Claude to codex coordinator

Claude to Codex Coordinator

Use coordination-core for classification and lifecycle envelopes, then map the result to OpenAI's official codex@openai-codex Claude Code plugin. Keep the portable contract independent of the plugin's commands and state strings.

Read references/openai-codex-plugin-cc.md before installing, upgrading, transferring a transcript, enabling the stop review gate, or relying on a specific command or runtime behavior.

Probe Before Delegating

When loaded as this Claude Code plugin, run:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/claude-to-codex-coordinator/scripts/probe.py"

In another skills-compatible harness, resolve scripts/probe.py from the installed directory that contains this SKILL.md; never resolve it from the active project's working directory.

The probe reports only sanitized availability facts. It omits install paths and raw authentication output. Stop and resolve any false prerequisite before delegating.

Do not copy raw /codex:setup output into a repository or handoff. It can contain account labels. Do not enable the optional stop review gate during setup unless the user explicitly requests it and will monitor the resulting Claude/Codex loop.

Resolve the Assignment

1. Apply the portable six-axis classifier.
2. Record whether the task is review-only or write-capable.
3. State the repository scope, definition of done, required tests, evidence, and external-action boundary.
4. Preserve an explicit user model or effort choice.
5. Otherwise leave --model and --effort unset so the Codex configuration remains authoritative.
6. Record the requested tier and effort even when the adapter cannot prove the actual resolved model.

Do not map routine, strong, or max to a concrete Codex model without a separate, versioned capability mapping. The plugin accepts explicit model and effort flags, but its normal result and status surfaces do not prove the actual resolved model. Record that field as unknown unless positive runtime evidence is available.

Choose the Official Operation

Intent	Claude Code operation
Verify setup	/codex:setup
Read-only implementation review	/codex:review
Challenge design and assumptions	/codex:adversarial-review
Investigate or implement	/codex:rescue
Continue the same completed rescue thread	/codex:rescue --resume
Force an independent thread	/codex:rescue --fresh
Inspect active or recent work	/codex:status
Read a finished result	/codex:result
Stop an active job	/codex:cancel
Import the current Claude session into Codex	/codex:transfer

Use --background for long or multi-step work and --wait for a small bounded task. Always state read-only intent explicitly when edits are not authorized; the upstream rescue subagent otherwise defaults to a write-capable Codex run.

Treat resume as between-turn follow-up steering. The official plugin does not expose arbitrary mid-turn steering. Cancel the active turn and start or resume a later turn when direction must change.

Supervise the Job

1. Preserve the Codex output verbatim at the plugin boundary.
2. Treat that output as worker evidence, not completion proof.
3. For background work, inspect /codex:status, then /codex:result after a terminal state.
4. Verify claimed edits against the working tree and run the required tests.
5. Use /codex:review or an independent reviewer for the exact diff.
6. Route one concrete follow-up through --resume only when continuity is useful.
7. Use /codex:cancel when the run is looping, out of scope, or no longer authorized.

Keep implementation, local validation, live validation, review, human approval, and release as separate ledger states. The plugin may edit the local workspace; it does not grant permission to push, open a pull request, publish, send a message, or perform another external action.

Protect Transcript Boundaries

/codex:transfer imports a Claude transcript from the active Claude projects directory into a persistent Codex thread. Use it only when the user explicitly wants that session transferred and the transcript belongs to the active project. Never transfer an unrelated session, raw credential, or private cross-project history.

The plugin's repository-scoped job state and logs are runtime evidence. Keep raw files outside the distributable package and admit only reviewed, redacted derivatives through coordination-conformance.

Produce the Handoff

Report:

• Claude session and Codex job or thread identifiers when safely available;
• requested tier and effort, requested model flags, and actual model or unknown;
• review-only or write-capable mode;
• repository scope and current working-tree or commit state;
• status, result, touched files, tests, and review evidence;
• whether the run was fresh, resumed, cancelled, or transferred;
• remaining approval or external-action gates;
• the precise next action and owner.