fix(litellm): keep a streamed thought signature sent without the call's name

[vetler]

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

• Closes: LiteLlm streaming drops a thought_signature sent on a tool-call delta without the call's name or arguments #7485
• Related: LiteLlm streaming drops Gemini thought_signature from tool calls; Gemini 3 then rejects the follow-up request #7438, fix(litellm): keep Gemini thought signatures on streamed tool calls #7441

This is the follow-up to #7441 for the case @surajksharma07 raised on #7438: a streamed thought signature that arrives on a tool-call delta with no name or arguments. #7441 landed without it (3abea3e); #7485 tracks it.

Problem:

_model_response_to_chunk drops a tool-call delta that has neither a name nor arguments as empty. A provider that sends the thought signature on such a delta, apart from the call's name and arguments, therefore still loses it, and Gemini 3 rejects the next request in the turn ("Function call is missing a thought_signature in functionCall parts").

Solution:

• A delta that carries a signature is no longer treated as empty, even without a name or arguments.
• In the streaming loop, such a delta attaches its signature to the call at its own index. The fallback index can already point past that call, so it is not used for the lookup.
• If that call has not started yet, the signature is held and set when the call's first named chunk arrives.
• The delta never opens a call of its own, which would reach the model as a call with no name, and it yields no partial response.

In live runs with gemini-3-flash-preview, neither Vertex AI's OpenAI-compatible endpoint nor LiteLLM's vertex_ai provider sent this shape: both put the signature on the delta that names the call (raw deltas below). So the unit tests are what exercise the new path; the live runs show the existing paths are unchanged.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

New tests in tests/unittests/models/test_litellm.py, all failing on main and passing with this change:

• test_model_response_to_chunk_keeps_a_signature_only_delta: a delta with only a signature is no longer dropped as empty.
• test_streaming_signature_after_its_call_signs_that_call: a signature-only delta that arrives after its call, with a second call in between, signs the first call and not the second.
• test_streaming_signature_before_its_call_signs_that_call: a signature-only delta ahead of its call opens no call of its own and signs the call that follows.

$ pytest tests/unittests/models/ tests/unittests/flows -q
2597 passed

$ tox -p auto   # pytest tests/unittests on Python 3.11-3.14
py311: 18307 passed, 4 failed
py312: 18298 passed, 4 failed
py313: 18298 passed, 4 failed
py314: 18298 passed, 4 failed

The four failures are the TestParseToolCallArguments logging tests, which depend on test order and fail the same way on main (fix in #7461).

pre-commit run --files passes on both changed files, and mypy reports no errors in lite_llm.py.

Manual End-to-End (E2E) Tests:

The same live check as on #7441: gemini-3-flash-preview through a Runner with RunConfig(streaming_mode=StreamingMode.SSE), asked for the weather and the local time in two cities so the model makes four parallel calls, with every raw tool-call delta logged. The script is in #7441's description (VERTEX_PROJECT=<project> ROUTE=openai|vertex_ai LOG_DELTAS=1 python e2e_parallel.py).

OpenAI-compatible endpoint:

[request 1 delta 1] index=0 id=call_2175506 name='get_weather' args_len=15 extra_content.google.thought_signature=yes provider_specific_fields.thought_signature=no finish=None
[request 1 delta 2] index=1 id=call_2175509 name='get_time' args_len=15 extra_content.google.thought_signature=no provider_specific_fields.thought_signature=no finish=None
[request 1 delta 3] index=2 id=call_2175517 name='get_weather' args_len=17 extra_content.google.thought_signature=no provider_specific_fields.thought_signature=no finish=None
[request 1 delta 4] index=3 id=call_2175518 name='get_time' args_len=17 extra_content.google.thought_signature=no provider_specific_fields.thought_signature=no finish=None
[assistant] function_call get_weather({'city': 'Oslo'}) thought_signature=<426 bytes>
[assistant] function_call get_time({'city': 'Oslo'}) thought_signature=None
[assistant] function_call get_weather({'city': 'Bergen'}) thought_signature=None
[assistant] function_call get_time({'city': 'Bergen'}) thought_signature=None
[assistant] function_response get_weather
[assistant] function_response get_time
[assistant] function_response get_weather
[assistant] function_response get_time
[assistant] text: In Oslo, it is currently 14:05 and the weather is rainy with a temperature of 9°C. [rest of the answer, about Bergen, trimmed]

LiteLLM vertex_ai provider:

[request 1 delta 1] index=0 id=<embedded sig> name='get_weather' args_len=16 extra_content.google.thought_signature=no provider_specific_fields.thought_signature=yes finish=None
[request 1 delta 2] index=1 id=call_3868561 name='get_time' args_len=16 extra_content.google.thought_signature=no provider_specific_fields.thought_signature=no finish=None
[request 1 delta 3] index=2 id=call_3868564 name='get_weather' args_len=18 extra_content.google.thought_signature=no provider_specific_fields.thought_signature=no finish=None
[request 1 delta 4] index=3 id=call_3868567 name='get_time' args_len=18 extra_content.google.thought_signature=no provider_specific_fields.thought_signature=no finish=None
[assistant] function_call get_weather({'city': 'Oslo'}) thought_signature=<658 bytes>
[assistant] function_call get_time({'city': 'Oslo'}) thought_signature=None
[assistant] function_call get_weather({'city': 'Bergen'}) thought_signature=None
[assistant] function_call get_time({'city': 'Bergen'}) thought_signature=None
[assistant] function_response get_weather
[assistant] function_response get_time
[assistant] function_response get_weather
[assistant] function_response get_time
[assistant] text: In Oslo, it is currently 14:05 and the weather is rainy with a temperature of 9°C. Similarly, in Bergen, it is 14:05 and the weather is also rainy with a temperature of 9°C.

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

🤖 Generated with Claude Code