Moonshot's _transform_messages unconditionally flattened content arrays
to plain text, dropping image_url blocks. Vision models like kimi-k2.5
accept the standard OpenAI content array format.
Now checks for image_url blocks before flattening — if any message
contains images the content array is preserved intact.
Fixes#20862
AgentCore MCP server endpoints require the Accept header to contain
both application/json and text/event-stream per the MCP specification
(Streamable HTTP transport). Without this header, requests are rejected
with a 406 Not Acceptable error (JSON-RPC code -32011).
Sets the Accept header at the top of sign_request() so both JWT/Bearer
and SigV4 authentication paths include it.
aspectRatio and imageSize were silently dropped because they weren't
listed in get_supported_openai_params(), so the validation layer filtered
them out before they could reach transform_image_generation_request().
Fixes#21070
logprobs, top_p, top_logprobs are only accepted by OpenAI when
reasoning_effort="none". Add validation matching the existing
temperature logic: raise UnsupportedParamsError or drop when
reasoning_effort is set to other values.
Gemini returns finishReason="STOP" even when tool calls are present,
and sends tool_calls and finishReason in separate streaming chunks.
The ModelResponseIterator now tracks tool_calls across chunks and
correctly maps finish_reason to "tool_calls" per the OpenAI spec.
Fixes#21041
Remove logit_bias, modalities, prediction, audio, web_search_options
from supported params for all GPT-5 reasoning models (OpenAI rejects
them). Add logprobs, top_p, top_logprobs for gpt-5.1/5.2 which support
them when reasoning_effort="none".
Related to #21572
gpt-5-search-api models were routed through OpenAIGPT5Config which
listed params like n, temperature, tools, reasoning_effort as supported,
but OpenAI rejects all of these for search models.
Fixes#21572
For multi-turn conversations, convert thinking_blocks on assistant
messages into content blocks prepended before the rest of the content,
so reasoning context is passed back to the hosted_vllm API.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The ChatGPT backend API sends non-spec-compliant streaming tool call chunks
where index is always 0 for parallel tool calls and id/name get repeated in
duplicate closing chunks. Add ChatGPTToolCallNormalizer to fix indices and
filter duplicates before they reach the consumer.
Fixes#21482
The adapter was injecting `summary: "detailed"` into the reasoning config
when routing Anthropic thinking requests to OpenAI's Responses API.
Per the OpenAI spec, reasoning.summary is opt-in — it should not be
added unless the user explicitly requests it.
The file had two unresolved git merge conflict markers from a merge of
litellm_oss_staging_02_17_2026 into main, causing a SyntaxError when
pytest tried to collect the test module.
Kept the instance-level mocking approach (from litellm_oss_staging) for
test_get_complete_url and test_validate_environment, which is consistent
with the rest of the file and avoids class-reference issues caused by
importlib.reload(litellm) in conftest.py.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The implementation correctly preserves tool_call order: existing results first
(call_1), then dummy results for missing ones (call_2). The test was asserting
the reverse order with incorrect comments. Fix the assertions to match the
actual correct behavior.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Use custom_endpoint=False so Databricks SDK auth fallback works
(custom_endpoint=True was blocking it). The api_base returned by
databricks_validate_environment is discarded since get_complete_url
builds the URL separately.
- Remove unused verbose_logger import
- Remove unused json import in tests
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Databricks supports the Responses API natively for GPT models, but litellm
was falling back to the completion transformation handler which converts
responses requests to chat completion calls, losing response schema enforcement.
This adds DatabricksResponsesAPIConfig that passes responses API requests
directly to Databricks' /responses endpoint for GPT models, while non-GPT
models (Claude, Llama, etc.) continue using the completion transformation path.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Read summary from the original thinking dict instead of hardcoding "detailed"
in _route_openai_thinking_to_responses_api_if_needed(). This preserves the
user's chosen summary value (e.g. "concise", "auto") for non-Claude models
routed through the Anthropic Messages adapter to OpenAI's Responses API.
Fixes#20998
- test_pillar_guardrails.py: Fix fixture to properly update module-level
litellm reference using global keyword and assignment from reload
- test_anthropic_experimental_pass_through_messages_handler.py: Add missing
assert keywords to kwargs comparison statements (lines 36, 60-62)
- test_proxy_server.py: Replace silent pytest.skip with explicit assertion
to catch router initialization regressions
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Fixes test failures that occur during parallel test execution (pytest -n 4)
due to module reloading issues with conftest.py reloading litellm.
Changes:
- Add module reload fixtures to ensure fresh references after conftest reloads
- Use patch.object and string-based patches instead of direct attribute assignment
- Use class name comparison instead of isinstance for reloaded modules
- Handle case where litellm is missing from sys.modules during parallel runs
- Move stream consumption inside patch contexts to avoid real API calls
- Mock litellm.acompletion instead of low-level HTTP handlers
- Add skipif decorator for enterprise-only test classes
Affected test files:
- test_container_integration.py
- test_responses_background_cost.py
- test_huggingface_embedding_handler.py
- test_vertex_ai_rerank_integration.py
- test_volcengine_responses_transformation.py
- test_pillar_guardrails.py
- test_litellm_pre_call_utils.py
- test_proxy_server.py
- test_converse_transformation.py
- test_chat_completions_handler.py
- test_aresponses_api_with_mcp.py
- test_anthropic_experimental_pass_through_messages_handler.py
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add _reset_litellm_http_client_cache autouse fixture (matching
test_vertex_gemma_transformation.py) to flush in_memory_llm_clients_cache
before each test. Without this, a cached real AsyncHTTPHandler from an
earlier test could bypass the class-level mock and cause real HTTP calls.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace instance-level patch.object(client, "post", side_effect=...) with
class-level patch of AsyncHTTPHandler and AsyncMock to reliably intercept
HTTP calls in CI where real Google credentials are available.
The old approach patched a specific instance's post method and passed
client=client to acompletion(). In CI, the mock wasn't intercepting actual
HTTP calls, causing 401 ACCESS_TOKEN_TYPE_UNSUPPORTED errors. The new
approach patches AsyncHTTPHandler at the class level so any instance
created internally by get_async_httpx_client() is also mocked.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two test files were reloading modules in setup_method/fixtures, which
caused class-reference staleness for subsequent tests in the same worker:
1. test_huggingface_embedding_handler.py reloaded
litellm.llms.custom_httpx.http_handler, creating a new HTTPHandler
class. Subsequent tests (e.g. hosted_vllm embedding) created
client = HTTPHandler() from the new class, but llm_http_handler.py
still held the old class reference. isinstance(client, HTTPHandler)
returned False, so a new unpatched client was used and
client.post was never called.
2. test_vertex_ai_rerank_integration.py reloaded
litellm.llms.vertex_ai.rerank.transformation in setup_method,
creating a new VertexAIRerankConfig class. The transformation test
file's module-level import still referenced the old class, so
@patch('...VertexAIRerankConfig._ensure_access_token') patched the
new class while self.config was an instance of the old class,
leaving the mock unapplied and hitting real Google credentials.
Fix: remove the reload calls. The module-level class references are
stable across tests within a worker; the reloads were solving a problem
that doesn't exist and actively created cross-test contamination.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fix isinstance() checks failing due to module reload in conftest.py.
The conftest.py fixture reloads the litellm module between test modules,
which causes class references imported at module-level to become stale.
When AsyncHTTPHandler is imported at the top of the file and then litellm
is reloaded by the fixture, the isinstance() check fails because the
returned instance is of the NEW AsyncHTTPHandler class while the test
is checking against the OLD class reference.
Solution: Import AsyncHTTPHandler locally within each test function that
uses isinstance() checks. This ensures we get the fresh class reference
after the module reload.
Fixed tests:
- test_session_reuse_integration
- test_get_async_httpx_client_with_shared_session
- test_get_async_httpx_client_without_shared_session
This resolves intermittent CI failures where parallel test execution
triggers the module reload behavior.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Add infrastructure for JSON-declared providers to support /v1/responses
via `supported_endpoints` field in providers.json. Simplify Perplexity
responses config from 410 to 40 lines by moving cost dict→float parsing
to generic validators in ResponseAPIUsage and Usage.
- Add `supported_endpoints` field to SimpleProviderConfig (default: [])
- Add `supports_responses_api()` to JSONProviderRegistry
- Create OpenAILikeResponsesConfig base class for responses API
- Add `create_responses_config_class()` with class caching
- ProviderConfigManager: Python classes take priority over JSON fallback
- Fix ResponseAPIUsage.cost field_validator to handle dict cost objects
- Fix Usage.__init__ to handle dict cost from chat completions
- Simplify PerplexityResponsesConfig with get_supported_openai_params guard
- Add 20 unit tests including Python-over-JSON priority test
- Rename _is_nova_lite_2_model → _is_nova_2_model to match all nova-2-* variants
- Add bedrock/converse/ routing prefix stripping in model detection
- Fix pre-existing test_get_supported_openai_params_bedrock_converse failure
- Remove thinking_blocks tests from Nova 2 test file (not Nova 2 behavior)
- Add end-to-end request, response, multi-turn, and model detection tests
- Parametrize key tests across both nova-2-lite and nova-2-pro model IDs
