oil-oil/codex-team-mode: 协调探索、执行和独立评审，让复杂任务按明确职责并行推进，由主 Agent 统一验收。

English · 简体中文

协调探索、执行和独立评审，让复杂任务按明确职责并行推进，由主 Agent 统一验收。

It is a value-based routing guide, not a mandatory pipeline.

The team

• Explorer（探索者）· Luna Medium · read-only — gathers evidence across current web sources, documents, datasets, codebases, schemas, APIs, logs, and configuration.
• Executor（执行者）· Luna High · workspace-write — completes clear, bounded work, including substantial bounded implementation, after scope, acceptance checks, and safety boundaries are clear.
• Reviewer（复审者）· Terra Medium · read-only — independently checks stable code, reports, plans, analyses, data, and other artifacts from fresh context.

Luna keeps discovery economical and gives bounded execution a higher reasoning margin. Terra provides fresh, independent review while the main thread retains architecture decisions and final acceptance.

The TOML sandbox is a profile default, not a guaranteed isolation boundary: a live parent permission override can be reapplied to children. Use the task-scoped usage report to verify each session's effective sandbox.

How routing works

• Use Team Mode when delegation, parallel work, context isolation, lower-cost execution, or independent review has clear value.
• Team Mode may use no subagents at all. The main thread handles straightforward work when an agent would add more coordination than value.
• Before every spawn, identify the material benefit and count briefing, inspection, waiting, and rework as coordination cost. Explicitly invoking Team Mode does not make a spawn mandatory.
• Give every child a dispatch packet with Outcome, Benefit, Sources, Scope, Checks, Stop when, and Return; keep the slice in the main thread if the packet is incomplete or the gain does not exceed coordination cost.
• When two or more independent slices are ready, prefer dispatching them in parallel. The team size is dynamic: there is no fixed number of agents and no required sequence.
• Give non-trivial read-only discovery to Explorer; the main thread can wait instead of repeating the same work.
• After discovery, the main thread chooses whether to continue directly or delegate.
• Use Executor for localized or substantial bounded implementation once architecture, acceptance, and safety decisions are clear. Keep novel architecture, weak or visual verification, export/compiler behavior, and high-consequence security or rollback judgment in the main thread.
• Use Reviewer only when fresh independent judgment has clear value. Start each new Reviewer with no inherited conversation and give it a concrete unresolved risk, exact evidence, checks already passed, and a bounded stop condition.
• Keep fan-out in the main thread; children do not create descendants under standard Team Mode.
• Parallelize only genuinely independent work and keep one writer per shared target.
• After a child error or interruption, inspect shared artifacts before retrying; recover usable work instead of automatically repeating it.
• The main thread inspects the actual sources, artifacts, changes, and verification before accepting delegated work.

Install

Install the Skill:

npx skills add oil-oil/codex-team-mode

The three working Agent profiles and the default-on default dispatch guard are separate from the Skill. Copy the four TOML templates in agents/ to ~/.codex/agents/ for personal use or /.codex/agents/ for one project. Onboarding runs only for first setup, missing profiles, or explicit repair and verification requests.

See Custom Agent Profiles for exact filenames, safe installation, validation, repair, and model customization. Open a new Codex task or restart Codex if newly installed profiles do not appear immediately.

After onboarding, Codex reports what was installed and how to disable only the guard. Disabling it is a recoverable move of default.toml outside the active agents directory; the three working profiles remain installed.

Use

The Skill can trigger automatically for substantial tasks, or you can invoke it directly:

Use $team-mode for this task. Choose the smallest useful team, prefer parallel dispatch for independent slices, and keep unresolved decisions and final acceptance in the main thread.

You do not need to name every agent yourself. The main thread chooses the smallest useful team, adapts it to the task's value, and remains responsible for the combined result.

调度与统计说明

复杂任务保留主线程关键判断，必要时在宿主支持且实际验证生效的条件下升级原角色模型，不增加常驻角色。一般变更默认一名Reviewer检查相关风险，三个Simplify视角不对应固定三人。用量诊断覆盖本地活动与归档日志，区分计数重复、最终回复与完成事件。

安装的自然语言入口、平台边界与测试命令见中文说明。

Customize

You can change model and model_reasoning_effort in agents/*.toml. Preserve the role boundaries: Explorer and Reviewer stay read-only, mutation permissions remain with Executor, new reviews use fresh context, and final acceptance stays with the main thread.

Repository layout

codex-team-mode/
├── agents/                  # Three working profiles plus one dispatch guard
├── assets/readme/           # README visuals and editable source layers
├── skills/team-mode/        # Installable Skill
│   ├── agents/openai.yaml
│   ├── references/          # Profile setup and evaluation guidance
│   ├── scripts/usage_by_model.py
│   └── SKILL.md
├── tests/                   # Agent, routing, and usage regression tests
├── LICENSE
└── README.md

MIT License

配置、依赖与使用边界

需要宿主提供可用的子 Agent 创建、消息与等待工具；自定义角色依赖宿主支持。无独立账号配置。

不把创建用户侧新任务与内部子 Agent 混为一谈。并行写代码须划分文件责任，汇总前核对真实产物与冲突。

使用示例：

用 team-mode 协作处理这个项目中互相独立的模块。