* feat(azure/prompt_shield.py): initial commit adding prompt shield guardrail + auto discovery mechanism for guardrails
reduces amount of code needed outside of guardrail integration for instrumentation
* feat(azure/prompt_shield.py): working azure prompt shield guardrail integration
Addresses https://github.com/BerriAI/litellm/issues/12254
* test: unit tests for prompt_shield
* fix(prompt_shield.py): add event hook validation for prompt shield guardrail
ensures prompt shield guardrail raises error if asked to run post_call (only runs on user prompt)
* feat(azure/): working text_moderation integration
* fix(text_moderation.py): suppress linting error
* test(test_azure_text_moderation.py): add unit test
* test(test_azure_text_moderation.py): add unit test for responses
* fix(text_moderation.py): return streaming error correctly
ensures error returned to user
* fix: fix linting error
* fix: fix linting check
* test: change mistral model
service tier exceeded
* fix(exception_mapping_utils.py): cover mistral in exception mapping
* Enhance Mistral API: Add support for parallel tool calls and refine name handling in tool messages. Plus, introduce a new test for parallel tool calls in the Mistral model.
* tests
* make mypy happy
* Refine name handling in Mistral chat transformation: clarify conditions for removing the 'name' field based on message role and content.
* refactor: streamline Mistral integration by removing deprecated references and adding a new handler
- Removed "mistral" from the list of compatible providers in constants.
- Updated the completion function in main.py to utilize the new Mistral handler.
- Deleted outdated Mistral chat and embedding files.
- Introduced a new handler for Mistral chat completions, implementing the llm_http_handler pattern.
- Added integration tests for the Mistral handler to ensure proper API base and key handling.
* lint
* fix: remove unneeded handler object
* add tests
* Addres PR comments
* fix(main.py): handle router custom azure model name for responses api bridge
* fix(responses/handler): ensure azure model name is stripped before sending to provider
Fixes model name error
* fix(google_genai/main.py): handle stream=true being set in kwargs
* docs: cleanup icons from sidebar
* fix(test-litellm.yml): add google-genai to test litellmyml
* fix(main.py): strip 'responses/' from bridge
* fix(main.py): fix linting errors
* fix(types/openai.py): allow item to be none
handle azure streaming response
* fix(base.py): allow extra fields + handle azure item = none value in response output item added event
* fix(main.py): correctly handle removing responses/
* test(test_main.py): add unit tests
When JSON_LOGS=True is set, error logs were not being formatted as JSON despite
the configuration. This was because the logging initialization code configured
individual loggers but failed to properly initialize all loggers with the JSON
formatter.
This fix ensures that when json_logs is enabled, the _initialize_loggers_with_handler()
function is called to:
- Configure all loggers (root, LiteLLM, Router, Proxy) with JSON formatter
- Disable logger propagation to prevent duplicate entries
- Set up exception handlers for JSON formatting
Fixes LIT-267
* Enhance Mistral API: Add support for parallel tool calls and refine name handling in tool messages. Plus, introduce a new test for parallel tool calls in the Mistral model.
* tests
* make mypy happy
* Refine name handling in Mistral chat transformation: clarify conditions for removing the 'name' field based on message role and content.
* handle mistral returning '' instead of None
* fix - tuple was never falsy so never triggered the exception
* test - add test suite for openmeter integration
* refactor - move tests for openmeter integration
* Move PANW Prisma AIRS test per feedback on PR #12116
- Move test to tests/test_litellm/proxy/guardrails/guardrail_hooks/
* Remove test file from old location
* Fix: Preserve full path structure for Gemini custom api_base (Fixes#11959)
This fix addresses an issue where custom api_base URLs (like Cloudflare AI Gateway)
were not working correctly with Google AI Studio (Gemini) models.
The problem was that the _check_custom_proxy method was simply appending the endpoint
to the custom base URL, resulting in malformed URLs like:
https://gateway.ai.cloudflare.com/v1/my-id/my-gateway/google-ai-studio:generateContent
Instead of the correct format:
https://gateway.ai.cloudflare.com/v1/my-id/my-gateway/google-ai-studio/v1beta/models/gemini-2.5-flash:generateContent
Changes:
- Modified _check_custom_proxy to preserve the full path structure from the original URL
- Extracts the path from the original Google AI URL and appends it to the custom base
- Maintains backward compatibility for Vertex AI models (unchanged behavior)
- Added comprehensive tests to verify the fix works correctly
Fixes#11959
* Fix: Update test to match actual Gemini URL format and fix double colon issue
- Fixed test expectation to include the full model path with 'gemini/' prefix
- Fixed double colon issue in Vertex AI URL construction when using custom api_base
- All tests now pass successfully
* fix(proxy_server.py): handle empty config yaml
Fixes https://github.com/BerriAI/litellm/issues/12163
* fix(gemini/common_utils.py): replace models/ as expected, instead of using 'strip'
Fixes https://github.com/BerriAI/litellm/issues/12160
* fix(anthropic/experimental_pass_through/messages/transformation.py): check for env var when selecting api key
* fix(anthropic/transformation.py): return tool_use content block start on anthropic bridge
Closes https://github.com/BerriAI/litellm/issues/12158
* fix(anthropic/streaming_iterator.py): fix setting index in block
ensure index is set just once and increments correctly when a new block is created
* fix(anthropic/adapters/handler.py): update logging obj with stream options value if set
* feat(anthropic/streaming_iterator.py): return usage from chat completion to messages bridge
enables usage tracking for non-anthropic models
Closes https://github.com/BerriAI/litellm/issues/12132
* fix(streaming_iterator.py): safely access usage chunk
* fix: suppress linting error
* test: update tests
* fix: fix streaming errors
The test_keys_delete_error_handling test was failing with:
- ConnectionError when the mock wasn't properly applied
- The test was checking str(result.exception) without first verifying exception exists
This fix adds an explicit check that result.exception is not None before
attempting to convert it to string, preventing potential AttributeError
and making the test more robust.
* Allow strings in calculate cost
Sometimes the cost per unit is a string (e.g.: If a value like "3e-7" was read from the config.yaml)
* Add comprehensive tests for string cost value handling
- Added test_string_cost_values() to test basic string cost conversion functionality
- Added test_calculate_cost_component_with_string_values() to test the calculate_cost_component function directly
- Added test_string_cost_values_edge_cases() to test mixed string/float costs and error handling
- Added test_string_cost_values_with_threshold() to test string costs with threshold pricing
- Enhanced _get_token_base_cost() to handle string-to-float conversion for base costs and threshold costs
- Enhanced generic_cost_per_token() to handle string-to-float conversion for audio and reasoning token costs
- All tests cover scientific notation (e.g., '3e-7'), decimal notation (e.g., '0.000001'), and error handling for invalid strings
- Maintains backward compatibility with existing float cost values
* Dry up code
* Fixed case where number was an integer
* Allowing None
---------
Co-authored-by: openhands <openhands@all-hands.dev>
* fix(rebuild-usage-object---ensure-cache_tokens-is-set): Ensures cache tokens is correctly set
Fixes https://github.com/BerriAI/litellm/issues/12149
* test(test_stream_chunk_builder_utils.py): add unit test to ensure cached tokens is part of stream chunk builder
Ensures standardized values are used
* fix(proxy_server.py): handle empty config yaml
Fixes https://github.com/BerriAI/litellm/issues/12163
* fix(gemini/common_utils.py): replace models/ as expected, instead of using 'strip'
Fixes https://github.com/BerriAI/litellm/issues/12160
* fix(anthropic/experimental_pass_through/messages/transformation.py): check for env var when selecting api key
* docs(config_settings.md): add api key to docs
* fix(team_endpoints.py): prevent overwriting current list of team models on new model add
* fix(networking.tsx): fix default proxy base url
* fix(proxy_server.py): include team only models when retrieving all deployments on `/v2/model/info` helper util
ensures team only models are shown to user
* fix(router.py): check model name by team public model name when team id given
Fixes issue where team member could not see team only models when clicking into that team on `Models + Endpoints`
* fix(team_member_view.tsx): fix rendering team member budget, when budget is set
* test: update tests
* test: update unit test
* fix(anthropic/experimental_pass_through): use given model name when returning streaming chunks
don't harcode model name on streaming
confusing for user
* fix(anthropic/streaming_iterator.py): remove scope of import
* feat(litellm_logging.py): allow admin to specify additional headers for using as spend tags
Closes https://github.com/BerriAI/litellm/issues/12129
* test(test_litellm_logging.py): add unit tests
* feat(openweb_ui.md): add custom tag tutorial to docs
* docs(cost_tracking.md): add tag based usage UI screenshot
* test: update test
* fix: fix import
* use common helper create_invitation_for_user
* use common util in proxy
* fix create_invitation_for_user
* refactor base email
* test_get_invitation_link_creates_new_when_none_exist
* fix code QA checks
* Add Azure OpenAI assistant features cost tracking
Implements cost tracking for Azure's new assistant features:
- File Search: $0.1 USD per 1 GB/Day (storage-based pricing)
- Code Interpreter: $0.03 USD per session
- Computer Use: $0.003 input + $0.012 output per 1K tokens
Features:
- Provider-specific pricing (Azure vs OpenAI)
- Model-specific pricing overrides via JSON config
- Environment variable configuration
- Backwards compatible with existing OpenAI pricing
* Add comprehensive tests for Azure assistant features cost tracking
- Unit tests for file search, code interpreter, computer use, vector store
- Integration tests for combined cost calculation
- Provider-specific pricing tests (Azure vs OpenAI)
- Model-specific pricing override tests
- Edge case handling (None inputs, zero values)
- All 17 tests passing
* Fix test and ensure all Azure assistant cost tracking tests pass
- Fixed integration test approach
- All 17 tests now passing
- Comprehensive coverage of Azure assistant features cost tracking
* Enhance cost tracking for Azure assistant features
- Safely convert and extract parameters for file search, computer use, and code interpreter sessions.
- Ensure model_info is consistently converted to a dictionary format.
- Improve error handling for input values to prevent type-related issues.
- Maintain compatibility with existing cost calculation methods.
* Refactor cost tracking for Azure assistant features
- Introduced separate methods for handling costs related to web search, file search, vector store, computer use, and code interpreter.
- Enhanced parameter extraction and conversion for file search and computer use.
- Improved error handling and type safety throughout the cost calculation process.
- Maintained compatibility with existing cost calculation methods while streamlining the overall structure.
* Fix user-team association issues in LiteLLM proxy
- Update list_team function to properly filter teams using user's teams array instead of only checking members_with_roles field
- Add Field descriptions and docstring to TeamMemberAddRequest and related models for better Swagger/OpenAPI documentation
- Maintain backward compatibility with fallback to members_with_roles if user lookup fails
This ensures users created with teams parameter appear correctly in team views and improves API documentation.
* Fix duplicate member checking in team_member_add endpoint
- Enhanced team_member_add_duplication_check to check both user_id and user_email
- Added additional duplicate prevention logic after user creation/lookup
- Fixed issue where users added by email could be duplicated in teams
- Added logging for debugging duplicate detection
This addresses the bug where adding the same user by email multiple times would create duplicate entries in the team's members_with_roles array.
* Improve duplicate member prevention in team_member_add
- Enhanced early duplicate check to handle both user_id and user_email
- Added late-stage duplicate prevention after user lookup/creation
- Fixed issue where users could be added multiple times by email
- Cleaned up debug logging
Note: There's still an edge case where the duplicate prevention may not work
correctly in all scenarios. This needs further investigation and testing.
* Refactor team_member_add endpoint for improved member management
- Split team_member_add functionality into smaller, dedicated functions for permission validation, member processing, and team member list updates.
- Enhanced permission checks to ensure only authorized users can add members.
- Streamlined member addition logic to reduce redundancy and improve readability.
- Maintained existing functionality while improving code structure and maintainability.
* Add tests for team_member_add helper functions
- Add test for _validate_team_member_add_permissions with admin user
- Add test for _validate_team_member_add_permissions with non-admin user
- Add test for _process_team_members with single member
- Add test for _process_team_members with multiple members
- Add test for _update_team_members_list with new member
- Add test for _update_team_members_list duplicate prevention
These tests ensure the refactored helper functions work correctly
after fixing the PLR0915 linting error.
* refactor(passthrough_endpoints-success-handler): refactor llm passthrough logging logic
isolate the llm translation work to enable cost tracking on sdk
* feat: initial implementation of passthrough SDK cost calculation
enables bedrock passthrough cost tracking to work
* feat(cost_calculator.py): working cost calculation for bedrock passthrough
* feat(litellm_logging.py): consider allm_passthrough in cost tracking
allows async calls (e.g. via proxy) to work
* feat(bedrock/passthrough): working event stream decoding for bedrock passthrough calls + logging instrumentation for passthrough sdk calls (log on stream completion)
Enables bedrock streaming cost calculation
* feat(litellm_logging.py): support streaming passthrough cost tracking
* feat(passthrough/main.py): working async streaming cost calculation
Closes https://github.com/BerriAI/litellm/issues/11359
* feat(proxy_server.py): fix passthrough routing when llm router enabled
* feat: further fixes
* feat(bedrock/): working bedrock passthrough cost tracking (non-streaming)
* feat(litellm_logging.py): working usage tracking for bedrock passthrough calls
ensures tokens are logged
* feat(bedrock/passthrough): add converse passthrough cost tracking support
* feat(base_llm/passthrough): remove redundant function
* refactor(litellm_logging.py): refactor function to be below 50 LOC
* test: update test
* test: remove redundant test
- Change verbose_logger.info() to verbose_logger.debug() for cost calculation messages
- Ensures cost calculation details are only shown when DEBUG logging is enabled
- Reduces log noise in production environments
Part 2/2 of fix for #9815
- Initialize verbose_logger, verbose_proxy_logger, and verbose_router_logger with the correct log level
- Ensures loggers respect the LITELLM_LOG environment variable setting
- Fixes part of issue #9815 where logs were shown regardless of log level setting
Part 1/2 of fix for #9815
* [Bug Fix] Anthropic - Token Usage Null Handling in calculate_usage (BerriAI/litellm#11920)
* [Fix] Missed a null check and used a cast instead by error
* feat: initial commit adding bedrock support via the new sdk passthrough logic
ensures correct sequencing of tasks (pre call checks etc. can run before signing request)
* fix(route_llm_requests.py): passthrough to allm_passthrough_route if no model found
* feat(bedrock/passthrough): working bedrock passthrough via sdk support
* fix(passthrough/main.py): re-add data and json
* feat(passthrough/main): support async passthrough calls to bedrock
* feat(passthrough/main.py): async streaming + completion support
* feat(llm_passthrough_endpoints.py): migrate bedrock passthrough calls to to new bedrock passthrough sdk
Enables calls to work correctly
* fix: fix linting errors
* test: update test
* fix(llm_http_handler.py): raise clearer error on anthropic unified route
Fixes https://github.com/BerriAI/litellm/issues/12063
* fix(key_management_endpoints.py): add new param `new_key` for setting the regenerated key value
user request
* test: add unit tests
* fix(pass_through_endpoints.py): use data instead of json for passthrough requests
fixes bedrock latency issue
* Revert "fix(pass_through_endpoints.py): use data instead of json for passthrough requests"
This reverts commit 021dfd9165.
Add reasoning content support when Responses API falls
back to completions API, enabling reasoning content for
all LLM providers (Anthropic, Vertex AI, etc.) since
OpenAI is currently the only native Responses API
provider.
* Add ReasoningSummaryTextDeltaEvent for streaming
reasoning deltas
* Update streaming iterator to detect and transform
reasoning content
* Extract reasoning content as separate output items in
transformations
* Support reasoning content alongside regular message
content
Closes https://github.com/BerriAI/litellm/issues/11302
The test was failing because the managed_files hook was not configured.
This fix adds a DummyManagedFiles class and mocks the create_file
function as an AsyncMock to ensure proper test execution.
