Operating Mode
- Run this skill as a Ralph-style closed-loop execution protocol adapted to Codex.
- Do not imitate Claude Code's exact stop-hook mechanics. Import the engineering discipline: explicit exit criteria, one high-value item per loop, durable state, and validation-backed continuation.
- Stay inside system, developer, repository, and safety constraints.
- Read references/ralph-loop.md when you need Ralph rationale, Codex adaptation boundaries, durable-state guidance, failure-mode handling, or copy-ready resume prompts after interruption.
Default Resume Prompt
Use this as the default Chinese resume prompt after a capacity error or dropped session:
使用 $ralph-loop。上一轮 Codex/CLI 因容量或连接问题中断。
不要依赖聊天记忆,先从当前仓库状态恢复:检查 git status、变更文件、最近 diff、失败的测试/日志,以及已有任务说明。
然后:
1. 用一句话重述目标。
2. 列出完成标准。
3. 区分哪些已经完成,哪些仍未完成。
4. 继续当前最高价值的下一步。
5. 完成后立刻做最小但有效的验证。
不要重复已经成功的编辑、安装、迁移或其他有副作用的命令,除非你先确认确实还需要执行。
除非任务已完成或遇到真实阻塞,否则继续闭环执行,不要停下来询问是否继续。
Set The Loop Contract
- Restate the real goal in one sentence before substantial work.
- Define explicit done criteria as early as possible.
- If the user did not provide done criteria, infer pragmatic ones from the request, repository conventions, and available validation.
- Pick the smallest loop unit that can be validated quickly.
- If the task is large, keep a lightweight ordered checklist in durable form when that materially improves continuity.
Run The Loop
- Re-read the goal, active files, open failures, and latest validation signal.
- Search the codebase before assuming something is missing, broken, or unimplemented.
- Pick the single highest-value unresolved item that moves the task toward done.
- Implement only that slice or fix only that failure class.
- Run the smallest high-signal validation for that slice.
- If validation fails, debug and retry instead of pausing for approval.
- If validation passes, update any durable artifacts needed for the next loop.
- Re-check the done criteria. If not done, continue immediately into the next loop.
Engineering Rules
- Default to one meaningful item per loop.
- Treat tests, linters, typecheckers, builds, runtime checks, and reproducible scripts as loop backpressure.
- Prefer durable repo state over long conversational reasoning.
- Prefer reversible assumptions over interruptions.
- Match established project conventions unless there is a clear technical reason not to.
- Preserve user changes and adapt to the current worktree instead of reverting unrelated edits.
- Do not stop at analysis if code can be changed and checked now.
- Do not ask "should I continue?" after completing a partial step.
Interruption Recovery
- Treat model-capacity errors, transient host failures, and dropped sessions as execution interruptions, not as evidence the task should stop.
- Resume from durable state first: current files,
git status, diffs, failing checks, task notes, and the latest validation signal matter more than chat memory. - When resuming, first reconstruct what is already complete and what remains open before making new edits.
- Do not blindly rerun commands with side effects. Re-run only after verifying necessity or idempotence.
- Prefer a structured resume prompt over a bare "continue" when recovering from capacity errors or a fresh session.
- Use the standard prompts in references/ralph-loop.md when the user asks for a safe Codex or CLI continuation after interruption.
Stop Conditions
- Stop when the explicit done criteria are satisfied with evidence.
- Stop when a real blocker remains after reasonable local attempts to resolve it.
- Real blockers include destructive or irreversible actions needing consent, missing secrets or approvals, legal or safety concerns, and irrecoverable ambiguity with materially different outcomes.
Communication
- Send short progress updates stating what is being done and what was learned.
- Report blockers only after trying reasonable local resolution paths.
- Keep the final handoff concise and outcome-focused.
- Include validation results, assumptions that mattered, and any residual risk.