codex-execlisted
Install: claude install-skill SeanoChang/seanock
# codex-exec
Codex runs on the user's OpenAI plan, so work sent here uses no Claude usage.
Claude leads: it writes the brief, reviews the result and commits. Codex
builds in its own git worktree, and sometimes gives a second review or an
adversary pass. Codex never merges and never commits.
## Route
- **Send to Codex:** work a command can grade (tests, build, lint, a diff check).
- **Keep on Claude:** design calls, shared contracts, bugs with an unclear
cause, refuting findings, research, writing in the user's style, changes to
the harness itself (`.claude/`, `.codex/`, hooks, CI), anything that needs
this conversation's context.
- **Not ready for Codex:** a task whose Done line you cannot write yet.
- **Too small for Codex:** the smallest run below read 92,000 input tokens
(most of them cached) and took 40 seconds. Do a few-line fix in the main
loop.
- **Whole task is routine:** tell the user to run it in the Codex TUI
themselves instead of driving it from Claude.
## Roles
A role is a Codex agent file in `~/.codex/agents/` (source:
`config/codex/agents/` in the harness repo). `codex-run --role <name>` sends
the role's instructions with the run and sets its model and effort.
| Role | Use for | Model, effort | Where it runs |
| ----------- | -------------------------------------------------------------- | ---------------------- | ----------------------------- |
| `builder`