero and tpm enforcement was
never applied across multiple requests.
Extend the guard condition to also trigger when the model group has a
deployment-default tpm or rpm limit, and import the two helpers at
module level.
Co-Authored-By: Claude (claude-sonnet-4-6) <noreply@anthropic.com>
Compute get_key_model_tpm/rpm_limit once before the guard condition
instead of calling each function twice (once to check non-None, once to
retrieve). Removes 2 extra llm_router.get_model_list() calls per request
when deployment defaults are active.
Co-Authored-By: Claude (claude-sonnet-4-6) <noreply@anthropic.com>
- Use min() across all matching deployments instead of first-wins when
resolving default_api_key_tpm/rpm_limit for a model group, so
load-balanced setups with different per-deployment limits always apply
the most conservative value
- Replace the global SensitiveDataMasker non_sensitive_overrides change
with a targeted excluded_keys set at the remove_sensitive_info_from_deployment
call site, avoiding unintended suppression of other fields
- Update the v1 parallel request limiter to pass model_name to
get_key_model_tpm/rpm_limit so deployment defaults apply there too
- Add 4 tests covering multi-deployment min semantics
Co-Authored-By: Claude (claude-sonnet-4-6) <noreply@anthropic.com>
Helicone (PRs #19288, #22603) and Langfuse (#22390) were present in the
v1.82.0-stable...v1.82.3-stable diff but omitted from the AI Integrations
logging section. Also updates the AI Integrations diff summary count from 2 to 4.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds `default_api_key_tpm_limit` and `default_api_key_rpm_limit` to
`GenericLiteLLMParams` so operators can set per-deployment rate limit
defaults in config.yaml. When a key has no model-specific tpm/rpm limit
configured, the proxy falls back to these deployment defaults (Case 2 in
spec). Key-level limits always take priority (Case 1).
- Extends `get_key_model_tpm_limit` / `get_key_model_rpm_limit` with a
`model_name` param and a priority-4 deployment-default fallback
- Passes `model_name=requested_model` in the parallel request limiter so
the fallback is triggered at enforcement time
- Adds `"limit"` to `SensitiveDataMasker` non-sensitive overrides so
`*_limit` fields are not masked in `/model/info` responses
- Adds 17 unit tests covering both spec cases and the `/model/info` path
Co-Authored-By: Claude (claude-sonnet-4-6) <noreply@anthropic.com>
- Add 'Contributing to Guardrails' category with links to:
- Generic Guardrail API (integrate without PR)
- Adding a New Guardrail Integration tutorial
- Adding Guardrail Support to Endpoints
- Add 'Team Bring-Your-Own Guardrails' link for team BYOG workflow
These docs existed but were only accessible from the 'LiteLLM AI Gateway'
sidebar. Now they're also accessible when browsing the 'Guardrail Providers'
section.
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
