[agent] Gemini 3 models degenerate under eve: thoughtSignature dropped on tool-call replay

[paulieb89]

Problem

Running the agent on google/gemini-3.5-flash (branch test/gemini-flash-model, deployed 2026-07-15) produced degenerate behaviour on the issue #1 dispatch: repeated tool-narration text ("Let's call read_file! Let's do that!" looped), four comments posted in ~20s, one 42KB of repetition.

Runtime logs showed the mechanism (AI SDK warning, repeated):

Replayed 2 functionCall part(s) for a Gemini 3 model without a thoughtSignature (tools: read_issue, bash). Injected the documented skip_thought_signature_validator sentinel to keep the request from failing with HTTP 400. The likely cause is application code that drops providerOptions.google.thoughtSignature when persisting or serializing assistant tool-call messages.

eve 0.22.5's message persistence drops providerOptions.google.thoughtSignature, breaking Gemini 3's reasoning replay contract. Compounded by #3 (no working-tree checkout), the model looped instead of working. Production was rolled back to anthropic/claude-sonnet-5 (main) the same night; the session terminated without intervention and the bot-comment filter prevented self-re-triggering.

Acceptance criteria

• Either eve persists thoughtSignature for Gemini 3 models (upstream fix or upgrade), or the repository documents that Gemini 3 model ids are unsupported for this agent.
• A dispatched turn on the chosen model completes without degenerate repetition.

Constraints and non-goals

• Do not work around by disabling tool-call persistence.
• Model changes are privileged self-modification and require maintainer review.

Verification

Runtime logs for a full dispatched turn show no thoughtSignature warnings and the agent produces a coherent comment/PR.