| name | submit-role |
| description | Submit a proof role prompt to a ChatGPT project via CDP and monitor for completion. Verifies Extended Pro (or toggles Deep Research for the literature role), sends, reports the chat URL, and provides recovery guidance for long-running roles. |
submit-role
Canonical content: .claude/commands/submit-role.md in this repo. Codex and
Claude Code share the same operational steps — edit the Claude command file,
not this stub, when changing the procedure.
TL;DR for Codex sessions
Use this when MathPipeProver pauses with waiting_external_agent and the
orchestrator needs to fulfill the pending request manually. The skill:
- Picks the right model mode based on the prompt file — Extended Pro for
analytical roles, Deep Research for the literature role
(
prompts/soft/02_literature_soft.md).
- Submits via
scripts/chatgpt_browser_agent.sh submit ... (preferred) or
scripts/chatgpt_browser_agent/cdp_submit.mjs (lower-level).
- Reports the chat URL.
- Suggests monitoring tools:
/inspect-chat (one-shot), /recover-chat
(harvest), /heartbeat <interval> (unattended orchestrator-pace loop).
Read .claude/commands/submit-role.md for the full step-by-step, CLI
argument shapes, and error-recovery guidance.
