Explicit code-review subagent model override is ignored

Describe the bug

The built-in code-review subagent is configured to use gpt-5.6-luna, but Copilot CLI starts it with gpt-5.6-sol.

The parent session was using gpt-5.6-luna, and the configured per-agent model was also gpt-5.6-luna. The configured value is silently ignored or replaced.

This matters because the selected subagent model can affect review quality, model availability, and AI Credit consumption.

Observed evidence

The effective configuration contains:

{
  "subagents": {
    "agents": {
      "code-review": {
        "model": "gpt-5.6-luna",
        "effortLevel": "high",
        "contextTier": "default"
      }
    }
  }
}

The parent assistant was using gpt-5.6-luna, but the task request and subagent event selected gpt-5.6-sol:

{
  "type": "tool.execution_start",
  "data": {
    "toolName": "task",
    "arguments": {
      "agent_type": "code-review",
      "model": "gpt-5.6-sol"
    },
    "model": "gpt-5.6-luna"
  }
}

{
  "type": "subagent.started",
  "data": {
    "agentName": "code-review",
    "model": "gpt-5.6-sol"
  }
}

Affected version

1.0.79

Steps to reproduce the behavior

The same behavior occurred in repeated reviews from the same session.

1. On Windows x64, configure the built-in code-review subagent to use gpt-5.6-luna through /subagents.
2. Start a Copilot CLI session and select gpt-5.6-luna.
3. Run /review and request a code review.
4. Inspect the session timeline or the corresponding file under ~/.copilot/session-state/*/events.jsonl.
5. Observe that the code-review subagent is started with gpt-5.6-sol.

Expected behavior

The code-review subagent should use the explicitly configured gpt-5.6-luna model.

If the requested model cannot be used, Copilot CLI should report the fallback clearly and identify the effective model instead of silently selecting another model.

Additional context

• Operating system: Windows 11 Pro
• Architecture: x64
• Shell: command prompt

The global settings file also contains a default model of gpt-5.6-terra, but the affected session was explicitly switched to gpt-5.6-luna. This is therefore not simply a mismatch with the global default model.

The behavior appears related to https://github.com/github/copilot-cli/issues/4377

Reproduisez le problème sur Windows x64 en utilisant /subagents et /review, puis examinez ~/.copilot/session-state/*/events.jsonl. Comparez le modèle de revue de code configuré avec les événements task et subagent.started. C’est terminé lorsque le modèle explicitement configuré est utilisé, ou que tout fallback est clairement signalé avec son modèle effectif.