feat: use Codex native review/start method / 使用 Codex 原生 review/start 方法

[raysonmeng]

Background / 背景

codex-plugin-cc 使用 Codex app-server 的原生 review/start RPC 方法来做 code review，而不是通过 turn/start 注入 review prompt。

What to borrow / 借鉴内容

• review/start 是 Codex 内建的专用 review 方法
• 支持 target（review 范围：PR diff、staged changes 等）
• 支持 structured output（schema-based review results）
• 支持 streaming（item/agentMessage/delta 实时输出）
• 参考：codex-plugin-cc/plugins/codex/scripts/lib/codex.mjs 的 runAppServerReview
• 参考：codex-plugin-cc/plugins/codex/scripts/lib/app-server-protocol.d.ts

Value / 价值

• 比注入 prompt 的方式更原生、更准确
• 可以利用 Codex 内建的 review 能力（adversarial review 等）
• Structured output 便于自动化处理 review 结果

[simonlpaige]

The shift to using review/start over prompt injection is necessary for reliable state management. Given that this method supports structured output and streaming via item/agentMessage/delta, we need to consider how AgentBridge will handle the intermediate parsing layer.

Specifically, when consuming the schema-based review results:

1. How should the bridge manage merging multiple streamed messages (delta) into a single, cohesive result object before forwarding it?
2. Since review/start accepts a specific target (e.g., PR diff), we must ensure that the calling process accurately captures and serializes the necessary context (the actual code range or patch) from Claude Code's current working session state to pass as the payload parameter, rather than just passing general instructions.

If the structured output schema is complex, implementing a dedicated message handler for validation failure/success states within daemon.ts will prevent downstream agents from receiving malformed data and failing silently.