Setup wizard sets `reasoning_effort: high` which causes CAPIError 400 on every prompt with Claude models

[Simon-McIntosh]

Describe the bug

During first launch, the Copilot CLI setup wizard prompts the user to configure reasoning_effort (offering options like "high"). When the user selects this, it is saved to ~/.copilot/config.json. However, reasoning_effort is an OpenAI-specific parameter that Claude models do not support. When the CLI sends this parameter to the API for any Claude model (including the default claude-sonnet-4.6), the API returns 400 Bad Request immediately — before any tool is called or any response is returned.

The result: the tool is completely non-functional from first launch, because the setup wizard itself configures it into a broken state.

Affected version

GitHub Copilot CLI 1.0.9

Steps to reproduce

1. Install Copilot CLI fresh
2. On first launch, when prompted about reasoning_effort, select "high"
3. The config is saved to ~/.copilot/config.json:

   {
     "reasoning_effort": "high",
     "streamer_mode": true
   }

4. Type any prompt and press Enter
5. Every single prompt fails immediately:

   ✗ Execution failed: CAPIError: 400 400 Bad Request
   (Request ID: FB0A:105CE9:33A8B7:3D6D1E:69BB8BD2)
   

Root cause (from logs)

The error occurs in getCompletionWithTools — the API rejects the request before any AI response is produced. The reasoning_effort parameter is being forwarded to Claude models that don't support it.

From ~/.copilot/logs/process-*.log:

[ERROR] error (Request-ID FB0A:105CE9:33A8B7:3D6D1E:69BB8BD2)
[ERROR] {
  "status": 400,
  "name": "CAPIError",
  "message": "400 400 Bad Request\n",
  ...
  at wst.getCompletionWithTools

Workaround

Manually remove reasoning_effort from ~/.copilot/config.json:

{
  "firstLaunchAt": "...",
  "banner": "never",
  "experimental": true
}

Expected behavior

Either:

1. The CLI should not offer reasoning_effort as a setup option when the default model is Claude, OR
2. The CLI should only forward reasoning_effort to models that support it (OpenAI o-series / GPT models), and silently ignore it for Claude/Gemini models

Additional context

• OS: WSL2 (Ubuntu), also reproduced on Linux
• Default model at time of failure: claude-sonnet-4.6
• The error is silent — there's no indication that the config is the cause; users see a generic "400 Bad Request" with no actionable guidance