Fixed three flaky tests that were intermittently failing in CI:
1. test_no_duplicate_spend_logs (test_litellm/responses/test_no_duplicate_spend_logs.py)
Problem: Used await asyncio.sleep(1) to wait for async logging completion,
which created race conditions. The async logging worker queues tasks
in the background, and sleep() doesn't guarantee completion.
Fix: Replaced sleep() with GLOBAL_LOGGING_WORKER.flush() which properly waits
for the logging queue to empty, ensuring all async logging tasks complete
before assertions run.
2. test_log_langfuse_v2_handles_null_usage_values (test_litellm/integrations/test_langfuse.py)
Problem: Used datetime.datetime.now() twice for start_time and end_time, which
could cause timing inconsistencies between test runs, especially in
CI environments with variable execution speeds.
Fix: Use fixed timestamps instead of datetime.now() to ensure consistent timing
across all test runs, eliminating timing-related flakiness.
3. test_watsonx_gpt_oss_prompt_transformation (test_litellm/llms/watsonx/test_watsonx.py)
Problem: Directly accessed mock_post.call_args without checking if it exists,
which could be None if the mock wasn't called or if an exception
occurred before the POST request. The test catches exceptions and
continues, making this a potential failure point.
Fix: Added proper assertions and use call_args_list[0] for safer access:
- Assert that call_args_list has at least one call
- Assert that call_args is not None
- Assert that 'data' key exists in kwargs
This ensures the test fails with clear error messages rather than
intermittent AttributeError exceptions.
All fixes maintain the original test intent while making them deterministic
and reliable in CI environments.
- Filter async_log_success_event calls by expected input message
- Bridge models (openai/codex-mini-latest) may make internal calls that also log
- Test now asserts exactly one call with the expected input 'Hey' instead of asserting total call count
- Makes test robust to bridge-related double logging while still validating core behavior
Reapplies the fix from commit a885e21543 that was
reverted in 6c9556be67.
The original revert was done because the test was flaky and giving false
negatives. This fix properly mocks the Langfuse client to ensure the test
can correctly verify that _log_langfuse_v2 converts None usage values to 0.
Changes:
- Add mock_langfuse_client.client attribute to prevent errors during init
- Add trace_id to mock_langfuse_generation for proper return value handling
- Remove redundant mock setup code
- Explicitly set logger.Langfuse to mock client after initialization
- Set logger.langfuse_sdk_version to ensure _supports_* methods work correctly
* Fix test_log_langfuse_v2_handles_null_usage_values test failure
The test was failing because the logger's Langfuse client wasn't properly
mocked. Even though sys.modules was mocked, the logger's __init__ method
creates its own Langfuse client instance that wasn't using the test's mock.
Changes:
- Explicitly set logger.Langfuse to the mock client after initialization
- Set logger.langfuse_sdk_version to ensure _supports_* methods work correctly
- Added mock_langfuse_client.client attribute to prevent errors during init
- Added trace_id to mock_langfuse_generation for proper return value handling
- Removed redundant mock setup code
This ensures the test can properly verify that _log_langfuse_v2 correctly
converts None usage values to 0 by allowing the mock's generation method
to be called and asserted.
Fixes: AssertionError: Expected 'generation' to have been called once. Called 0 times.
* feat(mcp): preserve tool metadata and full CallToolResult in MCP gateway
This PR fixes two issues that prevented ChatGPT from rendering MCP UI widgets
when proxied through LiteLLM:
1. Preserve Tool Metadata in tools/list
- Modified _create_prefixed_tools() to mutate tools in place instead of
reconstructing them, preserving all fields including metadata/_meta
- This ensures ChatGPT can see 'openai/outputTemplate' URIs in tools/list
and will call resources/read to fetch widgets
2. Preserve Full CallToolResult (structuredContent + metadata)
- Changed call_mcp_tool() and _handle_managed_mcp_tool() to return full
CallToolResult objects instead of just content
- Updated error handlers to return CallToolResult with isError flag
- Wrapped local tool results in CallToolResult objects
- This preserves structuredContent and metadata fields needed for widget rendering
Files changed:
- litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
- litellm/proxy/_experimental/mcp_server/server.py
Fixes issues where ChatGPT could not render MCP UI widgets when using
LiteLLM as an MCP gateway.
* feat(mcp): Preserve tool metadata and return full CallToolResult for ChatGPT UI widgets
- Preserve metadata and _meta fields when creating prefixed tools
- Return full CallToolResult instead of just content list
- Ensures ChatGPT can discover and render UI widgets via openai/outputTemplate
- Fixes metadata stripping that prevented widget rendering in ChatGPT
Changes:
- mcp_server_manager.py: Mutate tools in place to preserve all fields including metadata
- server.py: Return CallToolResult with structuredContent and metadata preserved
- Added test to verify metadata preservation
* fix: guard cost calculator when BaseModel lacks _hidden_params
---------
Co-authored-by: Afroz Ahmad <aahmad@Afrozs-MacBook-Pro.local>
Co-authored-by: Afroz Ahmad <aahmad@KNDMCPTMZH3.sephoraus.com>
* fix(responses): Add image generation support for Responses API
Fixes#16227
## Problem
When using Gemini 2.5 Flash Image with /responses endpoint, image generation
outputs were not being returned correctly. The response contained only text
with empty content instead of the generated images.
## Solution
1. Created new `OutputImageGenerationCall` type for image generation outputs
2. Modified `_extract_message_output_items()` to detect images in completion responses
3. Added `_extract_image_generation_output_items()` to transform images from
completion format (data URL) to responses format (pure base64)
4. Added `_extract_base64_from_data_url()` helper to extract base64 from data URLs
5. Updated `ResponsesAPIResponse.output` type to include `OutputImageGenerationCall`
## Changes
- litellm/types/responses/main.py: Added OutputImageGenerationCall type
- litellm/types/llms/openai.py: Updated ResponsesAPIResponse.output type
- litellm/responses/litellm_completion_transformation/transformation.py:
Added image detection and extraction logic
- tests/test_litellm/responses/litellm_completion_transformation/test_image_generation_output.py:
Added comprehensive unit tests (16 tests, all passing)
## Result
/responses endpoint now correctly returns:
```json
{
"output": [{
"type": "image_generation_call",
"id": "..._img_0",
"status": "completed",
"result": "iVBORw0KGgo..." // Pure base64, no data: prefix
}]
}
```
This matches OpenAI Responses API specification where image generation
outputs have type "image_generation_call" with base64 data in "result" field.
* docs(responses): Add image generation documentation and tests
- Add comprehensive image generation documentation to response_api.md
- Include examples for Gemini (no tools param) and OpenAI (with tools param)
- Document response format and base64 handling
- Add supported models table with provider-specific requirements
- Add unit tests for image generation output transformation
- Test base64 extraction from data URLs
- Test image generation output item creation
- Test status mapping and integration scenarios
- Verify proper transformation from completions to responses format
Related to #16227
* fix(responses): Correct status type for image generation output
- Add _map_finish_reason_to_image_generation_status() helper function
- Fix MyPy type error: OutputImageGenerationCall.status only accepts
['in_progress', 'completed', 'incomplete', 'failed'], not the full
ResponsesAPIStatus union which includes 'cancelled' and 'queued'
Fixes MyPy error in transformation.py:838
When Gemini image generation models return `text_tokens=0` with `image_tokens > 0`,
the cost calculator was assuming no token breakdown existed and treating all
completion tokens as text tokens, resulting in ~10x underestimation of costs.
Changes:
- Fix cost calculation logic to respect token breakdown when image/audio/reasoning
tokens are present, even if text_tokens=0
- Add `output_cost_per_image_token` pricing for gemini-3-pro-image-preview models
- Add test case reproducing the issue
- Add documentation explaining image token pricing
Fixes#17410
Fixes#17425
- Add length check for tool_calls in model_response.choices[0].delta
- Prevents empty tool call objects from appearing in streaming responses
- Add regression tests for empty and valid tool_calls scenarios
The previous implementation incorrectly used `thoughtSignature` as the criterion
to detect thinking blocks. However, per Google's docs:
- `thought: true` indicates that a part contains reasoning/thinking content
- `thoughtSignature` is just a token for multi-turn context preservation
(a part can have thoughtSignature without thought:true, e.g., function calls)
This caused functionCall data to leak into reasoning_content when using
Gemini 2.5 Pro with streaming + tools enabled.
Changes:
- _extract_thinking_blocks_from_parts now checks `part.get("thought") is True`
- Extract actual text content instead of json.dumps(part)
- Include signature only when present (optional in Gemini 2.5)
Refs:
- https://ai.google.dev/gemini-api/docs/thinking
- https://ai.google.dev/gemini-api/docs/thought-signatures
- Skip empty/whitespace text before calling Presidio API
- Handle error dict responses gracefully (e.g., {'error': 'No text provided'})
- Add defensive error handling for invalid result items
- Add comprehensive test coverage for empty content scenarios
Fixes crash in tool/function calling where assistant messages have empty content.
Fixes#17552
- Change Prisma include from 'users' to 'members'
- Use LiteLLM_OrganizationTableWithMembers type for membership validation
- Access organization.members instead of organization.users
- Add tests for membership validation
* fix async_log_success_event for _PROXY_DynamicRateLimitHandlerV3
* test_async_log_success_event_increments_by_actual_tokens
* fix redis TTL
* Potential fix for code scanning alert no. 3873: Clear-text logging of sensitive information
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
---------
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
* Fix Bedrock guardrail apply_guardrail method and test mocks
Fixed 4 failing tests in the guardrail test suite:
1. BedrockGuardrail.apply_guardrail now returns original texts when guardrail
allows content but doesn't provide output/outputs fields. Previously returned
empty list, causing test_bedrock_apply_guardrail_success to fail.
2. Updated test mocks to use correct Bedrock API response format:
- Changed from 'content' field to 'output' field
- Fixed nested structure from {'text': {'text': '...'}} to {'text': '...'}
- Added missing 'output' field in filter test
3. Fixed endpoint test mocks to return GenericGuardrailAPIInputs format:
- Changed from tuple (List[str], Optional[List[str]]) to dict {'texts': [...]}
- Updated method call assertions to use 'inputs' parameter correctly
All 12 guardrail tests now pass successfully.
* fix: remove python3-dev from Dockerfile.build_from_pip to avoid Python version conflict
The base image cgr.dev/chainguard/python:latest-dev already includes Python 3.14
and its development tools. Installing python3-dev pulls Python 3.13 packages
which conflict with the existing Python 3.14 installation, causing file
ownership errors during apk install.
* fix: disable callbacks in vertex fine-tuning tests to prevent Datadog logging interference
The test was failing because Datadog logging was making an HTTP POST request
that was being caught by the mock, causing assert_called_once() to fail.
By disabling callbacks during the test, we prevent Datadog from making any
HTTP calls, allowing the mock to only see the Vertex AI API call.
* fix: ensure test isolation in test_logging_non_streaming_request
Add proper cleanup to restore original litellm.callbacks after test execution.
This prevents test interference when running as part of a larger test suite,
where global state pollution was causing async_log_success_event to be
called multiple times instead of once.
Fixes test failure where the test expected async_log_success_event to be
called once but was being called twice due to callbacks from previous tests
not being cleaned up.
Add support for OpenAI's gpt-5.1-codex-max model, their most intelligent
coding model optimized for long-horizon agentic coding tasks.
- 400k context window, 128k max output tokens
- $1.25/1M input, $10/1M output, $0.125/1M cached input
- Only available via /v1/responses endpoint
- Supports vision, function calling, reasoning, prompt caching
* fix(github_copilot): preserve encrypted_content in reasoning items for multi-turn conversations
GitHub Copilot uses encrypted_content in reasoning items to maintain conversation
state across turns. The parent class (OpenAIResponsesAPIConfig._handle_reasoning_item)
strips this field when converting to OpenAI's ResponseReasoningItem model, causing
"encrypted content could not be verified" errors on multi-turn requests.
This override preserves encrypted_content while still filtering out status=None
which OpenAI's API rejects.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* chore: regenerate poetry.lock
* Revert "chore: regenerate poetry.lock"
This reverts commit 8796dc8f96.
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix: handle none content
* fix: defensive check on none value
* Fix test failures: Azure OCR skip, None content handling, PublicAI JSON config
- Skip aocr/ocr call types in Azure test (they don't use Azure SDK client)
- Handle None content in Responses API transformation (skip message creation)
- Update PublicAI tests to use JSON-based configuration system
- Add None check in PublicAI test fixture to fix type error
* fix(unified_guardrail.py): support during_call event type for unified guardrails
allows guardrails overriding apply_guardrails to work 'during_call'
* feat(generic_guardrail_api.py): support new 'tool_calls' field for generic guardrail api
returns the tool calls emitted by the LLM API to the user
* fix(generic_guardrail_api.py): working anthropic /v1/messages tool call response
send llm tool calls to guardrail api when called via `/v1/messages` API
* fix(responses/): run generic_guardrail_api on responses api tool call responses
* fix: fix tests
* test: fix tests
* fix: fix tests
