Stop hook should include recent messages to enable external memory systems

Summary

External memory systems (e.g. Mimir) need to persist the user/assistant exchange at the end of a turn. The current Stop hook only exposes last_assistant_message and a transcript_path, which forces integrations to parse the Codex JSONL transcript themselves.

Request

Include an inline messages array in the Stop and SubagentStop hook payloads, containing the recent user/assistant exchange. This would make Codex easier to integrate with external memory systems without requiring them to parse transcript_path.

This request mirrors the interface introduced in MoonshotAI/kimi-code#1299.

Proposed schema

{
  "messages": [
    { "role": "user", "content": "..." },
    { "role": "assistant", "content": "..." }
  ]
}

messages should be optional to preserve backward compatibility.

Reference implementation

I have implemented this on a fork at Liewzheng/codex in the branch mimir-stop-messages (see https://github.com/Liewzheng/codex/compare/main...mimir-stop-messages). The changes include:

• Added StopMessage { role, content } to codex-rs/hooks/src/schema.rs.
• Added an optional messages field to StopCommandInput and SubagentStopCommandInput.
• Populated messages from the last user/assistant entries in the JSONL transcript when the caller does not provide them inline.
• Updated codex-rs/core/src/hook_runtime.rs to thread messages into the Stop request.
• Regenerated schema fixtures for stop.command.input and subagent-stop.command.input.
• Added unit tests for the transcript fallback.

I'd be happy to open a PR, but I see this repository currently restricts pull requests to collaborators. Please let me know if you'd like me to provide the patch in another format or if there's a process for external contributors.