--- name: cf:specs description: "Turn a substantial or risk-elevated feature idea into a bounded, reviewed plan and executable flat task files, with human decisions at scope, findings, and completion. Use when work needs durable coordination or is not eligible for direct work; skip only when a change is clear, isolated, reversible, routine, and likely limited to one or two files." user-invocable: true argument-hint: "" metadata: author: haposoft version: "3.0.0" --- # Specs — process-first planning Quality comes from a short decision process; honesty comes from evidence. Keep the human at exactly three gates; durable output is editable Markdown. ## Step 0 — classify risk before routing Classify material risk before choosing a workflow; user wording never lowers an observed floor. - `critical`: auth/secrets/privacy; destructive/irreversible work or possible data loss/corruption; money/privilege/safety; production-state mutation. - `elevated`: cross-component contracts, compatibility, concurrency, external integration, or installed/runtime behavior. - `routine`: only when no elevated or critical signal applies. Work directly only when the cause and change are clear, isolated, reversible, `routine`, and likely limited to one or two files. Route unresolved user-owned observable choices to C1/C2 first; after they settle, route material competing technical designs through Brainstorm. Split three or more independent subsystems; otherwise use one Specs packet for any material work that does not qualify for direct work or Brainstorm-only exploration. A subsystem is independent only when its outcome, boundary, and verification/deployment path can move through the lifecycle separately. Do not create a plan merely because documentation mentions this skill; the user must invoke Specs or ask for durable planning. ## The three human gates | Gate | Timing | Human decision | |---|---|---| | C1 — Scope | before writing the plan | EXPAND, KEEP, or CUT the proposed minimum change | | C2 — Findings | after adversarial review | accept, reject, or revise each deduplicated finding | | C3 — Done | after execution | accept completion from current command output and receipts | Ask once at each gate. Do not ask for routine implementation choices between them. New evidence that invalidates scope returns to C1. ## Primary output layout Create one direct child of the repository's `specs/` directory: ```text specs// ├── plan.md ├── task-01-.md └── task-02-.md ``` Task files are flat beside `plan.md`. Do not create a nested task directory. Use safe lowercase slugs and read [`references/templates.md`](references/templates.md) first. ## Flow ### 1. Challenge scope, then open C1 Inspect the repository before drafting. Answer three questions with current `path:line` evidence: 1. What already exists and can be reused? 2. What is the smallest change set that delivers the requested outcome? 3. What signals expansion: more than eight touched files, more than two new services/classes, or three or more independently deliverable subsystems? Present the answers and ask the user to EXPAND, KEEP, or CUT. Record the chosen scope and explicit exclusions in `plan.md`. ### 2. Write the plan and task packets Keep `plan.md` as the short index: C1 decision, EARS criteria, exclusions, and task table. Give every criterion a stable ID mapped by task rows. For a Specs route, persist the canonical `## Coverage profile` from `references/templates.md`: one `CP-NN` row per observable outcome; tasks reference IDs instead of copying it. Each task has one usable outcome, normally at most about five owned files, explicit dependencies, measurable acceptance, and a runnable verification command; no preparation-only tasks. Author exact state: `pending` means semantically ready for the dependency-aware queue. Use `blocked` while a C1/C2 decision, accepted finding, or `UNKNOWN` closure remains open. Dependencies alone do not change `pending`; the resolver queues them. Promote only after current evidence closes every non-dependency blocker; keep pre-execution Receipts empty. `Required proof` is planned, not executed evidence. Unknown command/caller/environment reachability blocks `pending`; known unrun proof does not. Missing, failed, or unavailable required evidence blocks `done`/C3; levels never promote. Rederive affected CP rows after accepted scope, outcome, criteria, ownership, dependency, risk, or proof deltas, before status. Source/static checks prove the written contract, not live-model adherence. ### 3. Review adversarially, then open C2 Read [`references/review.md`](references/review.md) and run its fresh-context review. Require reproducible `path:line` plus a failure scenario; deduplicate, rank, and cap at 15. Ask the user to accept, reject, or revise each finding; apply only accepted decisions. After every edit, sweep all packet files and rederive every status. Dependencies use exact flat task basenames. Stop after two paper rounds; later findings require runtime evidence. ### 4. Execute one task at a time An implementation workflow owns execution. It selects one unblocked task with `Status: pending`, changes its single `Status:` field to `in_progress`, respects owned paths, runs the task's verification, and appends a real inline `## Receipt` only after the command finishes. Parallel work is allowed only for tasks with disjoint write ownership and satisfied dependencies. One file has one writer in a wave. The controller is the sole writer of task status and receipts. ### 5. Prove completion, then open C3 A task may say `Status: done` only when its inline receipt contains the exact command, `Exit: 0`, `Verification: PASS`, runtime-bound Base and Head values, and a non-empty fenced command-output block. Placeholder or remembered output is not evidence. Show the user the current evidence and unresolved limitations. The user decides at C3 whether the feature is done. A passing command proves only what it ran; it does not invent product approval, review independence, or runtime coverage. ## Ten operating laws 1. **A1 — Scope once.** Raise scope pressure at C1; reopen only with new proof. 2. **A2 — Small packets.** One task, one outcome, one owner, one proof command. 3. **A3 — One fact, one home.** Reference decisions; never copy mutable lists. 4. **A4 — Files are state.** Markdown is canonical; indexes are disposable. 5. **B1 — Cite the repository.** Use current `path:line` or `[UNVERIFIED]`. 6. **B2 — Review fresh.** The author does not impersonate the reviewer. 7. **B3 — Sweep edits.** Reconcile names, decisions, CP rows, ownership, dependencies, criteria, and proof. 8. **B4 — Stop paper churn.** Two pre-code rounds; later claims need runtime evidence. 9. **C1 — Derive state.** Status follows fresh blocker and proof evidence. 10. **C2 — Rules pay rent.** Require a cited failure; reassess for model changes. ## Machine boundary The shared workflow resolver recognizes only a regular `plan.md` plus one or more regular flat `task-*.md` files in one direct feature directory, and projects exact `Status:` values without inferring blockers from prose. The Stop gate re-reads every done Receipt, binding Base and Head to the live runtime until the task file is committed, unchanged, and the tree outside the specs root is clean. It detects drift between a receipt and the tree, not invention: a valid pair costs one command and no verification run, so a receipt for a command that never ran satisfies it in either mode. C3 is where a human weighs the evidence. These checks are a final safety net, not a substitute for C1-C3 judgment. Do not add a new schema, approval field, readiness bit, or review state to make the Markdown look more authoritative. ## Legacy compatibility Existing features that contain `spec.json`, nested task files, or separate receipt files stay on the installed legacy adapters. Do not migrate or rewrite them while authoring an unrelated v3 feature. New Specs output always uses the flat layout above and never requires the legacy kernel. ## Maintenance Keep this entrypoint focused on routing and invariants. Put templates and examples in `references/templates.md`; put review mechanics in `references/review.md`. Before adding guidance, identify the observed failure it prevents and remove duplicated wording elsewhere.