refactor(executor): modularize Codex runtime

docs/en/wegent/developer-guide/architecture.md

@@ -366,6 +366,8 @@ EXECUTOR_IMAGE: wegent-executor:latest # Executor image
 
 Rust executor is the only executor runtime implementation. Backend Chat shell work may still use an in-process path, while other tasks run through standalone/local executor. In Wework packaged App local-first mode, the app does not start a local Backend; it calls the executor sidecar directly over Tauri app IPC. Codex runtime control uses `codex app-server --stdio` JSON-RPC to create, continue, read, archive, and rename threads. The executor stores only the local task index and the required `localTaskId -> threadId` mapping.
 
+The Codex runtime is separated by responsibility under `executor/src/agents/codex/`: `home` manages the isolated Codex Home, authentication link, and configuration normalization; `interaction` routes user-input and MCP interaction responses; `run_state` reduces app-server events into turn outcomes; `diagnostics` truncates logs and summarizes sensitive output; and `tests` contains module-level regressions. `codex.rs` retains the public API, shared app-server lifecycle, and turn orchestration. New behavior should live in the matching responsibility module so configuration, protocol state, and diagnostics do not become coupled to the orchestration layer again.
+
 Wework's built-in browser MCP is provided by the Rust executor's `browser-mcp-server` subcommand and controls the right-side browser through a local bridge address allocated independently for each Tauri instance. The packaged app does not require Node.js or a separately deployed browser MCP server, and multiple instances do not share a fixed port.
 
 When Codex uses a shared app-server thread, cancelling an active turn must await acknowledgement of `turn/interrupt` before reporting cancellation to the caller. A retry can then start only after the previous turn has stopped, preventing an interrupt and a new request from interleaving and replaying cancelled input or dropping the retry message.

docs/zh/wegent/developer-guide/architecture.md

@@ -366,6 +366,8 @@ EXECUTOR_IMAGE: wegent-executor:latest # 执行器镜像
 
 Rust executor 是唯一的 executor 运行时实现。Backend 的 Chat shell 仍可走进程内路径，其他任务由 standalone/local executor 执行；Wework 打包 App 的 local-first 模式不启动本地 Backend，而是通过 Tauri app IPC 直接调用 executor。Codex 运行时通过 `codex app-server --stdio` 的 JSON-RPC 协议创建、继续、读取、归档和重命名线程，executor 只保存必要的本地任务索引和 `localTaskId -> threadId` 关联。
 
+Codex 运行时在 `executor/src/agents/codex/` 下按职责拆分：`home` 管理隔离的 Codex Home、认证链接和配置归一化，`interaction` 路由用户输入与 MCP 交互响应，`run_state` 将 app-server 事件归约为轮次结果，`diagnostics` 负责日志裁剪与敏感输出摘要，`tests` 保存模块级回归测试。`codex.rs` 保留对外 API、共享 app-server 生命周期和轮次编排。新增行为应进入对应职责模块，避免把配置、协议状态和诊断逻辑重新耦合到编排层。
+
 Wework 的内置浏览器 MCP 由 Rust executor 的 `browser-mcp-server` 子命令提供，并通过每个 Tauri 实例独立分配的本地桥接地址控制右侧浏览器。打包 App 无需安装 Node.js 或单独部署 browser MCP server，多实例也不会共享固定端口。
 
 Codex 使用共享 app-server 线程时，取消活动轮次必须先等待 `turn/interrupt` 的确认，再向调用方报告已取消。这样重试会在前一轮真正停止后创建，避免上一轮的中断和新轮请求交错，从而恢复已取消的输入或丢失重试消息。