autopilot

Run pre-commit or pre-merge quality checks: tests, completeness scan (no stubs/TODOs/mocks), code review. Use when: "quality gate", "quality checks", "run tests before merge", "check for stubs", "scan for completeness", "is this ready to commit?", "pre-merge review", "品質檢查", "準備好可以 commit 了嗎", "跑一下檢查". Not for: writing new tests (→ TDD), debugging CI failures, or receiving external review feedback.

Terse CEO front-door — Level 6: like /l5 (worktree-isolated hetero implementer + authoritative qc) but the VERIFICATION AUTHORING is also leaf-dispatched to a heterogeneous engine; depth-0 remains pure orchestration. Use when: "/l6 <goal>", "L6 <goal>", "全委", "全部派遣", "省 token 全外包", "delegate everything incl verification". Presets involvement=just-results, scope=Hold, project red lines plus -x additions (override --mode / --expand / --solo). Not for: /l5 when you still want to do verification yourself; /l4 all-Claude; /l3 inline.

Closing sequence forcing function — use at the END of any dev-flow workflow (L, H, Fix, S) to guarantee no step in the closing sequence gets silently compressed or skipped. On invocation, creates size-appropriate sub-tasks via TaskCreate so each step is individually trackable. MANDATORY for L-size (invoked at L-5) and H-size (invoked at step 9); optional for Fix/S. Use when: finishing L-size project, closing hotfix, "time to merge", "wrap this up", "跑完收尾", "收掉這個專案", "L-5 開始". Not for: mid-phase work, starting new work (→ dev-flow), authoring a plan doc (→ references/plan-template.md).

Bootstrap, organize, or archive project tracking docs. Use when: "archive this project", "bootstrap from plan", "set up project tracking", "clean up project docs", "update INDEX", "reorganize docs/projects/", "歸檔這個專案", "建專案", "整理專案文件", project is done and merged — needs cleanup. Not for: authoring a plan doc (→ references/plan-template.md), choosing what to work on (→ next), or running quality gates (→ quality-pipeline).

Engineering retrospective from git history — velocity, test ratio, focus score, commit patterns. Use when: "/retro", "retro on last N days", "how productive was this sprint", "analyze my commit history", "work patterns", "session analysis", "回顧", "分析工作模式", "這週做了什麼". Not for: viewing specific commit diffs, comparison audits, or debugging test coverage drops.

Pinned participatory pipeline — research industry best-practice → write a plan → bounded, rubric-frozen plan review (maximum two generations) → expand into a tracked project → execute per dev-flow. Each phase ends at a human approval gate. Use when: "research best practice on X then build it properly", "查業界 best practice 寫成 plan、有限 review 後展開成 project 照 dev-flow 跑", "把這個主題做成正式專案", "spec it out then ship it the rigorous way". Not for: full hands-off autonomy (→ ceo-agent), a quick fix / already-known implementation (→ dev-flow), research with no build (→ survey / deep-research), or a single irreversible decision (→ think-tank-dialectic).

Start here before writing any code — sizes task (S/L/H/Fix), sets up branch and session rules. Use when: "I'm starting on X", "quick fix for Y", "continuing from yesterday", "hotfix needed", "let's implement X", "skip to coding", "我要開始做 X", "快速修一下", "接續昨天的進度", resuming a feature branch, or any task that touches code. Not for: debugging (→ debug), authoring a plan doc (→ references/plan-template.md), pre-code design exploration (→ brainstorm), or code review (→ quality-pipeline).

Full delegation — you own the goal end-to-end, user just wants results. Triggers: "CEO mode", "get it done", "you decide", "handle everything", "I trust you", "take over", "full authority", "just do it", "搞定 X", "幫我處理", "全權處理", "你決定". Not for: research-only (→ survey), participatory planning (→ dev-flow), or parallel task dispatch.

Terse CEO front-door — Level 3: full autonomy, CEO executes inline on this thread, escalate only at the DOA boundary. Use when: "/l3 <goal>", "L3 <goal>", you want the "全權處理 / get it done" behavior as one command without the CEO startup Q&A. Presets involvement=just-results, scope=Hold, project red lines plus -x additions (override --mode / --expand). Not for: offloading to a background foreman (→ /l4), hetero impl engine (→ /l5), participatory planning (→ dev-flow), research-only (→ survey).

Terse CEO front-door — Level 4: dispatch ONE background, worktree-isolated sub-orchestrator "foreman" that runs dev-flow unattended and returns a verdict; CEO context stays clean and the authoritative qc verdict is held at depth 0. Use when: "/l4 <goal>", "L4 <goal>", you want a long autonomous run offloaded off the main thread. Presets involvement=just-results, scope=Hold, project red lines plus -x additions (override --mode / --expand / --solo). Not for: inline execution (→ /l3), hetero impl engine (→ /l5).

Terse CEO front-door — Level 5: like /l4 (background worktree-isolated foreman, depth-0 control loop + authoritative qc) but the IMPLEMENTER is orchestrated by the engine CLI and dispatched through the canonical `engine implement-review` path. Use when: "/l5 <goal>", "L5 <goal>", you want cost-arbitrage or a decorrelated second engine doing the mechanical impl. Presets involvement=just-results, scope=Hold, project red lines plus -x additions (override --mode / --expand / --solo). Not for: all-Claude run (→ /l4), inline (→ /l3).

Onboard a new heterogeneous engine into the autopilot lifecycle. Use when: "onboard a new engine", "qualify gpt-X / a new model as a reviewer", "is model Y good enough", "add an implementer engine", "add a planner engine", "add a verifier engine", "evaluate a model as orchestrator", "route a model by role", "new model for review/dispatch", "新增一個引擎", "驗證某模型夠不夠格", "這個 model 能不能用", "加一個 reviewer/implementer/verifier/orchestrator 模型". Not for: writing new scorecard scripts, inventing new routing policies, or deciding model-family domain fit.

Research what the industry does — dual-agent (researcher + skeptic) tech investigation. Use when: "research X", "survey", "investigate options/tradeoffs", "what do others use", "industry standard for X", "compare X vs Y vs Z", "I'm not sure which to pick", "業界怎麼做", "調研一下", "別人怎麼處理的". Not for: strategic priority decisions (→ think-tank), brainstorming designs, or creating implementation plans.

Multi-role debate (architect, ops, QA, product, UX, finance) for strategic decisions. Use when: "I want perspectives from different roles", "let's debate this", "tradeoff analysis", "which should we prioritize", "blast radius of X", "scope decision", "hear from architect/ops/product", "what's the impact?", "要辯論一下", "多角度分析", "幫我評估利弊". Not for: pure tech selection (→ survey), open brainstorming, or full delegation (→ ceo-agent).

Audit or refresh Autopilot harness capability state. Use when checking whether Codex, Claude Code, agy, Grok, MiniMax, Copilot CLI, or another harness currently supports skills, hooks, agents, status lines, headless dispatch, DI, gating, or runner roles; when platform facts may be stale; or before expanding cross-harness integrations.

Compare two existing implementations and find every difference. Use when: "compare X with Y", "check X against Y", "verify X matches Y", "flag anything missing", "比對 X 和 Y", "檢查有沒有漏掉", "驗證是否一致", feature parity review between old and new systems, an existing spec against an implemented target, or cross-system completeness checks. Not for future/unimplemented plan readiness, architecture-plan critique, debugging a single discrepancy, or writing comparison tests.

Scaffold a consuming project's autopilot config (`.claude/*-config.md` DI) from its detected reality — the bridge from "fresh repo" to "autopilot-calibrated repo". Use when: "onboard this repo to autopilot", "set up the .claude config", "scaffold autopilot for this project", "calibrate autopilot to this codebase", "把這個專案接上 autopilot", "建立 .claude 設定", "幫這個 repo 接 autopilot". Detects package manager / commands / coverage thresholds / doc convention / workspace layout, scaffolds the config set with autopilot-only (ecosystem-standalone) chains, then enriches the judgment-heavy configs (skill-routing, doc-drift domains, security surfaces) by reading the repo. Not for: bootstrapping project-tracking docs from a plan (→ project-lifecycle), authoring a plan (→ references/plan-template.md), or running quality gates (→ quality-pipeline).

Recommend what to work on next by scanning all projects, plans, backlog, and proposals. Use when: "/next", "what should I work on", "what's the highest priority", "what to pick up now", "I have N hours — what fits?", "下一步做什麼", "最高優先是什麼", "現在該做什麼", or right after archiving a completed project. Not for: creating plans, starting implementation, or running quality checks.

Evidence-first debugging for correctness issues. Invoke when diagnosing bugs, crashes, logic errors, data corruption, connectivity problems, intermittent failures (incl. flaky tests with environment divergence), or 'works on my machine' issues. For performance issues (slow / high latency / CPU / memory), use the profiling skill instead. For perf regressions with known change attribution, profiling is still primary — measure first, don't guess from the deploy diff.

Audit docs against actual code and find drift (wrong / stale / missing claims), report-only. Use when: "do the docs match the code", "check for doc drift", "audit docs vs code", "did this change make docs stale", "doc-sync", "文件跟 code 同步嗎", "查文件有沒有過時", "對照文件與實作", post-merge doc verification, periodic doc accuracy sweep. Two layers: a deterministic gate (reliable — links/fences/version/CLI-surface/roadmap; the real stopping condition, gate-able in CI) + an LLM sweep for discovery (scoped per-diff / full whole-repo; non-deterministic, never loop to zero). Not for: WRITING docs from scratch, fixing a single known typo, or build/test correctness (→ quality-pipeline). Report-only — surfaces findings; you triage + fix; mechanizable findings get demoted into the deterministic gate.

Write a standardized mid-work handoff doc for session continuation, and resume from one. Triggers: "寫 handoff", "寫 handover", "ctx 太滿", "context 快滿", "clear session 後繼續", "write a handoff", "hand off to the next session", "接手上個 session", "resume from handoff". Not-for: end-of-project closing (→ finish-flow), machine snapshot on /clear (→ the session-handoff hook, enable via ~/.autopilot/config.json handoff_inject), compaction recovery (→ state-checkpoint hook).

Testing strategy and baseline management — test pyramid placement, baseline守則, failure investigation funnel, regression scoping, flaky-test systemic handling. Invoke when validating changes or designing test approach. Not for: TDD red-green-refactor coding cycle (→ superpowers:test-driven-development if installed; this skill is orthogonal — see Coexistence section), specific test debugging (→ debug), or quality gate (→ quality-pipeline).

Distill your recurring procedures and corrections from local conversation history into personal custom skills, routed into YOUR own skill dirs — never into autopilot. Use when: "/distill", "what do I keep redoing", "turn my repeated workflow into a skill", "提煉我的重複流程", "把我反覆做的變成 skill", "我一直在重複做什麼", "把剛做完的專案蒸餾成 skill", "趁熱把這套流程收下來", "這個專案的方法論值得留", "distill this project/session". Not for: writing autopilot's own skills, project planning (→ dev-flow), or git-history productivity metrics (→ retro, the commit-history sibling).

Hegelian dialectic (Thesis → Antithesis → Synthesis) for irreversible or high-stakes decisions with genuine stalemate — NOT a "better think-tank", a different tool for a different situation. Use when: "this decision is hard to reverse", "both sides have real merit", "genuinely torn between X and Y", "irreversible decision", "real dilemma, not just tradeoffs", "need structured dialectic", "這個決定反悔成本很高", "兩邊都有道理", "在 X 和 Y 之間拉扯", "辯證一下", "不可逆決策", "逃不掉的 tradeoff". Always reach for `think-tank` first; escalate here only when think-tank's Decision Brief shows LOW consensus AND the decision is costly to reverse. Not for: everyday scope decisions (→ think-tank), pure tech selection (→ survey), reversible tasks where you can iterate, or reflexive "let's debate this" requests.

Pre-code Socratic design exploration — when the options don't exist yet. Use when: "I haven't figured out the approach", "help me think through X", "what should this even be / 還沒想清楚", "explore how to do X before we build", "我不確定要怎麼做", "尚未成形 / 沒有方向". DISCOVERS the options, surfaces 2-3 genuinely different approaches, and GATES implementation until you approve a design. Not for: deciding between options that ALREADY exist on the table (→ think-tank), external best-practice research (→ survey), an approach you've already chosen (→ references/plan-template.md to author the plan, then dev-flow), or a quick known fix (→ dev-flow).

Save a hard-won lesson or surprising fix so future sessions benefit. Use when: "save this to knowledge", "record this", "remember for next time", "/learn", "記下來", "存到 knowledge", "記住這個", user shares a gotcha or non-obvious fix, solution took 2+ attempts, environment-specific workaround discovered. Also handles: "knowledge health audit", "check MEMORY.md size", stale knowledge cleanup. Not for: active debugging, writing tests, or project status updates.

Evidence-first performance profiling — tool selection, methodology, interpretation. Use BEFORE guessing at performance issues. Tools first, logs second, code last. Covers: slow queries, high latency, CPU spikes, memory growth, throughput regressions (including 'got slower after deploy' — measure before assuming the deploy diff is the cause). Not for: crashes / logic errors (→ debug), tests being slow as test design issue (→ test-strategy).

Team allocation and dependency-aware parallelization — decide when to use teams, select roles, dispatch parallel work safely. Invoke for L-size tasks to evaluate whether team dispatch reduces wall-clock time.