Bug Description

When using google/gemini-2.5-flash or google/gemini-3-flash-preview as agent model, every request fails with:

GenerateContentRequest.contents[N].parts[0].function_response.name: Name cannot be empty.

Gemini API requires function_response.name to be non-empty, but Gateway's google-generative-ai adapter sends tool result messages with empty name field.

Reproduction

1. Set agent model to google/gemini-2.5-flash or google/gemini-3-flash-preview
2. Send any message in a channel that has tool definitions (skills)
3. Error occurs on every request, even after /new

Error Details

From gateway log:

{
  "error": {
    "code": 400,
    "message": "* GenerateContentRequest.contents[6].parts[0].function_response.name: Name cannot be empty.\n* GenerateContentRequest.contents[8].parts[0].function_response.name: Name cannot be empty.\n",
    "status": "INVALID_ARGUMENT"
  }
}

• Positions contents[6] and contents[8] are consistent across requests (before /new)
• After /new with LCM context injection, positions shift to contents[98] and contents[100]
• Both gemini-2.5-flash and gemini-3-flash-preview affected
• Other providers (anthropic, mm/MiniMax, moonshot) work fine with same session

Root Cause Hypothesis

Gateway converts internal OpenClaw message format to Gemini's GenerateContentRequest. Tool result messages (tool_result / function_response) need a name field matching the original function call. The converter appears to lose the function name during this translation.

Environment

• OpenClaw 2026.3.13
• API: google-generative-ai (v1beta)
• macOS Darwin 24.6.0