Gemini models return 400 Bad Request in Copilot CLI

**Request ID:** 8B9B:2E3309:6C8CC:19684A:6A593988
**Timestamp:** 2026-07-16T13:05:40Z
**Error:** CAPIError: 400 Bad Request

## Reproduction
1. Using Copilot with Gemini model selection (tried `gemini-3.1-pro-preview` and `gemini-3.5-flash`)
2. Sent plain text prompt with no attachments/tools
3. All advanced options at defaults (no custom reasoning/context overrides)
4. Extensions/tool-calling disabled

## Observed Behavior
- Request rejected with HTTP 400 before reaching model inference
- Error appears to be request-shape validation (not model timeout or API availability)
- Other models (GPT, Claude) work fine in same session

## Expected Behavior
Gemini models should accept and process the request like other available models

## Environment
- Copilot CLI (latest via app)
- Windows

Note: It appears to work if switched to "explore" mode

---
Request the backend to trace this request ID to identify which field/parameter was rejected.

首先，在 Windows 上使用列出的 Gemini 模型于 Copilot CLI 中复现该请求，使用请求 ID 8B9B:2E3309:6C8CC:19684A:6A593988，并比较 normal 和 explore 模式。追踪导致 HTTP 400 的请求字段；完成标准是普通提示词可以在 Gemini 中正常工作且无需更改模式，同时 GPT 和 Claude 的行为不受影响。

由索引模型根据 Issue 内容生成。