* fix(model_cost): add missing supports_system_messages and supports_tool_choice to bedrock/moonshotai.kimi-k2.5
* fix(streaming): ensure role=assistant is set on first streaming chunk via strip_role_from_delta
* fix(vertex_ai): ensure role=assistant on first streaming chunk for Llama models
Add VertexAILlama3StreamingHandler that injects role='assistant' into the
first streaming chunk delta when the Vertex AI Llama API omits it.
* docs: add reference to example_openai_endpoint repo for self-hosting fake OpenAI proxy (#21006)
- Updated benchmarks.md with a section on setting up fake OpenAI endpoints
- Updated load_test.md to mention the self-hosted option
- Updated load_test_advanced.md with a tip box about the example repo
Reference: https://github.com/BerriAI/example_openai_endpoint
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
* MCP fixes
* fix(oldteams.tsx): show policies when creating
* fix(proxy/_types.py): ensure mcp rest endpoints can be called by virtual key
ensures UI works with virtual key testing mcp endpoints
* refactor: migrate get object permissions table logic to happen in user api key auth - allows functions to trust user api key object they receive has what they need
* fix(rest_endpoints.py): filter for allowed tools based on what key has access to
* fix(mcp_server_manager.py): ensure only allowed MCP's are returned to the user, via rest endpoints
* Guardrails - add toxic/abusive content filter guardrails
* fix(streaming): preserve usage data from post-finish_reason chunks in OpenAI-compatible streaming
Fixes#16112
OpenRouter and other OpenAI-compatible providers send a usage chunk after
the finish_reason='stop' chunk when stream_options.include_usage is True.
The OpenAIChatCompletionStreamingHandler.chunk_parser() was not passing
the usage field to ModelResponseStream, causing real token counts from the
provider to be lost and falling back to inaccurate estimates.
* fix: resolve merge conflict in test file
- Fix typo in test method name (extra space)
- Move test_prompt_cache_key_in_optional_params to its own class
---------
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
* fix: allow Management keys to access user/daily/activity and team/daily/activity
* feat(vertex): surface trafficType via generic provider_specific_fields in Responses API
Extract Vertex AI's trafficType from usageMetadata in both streaming and
non-streaming paths, storing it in _hidden_params["provider_specific_fields"].
The Responses API transformation layer generically passes any
_hidden_params["provider_specific_fields"] dict to the ResponsesAPIResponse,
avoiding provider-specific logic in the bridge.
Also fix stream_chunk_builder to propagate _hidden_params from the last
streaming chunk to the rebuilt ModelResponse, ensuring provider metadata
survives the chunk→response rebuild.
---------
Co-authored-by: naaa760 <neh6a683@gmail.com>
Co-authored-by: yuneng-jiang <yuneng.jiang@gmail.com>
OAuth tokens (sk-ant-oat*) require Authorization: Bearer header per
Anthropic's OAuth specification, but were being sent via x-api-key
which Anthropic rejects with 'invalid x-api-key'.
- optionally_handle_anthropic_oauth: detect OAuth tokens in api_key
param (standard chat flow), not just Authorization header
- get_anthropic_headers: use Authorization: Bearer + required OAuth
headers for OAuth tokens, x-api-key for regular API keys
- Passthrough messages: skip x-api-key when Authorization is set
- Add oauth-2025-04-20 to beta headers whitelist config
* Add http support to custom code guardrails + Unified guardrails for MCP + Agent guardrail support (#20619)
* fix: fix styling
* fix(custom_code_guardrail.py): add http support for custom code guardrails
allows users to call external guardrails on litellm with minimal code changes (no custom handlers)
Test guardrail integrations more easily
* feat(a2a/): add guardrails for agent interactions
allows the same guardrails for llm's to be applied to agents as well
* fix(a2a/): support passing guardrails to a2a from the UI
* style(code-editor): allow editing custom code guardrails on ui + add examples of pre/post calls for custom code guardrails
* feat(mcp/): support custom code guardrails for mcp calls
allows custom code guardrails to work on mcp input
* feat(chatui.tsx): support guardrails on mcp tool calls on playground
* fix(mypy): resolve missing return statements and type casting issues (#20618)
* fix(mypy): resolve missing return statements and type casting issues
* fix(pangea): use elif to prevent UnboundLocalError and handle None messages
Address Greptile review feedback:
- Make branches mutually exclusive using elif to prevent input_messages from being overwritten
- Handle case where data.get('messages') returns None to avoid passing invalid payload to Pangea API
---------
Co-authored-by: Shin <shin@openclaw.ai>
* [Feat] MCP Gateway - Allow setting MCP Servers as Private/Public available on Internet (#20607)
* update MCPAuthenticatedUser
* add available_on_public_internet for MCPs
* update claude.md
* init IPAddressUtils
* init available_on_public_internet
* add on REST endpoints
* filter with IP
* TestIsInternalIp
* _extract_mcp_headers_from_request
* init get_mcp_client_ip
* _get_general_settings
* allowed_server_ids
* address PR comments
* get_mcp_server_by_name fix
* fix server
* fix review comments
* get_public_mcp_servers
* address _get_allowed_mcp_servers
* fixing user_id
* [Feat] IP-Based Access Control for MCP Servers (#20620)
* update MCPAuthenticatedUser
* add available_on_public_internet for MCPs
* update claude.md
* init IPAddressUtils
* init available_on_public_internet
* add on REST endpoints
* filter with IP
* TestIsInternalIp
* _extract_mcp_headers_from_request
* init get_mcp_client_ip
* _get_general_settings
* allowed_server_ids
* address PR comments
* get_mcp_server_by_name fix
* fix server
* fix review comments
* get_public_mcp_servers
* address _get_allowed_mcp_servers
* test fix
* fix linting
* inint ui types
* add ui for managing MCP private/public
* add ui
* fixes
* add to schema
* add types
* fix endpoint
* add endpoint
* update manager
* test mcp
* dont use external party for ip address
* Add OpenAI/Azure release test suite with HTTP client lifecycle regression detection (#20622)
* docs (#20626)
* docs
* fix(mypy): resolve type checking errors in 5 files (#20627)
- a2a_protocol/exception_mapping_utils.py: Fix type ignore comment for None assignment
- caching/redis_cache.py: Add type ignore for async ping return type
- caching/redis_cluster_cache.py: Add type ignore for async ping return type
- llms/deprecated_providers/palm.py: Add type ignore for palm.generate_text
- proxy/auth/handle_jwt.py: Add type ignore for jwt.decode options argument
All changes add appropriate type: ignore comments to handle library typing inconsistencies.
* fix(test): update deprecated gemini embedding model (#20621)
Replace text-embedding-004 with gemini-embedding-001.
The old model was deprecated and returns 404:
'models/text-embedding-004 is not found for API version v1beta'
Co-authored-by: Shin <shin@openclaw.ai>
* ui new buil
* fix(http_handler): bypass cache when shared_session is provided for aiohttp tracing
When users pass a shared_session with trace_configs to acompletion(),
the get_async_httpx_client() function was ignoring it and returning
a cached client without the user's tracing configuration.
This fix bypasses the cache when shared_session is provided, ensuring
the user's ClientSession (with its trace_configs, connector settings, etc.)
is actually used for the request.
Fixes#20174
---------
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
Co-authored-by: Shin <shin@openclaw.ai>
Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com>
Co-authored-by: yuneng-jiang <yuneng.jiang@gmail.com>
Co-authored-by: Alexsander Hamir <alexsanderhamirgomesbaptista@gmail.com>
Co-authored-by: shin-bot-litellm <shin-bot-litellm@users.noreply.github.com>
When an OpenAI-compatible upstream provider emits minimal streaming event
payloads that omit required fields (e.g. created_at, output, output_index,
content_index), Pydantic raises a ValidationError crashing the SSE stream
and returning HTTP 500.
Fall back to model_construct() on ValidationError, consistent with the
existing pattern in transform_response_api_response for non-streaming.
Fixes https://github.com/BerriAI/litellm/issues/20570
Signed-off-by: Varun Chawla <varun_6april@hotmail.com>
HuggingFace Text Embeddings Inference (TEI) returns embeddings as raw
arrays [[0.1, 0.2, ...]] instead of wrapped format {"embedding": [...]}.
This change handles both formats:
- Raw array: [[...]] (TEI, some HF models)
- Wrapped: {"embedding": [[...]]} (standard HF format)
Fixes SagemakerError: "HF response missing 'embedding' field" when using
TEI containers on SageMaker.
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
* fix:fix: prompt_cache_key OAI + Azure OpenAI
* test_prompt_cache_key_supported
* test_azure_openai_with_prompt_cache_key
* fix: remove unnecessary async from test_azure_openai_with_prompt_cache_key
Addresses Greptile feedback: litellm.completion() is synchronous, so
async def is unnecessary and would silently pass without running.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: remove unused filter_and_transform_beta_headers imports
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test_azure_openai_with_prompt_cache_key
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix: _should_use_api_key_header
* test_azure_ai_validate_environment_with_api_key
* fix: remove unused top-level RouteChecks import
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* docs: add missing env keys to config_settings reference
Add MODEL_COST_MAP_MIN_MODEL_COUNT, MODEL_COST_MAP_MAX_SHRINK_RATIO,
and MAX_POLICY_ESTIMATE_IMPACT_ROWS to the environment variables
reference table.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
1. Cross-account false match: Added _parse_arn_account_and_role_name()
helper that compares partition + account ID + role name (not just
role name) to prevent same-name-different-account false matches.
2. SSL verify: _is_already_running_as_role() now passes ssl_verify to
the STS client via self._get_ssl_verify(), consistent with all other
boto3 client creation in this module.
3. Overbroad AccessDenied fallback: The catch in _auth_with_aws_role now
only falls back to ambient credentials when _is_already_running_as_role
positively confirms the caller is the target role. Genuine trust-policy
or permission misconfigurations are re-raised.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
When aws_role_name is configured but the environment (ECS task role, EC2 instance
profile) is already running as that role, AssumeRole is unnecessary and can fail
with AccessDenied. This adds same-role detection for ECS/EC2 (extending existing
IRSA support) and a fallback to ambient credentials when AssumeRole returns
AccessDenied.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* feat: add opus 4.5 and 4.6 to use outout_format param
* generate poetry lock with 2.3.2 poetry
* restore poetry lock
* e2e tests, key delete, update tpm rpm, and regenerate
* Split e2e ui testing for browser
* new login with sso button in login page
* option to hide usage indicator
* fix(cloudzero): update CBF field mappings per LIT-1907 (#20906)
* fix(cloudzero): update CBF field mappings per LIT-1907
Phase 1 field updates for CloudZero integration:
ADD/UPDATE:
- resource/account: Send concat(api_key_alias, '|', api_key_prefix)
- resource/service: Send model_group instead of service_type
- resource/usage_family: Send provider instead of hardcoded 'llm-usage'
- action/operation: NEW - Send team_id
- resource/id: Send model name instead of CZRN
- resource/tag:organization_alias: Add if exists
- resource/tag:project_alias: Add if exists
- resource/tag:user_alias: Add if exists
REMOVE:
- resource/tag:total_tokens: Removed
- resource/tag:team_id: Removed (team_id now in action/operation)
Fixes LIT-1907
* Update litellm/integrations/cloudzero/transform.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix: define api_key_alias variable, update CBFRecord docstring
- Fix F821 lint error: api_key_alias was used but not defined
- Update CBFRecord docstring to reflect LIT-1907 field mappings
- Remove unused Optional import
---------
Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* Add banner notifying of breaking change
* Add semgrep & Fix OOMs (#20912)
* [Feat] Policies - Allow connecting Policies to Tags, Simulating Policies, Viewing how many keys, teams it applies on (#20904)
* 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>
* fix: type error & better error handling (#20689)
* [Docs] Add docs guide for using policies (#20914)
* 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>
* feat: add dashscope/qwen3-max model with tiered pricing (#20919)
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>
* fix linting
* docs: add Greptile review requirement to PR template (#20762)
* fix(azure): preserve content_policy_violation error details from Azure OpenAI
Closes#20811
Azure OpenAI returns rich error payloads for content policy violations
(inner_error with ResponsibleAIPolicyViolation, content_filter_results,
revised_prompt). Previously these details were lost when:
1. The top-level error code was not "content_policy_violation" but the
inner_error.code was "ResponsibleAIPolicyViolation" -- the structured
check only examined the top-level code.
2. The DALL-E image generation polling path stringified the error JSON
into the message field instead of setting the structured body, making
it impossible for exception_type() to extract error details.
3. The string-based fallback detector used "invalid_request_error" as a
content-policy indicator, which is too broad and could misclassify
regular bad-request errors.
Changes:
- exception_mapping_utils.py: Check inner_error.code for
ResponsibleAIPolicyViolation when top-level code is not
content_policy_violation. Replace overly broad "invalid_request_error"
string match with specific Azure safety-system messages.
- azure.py: Set structured body on AzureOpenAIError in both async and
sync DALL-E polling paths so exception_type() can inspect error details.
- test_azure_exception_mapping.py: Add regression tests covering the
exact error payloads from issue #20811.
- Fix pre-existing lint: duplicate PerplexityResponsesConfig dict key,
unused RouteChecks top-level import.
---------
Co-authored-by: Kelvin Tran <kelvin-tran@users.noreply.github.com>
Co-authored-by: yuneng-jiang <yuneng.jiang@gmail.com>
Co-authored-by: shin-bot-litellm <shin-bot-litellm@berri.ai>
Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Co-authored-by: Alexsander Hamir <alexsanderhamirgomesbaptista@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Harshit Jain <48647625+Harshit28j@users.noreply.github.com>
Co-authored-by: ken <122603020@qq.com>
Co-authored-by: Sameer Kankute <sameer@berri.ai>
* 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).
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.
Set cached tokenizer config directly and mock both sync and async
tokenizer functions to avoid race conditions when running with
parallel test execution (-n 16).
The issue was that parallel tests could populate the
litellm.known_tokenizer_config cache between clearing it and
when the code checked it. This caused the sync code path to be
used instead of the async path, bypassing the mocked async functions.
Fix:
1. Set cache directly instead of clearing it
2. Also mock sync versions _get_tokenizer_config and _get_chat_template_file
This ensures the test is deterministic regardless of test execution order.
