Three-Tier Agent Orchestrator

本 README 以中文為主，英文版請見文末。

English version skill is https://github.com/irons163/three-tier-agent-symphony.

一個 Codex skill，讓目前主線程承擔 GPT-6 Astra 的總指揮角色：理解目標、拆分任務、做架構決策、檢查結果並整合輸出；並依工作性質把獨立子任務交給 GPT-6 Sol Max 與 GPT-6 Luna Max。

角色分工

主要行為

• 目前主線程直接承擔 Astra 的統籌角色；skill 不自行判定、切換或要求切換主模型。
• 委派前確認 GPT-6 Sol Max 與 GPT-6 Luna Max subagent 是否可用。
• 啟動 subagent 時隔離主線程歷史：V2 使用 fork_turns: "none"，V1 使用 fork_context: false。
• V2 的 subagent 名稱依實際參數使用固定結尾：_sol_max 或 _luna_max。
• 任一必要模型或指定 reasoning 不可用時，完整停止工作，不以其他模型或較低 reasoning 替代。
• 能力齊全後，使用明確的目標、範圍、完成條件與驗證方式委派子任務。
• 由 GPT-6 Astra 檢查重要 diff、測試與證據，並負責最終整合。

必要條件

• Codex 支援 skills 與 subagents，且 subagent 工具可直接指定模型與 reasoning effort。
• 建議主線程選用 gpt-6-astra，但這不是 skill 的 hard-stop 條件。
• Sol Max subagent 使用 gpt-6-sol 與 max reasoning effort。
• Luna Max subagent 使用 gpt-6-luna 與 max reasoning effort。

不需要建立額外的 custom agent 設定檔。Skill 每次啟動 subagent 時都會直接傳入指定的模型、reasoning effort 與目前 schema 支援的上下文隔離欄位；實際 subagent 啟動結果才是能力判準。

V2 若支援 task_name，名稱會使用可辨識的工作前綴與固定角色結尾，例如 security_review_sol_max 或 run_tests_luna_max。Skill 會在啟動前確認名稱結尾與實際 model／reasoning 一致；V1 不傳入此欄位。

安裝

在 macOS 或 Linux 執行：

git clone https://github.com/irons163/three-tier-agent-orchestrator.git "${CODEX_HOME:-$HOME/.codex}/skills/three-tier-agent-orchestrator"

若目前 task 沒有重新載入 skill，請建立新 task；仍未出現時再重新啟動 Codex App。

使用

在 Codex prompt 明確啟用：

$three-tier-agent-orchestrator

也可以直接描述工作，例如：

使用 Three-Tier Agent Orchestrator 檢查這個專案：GPT-6 Astra 負責理解目標與整合，GPT-6 Sol Max 做架構與安全審查，GPT-6 Luna Max 執行測試與整理錯誤。

Repository 結構

.
├── README.md
├── SKILL.md
└── agents/
    └── openai.yaml

不需要額外的 custom agent 設定檔。

---

English Version

Three-Tier Agent Orchestrator

A Codex skill that assigns the current main thread the GPT-6 Astra orchestration role: understanding the objective, breaking down tasks, making architectural decisions, reviewing results, and integrating the final output while delegating independent subtasks to GPT-6 Sol Max and GPT-6 Luna Max based on the nature of the work.

Role assignments

Core behavior

• The current main thread assumes the Astra orchestration role; the skill does not detect, switch, or require switching the main model.
• Verify that the GPT-6 Sol Max and GPT-6 Luna Max subagents are available before delegation.
• Isolate main-thread history when spawning a subagent: use fork_turns: "none" on V2 and fork_context: false on V1.
• On V2, derive each subagent name from its actual launch arguments, using the fixed suffix _sol_max or _luna_max.
• If any required model or specified reasoning capability is unavailable, stop completely instead of substituting another model or using a lower reasoning level.
• Once all capabilities are available, delegate subtasks with explicit objectives, scope, completion criteria, and validation methods.
• GPT-6 Astra reviews important diffs, tests, and evidence, and is responsible for the final integration.

Requirements

• Codex supports skills and subagents, and the subagent tool can specify the model and reasoning effort directly.
• Using gpt-6-astra for the main thread is recommended but is not a hard-stop condition.
• The Sol Max subagent uses gpt-6-sol with max reasoning effort.
• The Luna Max subagent uses gpt-6-luna with max reasoning effort.

No additional custom agent configuration files are required. Whenever the skill launches a subagent, it directly passes the specified model, reasoning effort, and the context-isolation field supported by the current schema. The actual subagent startup result determines whether the capability is available.

When V2 supports task_name, the name uses a recognizable task prefix and a fixed role suffix, such as security_review_sol_max or run_tests_luna_max. Before launch, the skill verifies that the suffix matches the actual model and reasoning arguments. V1 does not receive this field.

Installation

Run the following on macOS or Linux:

git clone https://github.com/irons163/three-tier-agent-orchestrator.git "${CODEX_HOME:-$HOME/.codex}/skills/three-tier-agent-orchestrator"

If the current task does not reload the skill, create a new task. If it still does not appear, restart the Codex App.

Usage

Explicitly enable the skill in a Codex prompt:

$three-tier-agent-orchestrator

You can also describe the work directly, for example:

Use the Three-Tier Agent Orchestrator to review this project: GPT-6 Astra handles objective understanding and integration, GPT-6 Sol Max performs the architecture and security review, and GPT-6 Luna Max runs tests and organizes the errors.

Repository structure

.
├── README.md
├── SKILL.md
└── agents/
    └── openai.yaml

No additional custom agent configuration files are required.