Models like gemini-3.1-flash-lite-preview send the final streaming chunk
with empty content (text:"") alongside finishReason:"STOP", instead of
omitting content entirely. The existing fix (PR #21577) only handled
chunks without content, so this case was missed.
Now, after processing candidates, if tool_calls were seen in earlier
chunks and a choice has finish_reason="stop", it is overridden to
"tool_calls" to match the OpenAI spec.
Fixes#22900
- Add WARNING docstring to _get_shared_session_lock() about not resetting
the lock to None while coroutines may be in the recovery path
- Remove redundant proxy_server_module.shared_aiohttp_session assignment
in mock_init (add_shared_session_to_data overwrites it synchronously)
- Add try/except around _initialize_shared_aiohttp_session call to catch
and log exceptions (instead of letting them bubble to outer handler)
- Fix warning message when re-checked session is None (was incorrectly
logging closed session ID on a None session)
- Add debug logging to outer except handler instead of bare pass
- Add test for _initialize_shared_aiohttp_session raising exception
Strip internal logging ids from emulated sub-calls, dedupe included search_results by file_id, clean unused imports, and add unit coverage for dedupe behavior.
Made-with: Cursor
Address Greptile P1 review: tests that exercise the closed-session code
path need to reset the module-level lock to avoid RuntimeError on
Python < 3.10 when asyncio.Lock is reused across different event loops.
Include all function_call items when building emulated follow-up input and update tests to assert real emulated routing + Responses-format function tool structure.
Made-with: Cursor
When multiple requests detect a closed shared session simultaneously,
they would each create a new aiohttp.ClientSession, leaking intermediate
sessions and their TCP connectors. Added double-checked locking pattern
with asyncio.Lock to ensure only one coroutine recreates the session.
Added concurrent recreation test case.
Add Vertex batch cancellation support in LiteLLM batch APIs, route proxy cancel fallback using request provider headers, and return post-cancel batch state via retrieve to keep response shape compatible.
Made-with: Cursor
* fix(gemini): support images in tool_results for /v1/messages routing
convert_to_gemini_tool_call_result() dropped images in two cases:
- data-URL strings (data:image/...;base64,...) treated as plain text
- Anthropic image blocks in list content skipped
Add detection and convert both to Gemini inline_data BlobType so image
bytes are preserved.
Fixes#23712.
* fix(gemini): support images in tool_results for /v1/messages routing
convert_to_gemini_tool_call_result() dropped images in two cases:
- data-URL strings (data:image/...;base64,...) treated as plain text
- Anthropic image blocks in list content skipped
Add detection and convert both to Gemini inline_data BlobType so image
bytes are preserved.
Fixes#23712.
* fix(gemini): support images in tool_results for /v1/messages routing
convert_to_gemini_tool_call_result() dropped images in two cases:
- data-URL strings (data:image/...;base64,...) treated as plain text
- Anthropic image blocks in list content skipped
Add detection and convert both to Gemini inline_data BlobType so image
bytes are preserved.
Fixes#23712.
* fix(fireworks): skip #transform=inline for base64 data URLs
Closes#23583
* fix(fireworks): skip #transform=inline for base64 data URLs
Closes#23583
Appending #transform=inline to a data: URL corrupted the base64 payload,
causing binascii.Error (Incorrect padding) when Fireworks AI attempted to
decode the image. Data URLs are already inlined so the fragment is a no-op
anyway — guard both the str and dict image_url branches to skip the suffix
when the URL starts with "data:".
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(fireworks): skip #transform=inline for base64 data URLs
Closes#23583
* fix(fireworks): skip #transform=inline for base64 data URLs
Closes#23583
* fix(fireworks): skip #transform=inline for base64 data URLs
Closes#23583
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
* fix: Register DynamoAI guardrail initializer and enum entry
Fix the "Unsupported guardrail: dynamoai" error by:
1. Adding DYNAMOAI to SupportedGuardrailIntegrations enum
2. Implementing initialize_guardrail() and registries in dynamoai/__init__.py
The DynamoAI guardrail was added in PR #15920 but never properly registered
in the initialization system. The __init__.py was missing the
guardrail_initializer_registry and guardrail_class_registry dictionaries
that the dynamic discovery mechanism looks for at module load time.
Fixes#22773
Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
* Update litellm/proxy/guardrails/guardrail_hooks/dynamoai/__init__.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* Update litellm/proxy/guardrails/guardrail_hooks/dynamoai/__init__.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* test: Add tests for DynamoAI guardrail registration
Verifies enum entry, initializer registry, class registry,
instance creation, and global registry discovery.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Haiku 4.5 <noreply@anthropic.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Add disable_custom_api_keys UI setting that prevents users from specifying
custom key values during key generation and regeneration. When enabled, all
keys must be auto-generated, eliminating the risk of key hash collisions
in multi-tenant environments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When the shared aiohttp session closes (due to network interruption,
idle timeout, or Redis failover side effects), the proxy permanently
falls back to creating a new HTTPS connection per request, losing the
benefit of connection pooling for the entire pod lifetime.
Fix: make add_shared_session_to_data() async and recreate the session
when it is found closed, restoring connection pooling automatically.
Fixes#23806
When updating or regenerating a key without changing its key_alias, the
existing alias was being re-validated against current format rules. This
caused keys with legacy aliases (created before stricter validation) to
become uneditable. Now validation only runs when the alias actually changes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move the non-admin team validation into the existing get_team_object call
site to avoid an extra DB round-trip. The existing call already fetches
the team for limits checking — we now add the LIT-1884 guard there when
team_obj is None for non-admin callers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Internal users could exploit key/generate and key/update to create unbound
keys (no user_id, no budget) or attach keys to non-existent teams. This
adds validation for non-admin callers: auto-assign user_id on generate,
reject invalid team_ids, and prevent removing user_id on update.
Closes LIT-1884
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(test): add missing mocks for test_streamable_http_mcp_handler_mock
The test was missing mocks for extract_mcp_auth_context and set_auth_context,
causing the handler to fail silently in the except block instead of reaching
session_manager.handle_request. This mirrors the fix already applied to the
sibling test_sse_mcp_handler_mock.
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
* fix(ci): route OpenAI models through chat completions in pass-through tests
The test_anthropic_messages_openai_model_streaming_cost_injection test fails
because the OpenAI Responses API returns 400 for requests routed through the
Anthropic Messages endpoint. Setting LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES=true
routes OpenAI models through the stable chat completions path instead.
Cost injection still works since it happens at the proxy level.
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
* fix(ci): fix assemblyai custom auth and router wildcard test flakiness
1. custom_auth_basic.py: Add user_role='proxy_admin' so the custom auth
user can access management endpoints like /key/generate. The test
test_assemblyai_transcribe_with_non_admin_key was hidden behind an
earlier -x failure and was never reached before.
2. test_router_utils.py: Add flaky(retries=3) and increase sleep from 1s
to 2s for test_router_get_model_group_usage_wildcard_routes. The async
callback needs time to write usage to cache, and 1s is insufficient on
slower CI hardware.
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
* ci: retrigger CI pipeline
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
* fix(mypy): use LitellmUserRoles enum instead of raw string in custom_auth_basic
Fixes mypy error: Argument 'user_role' has incompatible type 'str'; expected 'LitellmUserRoles | None'
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
* fix: don't close HTTP/SDK clients on LLMClientCache eviction (#22926)
* 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
* [Fix] Install bsdmainutils for column command in security scans
The security_scans.sh script uses `column` to format vulnerability
output, but the package wasn't installed in the CI environment.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: handle string callback values in prometheus multiproc setup
When callbacks are configured as a plain string (e.g., `callbacks: "my_callback"`)
instead of a list, the proxy crashes on startup with:
TypeError: can only concatenate str (not "list") to str
Normalize each callback setting to a list before concatenating.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* bump: version 1.82.2 → 1.82.3
* fix(test): update test_startup_fails_when_db_setup_fails for opt-in enforcement
The --enforce_prisma_migration_check flag is now required to trigger
sys.exit(1) on DB migration failure, after #23675 flipped the default
behavior to warn-and-continue.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(cost_calculator): use model name for per-request custom pricing when router_model_id has no pricing
When custom pricing is passed as per-request kwargs (input_cost_per_token/output_cost_per_token),
completion() registers pricing under the model name, but _select_model_name_for_cost_calc was
selecting the router deployment hash (which has no pricing data), causing response_cost to be 0.0.
Now checks whether the router_model_id entry actually has pricing before preferring it.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>
Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Non-admin users (INTERNAL_USER) could call /key/block and /key/unblock on
arbitrary keys, and modify max_budget on their own keys via /key/update.
These endpoints are now restricted to proxy admins, team admins, or org admins.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Model-level guardrails (litellm_params.guardrails on a deployment) were
only merged into request metadata in the streaming post_call path
(async_post_call_streaming_hook) but not in the non-streaming path
(post_call_success_hook). This caused should_run_guardrail to skip the
guardrail because the guardrail name was never added to metadata.guardrails.
Add the same _check_and_merge_model_level_guardrails call to
post_call_success_hook before the guardrail callback loop, mirroring the
streaming path.
Fixes model-level guardrails silently not firing for non-streaming
post_call requests.
Add avideo_create_character, avideo_get_character, avideo_edit, and avideo_extension
to the skip condition since Azure video calls don't use initialize_azure_sdk_client.
Tests now properly skip with expected behavior instead of failing:
- test_ensure_initialize_azure_sdk_client_always_used[avideo_create_character] ✓
- test_ensure_initialize_azure_sdk_client_always_used[avideo_get_character] ✓
- test_ensure_initialize_azure_sdk_client_always_used[avideo_edit] ✓
- test_ensure_initialize_azure_sdk_client_always_used[avideo_extension] ✓
Made-with: Cursor
* fix: Fixes https://github.com/BerriAI/litellm/issues/23185
* fix(responses/main.py): ensure litellm metadata custom cost works
* refactor: move all logging updates to a common function, to have just 1 place to update logging kwarg updates
The eager tiktoken tests were clearing all litellm modules from
sys.modules and re-importing, creating new module objects with different
class identities. This broke unittest.mock.patch for all subsequent
tests on the same xdist worker. Running these tests in subprocesses
provides perfect isolation.
Fixes: test_metadata_passed_to_custom_callback_codex_models,
test_oidc_github_success, test_oidc_google_cached,
test_oidc_google_failure,
test_encrypted_content_affinity_bypasses_rpm_limits, and 5 others.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The isolate_litellm_state conftest fixture saved/restored litellm.callbacks
but never cleared it before each test, unlike the other callback lists. It
also didn't handle litellm.model_fallbacks. Leaked callbacks and fallback
config caused mocked tests to route through Router/fallback paths, hitting
real APIs with mock keys.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The 4 integration tests were flaky in CI because the AsyncHTTPHandler.post
mock was bypassed when aiohttp transport is used. Mock at the higher
BaseLLMHTTPHandler.async_response_api_handler level instead, which
bypasses the HTTP layer entirely while still exercising router deployment
selection, pre-call checks, and response post-processing (item ID rewriting).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
