* fix(streaming): map unknown finish_reason values to finish_reason_unspecified
Some LLM providers return non-standard finish_reason values that are not
in the OpenAIChatCompletionFinishReason Literal (e.g. ZhipuAI/GLM returns
'network_error' when a streaming error occurs mid-response).
Previously map_finish_reason() fell through with return finish_reason,
passing the unknown value directly to Choices.__init__() which calls
Pydantic validation. This caused a ValidationError that was caught by
stream_chunk_builder() and re-raised as the misleading:
litellm.APIError: Error building chunks for logging/streaming usage calculation
Fix: after all known provider-specific mappings, check if the value is in
the valid set (stop, length, tool_calls, content_filter, function_call,
guardrail_intervened, eos, finish_reason_unspecified, malformed_function_call).
Any value not in this set is mapped to 'finish_reason_unspecified' instead
of being returned as-is.
This is consistent with how other unknown stop reasons (e.g. Vertex AI's
FINISH_REASON_UNSPECIFIED) are already handled.
* refactor: use get_args(OpenAIChatCompletionFinishReason) for valid set
Per code review feedback: replace the hardcoded _valid_finish_reasons set
with a module-level frozenset derived dynamically from the source-of-truth
Literal type via typing.get_args(). This ensures the valid-reason check
stays in sync automatically when new finish reasons are added to the Literal,
and avoids recreating the set on every streaming chunk call.
* test(map_finish_reason): add unit tests and warning log for unknown finish reasons
- Add TestMapFinishReason class in test_core_helpers.py covering:
- All known OpenAI-native values pass through unchanged (parametrized)
- Provider-specific mappings: Anthropic, Cohere, Vertex AI
- Unknown/provider-specific values map to 'finish_reason_unspecified'
- Regression test for ZhipuAI/GLM-5 'network_error' case
- Add verbose_logger.warning() in map_finish_reason() when an unknown
finish_reason is encountered, so operators can track which providers
return non-standard values
* fix: add missing indexes for top CPU-consuming queries
Add indexes to eliminate full table scans on two of the top 5 queries
by CPU usage:
1. LiteLLM_VerificationToken(key_alias) — for ORDER BY key_alias ASC
queries when listing verification tokens
2. LiteLLM_SpendLogs(user, startTime) — for WHERE user = $1 AND
startTime BETWEEN $2 AND $3 GROUP BY queries on the spend logs
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use CREATE INDEX CONCURRENTLY to avoid table locks
Both indexes are now created with CONCURRENTLY and IF NOT EXISTS
to avoid blocking writes on large production tables.
Uses -- SkipTransactionBlock for Prisma migrate compatibility.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Gemini 2.0+ natively accepts JSON Schema in tool parameters, including
bare {} (TYPE_UNSPECIFIED), anyOf with null, and lowercase types. The
existing _build_vertex_schema pipeline was coercing {} to {"type": "object"},
breaking JsonValue/Any field semantics (issue #22391).
Add _build_vertex_schema_for_gemini_2() that only resolves $ref (which
Gemini doesn't support in tools) and filters unsupported fields. Use it
for Gemini 2.0+ models, keeping the full transform for Gemini 1.5.
Extend the "Proxy database access" section with guidelines to prevent
common DB performance issues, tailored to actual Prisma usage patterns
in the litellm codebase: N+1 queries, client-side processing, batching
writes, bounding result sets, select on wide tables, index coverage,
and schema file sync.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The test calls OpenAI's gpt-4o-audio-preview model which sometimes
doesn't return usage data in the streaming response. Fixed by:
- Adding @pytest.mark.flaky(retries=5, delay=2) for retry handling
- Fixing usage_obj loop to check chunk.usage is not None
- Skipping gracefully when OpenAI doesn't return usage data
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The test_create_skill test was consistently failing in CI with a 500 from
Anthropic because the SKILL.md frontmatter always used the same hardcoded
name (test-skill-litellm). Since test_delete_skill is permanently skipped,
skills accumulate in the CI account, and re-creating with a duplicate name
triggers an Internal Server Error on Anthropic's side.
Fix: pass a timestamp-based unique_suffix to create_skill_zip so each run
produces a distinct skill name in the zip's SKILL.md frontmatter.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add input_fidelity parameter ("high"/"low") to the image edit pipeline,
allowing users to control how much effort the model exerts to match
input image style and features. Fixes#22813.
The Claude Agent SDK sends max_tokens=32000 for unrecognized model names
(like "bedrock-nova-pro"), which exceeds Nova Pro's 10,000 limit. Enable
modify_params in the test proxy config so LiteLLM clamps max_tokens to the
model's actual limit. Also swap nova-premier to nova-pro since premier
requires provisioned throughput unavailable in CI.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The test was flaky under pytest-xdist parallel execution because it used
async acompletion (which runs completion() in a thread pool via
run_in_executor) and relied on shared global state (known_tokenizer_config,
iam_token_cache, module_level_client) that could be modified by other tests
running in parallel. Failures were silently swallowed by a broad try/except,
causing mock_post.call_count to remain 0.
Fix:
- Convert from async acompletion to sync completion, matching every other
test in the file. The test's intent is verifying prompt transformation,
not async behavior.
- Use monkeypatch.setitem for known_tokenizer_config to ensure proper
teardown isolation.
- Remove unnecessary mock layers (async template fetchers, iam_token_cache
pre-population, mock completion response) that were only needed for the
async code path.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When retrieving a batch via the unified batch ID path, only unified_batch_id
was set on _hidden_params but model_id was missing. The managed files hook
requires both to encode output_file_id into a managed ID.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The SearXNG search tests were failing in CI because they depend on a live
SearXNG instance that returns results. Since this provider is used by a
very small subset of customers, replace the flaky integration tests with
deterministic unit tests that validate request payloads, URL construction,
response parsing, and header configuration without requiring external infra.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The gemini-3.1-flash-image-preview model introduced a new pricing field
that was missing from the test's validation schema and cost_fields list.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Chocolatey v2.5.1 introduced interactive prompts that block CI. Add
--no-progress, --force flags and CHOCOLATEY_CONFIRM_ALL env var to
fully suppress user input in non-interactive environments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(agentcore): handle JSON responses from agents using sync return
BedrockAgentCoreApp agents that use synchronous `return` (instead of
async `yield`) respond with Content-Type: application/json instead of
text/event-stream. The streaming parser only handles SSE format, silently
discarding the JSON body and returning empty content to the client.
This adds Content-Type detection in both sync and async streaming
wrappers — when application/json is received, the response is parsed
and converted to a single-chunk stream. Also extends _parse_json_response
with a fallback chain supporting multiple agent response schemas (standard
AgentCore, Strands framework, plain string, raw JSON fallback).
* fix(agentcore): add dict-type guard to _parse_json_response
Prevent AttributeError when json.loads() returns a non-dict
(e.g. JSON array or primitive) by adding an isinstance check
at the top of _parse_json_response. Non-dict values fall back
to raw JSON string content.
* fix(agentcore): handle malformed JSON and split streaming chunks
- Wrap json.loads() in try/except in both sync and async streaming
wrappers so malformed JSON bodies raise a structured BedrockError
instead of a raw JSONDecodeError
- Split the JSON-fallback streaming path into two chunks (content
chunk with finish_reason=None, then stop sentinel with empty delta)
to match the SSE path convention
* fix(agentcore): catch IO errors in streaming JSON path + async error test
- Broaden except clause to catch both json.JSONDecodeError and IO-level
exceptions (httpx.ReadError, etc.) from response.read()/aread(), so
all failures surface as structured BedrockError
- Add async malformed-JSON test to mirror the sync test coverage
* feat(proxy): add Prisma DB pool and engine health metrics to Prometheus
Add a PrismaMetricsCollector that periodically queries pg_stat_activity
and the Prisma engine process to expose connection pool and engine health
as Prometheus gauges/counters. Auto-enabled when prometheus_system is in
service_callback.
New metrics:
- litellm_db_pool_active_connections (Gauge)
- litellm_db_pool_idle_connections (Gauge)
- litellm_db_pool_total_connections (Gauge)
- litellm_db_pool_waiting_connections (Gauge)
- litellm_db_engine_up (Gauge)
- litellm_db_engine_restarts_total (Counter)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: address Greptile review feedback
- Only increment engine_restarts counter on heavy reconnects (engine
actually dead), not lightweight network-blip reconnects
- Fix potential KeyError in _get_or_create_gauge/counter fallback path
when REGISTRY._names_to_collectors is absent
- Rename litellm_db_pool_waiting_connections to
litellm_db_pool_lock_waiting_connections to clarify it measures lock
contention, not pool slot queuing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: warn when prometheus_system enabled but watchdog disabled
Log a warning when users have prometheus_system in service_callback
but PRISMA_HEALTH_WATCHDOG_ENABLED=false, since DB pool and engine
metrics won't be collected in that configuration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* ci: retrigger CI checks
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor: use labeled gauge for DB pool connection metrics
Replace 3 separate pool gauges (active, idle, total) with a single
`litellm_db_pool_connections` gauge using a `state` label. This is more
Prometheus-idiomatic and exposes all pg_stat_activity states (active,
idle, idle in transaction, etc.) without ambiguity about what "total"
includes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: address Greptile review — stale labels and fallback re-registration
- Zero out known pg_stat_activity states that are absent from the current
query result, preventing stale gauge values from persisting.
- Simplify _get_or_create_gauge/counter by removing the fallback loop
that could re-register an already-registered metric (ValueError).
- Add test for stale label clearing across collection cycles.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: include "unknown" in _PG_STATES for stale label clearing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: collect immediately on start and consolidate into single query
- Move sleep to end of loop so metrics appear on /metrics immediately
after startup instead of after a 30s delay.
- Combine pool state and lock waiting queries into a single SQL query
using conditional aggregation, halving per-cycle DB overhead.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: prevent tight spin loop on collection error
Move asyncio.sleep outside the try/except so it always executes even
when _collect_engine_health() or _collect_pool_metrics() raises.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: add multiprocess_mode to _get_or_create_gauge initialization
- Include `multiprocess_mode` parameter to properly support multiprocessing in Gauge creation.
- Ensure consistent behavior for labeled and unlabeled Gauges.
* fix: handle invalid env var and document watchdog prerequisite
- Add try/except ValueError for PRISMA_METRICS_COLLECTION_INTERVAL_SECONDS
to prevent proxy startup crash on non-numeric values (e.g. "30s")
- Document that DB metrics require both prometheus_system callback and
PRISMA_HEALTH_WATCHDOG_ENABLED=true
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use defensive null coalescing for query_raw row values
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test: add invalid env var fallback test and fix mock signature
- Add test for non-numeric PRISMA_METRICS_COLLECTION_INTERVAL_SECONDS
- Add **kwargs to mock _patched_get_or_create_gauge for forward compat
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
When both enable_jwt_auth and enable_oauth2_auth are True, the proxy now
routes tokens based on their format:
- JWT tokens (3 dot-separated parts) -> JWT auth handler
- Opaque tokens -> OAuth2 auth handler
This enables using JWT for human users and OAuth2 for M2M (machine) clients
on the same LiteLLM instance. Previously, enabling OAuth2 would intercept
all tokens on LLM API routes before JWT auth could run.
When only one auth method is enabled, behavior is unchanged (backward compatible).
