* fix:Parse embedded JSON in the message field of logs
* Update litellm/_logging.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>
When transforming chat completion responses to Responses API format,
image_tokens from completion_tokens_details was not being included
in output_tokens_details. This affected Vertex AI/Gemini models that
return image token counts in candidatesTokensDetails with modality="IMAGE".
The fix adds image_tokens handling alongside existing reasoning_tokens
and text_tokens transformation.
* fix(responses): preserve streamed tool deltas when id is omitted
* fix(responses): guard ambiguous tool-call index reuse
* add missing indexes on VerificationToken table
* fix(bedrock): handle concatenated JSON in tool call arguments
When using Bedrock Claude Sonnet 4.5 with tools enabled, the model
sometimes returns multiple tool call arguments as concatenated JSON
objects in a single arguments string, e.g.
'{"command":["curl",...]}{"command":["curl",...]}{"command":["curl",...]}'
json.loads() fails on this with "Extra data", crashing the entire
request in _convert_to_bedrock_tool_call_invoke.
This commit:
- Adds split_concatenated_json_objects() helper in common_utils.py
that uses json.JSONDecoder.raw_decode() to walk a string and extract
each JSON object individually.
- Updates _convert_to_bedrock_tool_call_invoke() to catch JSONDecodeError
and attempt splitting concatenated objects into separate Bedrock
toolUse blocks (first block keeps original ID, subsequent blocks get
suffixed IDs).
- Fixes duplicate json.loads calls and a shadowed 'id' builtin.
- Adds 12 unit tests covering normal, empty, concatenated, and edge cases.
Fixes#20543
---------
Co-authored-by: Emerson Gomes <emerson.gomes@thalesgroup.com>
Co-authored-by: Sameer Kankute <sameer@berri.ai>
Co-authored-by: Carlo Alberto Ferraris <cafxx@mercari.com>
OIDC providers like Logto may return opaque (non-JWT) access tokens,
which caused jwt.decode() to raise DecodeError and crash the SSO
callback with a 500 error. Catch DecodeError and skip JWT-based
extraction gracefully, since user info is already available from
the UserInfo endpoint.
Fixes#20724
* Add chat completion support for websearch
* Add chat completion tool calls support and response transformation
* Add new methods in chat completion
* Add chat completion tool format
* Add callback for websearch in completion method
* Add test for web search
* Potential fix for code scanning alert no. 4046: Clear-text logging of sensitive information
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
* Update litellm/integrations/websearch_interception/tools.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix: empty guardrails/policies arrays should not trigger enterprise license check (#20567)
* fix: empty guardrails/policies arrays should not trigger enterprise license check (#20304)
The UI sends empty arrays for enterprise-only fields (guardrails, policies,
logging) even when the user has not configured these features. The backend
`is not None` check treated `[]` as a truthy intent to use the feature,
falsely requiring an enterprise license for basic team operations.
Backend: Add `and updated_kv[field] != [] and updated_kv[field] != {}`
guards in `_update_metadata_fields` so empty collections are skipped.
UI: Conditionally omit guardrails, logging, and policies from the
payload when empty instead of defaulting to `[]`.
Fixes#20304
* fix: allow clearing fields with empty collections while skipping enterprise check
Address PR review feedback:
1. Move the empty-collection guard into _update_metadata_field (singular)
so that empty lists/dicts skip only the premium license check but still
get written into metadata. This lets users intentionally clear a
previously-set field (e.g. guardrails: []) without being blocked, while
the UI's default empty arrays still don't trigger a false enterprise
error.
2. Remove sys.path hack from test file; use standard imports that work
with pytest discovery.
3. Add tests verifying that empty collections are moved into metadata
(field clearing works) even though they bypass the premium check.
Fixes#20304
* fix critical CVE vulnerabliltes (#20683)
* fix: add hook to handle db case (#20635)
* Add team policy mapping for zguard (#20608)
* support policy mapping on team key level
* update document
* update document
* address comments
* update document
* add unit test for new feature
* add more test case
* feat: add support for anthropic_messages call type in prompt caching (#19233)
* feat: add support for anthropic_messages call type in prompt caching
* test: move anthropic_messages prompt caching test to main router test file
* add tutorial on using claude code with prompt cache routing
* docs: add SDK proxy authentication (OAuth2/JWT auto-refresh) documentation (#20680)
Adds documentation for the litellm.proxy_auth feature that automatically
obtains and refreshes OAuth2/JWT tokens when connecting to a LiteLLM Proxy.
* Fixes#20582 (#20663)
* fix: show error details instead of Data Not Available for failed requests (#20656)
* fix(ui): add null guard for models in API keys table (#20655)
The VirtualKeysTable crashed when rendering keys with null or undefined
models field. The className expression tried to access .length on null,
throwing a TypeError that broke the entire keys table.
Added Array.isArray() guard before accessing .length on the models value.
Fixes#20611
* Fix: Spend logs pickle error with Pydantic models and redaction (#20685)
* docs: add callback registration optimization to v1.81.9 release notes (#20681)
* docs: add callback registration optimization to v1.81.9 release notes
* Update v1.81.9.md
---------
Co-authored-by: Alexsander Hamir <alexsanderhamirgomesbaptista@gmail.com>
* Fix spend logs pickle error with Pydantic models
Replace copy.deepcopy() with Pydantic-safe serialization to avoid
"cannot pickle '_thread.RLock' object" errors when request/response
redaction is enabled.
Changes:
- Add _convert_to_json_serializable_dict() helper that uses
model_dump() for Pydantic models instead of pickle
- Replace copy.deepcopy() calls in request and response redaction
paths with the new helper function
- Recursively handles nested dicts, lists, and Pydantic models
Root cause: Pydantic v2 BaseModel instances contain internal
_thread.RLock objects for thread-safety. When copy.deepcopy()
attempts to pickle these objects, it fails because threading
primitives cannot be pickled.
Fixes#20647
* chore: remove unused copy import
Remove unused copy import that was causing lint failure. The copy.deepcopy()
calls were replaced with _convert_to_json_serializable_dict() helper function
in the previous commit, making the copy module no longer needed.
---------
Co-authored-by: ryan-crabbe <128659760+ryan-crabbe@users.noreply.github.com>
Co-authored-by: Alexsander Hamir <alexsanderhamirgomesbaptista@gmail.com>
* fix(vertex_ai): propagate extra_headers anthropic-beta to request body (#20666)
Vertex AI requires Anthropic beta flags in the request body
(anthropic_beta array), not as HTTP headers. The Bedrock handler
already extracts user-specified beta headers from the headers dict,
but the Vertex handler was missing this, causing extra_headers like
interleaved-thinking-2025-05-14 to be silently dropped.
This extracts anthropic-beta values from optional_params extra_headers
and merges them into the anthropic_beta request body field, and also
removes extra_headers from the request body since the parent's
transform_request spreads optional_params into data.
* fix(streaming): preserve interleaved thinking/redacted blocks
* test(streaming): build thinking chunks with typed Delta/StreamingChoices
* Fix video list pagination cursors not encoded with provider metadata
first_id and last_id in the video list response were returned as raw
provider IDs while data[].id was properly wrapped with
encode_video_id_with_provider(). This caused pagination to break when
clients passed unencoded cursors back as the `after` parameter.
- Encode first_id/last_id in transform_video_list_response
- Decode the `after` param in transform_video_list_request via
extract_original_video_id()
- Add 6 unit tests covering encoding, decoding, passthrough, and
full round-trip pagination
Fixes#20708
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(responses): preserve streamed tool deltas when id is omitted
* fix(responses): guard ambiguous tool-call index reuse
* Add compaction for vertex ai
* Add all new feat for v1/messages
* Add inference_geo as supported messages param
* Add inference based costing
* Add inference_geo as supported messages param
* Add support for fast param
* Add fast mode for other providers
* Add documentation for Fast Mode
* add missing indexes on VerificationToken table
* Fix structured response of tool call
* Add tests for WebSearch interception with chat completions API
* Add doc for chat completion web search
* Fix: is_web_search_tool_chat_completion
* Fix double json import
* Add new vercel ai anthropic models
* Fix: base_model name for body and deplyment name in URL
* Add output_config as supported param
* Add response schema for vercel ai sonnet 4.5
* handle when litellm_parrams might be none
* Fix : litellm/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py
* fix: Missing return statement for async streaming
* Fix: get_supported_anthropic_messages_params
* Fix mypy issues
* Fix mypy issues
* Add support for extra fields in Generic SSO via GENERIC_USER_EXTRA_ATTRIBUTES
Enables extraction of additional fields from the Generic SSO userinfo endpoint response beyond the standard 8 fields (id, email, name, etc.). Custom handlers can now access these fields via CustomOpenID.extra_fields dict.
Changes:
- Add extra_fields: Optional[Dict[str, Any]] to CustomOpenID type
- Add GENERIC_USER_EXTRA_ATTRIBUTES env var (comma-separated field names)
- Extract specified fields using get_nested_value() with dot notation support
- Add 4 test cases covering basic, nested, and missing field scenarios
- Update custom_sso.py example showing how to access extra_fields
Backward compatible: extra_fields is None when env var not set
* docs: Add documentation for GENERIC_USER_EXTRA_ATTRIBUTES
Document the new GENERIC_USER_EXTRA_ATTRIBUTES environment variable for Generic SSO
- Add to admin_ui_sso.md: explanation and usage examples
- Add to config_settings.md: environment variable reference
- Add to custom_sso.md: code example showing how to access extra_fields
- Includes examples for nested field paths with dot notation
---------
Co-authored-by: Sameer Kankute <sameer@berri.ai>
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Co-authored-by: Varun Chawla <34209028+veeceey@users.noreply.github.com>
Co-authored-by: Harshit Jain <48647625+Harshit28j@users.noreply.github.com>
Co-authored-by: jwang-gif <j.wang@zscaler.com>
Co-authored-by: nuernber <benjamin.nuernberger@jpl.nasa.gov>
Co-authored-by: Cesar Garcia <128240629+Chesars@users.noreply.github.com>
Co-authored-by: John Lathouwers <john.lathouwers@oracle.com>
Co-authored-by: ryan-crabbe <128659760+ryan-crabbe@users.noreply.github.com>
Co-authored-by: Alexsander Hamir <alexsanderhamirgomesbaptista@gmail.com>
Co-authored-by: Elias Högbom Aronsson <elias.aronson@gmail.com>
Co-authored-by: Emerson Gomes <emerson.gomes@thalesgroup.com>
Co-authored-by: tshushan <tshushan@outbrain.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Carlo Alberto Ferraris <cafxx@mercari.com>
Fixes#17953
The `is_premium()` method logs two DEBUG messages on every call.
Since this method is called on every proxy request, the logs get
spammed with repeated "ENTERING IS_PREMIUM" and "Updated
license_str" messages, making it very difficult to debug actual
issues.
Add a `_premium_check_logged` flag so that these debug messages
are only emitted once (on the first call). The license check
itself continues to run on every call — only the verbose logging
is suppressed after the first invocation.
When MCP SDK hits root-level /register, /authorize, /token without
server name prefix, auto-resolve to the single configured OAuth2
server. Also fix WWW-Authenticate header to use correct public URL
behind reverse proxy.
* fix(prometheus): sanitize label values to prevent metric scrape failures
Unicode characters like U+2028 (Line Separator) in Prometheus label values
break the text exposition format, causing scrapers (e.g. Datadog) to fail
parsing the entire /metrics endpoint. One bad label value causes ALL metrics
to be lost, not just the affected metric.
Add _sanitize_prometheus_label_value() and apply it in prometheus_label_factory()
and all direct .labels() call sites.
* fix(prometheus): handle non-string label values in sanitization
Coerce non-string values (int, bool, float) to str before applying
sanitization, preventing AttributeError on .replace() calls.
* fix(prometheus): run sanitization on coerced non-string values
Non-string values should be coerced to str and then sanitized (not
returned early), so their string representations also get cleaned.
* fix(prometheus): widen type hint to Optional[Any] for label value sanitization
* fix(callbacks): allow MAX_CALLBACKS override via env var (#20778)
* fix(callbacks): allow MAX_CALLBACKS override via env var
- Move MAX_CALLBACKS from logging_callback_manager.py to constants.py
- Add LITELLM_MAX_CALLBACKS env var override (default: 30)
- Add troubleshooting doc explaining the limit and override
Fixes issue where large deployments with 60+ teams using guardrails
would hit the hardcoded MAX_CALLBACKS=30 limit and fail to start.
* docs: add max_callbacks to sidebar navigation
---------
Co-authored-by: shin-bot-litellm <shin-bot-litellm@users.noreply.github.com>
* fix callbacks issue
---------
Co-authored-by: shin-bot-litellm <shin-bot-litellm@berri.ai>
Co-authored-by: shin-bot-litellm <shin-bot-litellm@users.noreply.github.com>
