菜单
Recover from failed, stalled, or unclear Codex subagent runs without over-reading unstable transcripts. Use when a Codex subagent failed, stalled, drifted, returned an insufficient result, or needs steering, closing, or a follow-up recovery prompt.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Install, update, audit, or repair Promode for Codex in a project. MUST be loaded when the user asks to set up promode-codex, install Promode for Codex custom agents, check the Codex Promode hook or setup, update the project-scoped Promode agents, scaffold Promode project tracking files, or verify that Promode is correctly adapted for Codex.
Audit how well a repository aligns with the Promode methodology in Codex, then produce a prioritised improvement plan. Use when the user asks to assess Promode alignment, audit a repo against Promode practices, improve tests, feedback loops, traceability, agent knowledge, runbooks, or architecture, or bring a codebase into Promode shape. Also flags stale Claude Promode install leftovers; use managing-promode-codex for Codex setup audits.
Put the bulk of acceptance coverage below the UI through a fast, deterministic headless client driving an operator seam, and reserve a surgical UI state-graph tier for defects that only manifest through the real GUI. Use when designing test/QA or acceptance-testing strategy, automating acceptance, end-to-end (E2E), or QA testing of a running app, deciding what to cover with fast headless tests vs slow UI/E2E, building agent-driven exploration or automation of a running app, building a below-UI operator seam (interaction layer) or headless client, or crystallising agent-discovered knowledge into reusable deterministic artifacts (maps, graphs, scripts, tests). Covers the Discovery⇄Determinism flywheel, the operator-seam architecture (one seam serving both a headless test client and AI-agent tools), and layered headless-first acceptance testing with a surgical UI state-graph tier for GUI-only defects.
Write a handoff document so a fresh Codex agent can continue this work after the conversation ends. Invoke when the user is about to clear or compact context, when context is filling up, or when the user asks to hand off, checkpoint, pause, or resume work later. Optional user text may specify what the next session should focus on.
| name | recovering-subagents |
| description | Recover from failed, stalled, or unclear Codex subagent runs without over-reading unstable transcripts. Use when a Codex subagent failed, stalled, drifted, returned an insufficient result, or needs steering, closing, or a follow-up recovery prompt. |
<core_rule>
Do not treat Codex transcript files as a stable API. Current Codex docs expose
transcript_path and agent_transcript_path for convenience, but explicitly
warn that transcript format may change. Prefer tool-visible agent status,
notifications, final messages, changed files, and explicit reported evidence.
</core_rule>
<recovery_process>
promode_debugger if installed.<codex_tools>
Use whatever Codex multi-agent controls are available in the current harness:
wait_agent, send_input, resume_agent, and close_agent when present. In
CLI sessions, /agent can switch between active agent threads.
</codex_tools>
<recovery_prompt_template> Use a prompt like:
You drifted from the assigned scope. Stop adjacent work. Report only:
1. What you completed, with file paths.
2. What failed or remains uncertain.
3. The single next action you recommend.
Do not make further edits unless explicitly asked.
</recovery_prompt_template>
<success_criteria> Recovery is complete when the parent agent knows whether to integrate, steer, respawn, dispatch a debugger/reviewer, or close the subagent. Do not keep polling a stalled agent when a smaller replacement task would be clearer. </success_criteria>