Bug type

Behavior bug (incorrect output/state without crash)

Summary

[Bug] Gemini 3 function calling softlocked via Ollama Cloud (missing thought_signature)

Steps to reproduce

1. Configure OpenClaw to use Ollama Cloud as the provider.
2. Set model to gemini-3-flash-preview:cloud.
3. Enable tools/function calling.
4. Send a prompt requiring a tool.
5. Crash with 400 error missing thought_signature.

Expected behavior

OpenClaw must preserve the thought_signature metadata for Gemini 3 during the tool calling loop, even when the traffic is routed through an OpenAI-compatible proxy like Ollama Cloud.

Actual behavior

Describe the bug
When using the Gemini 3 Flash model through Ollama Cloud (OpenAI-compatible endpoint), function calling completely crashes. The API returns a 400 error: Function call is missing a thought_signature...

Why this is critical:
Users who rely on Ollama Cloud cannot simply switch to the native Google API provider as a workaround. The recent fix for Anthropic thinking blocks (Issue #29618) does NOT resolve this for Gemini, as Gemini's reasoning metadata is still being stripped by OpenClaw's sanitization or the OpenAI wrapper before reaching the API. We are essentially softlocked from using tools.

OpenClaw version

2026.3.2

Operating system

Ubuntu 24.04

Install method

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response