v1.3.0 — Resume after interruptions, fuller handoffs, and a new overview

Subagent Manager is a Codex Skill that lets Codex lead a team of AI subagents: it decides what to delegate, picks the right model for each piece, hands over the context it needs, and checks every result. v1.3.0 makes long delegated work survive interruptions, gives delegates fuller context, and narrows each specialist to the job it does best.

Highlights

• Picks up after interruptions. After an interruption or context compaction, the lead checks whether each delegate can continue. It reuses delegates that can, keeps finished and verified work, and reassigns only what is left, without guessing at missing progress.
• Always the latest rules. The Skill now carries a rule version (2026-10-14.1). The agent checks it at each new turn and after an interruption, and rereads the rules only when they have changed.
• Fuller context handoffs. With fork_turns: "all", a Sol delegate inherits the whole conversation and the lead's model and effort. Independently configured delegates get an explicit model and effort plus a sufficient brief, and important consultations receive the original material rather than a short summary.
• Sharper specialist roles. GPT-6 Astra returns as a selective expert for consultation, independent review, and difficult execution (mainly Max, Xhigh as needed). Luna Max takes over routine text work. Gemini via AGY now focuses on video and related frame analysis, with an explicit model and effort and clear notes on video sampling.
• Grok research encouraged. Use Grok whenever search helps. Daily use above 30% of the weekly allowance now triggers a single reminder; it no longer pauses research or stops new Grok tasks.

Also in this release

• The Skill and all reference guides are now in English; the agent still replies in your language.
• New homepage in English and Chinese, with an overview diagram. The diagram's source and logo provenance are in assets/source/.
• Clearer Grok quota ledger guidance, including converting a "used" percentage to "remaining" before recording it.
• Companion links now point to agent-search-booster.

Upgrade

Replace your copy of skills/subagent-manager/ in $HOME/.agents/skills/ or your project's .agents/skills/ with the folder from this release. If you customized the rules, compare them with SKILL.md and references/ first. Restart Codex if the updated Skill does not load.

Available models, effort levels, and external CLIs depend on your own Codex setup and accounts. Released under the MIT license.

v0.2.0 — 6.1模型路由与嵌套委派规则

6.1 Sol成为复杂子任务的主要候选；主代理已是6.1 Ultra时，委派重点是独立交付、上下文隔离和审查。Astra退出日常默认分工，保留为少量高档对照或兜底。

本次更新：

• 6.1子代理按Low—Max选择，复杂任务通常High，难以判断时Xhigh；Ultra保留可选但默认不选。
• Astra确需时以Max为主、Xhigh按需；Advisor是一种职责，不自动绑定Astra。
• 需要下一层委派时明确交代职责，并核对工具、权限和并发条件；每一层都显式指定模型与思考强度。
• Max也可以在明确要求下委派。最小测试确认Luna Max能创建下一层代理，6.1 Ultra子代理能完成任务，但没有证明Ultra子代理更主动。
• 模型评测注明日期与适用边界；公开价格和credits不用于推算本案例使用者的订阅包含额度。

验证：Skill格式与Markdown相对链接检查通过，公开候选稿未检出本机路径、固定代理端口或常见令牌模式，三类调用场景未发现实质规则冲突。模型档位、外部CLI和工具权限仍以使用者实际环境为准。

仓库中的AGY/Grok接入及Windows代理模块均可按需使用，安装和登录方法见README与适配说明。