Gemini 3.x unusable: client omits thought_signature on function-call parts, investigations fail on turn 2

[Smana]

Summary

Any investigation using a Gemini 3.x model fails on the second model turn with 400 INVALID_ARGUMENT. Gemini 3 requires a thought_signature to be replayed on functionCall parts in conversation history; internal/model/gemini never captures or replays one.

This blocks the entire current Gemini generation (3.x). Gemini 2.5 is unaffected.

Error

demo: investigation failed: model: streamGenerateContent status 400 (request-id ""):
INVALID_ARGUMENT: Function call is missing a thought_signature in functionCall parts.
This is required for tools to work correctly, and missing thought_signature may lead to
degraded model performance. Additional data, function call `default_api:what_changed`,
position 2.

Reproduction

# /tmp/gem.yaml
model:
  provider: gemini
  model: gemini-3.5-flash
  api_key_env: GEMINI_API_KEY

lore demo investigate --scenario harbor-chart-bump --config /tmp/gem.yaml

model	result
gemini-3.5-flash	first tool call (what_changed) succeeds; the next turn — which replays that call in history — fails with the 400 above. Investigation aborts, exit 1
gemini-2.5-flash-lite	completes normally: 4 model calls, verify pass included, findings rendered

The failure is at position 2, i.e. the first turn that sends a prior functionCall back as history — consistent with the signature being required on replay, not on the initial response.

Cause

grep -c 'thought_signature\|thoughtSignature' internal/model/gemini/*.go → 0.

The client neither reads the field from streamed functionCall parts nor writes it back when reconstructing history. providers.Message already carries an Opaque json.RawMessage field for exactly this class of provider-specific replay data — it is documented as "provider-specific content the client must replay verbatim", and the Anthropic client uses it to round-trip thinking blocks with their signatures. The Gemini client leaves it empty.

Impact

• Every Gemini 3.x model is unusable for investigations — hard failure, not degradation.
• Affects lore investigate, lore serve, lore demo investigate, and any eval run configured with a Gemini 3.x model.
• Gemini 2.5 models still work, but they are the previous generation.

Suggested fix

Mirror the Anthropic client's approach: capture the thoughtSignature from streamed functionCall parts into Message.Opaque, and replay it verbatim when rebuilding request contents. Add a round-trip test asserting a signature survives response → history → next request.

Verified 2026-08-02 against generativelanguage.googleapis.com/v1beta.
Related: #391 (a different provider-compatibility failure, glm-4.6 + forced tool_choice).