A Fable of Codexes

A Claude Code skill that makes Claude the conductor of a fleet of OpenAI Codex workers. Claude Fable 5.1 plans the work, dispatches parallel Codex workers on GPT-6.1 Sol to implement it, and verifies every branch before it merges. Each Sol worker can run its own team of Sol and Luna agents, sized to the task, and wide GPT-6 Luna sweeps take bulk read-only work. Claude Opus 5.5 agents handle design and lead squads of Sol workers, and Claude Sonnet 5.5 agents survey the codebase.

The conductor's context stays free for planning and judgment because workers spend their own context on implementation and routine verification, and the conductor reads their reports rather than their transcripts. Git worktrees let many writers work in parallel without collisions, and campaign state lives in the repo, so a later session resumes mid-campaign without setup.

The fleet

Role	Model and effort	Job
Conductor	Claude Fable 5.1, high (or Opus 5.5, high)	Plans, dispatches, integrates, and verifies. Writes no feature code.
Worker	Codex gpt-6.1-sol, high	Implements a task in its own worktree, alone or with a leaf team sized to the task.
Leaf	feature: Sol high; critic: Sol high, read-only; grunt: gpt-6-luna xhigh	Takes one slice of a Sol worker's task: judgment work, review of a sibling's diff, or mechanical work.
Design agent and squad lead	Claude Opus 5.5, high or xhigh	UI/UX and design review; leads a squad of Sol workers toward one integration branch.
Consultant and arbiter	Codex gpt-6.1-sol, high or xhigh, read-only	Reviews Claude diffs, answers architecture questions, judges bake-offs.
Deep consult	Codex gpt-6-astra, max, read-only	Runs only when you ask for an exceptional effort on a rare, very hard question.
Scout	Claude Sonnet 5.5, high or xhigh, or Codex gpt-6-luna: max for one deep scout, xhigh for many in parallel; read-only	Surveys the codebase, sweeps it package by package, and researches current APIs.
Other-family reviewers (optional)	agy Gemini 3.8 Flash high, grok Grok 4.7 xhigh, muse Muse Spark 1.3 Contributor max	Read-only second opinions from another model family.

Without Codex, Sonnet 5.5 workers take the implementation role, Opus 5.5 takes tasks that need design judgment, and every other part of the campaign stays the same.

flowchart TD
    C[Fable 5.1 conductor] -->|task briefs| W["Sol workers ×N<br>with leaf teams,<br>one worktree each"]
    C -->|sub-goal briefs| L[Opus 5.5 squad lead]
    L --> LS["Sol workers ×N<br>with leaf teams,<br>one worktree each"]
    LS --> I[campaign/&lt;squad&gt;<br>integration branch]
    W --> V[Conductor merges,<br>re-verifies]
    I --> V
    V --> M[(main)]

What the skill does

The skill is campaign-conductor.

• Bootstrap. First use in a repo copies templates into docs/campaign-hq/ and the three leaf role files into .codex/agents/, then adds a pointer to the project's CLAUDE.md so later sessions resume from repo state.
• Routing. Defaults follow the fleet table. Say "send tests to Opus" or "run this wave at xhigh" and the conductor writes it to preferences.md, where it persists across sessions.
• Sizing. Small projects get a plan written directly. Large or unfamiliar ones get a parallel read-only survey first, and you sign off on the plan before any writer starts.
• Parallel waves. Each writer gets one worktree and branch. A fleet table tracks every dispatch with its session id, and each wave runs dispatch, collect, integrate, verify. A finished Codex session resumes with its context for small corrections.
• Leaf teams. A Sol worker spawns feature (bounded judgment), critic (read-only review), and grunt (mechanical, on Luna) leaves by role name. The role files carry the model and effort, so briefs name roles and a new model generation changes one line per file.
• Squads. An Opus 5.5 lead runs several Sol workers toward one integration branch. Depth stops at conductor, lead, Sol worker, leaf; leaves never spawn.
• Review gates. Every worker returns a fixed-schema JSON report. Codex reviews Claude diffs and Claude reviews Codex diffs, and high-stakes tasks can run as a bake-off judged on diffs and test output.
• Permissions. The Codex sandbox, network access, and Claude permission mode are set once at kickoff and recorded, so a wave never stops on a prompt.
• Memory. Dispatch outcomes and your corrections go into LEARNINGS.md and are compacted into standing rules before the file grows expensive.

examples/campaign-hq/ shows the state files mid-campaign, including a worker brief, a report, and the routing file.

Campaign state lives in the project:

docs/campaign-hq/
├── CAMPAIGN.md      plan, phases, fleet table
├── LEARNINGS.md     standing rules + dispatch log
├── preferences.md   worker routing, permission envelope
├── briefs/          one file per dispatch
├── out/             collected worker reports (event streams gitignored)
└── schemas/         worker-result.json

Install

npx skills add jvogan/a-fable-of-codexes --skill campaign-conductor

or manually:

git clone --depth 1 https://github.com/jvogan/a-fable-of-codexes.git /tmp/afoc
cp -r /tmp/afoc/skills/campaign-conductor ~/.claude/skills/

Use

In any project, tell Claude:

start a campaign

Claude creates docs/campaign-hq/, sizes the plan to the project, and starts dispatching workers. Every later session in that repo resumes the campaign. Direct it in plain language:

• "add a phase for the billing migration"
• "send tests to Opus from now on" (saved to preferences.md)
• "think as hard as you can and get an Astra second opinion on this race condition" (one read-only gpt-6-astra run at max)
• "status" (reads the plan and fleet table)

Requirements

• Claude Code, with Fable 5.1 or Opus 5.5 as the session model. The skill uses the Agent and Workflow tools and runs Claude subagents on Opus 5.5 and Sonnet 5.5.

• OpenAI Codex CLI (github.com/openai/codex). Install with npm install -g @openai/codex or brew install codex, then run codex login. ChatGPT-plan auth draws on plan usage, and limits vary by plan; API-key auth is billed per token. Size waves to your limits. The skill passes model and effort on every dispatch; to match it outside campaigns, set the default in ~/.codex/config.toml:

  model = "gpt-6.1-sol"
  model_reasoning_effort = "high"
  

  Effort runs low, medium, high, xhigh, max, ultra, and the skill routes nothing below high. Luna stops at max. ultra delegates to subagents on its own, so the skill uses it only when you ask.

• Other agent CLIs (optional). Each adds a model family for read-only second opinions and signs in with OAuth: the Antigravity CLI agy (antigravity.google) on gemini-3.8-flash-high, Grok Build grok on grok-4.7 at xhigh, and Meta's Muse Code muse on muse-spark-1.3-contributor at max. Without them, cross-model review runs between Claude and Codex.

• Codex plugin for Claude Code (optional, github.com/openai/codex-plugin-cc). Adds /codex:review, /codex:adversarial-review, and background-delegation commands for one-off tasks outside a campaign:

  /plugin marketplace add openai/codex-plugin-cc
  /plugin install codex@openai-codex
  

Changing models

Model names live in four places. When a new generation ships, update them together:

1. skills/campaign-conductor/SKILL.md: the frontmatter and the routing table.
2. skills/campaign-conductor/references/: the shape tables and command examples.
3. skills/campaign-conductor/assets/: the role files and preferences.md.
4. examples/campaign-hq/ and this README.

A campaign already running keeps its own copies in .codex/agents/ and docs/campaign-hq/preferences.md; edit those to move it to new models.

Validation

python3 scripts/validate.py
npx --yes skills add . --list

validate.py checks each skill against the Agent Skills spec (frontmatter fields, name format, description length), plus this repo's 500-line body limit, relative links, and the required keys in every Codex role file. The skills command confirms the installer can find the package. CI runs both on every push and pull request.

License

MIT