菜单
Run the machine-verifiable gate (typecheck/test/lint, plus a screenshot for UI) so an agent's diff is "review-ready" before it reaches a human. Triggers "proof of work", "is this review-ready", "verify before review", "prove it is green", or after an agent finishes a diff-producing task.
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。
Lock file edits to one directory for the session — Edit/Write outside it are blocked, not just warned. Use when debugging to stop accidental edits to unrelated code, or to fence a parallel agent to one module. Triggers "freeze edits", "lock editing scope", "restrict edits to", "only edit this folder", "unfreeze", "lift the freeze".
Lighthouse audit + improvement loop until targets met. Triggers "lighthouse", "performance audit", "page speed", "improve scores", "LCP", "CLS", "INP", "core web vitals".
Extremely strict whole-codebase maintainability audit — checks structural quality, 1k-line file sprawl, thin wrappers, leaked logic, spaghetti growth, and dependency freshness + usage quality via context7. Pushes "code-judo" moves that delete whole branches instead of rearranging them. When fixes land, the same pass updates CHANGELOG, MANUAL, and derived schemas so the docs don't drift. Triggers "nuclear review", "thermonuclear review", "thermo-nuclear code quality review", "code judo", "deep code quality audit", "harsh maintainability review", "whole codebase review". Sibling to `/review` (per-diff Darkroom checklist) and `/zero-tech-debt` (rework patch to end-state).
Adversarial whole-repo audits in three goal-spec modes — hunt defects, drift, and dead ends rather than validate. Triggers "adversarial audit", "fable audit", "audit the codebase", "expectation gaps" (codebase mode - correctness, incoherences, affordance mismatches); "audit the docs", "docs audit", "documentation audit", "doc drift" (docs mode - accuracy vs code, inverted pyramid, doc sizing, diagram backlog); "process audit", "audit the workflows", "walk the journeys", "end-to-end audit" (process mode - empirical journey walking, state-machine map, dead ends, re-run semantics). Sibling to /nuclear-review (maintainability + code-judo deletion) - this skill audits whether the system does what it promises, not whether it should exist.
CEO/founder-mode plan review. Stress-tests plans through a product lens. Use when: - User says "ceo review", "founder review", "product review" - User wants to challenge whether they're solving the right problem - User asks "is this the right approach?", "should we even build this?" - Before committing to a large feature or architectural change - User wants to evaluate a plan from a product/business perspective
Local pre-commit review of unstaged/staged diff against the Darkroom quality checklist (TypeScript, React, a11y, perf, security). Also summarizes inbound PR review comments on the active PR. Distinct from native /review which inspects open PRs. Triggers "review my changes", "check this diff", "feedback on this code", "PR comments", "what did reviewers say", "summarize PR feedback", post-implementation self-review before commit.
基于 SOC 职业分类
| name | proof-of-work |
| description | Run the machine-verifiable gate (typecheck/test/lint, plus a screenshot for UI) so an agent's diff is "review-ready" before it reaches a human. Triggers "proof of work", "is this review-ready", "verify before review", "prove it is green", or after an agent finishes a diff-producing task. |
The Amdahl-shrink move from the Orchestration Tax: human review is the serial bottleneck, so don't spend it confirming what a machine can verify. An agent's diff is review-ready only when the machine-verifiable battery is green — types, tests, lint (and a screenshot for UI). What a machine can prove shouldn't cost a human's attention.
Run the battery on the current working tree:
bun run proof
It detects typecheck / test / lint from package.json, runs them cheapest-first, and prints one verdict:
review-ready ✓NOT review-ready ✗ — fix the failing gate before a human looksProjects can opt into advisory probes by depending on the tool — the gate then runs the project's pinned binary: react-doctor (React render/quality score, telemetry off) and/or deslop (framework-agnostic cross-file dead-code count). Advisory results are reported but never flip the verdict — deterministic signals alongside the hard gates, not blockers. Silent for projects that don't depend on them.
For UI changes, attach a screenshot (/qa or the chrome-devtools MCP) as the visual half of the proof — tests can't prove "looks right".
The mechanical battery proves the diff is self-consistent — it compiles, tests pass, lint is clean. It cannot prove the diff is correct: a bug that typechecks and passes the tests you wrote sails straight through. When the Codex bridge is available, add a semantic probe from a different model family on top of the mechanical gate:
bun "$HOME/.claude/src/scripts/codex-run.ts" review
Treat it exactly like react-doctor and deslop: advisory — reported alongside the verdict, never flips it. A green mechanical gate stays review-ready even if Codex raises a finding; surface the finding for the human to weigh, don't block on it. The bridge is gated and fails open — silent when Codex isn't installed, authed, or has quota.
Keep it out of bun run proof itself. That gate is cheapest-first and runs constantly; a remote model call would make every proof slow. Run this probe deliberately on non-trivial diffs — not on every typo fix — which also keeps it cheap regardless of how roomy the Codex window is.
the review-queue branch of tool-cadence.ts) limits how many unproven diffs pile up; this gate makes each one cheaper to close.Pure-research or read-only agent output (explore, oracle) has no diff to prove — proof-of-work is for changes, not findings.