When scanning backward over counted messages, preserve old behavior for
adjacent assistant turns by inserting user_continue if the immediate
previous raw message is assistant. This handles malformed
assistant(tool_calls)->assistant(no-tool-calls) inputs without splitting
valid assistant(tool_calls)->tool chains.
Made-with: Cursor
- aresponses() now pops prompt_id from kwargs after the async hook runs
and passes merged_optional_params via _async_prompt_merged_params.
responses() checks for this internal kwarg first and skips the sync
hook entirely when present — eliminating double-merge of template
messages.
- merged_optional_params from async_get_chat_completion_prompt is no
longer discarded (_); it flows through to local_vars in responses().
- Async tests now assert get_chat_completion_prompt.assert_not_called()
to directly detect any double-execution regression.
Made-with: Cursor
- Populate container_id on streaming code_interpreter_results by
re-emitting at message_delta when container info arrives
- Reconstruct Pydantic OutputCodeInterpreterCall objects from plain
dicts in _extract_tool_result_output_items so responses_output
has uniform types across streaming and non-streaming paths
- test_non_bash_tool_result_skipped: verifies text_editor results produce
zero code_interpreter_call items
- test_end_to_end_streaming_chunks_to_code_interpreter_output: exercises
full path from Anthropic SSE chunks through ModelResponseIterator,
stream_chunk_builder, and _extract_tool_result_output_items without
a live server
- Empty stdout/stderr now produces outputs=None (matching OpenAI parity)
instead of outputs=[{logs:""}], in both streaming and non-streaming paths
- Fix test fixture to use real Anthropic type "bash_code_execution_tool_result"
instead of "code_execution_tool_result"
- Add test for empty-output → outputs=None behavior
- Add unit tests for _extract_tool_result_output_items: Pydantic objects,
plain dicts (post-model_dump), empty/missing provider_specific_fields,
and in-place substitution preserving output ordering
stream_chunk_builder uses "last value wins" for list-valued
provider_specific_fields keys. _build_code_interpreter_results was
emitting only new items (incremental), so earlier results were silently
dropped when multiple sequential code executions occurred.
- Emit cumulative list from _build_code_interpreter_results, matching
web_search_results pattern
- Assemble server_tool_use input from input_json_delta deltas at
content_block_stop (Anthropic streams input: {} in start block)
- Handle dict items in _extract_tool_result_output_items after
model_dump() serialization in stream_chunk_builder
- Simplify _merge_provider_specific_fields to last-value-wins for lists,
matching stream_chunk_builder semantics
PR #18945 added support for capturing Anthropic server-side tool results
(bash_code_execution_tool_result, etc.) in provider_specific_fields, but
the data never reached the Responses API output because:
1. Non-streaming: provider_specific_fields wasn't copied into _hidden_params
2. Streaming: chunk delta's provider_specific_fields wasn't accumulated
3. Tool results weren't mapped to standard output items
This fix:
- Copies provider_specific_fields to _hidden_params in transform_response()
- Accumulates provider_specific_fields from streaming chunk deltas
- Maps bash_code_execution_tool_result to code_interpreter_call output items
with code and outputs (matching OpenAI's native shape)
- Removes redundant function_call items for server-side tools
- Adds OutputCodeInterpreterCall type to the output union
Restore backward-compatible behavior: only insert assistant_continue
between directly adjacent user messages, not across tool-call chains.
The _counts_for_alternation skip logic was a silent behavioral change
for [user, assistant(tc), tool, user] sequences.
Made-with: Cursor
- Fix async path: call async_get_chat_completion_prompt in aresponses()
before executor dispatch, mirroring acompletion() in main.py. Discard
merged_optional_params in async path (sync responses() handles them
via local_vars), avoiding TypeError from duplicate kwargs in partial().
- Fix provider re-resolution: replace "/" in model heuristic with
model != original_model comparison so bare model names are handled.
- Add 3 async tests covering hook invocation, optional param
propagation, and non-message item filtering in aresponses().
Made-with: Cursor
- Restore backward compat for leading assistant(tool_calls) — always prepend user_continue
- Replace partial assertions with full list assertion in trailing tool-call test
Made-with: Cursor
Avoid inserting assistant continue messages in the middle of assistant tool_call->tool chains by inserting before the next counted user turn, and add regression coverage for this edge case.
Made-with: Cursor
These were only used in block_key/unblock_key for cache refresh,
which now uses _delete_cache_key_object instead.
Co-authored-by: yuneng-jiang <yuneng-jiang@users.noreply.github.com>
- test_block_key_nonexistent_key_returns_404: verifies block_key returns
404 (not misleading 401) when the key doesn't exist in the DB
- test_unblock_key_nonexistent_key_returns_404: same for unblock_key
- test_block_key_existing_key_succeeds: verifies block_key succeeds and
invalidates cache for existing keys
- Update test_unblock_key_supports_both_sk_and_hashed_tokens to reflect
the new cache invalidation pattern (_delete_cache_key_object instead
of get_key_object + _cache_key_object)
Co-authored-by: yuneng-jiang <yuneng-jiang@users.noreply.github.com>
- Promote _fetch_managed_vector_stores_by_uuids from @staticmethod to a module-level
async helper get_managed_vector_store_rows_by_uuids, following the same standalone
helper pattern as get_team_object / get_key_object so the hot-path DB read is a
named importable function rather than an inline prisma_client.db.* call
- Pass no-log=True to both inner _call_aresponses sub-calls so they do not fire
independent billing/monitoring callbacks; cost is accumulated in the synthesized
response's _hidden_params for the outer responses() call
- Add test_H11b covering the primary queries (plural array) function-tool schema,
complementing H11 which exercises only the backward-compat singular query path
Made-with: Cursor
On /key/block, /key/unblock, and /key/update, the request body 'key'
field could contaminate the api_key Security dependency, causing the
auth layer to authenticate against the target key instead of the
caller's bearer token. This returned 401 for a nonexistent body key
even when the Authorization header contained a valid master key.
Added a guard in user_api_key_auth that re-reads the Authorization
header directly from the request, ensuring the header is always the
authoritative source for authentication.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Re-add should_use_emulated_file_search() to emulated_handler.py so H5/H6/H7/H13 tests don't fail with ImportError
- Remove per-file-id deduplication from _build_search_results_for_include so all chunks are returned (matching OpenAI native file_search behaviour); update test_H14 to assert 2 results
- Extract raw prisma DB query in check_vector_store_ids_access into a static _fetch_managed_vector_stores_by_uuids helper so the hot request path uses a named, testable function instead of an inline prisma_client.db.* call
- Remove developer-local path from test module docstring
Made-with: Cursor
- Document intentional legacy-matching behavior: when user_id is
provided to an org admin, no org filter is applied (returns all of
that user's teams across all orgs, same as legacy endpoint)
- Fix two existing security tests to properly patch user_api_key_cache,
proxy_logging_obj, and get_user_object instead of relying on
incidental error handling
- Add three new org admin test cases:
- Org admin sees org-scoped teams (200 with correct where clause)
- Org admin rejected when filtering by other org (403)
- Org admin with user_id filter returns target user's teams
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Replace raw find_unique with get_user_object in
_build_team_list_where_conditions for cache/metrics consistency
- Remove over-complex OR clause for org admin + user_id: when user_id
is provided, filter by that user's direct team memberships (same as
regular users) since the access control gate already verified the
org admin's authority
- Preserve caller-supplied organization_id instead of overwriting with
org_admin_org_ids
- Update test mock to match get_user_object call path
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add _is_reasoning_effort_level_explicitly_disabled to use opt-out semantics
for minimal effort: unknown/unlisted models pass through, only blocked when
the model map explicitly sets supports_minimal_reasoning_effort=false.
xhigh keeps opt-in semantics (must be explicitly supported).
Adds test for unknown-model passthrough and explicit-disabled detection.
Made-with: Cursor
- Add try/except httpx.HTTPStatusError blocks in _async_cancel_batch for
both POST cancel and GET retrieve calls, with verbose_logger error logging
- Fix endpoint extraction inconsistency: compute endpoint from URL without
:cancel suffix so it matches behaviour of create_batch/retrieve_batch
- Add explicit validation that api_base ends with ':cancel' before
stripping it, raising a descriptive error for unsupported custom proxy
URL rewriting scenarios
- Use string-based patch() in test instead of patch.object() for robustness
against import order changes
Made-with: Cursor
- Fix bug where only last text item's annotations were preserved when
thread.message.completed contained multiple text content items
- Accumulate annotations via extend() instead of overwriting
- Add test_azure_ai_agents_streaming_annotations_from_completed_message
- Add test_azure_ai_agents_streaming_accumulates_annotations_from_multiple_text_items
Addresses Greptile review on PR #23849
Made-with: Cursor
_get_token_and_url_context_caching() was hardcoding model=None when
calling _check_custom_proxy(), which raises ValueError when api_base
is set because Gemini proxy URLs need the model name:
{api_base}/models/{model}:cachedContents
Fixes#23846
Azure GPT-5.4+ models now get the same auto-routing treatment as OpenAI
when both `reasoning_effort` and `tools` are used in `litellm.completion()`.
Previously, `reasoning_effort` was silently dropped for Azure; now the
request is bridged to the Responses API which supports both parameters.
Fixes#23914
Fixes#23890 — Mistral's Voxtral transcription with `diarize=true` returns
`segments` (with speaker_id, timestamps) and `language`, but these fields
were dropped when mapping the response to TranscriptionResponse.
Address review feedback from greptile — use new_callable=AsyncMock
on the concurrent test's patch.object to ensure the mock is properly
typed as async, even though side_effect already handles the coroutine.
