[bug] Gemini API rejects MCP tool schemas with properties/required on non-object types

[ChickenBreast-ky]

Description

When using MCP servers (e.g., Notion MCP) with Gemini models, the API returns a 400 error because some tool schemas have properties or required fields on non-object types.

Error Message

Bad Request: {
  "error": {
    "code": 400,
    "message": "* GenerateContentRequest.tools[0].function_declarations[80].parameters.properties[data].properties: only allowed for OBJECT type
* GenerateContentRequest.tools[0].function_declarations[80].parameters.properties[data].required: only allowed for OBJECT type",
    "status": "INVALID_ARGUMENT"
  }
}

Steps to Reproduce

1. Enable Notion MCP (or other MCP servers with similar schema patterns)
2. Use a Gemini model (e.g., gemini-3-flash-preview)
3. Send any message
4. Error occurs on requests with larger tool sets

Expected Behavior

Tool schemas should be sanitized to remove properties and required from non-object types before sending to Gemini API.

Current Behavior

The sanitizeGemini function in transform.ts handles enum conversion and required field filtering, but doesn't strip properties/required from non-object types.

Proposed Solution

Add to sanitizeGemini:

if (result.type && result.type !== "object") {
  delete result.properties
  delete result.required
}

Related Issues

• Similar to AI_APICallError: GenerateContentRequest.tools[0].function_declarations[32].parameters. required[0]: property is not defined with Gemini API #3140, Schema validation error with Google/Gemini models #8036

Environment

• opencode version: latest
• MCP: Notion MCP
• Model: gemini-3-flash-preview

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• AI_APICallError: GenerateContentRequest.tools[0].function_declarations[32].parameters. required[0]: property is not defined with Gemini API #3140: AI_APICallError with Gemini API - properties/required validation issue
• sanitize tools for Gemini 3 compatibility #4512: sanitize tools for Gemini 3 compatibility (addresses same root cause)
• Error with Gemini models when MCP enabled #5832: Error with Gemini models when MCP enabled
• Strict JSON Schema validation fail on backends like SGLang (Missing required field in empty tool parameters) #8184: Strict JSON Schema validation fail - missing required field in empty tool parameters
• LiteLLM Proxy: Vertex AI Gemini models fail for tools with empty parameters #9233: LiteLLM Proxy: Vertex AI Gemini models fail for tools with empty parameters

These issues all relate to the same root cause: Gemini API's strict validation that only allows properties and required fields on OBJECT type parameters.

Feel free to ignore if your specific case differs from these.