菜单
Land reviewed, CI-green Hermit changes before human review and mark them for follow-up. Use as the default autonomous landing discipline.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/facebookexperimental/hermit --skill post-facto-review命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Apply a disciplined read, plan, execute, and adversarially verify workflow. Use for complex research, architecture, implementation, or audit tasks.
Debug hermit/detcore runs (nondeterminism, hangs, syscall gaps, scheduling) using hermit's built-in logging and log-diff FIRST, before reading source. Use whenever a guest program behaves unexpectedly, --verify fails, or a run hangs under hermit.
Gate key Hermit changes on adversarial review and explicit human approval. Use only when the user explicitly requests human-review-first mode.
基于 SOC 职业分类
正在显示 SKILL.md
| name | post-facto-review |
| description | Land reviewed, CI-green Hermit changes before human review and mark them for follow-up. Use as the default autonomous landing discipline. |
Every PR description and comment created under this workflow MUST start with the applicable role tag:
[impl agent, MODEL] for implementation agents[adversarial-reviewer agent, MODEL] for review agents[coordinator, MODEL] for coordinator agents[Human] for the human ownerExamples: [impl agent, gpt-5.6-sol],
[adversarial-reviewer agent, opus-4.8].
The currently-active landing discipline for autonomous multi-agent work.
Changes land as soon as they are reviewed and CI-green; the human reviews them
after they are on main and fixes forward. This is the fast counterpart to
human-review-first, which is dormant and gated on an
explicit human request.
Status: ON (default). This is how the repo runs today. To switch to the cautious gate, the user must explicitly ask for human-review-first mode.
Optimize for merge velocity while keeping a real quality bar. Autonomy is not an excuse to skip review — key changes are still adversarially reviewed. The difference from human-review-first is ordering: the human's review happens after landing, and mistakes are corrected by follow-up commits rather than by blocking the queue.
The "key change" definition is identical to human-review-first: new syscalls, major Reverie API changes (small additive extensions are OK), scheduler/determinism-model changes, record/replay format changes.
Before landing a key change:
human-review — marks a PR the human still wants to look at. Never
auto-close or auto-land a human-review PR. Under post-facto mode these
stay open for the human even though other work lands around them.post-facto-review — marks a PR that landed autonomously and is awaiting the
human's after-the-fact review.human-approved. That label means a human actually approved,
and only a human may apply it. A bot claiming approval is a defect.locally-validated — the legitimate substitute for green CI when the CI lane
cannot go green for environmental reasons: run the checks the PR can affect
locally, prove any residual failure is baseline/environmental, then label +
merge on real GitHub-hosted green where possible (avoid --admin over red CI).Autonomously-landed code carries in-source breadcrumbs so a human reviewing post-facto can find exactly what a bot wrote and what still needs eyes:
// AUTONOMOUS-BOT-IMPLEMENTED — this code was written and landed by a bot
without prior human review.// TODO-HUMAN-REVIEW(PR-id) — a specific spot the human should scrutinize;
include the PR number, e.g. // TODO-HUMAN-REVIEW(#206).Keep markers at the smallest meaningful scope (the function/block that is novel), not blanketed across untouched code.
Once a key change survives adversarial review and CI is green, land it —
squash-merge to main. Do not wait for a human.
main is
unprotected); a red self-hosted lane does not block landing.--admin, it should
only be bypassing the known-environmental self-hosted lane, not a genuine
red on GitHub-hosted or on a meaningful check.main (see the PR DAG
section of human-review-first).The human reviews landed changes after the fact (aided by the labels and code
markers above). Corrections are made by follow-up commits/PRs, not by
reverting the queue — fix forward. If a human review finds a real defect, open a
fix PR that removes the relevant // TODO-HUMAN-REVIEW marker once addressed.
Switch to human-review-first when the user explicitly asks for it. Announce the switch; from that point every key change waits for human approval before landing.