Codex to Claude Coordinator

A bidirectional cross-harness plugin, distributable skills, and research project for coordinating agent work from delegation through reviewed, tested handoff.

It provides a repeatable process for:

• resolving a portable capability tier and reasoning effort to an available Claude worker using explicit evidence;
• starting or reusing the right Claude session;
• delegating Claude Code work back to Codex through OpenAI's official codex-plugin-cc;
• steering stalled work without losing project boundaries;
• requiring /review, self-review, available pull-request review agents, tests, and demos;
• separating local readiness from approval, release, and other external gates;
• reporting real blockers and precise next actions.

Codex is the first coordinator and Claude is the first delegated harness, but the supported path is now bidirectional. The portable coordination-core skill contains no vendor commands or model names; coordination-conformance records and validates capability evidence; codex-to-claude-coordinator adapts Codex to Claude; and claude-to-codex-coordinator composes the same lifecycle with OpenAI's official Claude Code plugin.

Discover

• Complete package on skills.sh
• codex-to-claude-coordinator
• claude-to-codex-coordinator
• coordination-core
• coordination-conformance

Model and effort policy

An explicit user model choice always wins. Otherwise, classify the task into six independent axes using the portable core: difficulty, risk, duration, coordination shape, requested capability tier, and reasoning effort.

The Claude adapter maps requested tiers as follows:

Requested tier	Selection
Routine	Sonnet 5
Strong	Opus when selectable; otherwise Sonnet 5
Max	Fable only when selectable and an explicit Fable consumed meter is below 25%; otherwise Opus, then Sonnet 5

Model availability is probed before session creation. Shared, stale, or unlabeled usage bars never become per-model percentages. A steer keeps the session's actual model unless the adapter proves that it can switch; otherwise the coordinator may start a replacement session. Effort is selected separately.

The Claude-to-Codex adapter records the same tier and effort intent but leaves the official plugin's --model and --effort flags unset unless the user explicitly selects them. The configured Codex default remains authoritative, and the actual resolved model stays unknown unless the runtime exposes it.

Install

Install the complete skill set into a project for Codex:

npx skills add wildcard/codex-to-claude-coordinator \
  --skill '*' \
  --agent codex \
  -y

Add -g for a CLI-managed user-level install. For direct authoring, Codex's official locations are repository .agents/skills and user $HOME/.agents/skills.

Install the same source for several skills-compatible harnesses:

npx skills add wildcard/codex-to-claude-coordinator \
  --skill '*' \
  --agent codex claude-code cursor github-copilot goose openhands \
  -y

Install for every agent recognized by the current skills CLI:

npx skills add wildcard/codex-to-claude-coordinator --all

The 0.2 package contains all four skills, including the bidirectional claude-to-codex-coordinator adapter.

For Claude Code plugin development, validate and load the repository directly:

claude plugin validate --strict .
claude --plugin-dir .

To persist a local checkout in Claude Code without publishing it:

claude plugin marketplace add /path/to/codex-to-claude-coordinator --scope user
claude plugin install \
  codex-to-claude-coordinator@codex-to-claude-coordinator \
  --scope user

Start a fresh Claude Code session after installing. This local marketplace registration is machine-local and does not publish the candidate.

Marketplace publication remains approval-gated. Restart or reload a harness if a newly installed skill does not appear immediately.

To give Claude Code the official runtime that delegates back to Codex:

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup

This repository does not vendor that runtime. It supplies the coordinator policy and evidence contract around the official plugin.

Use

Invoke the skill explicitly:

Use $codex-to-claude-coordinator to delegate and supervise this task in Claude.

It also activates for requests to start, steer, monitor, or audit Claude sessions.

Coordinate in the reverse direction:

Use $claude-to-codex-coordinator to delegate and supervise this Codex task from Claude Code.

When loaded as a Claude Code plugin, invoke it as:

/codex-to-claude-coordinator:claude-to-codex-coordinator

The adapter then selects an official operation such as /codex:rescue, /codex:status, /codex:result, or /codex:cancel.

Use the portable core without a vendor adapter:

Use $coordination-core to classify this task and create a coordination envelope stream.

When the repository is loaded as a Claude Code plugin, invoke the same skill through its plugin namespace:

/codex-to-claude-coordinator:coordination-core

Test a harness capability without inferring missing support:

Use $coordination-conformance to record and validate this local experiment.

Generate a one-run prompt and temporary support-file bundle without installing the skill:

npx skills use wildcard/codex-to-claude-coordinator@coordination-core

The remote form uses the current public package. For a local checkout, use npx skills use . --skill coordination-core.

Privacy

Keep delegated context scoped to the active task. Do not copy credentials, unrelated transcripts, raw plugin setup output, or private project material into a different project or published artifact. Use /codex:transfer only for an explicitly authorized transcript from the active project.

Research

• Portable architecture
• Distribution plan
• Launch readiness
• Harness landscape
• Research roadmap
• Claude advisor protocol
• First Claude advisory memo
• Changelog

License

MIT