feat: stop-gate review hook / 会话结束前自动 review

[raysonmeng]

Background / 背景

codex-plugin-cc 有一个 stop-gate hook：Claude Code 准备关闭时自动触发一轮 Codex review，发现问题可以阻止退出。

What to borrow / 借鉴内容

• Claude Code 的 Stop hook 触发时，自动调用 Codex 做一轮 review
• 如果发现未解决的问题，返回 { decision: "block", reason: "..." } 阻止退出
• 用户可以修复后重试，或选择强制退出
• 参考：codex-plugin-cc/plugins/codex/scripts/stop-review-gate-hook.mjs

Value / 价值

• 自动安全网，防止未经 review 的代码被遗留
• 特别适合长时间编码会话结束前的质量检查
• 实现简单（只需一个 hook + 一次 review 调用）

[simonlpaige]

Good approach. A few considerations from looking at how this would integrate with the current architecture:

1. Daemon state coupling — The stop-gate hook in bridge.ts needs to check whether the daemon is still alive before calling the review. If the daemon has already shut down (e.g., user killed it manually), the review call will hang. Suggest adding a health check on the daemon's Unix socket before initiating the review, with a fallback that allows the stop to proceed if the daemon is unreachable.

2. Review result propagation — The MCP protocol Claude Code uses for the Stop hook expects a synchronous boolean response. But Codex review is async (could take seconds). You'll need to either block the event loop (not ideal) or use the setTimeout trick to defer the actual stop while the review runs. The codex-plugin-cc implementation handles this with a promise that resolves once the review completes.

3. Edge case: force-quit — If the user sends SIGTERM/SIGKILL directly to Claude Code's process, the hook never fires. Not a blocker, but worth documenting in the README under limitations so users don't treat it as a hard safety guarantee.

4. Reconnection scenario — If Claude Code restarts and reconnects (exponential backoff path), the daemon should clear any stale "block" state from a previous session's stop-gate. Otherwise a block from yesterday's session could prevent today's exit.

I can draft a minimal implementation for bridge.ts that handles points 1-2 if that's useful.