codex-offload-mcp: Command for Claude Code

The user wants the plan→execute collaboration: you do the design thinking, Codex does the typing.

Task: $ARGUMENTS

1. Plan it yourself, in this conversation. Work out the approach: which files change, what each change is, the order to do them in, and how to tell it worked (tests to run, behaviour to check). Investigate the codebase as much as you need — this is the part only you can do.
2. Write the plan as a self-contained, ordered list. Codex cannot see this conversation, so name every file and state every step and acceptance check in the plan text itself.
3. Hand it off with codex_execute_plan(plan, cwd=<repo root>). Since the design is done, consider a lower reasoningEffort unless individual steps are themselves subtle. You get a jobId back immediately.
4. Keep working or verify — don't sleep-poll. The job runs detached; do real work meanwhile and collect with codex_result when you're ready, which checks Codex's report against git. Don't narrate a wait and fire sleep — foreground sleep is blocked by the harness, so the wait never happens and you just re-poll seconds later. If you genuinely have nothing else to do, block on completion by running the returned watchCommand via Bash with run_in_background; it gives one notification when the job finishes. Don't use Monitor here: it is for repeated events, not a one-shot completion wait. Also don't narrate a wait and fire nothing — ending the turn on "I'll check back shortly" leaves nothing to bring you back; either arm watchCommand now or keep working now. If Codex reports a blocker, revise the plan and resume with codex_reply.

If the task is exploratory (each step changes what you'd do next) or needs this conversation's context to execute, say so and do it in-process instead — plan→execute only fits work whose shape is known before it starts.