fix(plugins-google): thought_signature dropped for version-less Gemini model aliases

[ngoanpv]

Problem

Multi-turn function calling fails with 400 INVALID_ARGUMENT: "Function call is missing a thought_signature in functionCall parts" when the LLM model is a version-less Gemini alias such as gemini-flash-latest or gemini-flash-lite-latest (these resolve to Gemini 3 server-side). It works on the first turn and breaks on the first follow-up that echoes a function call back.

Root cause

_requires_thought_signatures(model) gates BOTH storing (_parse_part) and resending (_run) of thought_signature, but only matches literal gemini-2.5* / gemini-3* strings. Version-less aliases return False, so the signature is never stored or resent — and Gemini 3 requires it echoed back.

Fix

Make signature handling response-driven instead of model-name-driven:

• store a thought_signature whenever the API returns one (_parse_part);
• resend whatever was stored, whenever present (_run).

Correct for every signature-emitting model + alias; a no-op for models that never emit them (nothing stored → nothing resent). Signature handling at runtime no longer depends on the _requires_thought_signatures model-name heuristic. The helper itself is retained (its alias handling corrected so gemini-flash-latest / gemini-flash-lite-latest also return True) because the existing unit tests in tests/test_google_thought_signatures.py import and assert on it.

Model coverage note

The current latest flash models are gemini-3.5-flash and gemini-3.1-flash-lite. Both already match the gemini-3 detection, so — unlike the version-less -latest aliases — they were already storing and resending signatures before this change; the response-driven fix keeps them correct. The fix also covers the -latest aliases automatically (they no longer depend on the name heuristic at runtime), while the retained helper additionally recognizes them so the unit tests stay accurate. The dropped-signature 400 was specific to the version-less -latest aliases.

Test

Verified live against the real Gemini API, driving the plugin's llm.LLM through a 2-turn function-calling exchange with a single save_answer(answer: str) tool (turn 1 forces the tool call; turn 2 echoes the FunctionCall + FunctionCallOutput back, reusing the harvested call_id).

gemini-flash-lite-latest (the version-less alias bug)

Before (unpatched): turn 1 stores nothing (LLM._thought_signatures is empty), turn 2 raises:

APIStatusError status_code=400 INVALID_ARGUMENT
"Function call is missing a thought_signature in functionCall parts. This is required
for tools to work correctly... function call `default_api:save_answer`, position 2."

After (patched): turn 1 stores the signature (LLM._thought_signatures keys = ['<call_id>']), turn 2 succeeds with a normal assistant response — no 400. Same script, same model, only the plugin changed.

gemini-3.5-flash (latest flash)

Also verified live for gemini-3.5-flash. Because this model matches the gemini-3 detection in both the original and patched code, the runtime gate is True either way, so it never hit the dropped-signature 400. With the patched plugin, multi-turn function calling succeeds: turn 1 stores the signature (LLM._thought_signatures keys = ['<call_id>']), turn 2 returns a normal assistant response — no 400. The response-driven fix keeps gemini-3.5-flash (and gemini-3.1-flash-lite) correct.

Unit tests

The existing unit cases in tests/test_google_thought_signatures.py still pass unchanged. Added _requires_thought_signatures → True cases for the current latest models gemini-3.5-flash and gemini-3.1-flash-lite, plus the version-less aliases gemini-flash-latest and gemini-flash-lite-latest (38 passed).

---

Follow-up fix — model-detection gaps in thinking_config and flash defaults

The same fragile string matching affected two sibling helpers in llm.py, surfaced during review:

• _is_gemini_3_model / _is_gemini_3_flash_model only matched literal gemini-3* / gemini-3-flash*. They returned False for the version-less aliases gemini-flash-latest / gemini-flash-lite-latest (which resolve to Gemini 3.x flash server-side) and mis-handled gemini-3.5-flash.
• Consequence 1: the thinking_config block called _is_gemini_3_model(model); for the aliases it took the "Gemini 2.5 and earlier" branch and raised ValueError("does not support thinking_level"), even though those aliases support thinking_level.
• Consequence 2: _is_gemini_3_flash_model("gemini-3.5-flash") returned False, so 3.5-flash missed the "minimal" flash thinking default and fell back to "low".

Fix

A shared _GEMINI_3_FLASH_ALIASES constant plus alias/3.x-aware helpers:

• _is_gemini_3_model now matches any gemini-3 substring (covers gemini-3.5-flash, gemini-3.1-flash-lite) and the version-less aliases.
• _is_gemini_3_flash_model is True for any 3.x flash model and both aliases.
• The thinking_config block is unchanged — correcting the helpers makes the aliases take the Gemini-3 branch (no ValueError) and gives gemini-3.5-flash the "minimal" flash default.

Tests

Expanded the parametrized cases in tests/test_google_thought_signatures.py for all three helpers to cover the aliases, gemini-3.5-flash, and gemini-3.1-flash-lite (e.g. _is_gemini_3_model("gemini-flash-latest") → True, _is_gemini_3_flash_model("gemini-3.5-flash") → True, plus gemini-3-pro-preview → False to keep pro models out of the flash path).

[CLAassistant]

All committers have signed the CLA.

[ngoanpv]

The latest commit (ea985d9) resolves the flagged model-detection findings in llm.py:

• _is_gemini_3_model / _is_gemini_3_flash_model now recognize the version-less aliases gemini-flash-latest / gemini-flash-lite-latest (via a shared _GEMINI_3_FLASH_ALIASES constant) and the gemini-3.5-flash / gemini-3.1-flash-lite point releases.
• As a result, the thinking_config block no longer raises ValueError("does not support thinking_level") for those aliases — they now correctly take the Gemini-3 branch — and gemini-3.5-flash gets the "minimal" flash thinking default instead of falling back to "low".

Expanded the parametrized cases in tests/test_google_thought_signatures.py for all three helpers to cover the aliases, gemini-3.5-flash, and gemini-3.1-flash-lite.

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 3 additional findings in Devin Review.

[devin-ai-integration]

🚩 _thought_signatures dict grows unboundedly across chat sessions

The _thought_signatures dict at llm.py:249 lives on the LLM instance and accumulates entries from every _parse_part call across all LLMStream instances. There is no eviction mechanism. For long-running agents with many multi-turn function-calling interactions, this dict could grow unboundedly. This is a pre-existing issue (not introduced by this PR) but worth noting since the PR's removal of the model-name guard technically makes it possible for any model that emits thought_signatures to contribute entries (in practice, only the same set of models as before).

(Refers to line 249)

---

Was this helpful? React with 👍 or 👎 to provide feedback.