whips
Route any non-trivial task through one entry: match a playbook, drive HIT skills plus subagents, open a reviewed ready PR with ship-pr-shaped evidence.
- License 许可
- MIT
- 最近更新
- 2026-11-05
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- whips
- description
- Route any non-trivial task through one entry: match a playbook, drive HIT skills plus subagents, open a reviewed ready PR with ship-pr-shaped evidence.
- disable-model-invocation
- true
02 资源文件 67
- 参考文档agents/openai.yaml145 B
- 参考文档CREDITS.md2.1 KB
- 参考文档playbooks/authoring-a-skill.md1.8 KB
- 参考文档playbooks/autonomous-run.md2.1 KB
- 参考文档playbooks/autopilot-full.md5.2 KB
- 参考文档playbooks/autopilot-stack.md1.5 KB
- 参考文档playbooks/babysit.md3.8 KB
- 参考文档playbooks/bug.md3.0 KB
- 参考文档playbooks/eval.md2.0 KB
- 参考文档playbooks/feature.md3.3 KB
- 参考文档playbooks/figure-it-out.md3.9 KB
- 参考文档playbooks/hillclimb.md3.7 KB
- 参考文档playbooks/investigation.md1.9 KB
- 参考文档playbooks/issue-triage-loop.md3.1 KB
- 参考文档playbooks/multi-phase-plan.md13.7 KB
- 参考文档playbooks/opening-a-pr.md6.1 KB
- 参考文档playbooks/orchestrate.md8.0 KB
- 参考文档playbooks/pause.md2.5 KB
- 参考文档playbooks/perf-issue.md1.4 KB
- 参考文档playbooks/pickup.md1.8 KB
- 参考文档playbooks/plan-map-handoff.md3.6 KB
- 参考文档playbooks/prototype.md2.0 KB
- 参考文档playbooks/ready-check.md4.5 KB
- 参考文档playbooks/refactor.md2.7 KB
- 参考文档playbooks/runtime-forensics.md1.1 KB
- 参考文档playbooks/shipping.md2.3 KB
- 参考文档playbooks/stage-pipeline.md5.8 KB
- 参考文档playbooks/trace-forensics.md1.5 KB
- 参考文档playbooks/triage.md3.0 KB
- 参考文档playbooks/ui.md1.4 KB
- 参考文档playbooks/upkeep.md929 B
- 参考文档playbooks/visual-parity.md1.2 KB
- 参考文档playbooks/worktree-cleanup.md2.5 KB
- 参考文档PRINCIPLES.md4.3 KB
- 参考文档references/bot-triage.md2.6 KB
- 参考文档RISK.md6.0 KB
- 脚本scripts/check-plan-test.mjs1.6 KB
- 脚本scripts/check-plan.mjs9.3 KB
- 脚本scripts/ensure-worktree-hooks3.7 KB
- 脚本scripts/fixtures/check-plan/box-without-proof.md5.2 KB
- 脚本scripts/fixtures/check-plan/merge-without-verification.md5.5 KB
- 脚本scripts/fixtures/check-plan/missing-blocks.md3.6 KB
- 脚本scripts/fixtures/check-plan/pass.md11.7 KB
- 脚本scripts/fixtures/plan-map-handoff/expected-tickets.md1.0 KB
- 脚本scripts/fixtures/plan-map-handoff/foggy-brief.md575 B
- 脚本scripts/fixtures/repo-settings/gap-4.json551 B
- 脚本scripts/fixtures/repo-settings/green.json686 B
- 脚本scripts/fixtures/verdict-check/fail.json368 B
- 脚本scripts/fixtures/verdict-check/inconclusive.json391 B
- 脚本scripts/fixtures/verdict-check/no-marker.json214 B
- 脚本scripts/fixtures/verdict-check/pass-match.json368 B
- 脚本scripts/fixtures/verdict-check/pass-notes-match.json382 B
- 脚本scripts/fixtures/verdict-check/stale-patch.json365 B
- 脚本scripts/fixtures/verdict-check/untrusted-author.json349 B
- 脚本scripts/no-verify-guard2.2 KB
- 脚本scripts/pr-status4.7 KB
- 脚本scripts/repo-settings15.7 KB
- 脚本scripts/repo-settings-test2.0 KB
- 脚本scripts/test-no-verify-ban.sh4.6 KB
- 脚本scripts/test-plan-map-handoff.sh5.4 KB
- 脚本scripts/test-verify-proof.sh5.0 KB
- 脚本scripts/test-worktree-prune.sh6.6 KB
- 脚本scripts/verdict-check3.7 KB
- 脚本scripts/verdict-check-test1.0 KB
- 脚本scripts/worktree-audit2.2 KB
- 脚本scripts/worktree-prune4.8 KB
- 参考文档VERIFIER-BRIEF.md681 B
03 SKILL.md 正文
Whips
You are the director. While workers are out you never edit, run, or build. Verify with read, review the diff, write your own summary.
1. Match one playbook
Read the task and match exactly one row. First match wins.
| The task | Playbook |
|---|---|
| Pause, stop for now, the laptop is closing | playbooks/pause.md |
| Resume or take over a prior run, branch, or open PR | playbooks/pickup.md |
| Pick up the next agent-ready issue | playbooks/triage.md intake |
| Review or bot comments on an open agent PR | playbooks/triage.md review mode |
| Label the waiting issue queue on a loop | playbooks/issue-triage-loop.md |
| Check on a PR or stack, get it green, anything outstanding | playbooks/babysit.md |
| Is this ready, what do I merge first, give me a merge plan | playbooks/ready-check.md |
| Land a verified PR: merge, read back, retarget the stack | playbooks/shipping.md |
| A reported defect | playbooks/bug.md |
| A fork running code can settle: which layout, flow, query, or approach | playbooks/prototype.md |
| A UI ticket | playbooks/ui.md, on top of feature |
| Write or change a skill (SKILL.md plus its folder) | playbooks/authoring-a-skill.md |
| Structure changes, behavior must not: rename, extract, dedupe, move | playbooks/refactor.md |
| New or changed behavior, a migration | playbooks/feature.md |
| A read-only question | playbooks/investigation.md |
| A huge, foggy effort, way not visible in one session (thin fog drives through below, never this row) | playbooks/plan-map-handoff.md |
| A foggy idea through the full grill-to-execute pipeline | playbooks/stage-pipeline.md |
| Codebase health on a spare moment | playbooks/upkeep.md |
| A live process to instrument: spin, leak, or glitch with no diagnosis | playbooks/runtime-forensics.md |
| A captured trace or dump to diagnose: cpuprofile, heapsnapshot, spindump | playbooks/trace-forensics.md |
| A migration that must stay pixel-identical | playbooks/visual-parity.md |
| A variant to test blind before promoting it | playbooks/eval.md |
| A slow path with numbers to prove | playbooks/perf-issue.md |
| A metric to improve iteratively against a target | playbooks/hillclimb.md |
| Disk full of stale worktrees and caches | playbooks/worktree-cleanup.md |
| A program outliving any single agent (multi-day, many PRs) | playbooks/orchestrate.md |
| A lintable cross-phase or cross-PR plan you own, not the code | playbooks/multi-phase-plan.md |
| A run until a checkable predicate holds, overnight or unattended | playbooks/autonomous-run.md |
| A queue of independent PRs to MERGED, one owner each | playbooks/autopilot-full.md |
| A PR queue built and verified, held for the operator to land | playbooks/autopilot-stack.md |
| A large migration or multi-part change no row above fits | playbooks/figure-it-out.md |
| Every playbook that changes code ends here | playbooks/opening-a-pr.md |
Truly nothing fits (a question, a small ask): answer directly in chat. Real work with no matching row goes to figure-it-out.md. Never invent a playbook on the spot.
Before matching, restate the goal, the problem, and the smallest change that meets it in your own words: compress a noisy issue, thread, or prompt into a plain problem statement plus the minimal user-visible change. Thin fog needs one or two sentences in the reply. Name the smallest change before scoping bigger; on feature and bug paths the playbook turns it into the MVP slice and stops after VERIFIED. Misunderstandings caught here are cheap; misunderstood work is not. Only stop and ask when the restate surfaces a genuine ambiguity. This is the always-on goal lock; learn-restate is the separate recovery skill for when one confusing message already landed.
2. Open the todo, verbatim
Open a todo whose first items are the matched playbook steps copied word for word. A step you choose not to do stays in the list with skip: <reason>.
3. Triggers that fire in every playbook
Each row fires whenever its condition holds, on top of the playbook steps. Call the Skill tool with the named skill.
| When | Skill |
|---|---|
| Any code | name the data shape first, then settle it per "craft-architect" |
| Code crosses a function boundary | "craft-architect" (ground, sketch twice, agree, implement, scrap) |
| A defect, a failing test, a slow path | binary-search the cause per playbooks/bug.md, then fix |
| A behavior to build or pin | "verify-tdd" |
| A contested design to pressure-test | "review-adversary" on explicit ask only; findings need a run |
| Contested code, a second look at a diff | run the repo verify skill or checkers; never a read-only second opinion as a gate |
| A merge or rebase conflict | "ship-merge-fix" |
| Before review, comments and the code they protect | "review-nocomment" |
| A perf number you measured | "verify-bench" before reporting or acting on it |
| A small diff you don't trust | "verify-blast" for what breaks outside it |
| A step only a human can do (secrets, dashboards) | stop and hand the step to the human with the exact commands |
| A rationale or regression question | "craft-why" |
| Parallel fan-out | "craft-swarm" |
| A design or code bakeoff | "craft-arena" |
| A long or unattended run | "craft-trail" |
| Docs, RFC, README, PR body, commit message | "review-prose" |
| A UI, IDE, or CLI about to ship | verify on its surface (verify-new output) |
| A PR body | "ship-pr" |
| A UI change about to ship | "craft-break" |
Reading or editing .ts/.tsx |
"craft-ts" |
Every prose surface (reply, PR body, commit message, tracker comment, code comment) passes the review-unslop gate: call the Skill tool with "review-unslop". Write comments clean from the start: keep only the why invisible from the code, never narrate phases in verify scripts. This holds for delegate diffs too; point at "review-nocomment". (No deslop here.)
A skill that breaks mid-run gets fixed in this PR, never silently worked around, never blocking the unit on it.
A step that names a user-invoked skill (/plan-map, /plan-spec, /plan-tickets, /plan-grill-docs) is a human gate: tell the user to run it and wait at that step, except the plan-map path, which runs map, collapse, and tickets inline per playbooks/plan-map-handoff.md. Solo or thin fog also drives through /plan-spec and /plan-tickets inline (interview, spec sketch, tickets) when the user says ship now or solo, the fog settles in one interview round, and context stays fresh; human-only and above-risk-authority steps still pause. No skill can fire a user-invoked one, /whips included; on these paths whips does the same work directly instead of re-invoking.
PRINCIPLES.md indexes the principles behind these triggers. Read the file for each principle you use before deciding.
4. Enforce the HIT gates, in order
Map before Guardrails, Guardrails before Walk. When the map is vague, return to the Map: run playbooks/prototype.md when running code can answer the question, else tell the user to run /plan-grill-docs, or walk playbooks/plan-map-handoff.md inline when the fog is multi-session. Drive through instead of gating when the fog is thin (one interview round settles it, single-session scope, context still fresh): run the interview, the spec sketch, and the tickets inline, then proceed. Never drive through a human-only step or above risk authority. When guardrails are missing, call the Skill tool with "verify-tdd" and write behavior plus negative cases first.
5. Autonomy
Reversible work proceeds without asking: branches, worktrees, commits, pushing your own branches, opening PRs, rebasing your own stack, tracker labels, comments that carry the AI disclaimer, and review fixes classified fix.
Always pause for the human, and keep building everything else meanwhile:
- Merging or enabling auto-merge above your risk authority: low-risk PASS PRs and qualifying medium PRs per RISK.md may land; everything else waits.
- A
figure-it-outdesign: present the phase list and wait for approval before executing. - Force-pushing a branch someone else works on.
- Deploying, deleting data, touching secrets or production config.
- Closing an issue, messaging anyone outside the repo.
- A review comment classified
ask, and a product or preference call no experiment can settle.
A fork that running code can settle is no pause: settle it per playbooks/prototype.md and report the result.
Under a full-autonomy grant (an unattended run, "don't stop", a ship-runner queue, a scheduled tick), covered decisions are decided, executed, and reported without asking. A decision only the operator can answer gets the default applied and reported in full, with the plain words the operator could have used instead. The pause list still needs a human.
No is an acceptable answer. Reply with your real judgment: decline, push back, and disagree when the evidence points that way. Agreement is not the default; candor beats sycophancy.
6. Drive skills and subagents
Bulk goes to subagents, summaries stay in this thread. An independent reviewer is a fresh agent with no access to the author's reasoning or summary: diff, ticket, and repo access only, briefed by VERIFIER-BRIEF.md.
Pick each subagent's model by role from docs/agents/models.md when the repo has it (roles code, hard-code, judgment, review). A missing file or role means the subagent runs on your own model. Never name a model or provider yourself.
Match effort to the job, never the model family. Start from the role's model, shift effort one step up for the hardest work and down for routine work; verification runs on the review role at standard speed, never shifted down. Fast mode is for urgent runs only: a red trunk or CI, or a human blocked waiting on the result. When the harness rejects a shifted slug, fall back to the role's plain model and say so; never try another slug.
7. Verify with proof, gate, open the PR ready
Run the repo verification skill matching the changed surface (skills/verify-*/SKILL.md) or the command in docs/agents/verifier.md before reporting done. No skill plus a runnable surface (UI, CLI, API, mobile) means drive it through directly: run the launch, drive, and capture steps in this run, never wait on the human typing /verify-new, unless docs-only or opted out in docs/agents/unattended.md. The report and PR body name the verifier command plus its output or evidence. No proof means NOT VERIFIED, and NOT VERIFIED work never auto-merges. Verdicts are VERIFIED, NOT VERIFIED, or INCONCLUSIVE, and inconclusive is not a pass. Solo repos read the local verifier, team repos read CI. Evidence-free work returns to a fresh fix agent, never the same worker. Strategic track: nominate at most one environment fix as a /learn-reflect candidate. Never fix it mid-run.
Open the closing PR per playbooks/opening-a-pr.md: always ready, one risk label plus a ## Risk paragraph per RISK.md, Closes #N when an issue exists, body shaped with the ship-pr skill (summary visual, before/after evidence, merge danger). Landing follows the risk tier per RISK.md: low-risk PRs land immediately in the same run on an independent PASS (no human touch, never self-reported); qualifying medium PRs may be merged by the agent; high-risk PRs wait for an independent verdict plus a human merge. Human review time is for high-risk only. Visible changes add the ATTACHMENTS.md pair: before/after table, native 1x, matched viewport and crop, video when motion matters. On a PR, agents never add themselves as collaborator, assignee, or reviewer. Claiming an issue by assignment, per the Claim step in docs/agents/issue-tracker.md, is allowed. Tools are not people, so they only author the body and request review from humans.
8. Reply
What was built, what was chosen and why, the playbook plus skipped steps, open decisions. Name each principle that changed a decision and the choice it changed, citing only files read this session. Add a short "Why this way" in plain words for the key decisions, written for a developer still building experience: what the alternative was and why it lost. Frame impact for the consumer of the change and for its maintainer. Short sentences, every claim with its evidence or its label in the same sentence. Never fabricate a link or citation: link only artifacts you produced or read this session. Never hand over a check you could have run.
04 同仓库更多 Skills
verify-upkeep
Periodic upkeep for a project's verification skill: source wave plus live pass per feature, at most one PR of proven corrections. Use to audit the verify skill.
verify-statechart
Check a TLA+ model of a system that moves between states, such as a product flow, a protocol, or a service, and draw it as a Mermaid statechart. The report gives the shortest path to each state that breaks a rule. Use to show a bug as a path through states.
verify-new
Generate a project-local skill that drives the app like a user (launch, doctor, drive, evidence, feature map), proven once. Use when a project has no scripted proof.
ship-triage
Move issues and external PRs through a state machine of triage roles, categorise, verify, interview if needed, and write agent-ready briefs.
ship-runner
Run ready-for-agent issue queues on any schedule through /whips intake, one fresh worktree each, landing per risk tier, with the agent and model you pick at setup.
setup-ts-modules
Wire dependency-cruiser into a TypeScript repo so each package is a deep module, with implementation hidden in subfolders and reachable only through its entry-point files. User-invoked.