English | 简体中文
A Codex Skill for coordinating specialized AI agents: choose suitable delegates, transfer context, divide file responsibilities, and combine checked results.
This repository collects a workflow I have refined through everyday use. It is intended for anyone using Codex for research, writing, data extraction, or programming who wants a clearer division of labor. The core is a Skill—a set of working rules Codex reads during a task—accompanied by model recommendations, external CLI notes, and handoff guidelines.
The homepage is bilingual. The Skill and its detailed reference documents are written in English; the agent continues to reply in your preferred language.
A capable main agent, such as 6.1 Sol Ultra, can handle much of the work directly. Delegate when a separate task or context helps; account for the time needed to prepare materials and review the result.
Delegation offers four common benefits:
- Context isolation: Keeping extensive reading, data extraction, or intermediate logs in a separate context helps the main agent remain focused on core analysis.
- Parallel progress: Self-contained tasks with clear boundaries can proceed alongside the main train of thought.
- Independent review: Reviewing a proposed plan in a separate context can reveal unstated assumptions, edge cases, and omissions. It does not remove bias or replace evidence.
- Context continuity: When a specialist needs the ongoing project's background, a supported full-history fork carries available conversation context into the assignment.
Short tasks can be worth delegating if the deliverable is clear. Conversely, tasks that depend heavily on the ongoing dialogue or require frequent back-and-forth are usually better handled directly by the main agent.
Consider checking a technical plan alongside 20 downloaded reference documents. One practical division of labor is:
| Role | Responsibility | Deliverable |
|---|---|---|
| Main agent (e.g., 6.1 Sol Ultra) | Clarify goals and constraints, guide core analysis, coordinate tasks, and assemble the final result | Final assessment and integrated findings |
| Luna Max | Read the 20 documents, extract specified fields, and verify citations against the text | Structured data table, source references, questions, and unresolved items |
| 6.1 Sol (if helpful) | Analyze the plan using inherited project context or a prepared handoff, depending on the task | Key assumptions, concrete concerns, and suggested checks |
When handing off work, the main agent should provide the objective, necessary materials, constraints, and acceptance criteria. Returned work should be checked against original sources or tests; agreement between models is not proof of fact.
These default candidates reflect practical tradeoffs between task requirements and tool availability. Model and effort availability depends on your Codex setup and account tier.
| Candidate | Primary use | Selection considerations |
|---|---|---|
| GPT-6 Luna Max | Routine extraction, locating code, small edits, and checks | Tasks with clear criteria and straightforward verification, especially when relying on Codex Skills, MCP tools, search, or project tools |
| Gemini 3.8 Flash High (via optional AGY CLI) | Video understanding and related frame analysis | Explicit evidence, timestamps, and coverage are requested. AGY has its own context and tools; sampling settings must be verified rather than assumed from the prompt |
| GPT-6.1 Sol | Complex analysis, implementation, and substantive review | Prefer full-history inheritance when the lead's model and effort fit the assignment; independent configuration remains available |
| GPT-6 Astra | Selective expert consultation, independent review, and difficult execution | Use when there is a concrete reason to expect added value, mainly Max with Xhigh as needed. These responsibilities are not special vendor tool modes |
| Grok CLI (optional) | Research on X posts and threads, plus useful web research and verification | Strongly encouraged whenever search enhancement helps the task. Daily use above 30% of the weekly allowance triggers a reminder that does not pause research or restrict further dispatch |
For a suitable Sol task, prefer fork_turns: "all" when the intended model and effort match the lead. Under the current interface, this deliberately inherits both settings, so incompatible overrides are omitted. For independently selected models or effort, use explicit invocation parameters and a sufficient none handoff. Mentioning a model in prompt text does not configure it.
Even with inherited history, each assignment states its goal, necessary inputs, file responsibility, and expected result. Full-history forks do not synchronize later messages, and none does not isolate files. The main agent usually coordinates one level; further delegation is available when it has a concrete benefit. Ultra is allowed, including inheritance from a Sol Ultra lead, but does not guarantee further delegation.
See model selection rules for full reasoning tiers, older-model fallbacks, and the limits of benchmark evidence.
You can start with Codex's built-in subagents if your environment supports them:
- Copy the
skills/subagent-manager/folder into your personal$HOME/.agents/skills/directory or your project's.agents/skills/directory. Choose one location. If a Skill with the same name already exists, review your customized rules before replacing it. - Mention
$subagent-managerin a task, or let Codex select it based on context. If the Skill does not appear, restart Codex. See the official installation guide. - Try an initial task and let Codex determine what is worth delegating:
Use $subagent-manager to help review this technical plan.
The background materials are in ./materials/.
Extract the required fields and verify their source locations.
Determine which parts are worth delegating to subagents, and specify the
appropriate model, effort, and context handoff for each assignment.
The main agent should continue the core analysis and handle final review.
The Skill provides working rules, rather than installing tools or granting model access. If a model or interface is unavailable, Codex should report the limitation and choose a workable path.
Add the following integrations as needed:
| Extension | Purpose | Requirements |
|---|---|---|
| AGY / Gemini | Analyze video and related frames, returning evidence to the lead | AGY CLI installed and logged in; PowerShell 7 for the wrapper script |
| Grok | Investigate original posts, discussions, and web sources | Grok CLI installed and logged in, with applicable free or subscription usage allowances; this workflow uses only authorized included usage |
| Grok quota rules | Check usage after grouped research units; notify above 30% daily consumption of the weekly allowance | Default: 5 substantive research units per check; Python 3 for the snapshot ledger |
| Windows proxy module | Set a process-level proxy for a single CLI invocation | Direct connections do not need this; if needed, supply your client's HTTP(S) or Mixed port |
Keep these Grok usage metrics distinct:
/usagein interactive mode shows weekly account usage and reset time.grok usage <SESSION_ID>shows token counts and costs for a specific session.- The included ledger script saves account snapshots and records daily usage. The main agent can query usage and pass snapshots to the ledger, but the script itself does not query the account.
Adjust model tiers, effort levels, and Grok query budgets according to your actual account permissions, tools, and schedule. Public benchmarks are a reference, not a direct predictor of your available task volume.
| Topic | Reference |
|---|---|
| Delegation decisions, handoffs, and review rules | SKILL.md |
| Model candidate details, effort levels, and benchmark limits | Model selection |
| Expert consultation, execution, independent review, and nesting | Task handoffs |
| Environment verification and configurable defaults | Customization |
| Practical notes and background on revisions | Experience notes |
This Skill can be used alone or alongside agent-search-booster for multi-channel research. See Releases for updates. Released under the MIT license.