wdi-method 0.6.18 → 0.6.24

package/kit/skills/wdi-build/SKILL.md

@@ -1,393 +1,404 @@
----
-name: wdi-build
-description: Use at G5 Release — one spec from open to closed in one supervised run. Opens the spec, hands the owner to-spec and to-tickets, ships every ticket to a green PR through a five-step pipeline, then closes the spec. One invocation, not four.
----
-
-# WDI Build
-
-One unit of work, one invocation. A spec used to need four calls — open it, produce the contract, ship each
-ticket, close it — and three of those were bookkeeping. They are all in here now, because a unit of work that
-needs four invocations gets three of them skipped.
-
-**The engine layer below this gate is not BMad's, and it MUST be this one.** `to-spec`, `to-tickets`,
-and `implement` produce the contract, the tickets, and the code; `tdd` and `code-review` do the work
-inside them. All five are `mattpocock/skills`, installed **in this repo** — the installer refuses
-without them. Thirteen BMad skills are **retired** at this gate and MUST NOT be used anywhere:
-`bmad-spec`, `bmad-build`, `bmad-build-auto`, `bmad-code-review`, `bmad-retrospective`,
-`bmad-agent-dev`, `bmad-create-epics-and-stories`, `bmad-create-story`, `bmad-dev-story`,
-`bmad-dev-auto`, `bmad-quick-dev`, `bmad-sprint-planning`, `bmad-sprint-status`. That is enforced, not
-requested: `install` and `update` lock each one out of model invocation and add a `Skill()` deny rule.
-`bmad-skill-register.md` carries the list and the criterion behind it.
-
-**All five engines are INVOKED, by this skill, through the Skill tool.** Upstream ships `to-spec`,
-`to-tickets` and `implement` with `disable-model-invocation: true`; `wdi-method` strips it from the
-copies this repo owns, so there is no command to hand to the owner and no reading-and-following to do.
-Where an engine needs a decision — the seams, the `to-tickets` quiz — that decision is made before the
-invocation and passed IN it, because an engine that stops to ask inside an unattended run is a run
-that stalls with nobody there to answer.
-
-**If an engine will not invoke, stop and say why.** `disable-model-invocation` back in its frontmatter
-is what `npx skills update` does, and the fix is one command: `npx wdi-method engines --fix`, or the
-`wdi-init` / `wdi-upgrade` skill. MUST NOT work around it by pasting the engine's process inline: the
-engine's rules are its own, and a paraphrase of them is not the engine.
-
-**Under an active mandate the owner's part is `wdi-autopilot`'s.** A `DEC-` of `type: mandate` at
-`status: accepted`, unexpired, moves **every** "the owner runs" and "the owner decides" in this skill to the
-coordinator — G5's checklist and Step 2's *stop and reach the owner* included, which reach the coordinator and
-not a person. Three of them change **shape** as well as owner: the engines run by **read-and-follow** — a builder brief
-that names the engine's `SKILL.md` path and carries out its process — or from a copy in the repo, and the ledger
-names which; the seams, the `to-tickets` quiz, and § When the code turns out to be right are decided by the
-coordinator and written to the ledger, one row each; and whatever the mandate lists as `parked` still stops,
-reported for the owner rather than decided. One thing changes **shape** rather than owner: a mandate is one
-unit of work and reaches `main` through **one PR**, so Step 4 commits the ticket to the run branch instead of
-opening a PR per ticket, and Step 5's push and CI watch happen at every spec close on that one PR — the
-checklist, the head-SHA rule, and the caps are unchanged. Nothing else in this skill relaxes. `wdi-autopilot` owns the
-mandate, the ledger, and the loop; this skill MUST NOT restate them.
-
-**No orchestration tool is required, and this skill MUST NOT name one.** How the work is carried out is the
-session's own arrangement: an orchestration skill, this CLI's native subagent tooling, or the coordinator doing
-a step itself. Which CLI, which model, and which effort staff a step come from the local Agent Rules, and MUST
-NOT be restated here. Every rule below holds whatever the mechanism — they are about the artifact and who
-judges it, not about how an agent is launched.
-
-Two roles exist regardless. The **coordinator** — this session — holds every judge call, the registries, and
-every remote git action. The **builder** of a step holds the spec and the code.
-
-When the coordinator is also the builder, the separation the pipeline rests on is absent. That MUST be reported
-as absent rather than worked around: a step judged by whoever wrote it is a self-report. One consequence is
-hard — see Step 3: at `risk_accepted: low` the panel requires reviewers who are not the builder, and a session
-that cannot provide them is **blocked**, not excused.
-
-## Precondition, and the one that is easy to miss
-
-| Check | When it fails |
-|---|---|
-| Every component this spec touches has passed G4, **or** sits at `mode: catalog` | Route to `wdi-component`. `spec-after-g4` checks it, and `catalog` skipping G4 is by design, not an exception |
-| An isolated worktree | Isolate first. MUST NOT run in a shared checkout |
-| Every `prd` slug names a real `.what/_prd/<initiative>/` folder | A spec without a promise covering it is a spec nobody agreed to (`spec-names-release-prd`) |
-
-The repo commits straight to `main` and opens a PR only when asked. **Invoking this skill is that ask**, for
-this spec only; it MUST NOT be read as standing permission for the next change.
-
-## Phase 1 — Open the spec
-
-Record it in `.control/registry/specs.yaml`. The frame opens here; the `tickets` rows land in Phase 2, because
-that is where they are born.
-
-| Field | Rule |
-|---|---|
-| `id` | `SPEC-<N>`, monotonic, never reused. A `W<N>` id in a frozen record is a **retired alias** and MUST NOT be rewritten |
-| `release` | MUST be stated. The release↔spec cadence is situational and MUST NOT be inferred from numbering (`spec-names-release-prd`) |
-| `prd` | MUST be stated: which initiative PRDs this spec delivers against. MUST NOT be derived from `release` |
-| `fr` | The `FR` this spec satisfies. Ideally one — an `FR` is human-testable from birth |
-| `size` | `S` · `M` · `L`. MAY be raised mid-flight; MUST NOT be lowered |
-| `depends_on` | At **spec** level. A spec declaring none runs in parallel with its neighbours |
-| `spec_folder` | One per spec, not one per spec × component |
-| `tickets` | Flat, one row per ticket: `id` · `component` · `satisfies: [UC]` · `blocked_by` · `touches` · test names |
-| a ticket `id` | `<spec-id>-<NN>` — `SPEC-3-01`. The engine numbers its files from `01` per feature, which is unique only inside one spec; the RTM needs a key that is unique across the corpus |
-
-**`tickets` is an index, not a store.** The ticket's prose lives where the tracker put it; the row carries only
-what RTM and the validators read. And ticket **status** MUST NOT be copied here — it is read from the ticket
-itself when `.control/generated/` regenerates (`ticket-status-one-home`). Two homes for one fact is how registries start lying.
-
-`ticket-status-one-home` finds the file at `{spec_folder}/issues/<NN>-*.md`, from the number at the tail of the id, and reads its
-status from either a `**Status:**` body line — what the engine writes, because a ticket file is a tracker
-payload and trackers do not read YAML — or `status:` in frontmatter.
-
-The `epics` nesting is **repealed**. A ticket names its `component` directly; an intermediate level that only
-grouped rows bought nothing and cost a lookup.
-
-Size does not choose which gates are active — that is `mode`'s job. It decides two things: session merging
-(`S` merges G4 and G5 into one 20-minute session) and whether `SPEC.md` is written at all.
-
-## Phase 2 — The contract, and the tickets
-
-Two engines, and **this skill invokes both.** Invoke, then verify and land.
-
-| Size | What this skill invokes | What lands |
-|---|---|---|
-| `M` · `L` | `to-spec`, then `to-tickets` | `SPEC.md` in `spec_folder` · ticket files · the `tickets` index rows |
-| `S` | `to-tickets` only | ticket files · the `tickets` index rows. **No `SPEC.md`** |
-
-Each invocation carries what the engine would otherwise stop to ask: the tracker is already configured
-in `docs/agents/issue-tracker.md`, the seams and the testing decisions are settled above, and the
-`to-tickets` quiz — granularity and blocking edges — is answered from the size table in
-`delivery-flow-guide.md` plus `depends_on` and `touches`. Under a mandate every one of those answers is
-also a ledger row.
-
-At `S` the tickets **are** the contract. `to-tickets` accepts a conversation directly, so a middle document
-buys nothing there. From `M` up it is written first, because two things have to be settled **before** tickets
-are cut and neither survives being decided afterwards: the **seams** the feature will be tested at — fewest
-possible, highest possible, agreed with the owner — and the **testing decisions**, which say what a good test
-is here and name the prior art.
-
-### What you verify before landing anything
-
-- **Nothing new.** `SPEC.md` is a **projection** of `.what/` + `.how/`. It MAY restate a promise in its own
-  words — a machine contract is more useful self-contained — but **every restatement MUST refer to the live
-  corpus document it came from.** A reference that does not resolve is not a wording problem: the gap is
-  upstream. Route to `wdi-component` or `wdi-blueprint`, and do not let the contract invent it.
-- **Every user story resolves to an id.** A user story with no `FR` or `UC` behind it is a **new promise**,
-  and a new promise is `wdi-product`'s, never a spec's.
-- **Every ticket names what it satisfies.** No `satisfies`, no landing: without it the chain
-  `FR → UC → ticket → test` breaks and RTM cannot say which promise went green. This is the one field
-  `to-tickets` does not ask for on its own, so it is the one most likely to be missing.
-- **Every ticket is vertical.** A slice of one layer is not a ticket. The exception is a wide refactor,
-  sequenced expand → migrate in batches → contract; `delivery-flow-guide.md` owns that rule.
-- **Ticket files land under `spec_folder`, and `spec_folder` is `.scratch/<spec-id>-<slug>/`.** Their
-  **shape** is the engine's — one file per ticket, numbered in dependency order, blocking edges declared
-  — and the location is ours, written in `docs/agents/issue-tracker.md` where the engines read it. The
-  id in front of the slug is not decoration: four live repos wrote that leaf four different ways, one of
-  them all four inside a single repo, and a folder nothing can trace back to a row in `specs.yaml` is
-  how a spec goes missing. A ticket at the repo root, or under `docs/`, or in a `.scratch/` directory
-  with no row in `specs.yaml`, is drift.
-
-`SPEC.md` and ticket files **are not read by humans.** Both are machine contracts, and no review burden MAY be
-moved onto them. `wdi-review` MAY still be dispatched over the contract; its trace lands on the spec in
-`specs.yaml` — and where there is no `SPEC.md`, one trace covers the **ticket set as one artifact**, never one
-review per ticket.
-
-## Phase 3 — Ship each ticket
-
-Work the **frontier**: the tickets whose blockers are all closed. A ticket whose `blocked_by` is not yet
-satisfied MUST NOT be started, however ready it looks.
-
-| # | Step | Engine | Exit condition |
-|---|---|---|---|
-| 1 | Encode | invoke `tdd` | **Failing tests exist that encode this ticket's acceptance criteria** |
-| 2 | Build | invoke `implement` — it uses `tdd` at the agreed seams and calls `code-review` itself | Those tests green, typecheck clean, full suite green once |
-| 3 | Panel | invoke `code-review`, as a **separate** dispatch by a **different** agent | Panel adjudicated, zero unresolved must-fix |
-| 4 | Publish | — | Branch pushed, PR open, ticket-closing checklist answered |
-| 5 | CI | — | All checks conclude green on the pushed head SHA |
-
-**Step 1 is the change that matters most.** It used to be a plan, judged from a frontmatter field a builder
-wrote about itself. Now it is a **failing test suite** — the acceptance criteria, encoded, and demonstrably
-red. That is evidence rather than a claim, and it is what `ticket-has-test` and G5's ★2 have always been asking for:
-*acceptance criteria proven by a test, not by an agent's statement.* Under TDD the test exists **before** the
-code, so the proof is not retrofitted.
-
-### Engine rules
-
-- **MUST judge Steps 1 and 2 from the test suite**, not from a report and not from a status field. Red at the
-  right assertions closes Step 1; green with a clean typecheck closes Step 2. A builder's chat report MUST NOT
-  settle either.
-- Ticket **status** is read from the ticket itself. It MUST NOT be copied into `specs.yaml` (`ticket-status-one-home`).
-- Every step MUST start from a **fresh context**, and a fix round MUST NOT go back to whoever produced the
-  code. The ticket carries everything the next builder needs, and inherited context is how a step stops judging
-  the artifact on its own merits. Where one session runs consecutive steps itself, it MUST re-read the artifact
-  rather than trust what it remembers writing.
-- A step handed to a separate agent MUST be launched so it cannot sit waiting for input nobody will give, and
-  its start MUST be confirmed from observed activity — never from a readiness match alone.
-- A step that fails because the agent could not spawn what it needed is a **capability** failure, not a ticket
-  failure. MUST retry it somewhere that can, and MUST NOT record it against the ticket.
-- MUST NOT reorder or drop a step. A step with nothing to do MUST be reported as such, not skipped silently.
-- A ticket returned by the panel MUST have its acceptance criteria amended and its status reset before it is
-  picked up again — see Step 3.
-
-### What every builder brief MUST carry
-
-Three rules this corpus adds, and `/implement` knows none of them. All three MUST reach whoever writes code in
-this spec — in the dispatch when a step is dispatched, and in the session's own working instructions when it
-is not.
-
-- **Debugging is conditional, never a phase.** When a test or build fails and the cause is not known, the
-  builder MUST run `wdi-systematic-debugging` before proposing any fix. A third failed fix attempt is the signal
-  to escalate, not to try a fourth.
-- **The corpus is not the builder's to change.** A builder MUST NOT edit `.what/`, `.how/`, or an `applied`
-  `DEC-`. A deviation from the SDD or an `AD-N` is **reported**, and it becomes a `DEC-` through
-  `wdi-decision` — never absorbed as a code patch.
-- **Verification is run, not assumed.** The commands are this product's, and they live in
-  `.constitution/project/codebase-stack-guide.md` — build, test, and whatever the front end needs, each with
-  the directory it runs from. A skill MUST NOT carry one product's build line. A green registry workflow
-  MUST NOT be reported as proof the code compiles; they answer different questions.
-
-### Step 1 — encode the acceptance criteria as failing tests
-
-- The tests MUST be written **at the seams the spec agreed** — existing seams preferred, highest possible,
-  fewest possible. Inventing a new seam here, after the agreement, is a finding: it means Step 1 is redesigning
-  what Phase 2 settled.
-- Every acceptance criterion MUST have at least one test. A criterion no test can express is not an acceptance
-  criterion — it goes back to the ticket, or to the owner as an intent gap.
-- The tests MUST be **seen red, at the right assertions.** A test that passes before the code exists is testing
-  nothing, and a suite that fails for the wrong reason — an import error, a missing fixture — has not encoded
-  anything yet.
-- A test MUST NOT assert a literal where the behaviour is what matters. The panel treats that as a test that
-  cannot fail, and returns it.
-
-### Step 2 — build
-
-- The owner runs `/implement`, and it MUST be given the ticket and the three brief rules above.
-- It commits to the current branch and **never pushes**. That is its own behaviour and it is what we want; the
-  coordinator is the hand that pushes.
-- **`/implement` calls `/code-review` itself, and that call does NOT satisfy Step 3.** It is the builder
-  reviewing its own work — self-review by construction. Step 3 stands as a separate dispatch regardless of what
-  ran inside Step 2.
-- The full suite MUST be run green once at the end, not only the tests this ticket touched. A ticket that
-  passes its own tests and breaks a neighbour's has not finished.
-- Where the work turns out to need something the ticket does not authorise, it stops and reaches the owner with
-  the question verbatim. The builder MUST NOT widen its own scope.
-
-### Step 3 — panel, then judge
-
-Panel composition follows `risk_accepted`: at `low` a two-reviewer panel is **required** on the code; at
-`medium` and `high` it is available and SHOULD be used when the diff touches money, personal data, or a third
-party. The local Agent Rules govern which CLIs and models staff the panel. **A reviewer MUST be a different
-agent from the builder** — the builder's own review layers, `/implement`'s internal `/code-review` included,
-are self-review by construction and never satisfy the panel. This is the one separation in the pipeline that
-MUST NOT be collapsed: where the session cannot provide it and `risk_accepted` is `low`, the ticket is blocked
-and the owner MUST be told, because the field they set is what makes the panel required.
-
-`code-review` reviews along two axes — **Standards**, against this repo's documented conventions, and
-**Spec**, against what the ticket asked for. Both MUST run. Reporting one axis as the panel is reporting half
-a review.
-
-- MUST adjudicate every contested finding by reading the cited lines. Votes MUST NOT settle a finding. A finding
-  neither reviewer can locate in the diff is dismissed with that reason stated.
-
-**MUST return to Step 2:**
-
-- Breaks a ticket's acceptance criterion, or contradicts the contract, the SDD, an `AD-N`, or an `applied` `DEC-`
-- Wrong behaviour, crash, or data loss reachable from the running app
-- Corpus drift: an `LC` touched but not registered, a contract changed in code but not in `02-contracts/`, a
-  screen added without its `01-ux/` entry
-- An enum value rendered straight to the screen instead of read off its label map
-- Payment or private data reaching a tracked file
-- A weakened guard, or a test that cannot fail — including one asserting a literal instead of the behaviour it
-  claims to cover
-
-**MUST record as follow-up and MUST NOT return to Step 2:** style or naming with no behaviour delta · a
-refactor outside this ticket's scope · a pre-existing defect this ticket did not touch · a speculative risk
-with no reachable path.
-
-A must-fix MUST return as a **ticket amendment**, never a chat instruction: amend what the ticket asks for,
-note the finding that caused it, reset the ticket's status to `ready-for-agent`, then pick it up again. **What
-a ticket `satisfies` is not amendable here** — that is the promise, and a must-fix rooted in the promise is an
-intent gap that goes to the owner through `wdi-product`. Changing the `FR` a ticket serves in order to make the
-code pass is how a corpus starts agreeing with whatever was built.
-
-Cap: **2 return trips.** MUST re-run the whole panel after each fix round — a fix introduces defects. On hitting
-the cap MUST escalate and MUST NOT open a PR carrying an unresolved must-fix.
-
-### When the code turns out to be right and the document wrong
-
-This is normal during G5 and it is **not drift**. Building a thing is how you find out what the thing is.
-
-The builder still MUST NOT edit `.what/`, `.how/`, or an `applied` `DEC-` to make its own code fit — that is
-the corpus learning to agree with whatever was built, and it is a different failure. What happens instead:
-
-1. **Say it once.** Name what the code does, which promise it contradicts, and what that costs. One place,
-   one time. An `AD-N` is the one contradiction that **stops** — `decision-guide.md` owns it.
-2. **The owner decides.** If they adopt the code, that survey is spent.
-3. **The owning skill edits**, in the present tense, as if the design had always said this — `wdi-product`
-   for an `FR`, `wdi-component` for behaviour or design, `wdi-blueprint` for a cross-component rule.
-
-What MUST NOT happen after step 2: raising the same conflict again in a later pass, opening an `OQ-` for it,
-dispatching a review over it, or writing anywhere in the corpus that the change arrived late. The commit is
-that record. `corpus-guide.md` § The corpus is written in the present tense is the binding rule.
-
-### Step 4 — ticket-closing checklist, then push and PR
-
-The checklist is **three items**, and it MUST be answered before the PR opens:
-
-1. Something the next person needs to know? → **into the document that carries it**, and that is almost
-   always where it ends. It reaches `wdi-decision` only when no design document has a home for it —
-   `decision-guide.md` § A decision's first home. A ticket contradicting an `AD-N` **stops** rather than
-   closing; that is the one case where recording is mandatory.
-2. A trap for the next agent? → recorded where the next agent will read it.
-3. Test names matching what `specs.yaml` records?
-
-The five items that left this list moved to Phase 4, where the information actually exists.
-
-- MUST run the repository's commit/push audit before `git push`: refuse the forbidden paths, run the guard test,
-  fix content on failure. A failing guard is a finding about the content — MUST NOT weaken the guard or the test.
-- MUST NOT push to `main`/`master`, MUST NOT force-push, MUST NOT merge.
-- The coordinator MUST be the hand that pushes and opens the PR.
-
-### Step 5 — watch CI, then judge
-
-- MUST wait for every check to conclude, then confirm the checks belong to the **pushed head SHA**. A green
-  report from a stale run is a false report.
-- `korpus.yml` validates the corpus, not the code. Build and test evidence comes from Step 2's own runs.
-- Classify each failure before acting: a defect from this change → Step 1 if the test was missing, Step 2 if
-  the code was wrong, with `wdi-systematic-debugging` when
-  the cause is unknown · infrastructure or flake → re-run **once**, and MUST NOT patch code to mask it; a second
-  identical failure is a defect · a guard failure → fix the content.
-- Cap: 2 return trips. On hitting the cap MUST report red honestly rather than keep pushing.
-
-### Parallel tickets
-
-Tickets with no blocking edge between them MAY run at once — that is what the frontier is — but four
-conditions MUST hold: each concurrent builder in its own worktree; `parallel-tickets-blocked` green for every pair released together;
-the first ticket that establishes a component's shape already closed, so later tickets inherit its code map;
-and no shared registry write in flight.
-
-**A wide refactor is the exception, and it inverts the rule.** Its batches MUST run in sequence, not in
-parallel, because each keeps CI green only while the expand still stands. Where even a batch cannot stay green
-alone, they share an integration branch and green is promised only at the final integrate-and-verify ticket.
-
-The pattern that MUST be preferred: run the biggest blocker alone first, let its shape decisions land, then fan
-out.
-
-## Phase 4 — Close the spec
-
-**Six steps.** Run in this order and stop at the first failure.
-
-1. **Registry catch-up.** Every `LC` the spec's design named is registered in `components.yaml`, and every
-   `touches` value resolves — `lc-registered`. This is the moment those questions have answers.
-2. **Inventories refreshed from code.** Run `.constitution/method/scripts/inventory.py`. The plan-versus-reality
-   difference is reported as a finding; it MUST NOT be patched into agreement by hand.
-3. **Structure maps refreshed** through `wdi-init` intent `structure`, if a base folder was born or removed or a
-   key file moved.
-4. **Distillation.** Every applicable row of the ownership table in `corpus-guide.md` has been landed by its
-   owner. Anything durable in the spec folder leaves it now, or dies with it — **the ticket files included.**
-   Their prose is working output; what survives is the index in `specs.yaml` and whatever the checklist routed.
-5. **RTM green.** Every traceability row for this spec is closed. New risks are in the risk register with an
-   owner.
-6. Mark the spec `status: closed` in `specs.yaml`.
-
-The retrospective step is **repealed**, and `RTR-` with it. It was the only thing size `L` decided, and the
-only thing `V19` checked.
-
-- You MUST NOT close a spec with an open distillation row. The whole point of an ephemeral working layer is that
-  durable truth leaves it first.
-- You MUST NOT reopen a closed spec to add scope. Scope arriving late opens a new spec, or goes through
-  `wdi-decision` if it invalidates what is already planned.
-- A change invalidating more than 30% of a spec's tickets MUST go through `wdi-decision`, not a patch to
-  `specs.yaml`.
-- Fast Path work is recorded as `fastpath` with no gates. If an `FR` turns out to be touched, the work MUST stop
-  and be raised to a spec `S`.
-
-## Red Flags — STOP
-
-- "One reviewer approved, good enough"
-- Judging Step 1 or Step 2 from a report instead of from the test suite
-- **Counting `/implement`'s own `/code-review` as the panel** — that is the builder reviewing itself
-- Writing `SPEC.md` at size `S`, or skipping it at `M` and up
-- Landing a ticket with no `satisfies` — the RTM chain breaks silently and nothing else notices
-- Amending what a ticket `satisfies` to make a must-fix go away
-- A ticket that slices one layer instead of cutting through all of them, outside a wide refactor
-- Running a wide refactor's batches in parallel
-- Claiming this skill invoked `to-spec`, `to-tickets`, or `implement` — it cannot; the owner runs them, or under
-  a mandate a builder reads and follows them, and the ledger says so
-- A builder editing `.what/`, `.how/`, or an `applied` `DEC-` to make its code fit
-- Fixing a failing test without knowing why it failed
-- Opening a PR with an unresolved must-fix, or before the ticket-closing checklist is answered
-- Editing a guard, a test, or an assertion to turn something green
-- Reporting green without checking the head SHA, or reading green `korpus.yml` as a passing build
-- Leaving a dispatched step able to stall on a question nobody is there to answer
-- Naming an orchestration tool as this skill's requirement, or restating a CLI/model mapping the Agent Rules own
-- Closing the spec without the registry catch-up in Phase 4 — that is where five checklist items now live
-- Letting the contract state something `.what/` and `.how/` do not, or restate it without a reference
-
-**Each of these means: stop, return to the step or phase that owns it, or escalate to the owner.**
-
-## Output
-
-MUST follow the Agent Rules `Answer Closing` block, carrying these additions:
-
-- **what was done** — spec, `FR` satisfied, every ticket with its branch and PR URL, what the code now does
-- **what blocked it** — every step that looped, every dismissed finding and why, anything left red, and every
-  Phase 4 item that did not pass
-- **what comes next** — every finding recorded as follow-up, every entry in a contract's `deferred` list, every
-  ticket-closing item routed to another skill, and the plan-versus-code inventory differences. A follow-up
-  absent here is a lost finding.
+---
+name: wdi-build
+description: Use at G5 Release — one spec from open to closed in one supervised run. Opens the spec, hands the owner to-spec and to-tickets, ships every ticket to a green PR through a five-step pipeline, then closes the spec. One invocation, not four.
+---
+
+# WDI Build
+
+One unit of work, one invocation. A spec used to need four calls — open it, produce the contract, ship each
+ticket, close it — and three of those were bookkeeping. They are all in here now, because a unit of work that
+needs four invocations gets three of them skipped.
+
+**The engine layer below this gate is not BMad's, and it MUST be this one.** `to-spec`, `to-tickets`,
+and `implement` produce the contract, the tickets, and the code; `tdd` and `code-review` do the work
+inside them. All five are `mattpocock/skills`, installed **in this repo** — the installer refuses
+without them. Thirteen BMad skills are **retired** at this gate and MUST NOT be used anywhere:
+`bmad-spec`, `bmad-build`, `bmad-build-auto`, `bmad-code-review`, `bmad-retrospective`,
+`bmad-agent-dev`, `bmad-create-epics-and-stories`, `bmad-create-story`, `bmad-dev-story`,
+`bmad-dev-auto`, `bmad-quick-dev`, `bmad-sprint-planning`, `bmad-sprint-status`. That is enforced, not
+requested: `install` and `update` lock each one out of model invocation and add a `Skill()` deny rule.
+`bmad-skill-register.md` carries the list and the criterion behind it.
+
+**All five engines are INVOKED, by this skill, through the Skill tool.** Upstream ships `to-spec`,
+`to-tickets` and `implement` with `disable-model-invocation: true`; `wdi-method` strips it from the
+copies this repo owns, so there is no command to hand to the owner and no reading-and-following to do.
+Where an engine needs a decision — the seams, the `to-tickets` quiz — that decision is made before the
+invocation and passed IN it, because an engine that stops to ask inside an unattended run is a run
+that stalls with nobody there to answer.
+
+**If an engine will not invoke, stop and say why.** `disable-model-invocation` back in its frontmatter
+is what `npx skills update` does, and the fix is one command: `npx wdi-method engines --fix`, or the
+`wdi-init` / `wdi-upgrade` skill. MUST NOT work around it by pasting the engine's process inline: the
+engine's rules are its own, and a paraphrase of them is not the engine.
+
+**Under an active mandate the owner's part is `wdi-autopilot`'s.** A `DEC-` of `type: mandate` at
+`status: accepted`, unexpired, moves **every** "the owner runs" and "the owner decides" in this skill to the
+coordinator — G5's checklist and Step 2's *stop and reach the owner* included, which reach the coordinator and
+not a person. Three of them change **shape** as well as owner: the engines run by **read-and-follow** — a builder brief
+that names the engine's `SKILL.md` path and carries out its process — or from a copy in the repo, and the ledger
+names which; the seams, the `to-tickets` quiz, and § When the code turns out to be right are decided by the
+coordinator and written to the ledger, one row each; and whatever the mandate lists as `parked` still stops,
+reported for the owner rather than decided. One thing changes **shape** rather than owner: a mandate is one
+unit of work and reaches the active development branch (`policy.development_branch`, default `main`) through **one PR**, so Step 4 commits the ticket to the run branch instead of
+opening a PR per ticket, and Step 5 splits: the coordinator pushes the run branch at every spec close, but
+**the cloud run happens once, at `wdi-autopilot` § Finish** — every intermediate push starts nothing, and
+what a spec close is judged on until then is the run branch's own full suite, run locally. The checklist,
+the head-SHA rule, and the caps are unchanged. Nothing else in this skill relaxes. `wdi-autopilot` owns the
+mandate, the ledger, and the loop; this skill MUST NOT restate them.
+
+**No orchestration tool is required, and this skill MUST NOT name one.** How the work is carried out is the
+session's own arrangement: an orchestration skill, this CLI's native subagent tooling, or the coordinator doing
+a step itself. Which CLI, which model, and which effort staff a step come from the local Agent Rules, and MUST
+NOT be restated here. Every rule below holds whatever the mechanism — they are about the artifact and who
+judges it, not about how an agent is launched.
+
+Two roles exist regardless. The **coordinator** — this session — holds every judge call, the registries, and
+every remote git action. The **builder** of a step holds the spec and the code.
+
+When the coordinator is also the builder, the separation the pipeline rests on is absent. That MUST be reported
+as absent rather than worked around: a step judged by whoever wrote it is a self-report. One consequence is
+hard — see Step 3: at `risk_accepted: low` the panel requires reviewers who are not the builder, and a session
+that cannot provide them is **blocked**, not excused.
+
+## Precondition, and the one that is easy to miss
+
+| Check | When it fails |
+|---|---|
+| Every component this spec touches has passed G4, **or** sits at `mode: catalog` | Route to `wdi-component`. `spec-after-g4` checks it, and `catalog` skipping G4 is by design, not an exception |
+| An isolated working tree | Isolate first. MUST NOT run in a shared or dirty checkout. Permitted isolation models: an isolated linked worktree (`git worktree add`), or an exclusive primary working tree checked out to the task branch with a clean status dedicated to this task (`.constitution/method/branch-guide.md`). Standing exception: Phase 1 (Open the spec) and Phase 2 (The contract and the tickets) authoring runs directly on `development_branch` (`policy.development_branch`, default `main`) with no code changes; the isolated working tree requirement binds Phase 3 (Ship each ticket) onward |
+| Every `prd` slug names a real `.what/_prd/<initiative>/` folder | A spec without a promise covering it is a spec nobody agreed to (`spec-names-release-prd`) |
+| Development branch verified locally or on remote (`refs/heads/<branch>` or `refs/remotes/origin/<branch>`) | Branch missing. Stop immediately and report to maintainer; MUST NOT guess or silently fall back to `main` (`.constitution/method/branch-guide.md`) |
+
+The repo commits straight to the active development branch (`policy.development_branch`, default `main`) during Phase 1 & 2 authoring and opens a PR only when asked. **Invoking this skill is that ask**, for
+this spec only; it MUST NOT be read as standing permission for the next change.
+
+## Phase 1 — Open the spec
+
+Record it in `.control/registry/specs.yaml`. The frame opens here; the `tickets` rows land in Phase 2, because
+that is where they are born.
+
+| Field | Rule |
+|---|---|
+| `id` | `SPEC-<N>`, monotonic, never reused. A `W<N>` id in a frozen record is a **retired alias** and MUST NOT be rewritten |
+| `release` | MUST be stated. The release↔spec cadence is situational and MUST NOT be inferred from numbering (`spec-names-release-prd`) |
+| `prd` | MUST be stated: which initiative PRDs this spec delivers against. MUST NOT be derived from `release` |
+| `fr` | The `FR` this spec satisfies. Ideally one — an `FR` is human-testable from birth |
+| `size` | `S` · `M` · `L`. MAY be raised mid-flight; MUST NOT be lowered |
+| `depends_on` | At **spec** level. A spec declaring none runs in parallel with its neighbours |
+| `spec_folder` | One per spec, not one per spec × component |
+| `tickets` | Flat, one row per ticket: `id` · `component` · `satisfies: [UC]` · `blocked_by` · `touches` · test names |
+| a ticket `id` | `<spec-id>-<NN>` — `SPEC-3-01`. The engine numbers its files from `01` per feature, which is unique only inside one spec; the RTM needs a key that is unique across the corpus |
+
+**`tickets` is an index, not a store.** The ticket's prose lives where the tracker put it; the row carries only
+what RTM and the validators read. And ticket **status** MUST NOT be copied here — it is read from the ticket
+itself when `.control/generated/` regenerates (`ticket-status-one-home`). Two homes for one fact is how registries start lying.
+
+`ticket-status-one-home` finds the file at `{spec_folder}/issues/<NN>-*.md`, from the number at the tail of the id, and reads its
+status from either a `**Status:**` body line — what the engine writes, because a ticket file is a tracker
+payload and trackers do not read YAML — or `status:` in frontmatter.
+
+The `epics` nesting is **repealed**. A ticket names its `component` directly; an intermediate level that only
+grouped rows bought nothing and cost a lookup.
+
+Size does not choose which gates are active — that is `mode`'s job. It decides two things: session merging
+(`S` merges G4 and G5 into one 20-minute session) and whether `SPEC.md` is written at all.
+
+## Phase 2 — The contract, and the tickets
+
+Two engines, and **this skill invokes both.** Invoke, then verify and land.
+
+| Size | What this skill invokes | What lands |
+|---|---|---|
+| `M` · `L` | `to-spec`, then `to-tickets` | `SPEC.md` in `spec_folder` · ticket files · the `tickets` index rows |
+| `S` | `to-tickets` only | ticket files · the `tickets` index rows. **No `SPEC.md`** |
+
+Each invocation carries what the engine would otherwise stop to ask: the tracker is already configured
+in `docs/agents/issue-tracker.md`, the seams and the testing decisions are settled above, and the
+`to-tickets` quiz — granularity and blocking edges — is answered from the size table in
+`delivery-flow-guide.md` plus `depends_on` and `touches`. Under a mandate every one of those answers is
+also a ledger row.
+
+At `S` the tickets **are** the contract. `to-tickets` accepts a conversation directly, so a middle document
+buys nothing there. From `M` up it is written first, because two things have to be settled **before** tickets
+are cut and neither survives being decided afterwards: the **seams** the feature will be tested at — fewest
+possible, highest possible, agreed with the owner — and the **testing decisions**, which say what a good test
+is here and name the prior art.
+
+### What you verify before landing anything
+
+- **Nothing new.** `SPEC.md` is a **projection** of `.what/` + `.how/`. It MAY restate a promise in its own
+  words — a machine contract is more useful self-contained — but **every restatement MUST refer to the live
+  corpus document it came from.** A reference that does not resolve is not a wording problem: the gap is
+  upstream. Route to `wdi-component` or `wdi-blueprint`, and do not let the contract invent it.
+- **Every user story resolves to an id.** A user story with no `FR` or `UC` behind it is a **new promise**,
+  and a new promise is `wdi-product`'s, never a spec's.
+- **Every ticket names what it satisfies.** No `satisfies`, no landing: without it the chain
+  `FR → UC → ticket → test` breaks and RTM cannot say which promise went green. This is the one field
+  `to-tickets` does not ask for on its own, so it is the one most likely to be missing.
+- **Every ticket is vertical.** A slice of one layer is not a ticket. The exception is a wide refactor,
+  sequenced expand → migrate in batches → contract; `delivery-flow-guide.md` owns that rule.
+- **Ticket files land under `spec_folder`, and `spec_folder` is `.scratch/<spec-id>-<slug>/`.** Their
+  **shape** is the engine's — one file per ticket, numbered in dependency order, blocking edges declared
+  — and the location is ours, written in `docs/agents/issue-tracker.md` where the engines read it. The
+  id in front of the slug is not decoration: four live repos wrote that leaf four different ways, one of
+  them all four inside a single repo, and a folder nothing can trace back to a row in `specs.yaml` is
+  how a spec goes missing. A ticket at the repo root, or under `docs/`, or in a `.scratch/` directory
+  with no row in `specs.yaml`, is drift.
+
+`SPEC.md` and ticket files **are not read by humans.** Both are machine contracts, and no review burden MAY be
+moved onto them. `wdi-review` MAY still be dispatched over the contract; its trace lands on the spec in
+`specs.yaml` — and where there is no `SPEC.md`, one trace covers the **ticket set as one artifact**, never one
+review per ticket.
+
+## Phase 3 — Ship each ticket
+
+Work the **frontier**: the tickets whose blockers are all closed. A ticket whose `blocked_by` is not yet
+satisfied MUST NOT be started, however ready it looks.
+
+| # | Step | Engine | Exit condition |
+|---|---|---|---|
+| 1 | Encode | invoke `tdd` | **Failing tests exist that encode this ticket's acceptance criteria** |
+| 2 | Build | invoke `implement` — it uses `tdd` at the agreed seams and calls `code-review` itself | Those tests green, typecheck clean, full suite green once |
+| 3 | Panel | invoke `code-review`, as a **separate** dispatch by a **different** agent | Panel adjudicated, zero unresolved must-fix |
+| 4 | Publish | — | Branch pushed, PR open, ticket-closing checklist answered |
+| 5 | CI | — | All checks conclude green on the pushed head SHA |
+
+**Step 1 is the change that matters most.** It used to be a plan, judged from a frontmatter field a builder
+wrote about itself. Now it is a **failing test suite** — the acceptance criteria, encoded, and demonstrably
+red. That is evidence rather than a claim, and it is what `ticket-has-test` and G5's ★2 have always been asking for:
+*acceptance criteria proven by a test, not by an agent's statement.* Under TDD the test exists **before** the
+code, so the proof is not retrofitted.
+
+### Engine rules
+
+- **MUST judge Steps 1 and 2 from the test suite**, not from a report and not from a status field. Red at the
+  right assertions closes Step 1; green with a clean typecheck closes Step 2. A builder's chat report MUST NOT
+  settle either.
+- Ticket **status** is read from the ticket itself. It MUST NOT be copied into `specs.yaml` (`ticket-status-one-home`).
+- Every step MUST start from a **fresh context**, and a fix round MUST NOT go back to whoever produced the
+  code. The ticket carries everything the next builder needs, and inherited context is how a step stops judging
+  the artifact on its own merits. Where one session runs consecutive steps itself, it MUST re-read the artifact
+  rather than trust what it remembers writing.
+- A step handed to a separate agent MUST be launched so it cannot sit waiting for input nobody will give, and
+  its start MUST be confirmed from observed activity — never from a readiness match alone.
+- A step that fails because the agent could not spawn what it needed is a **capability** failure, not a ticket
+  failure. MUST retry it somewhere that can, and MUST NOT record it against the ticket.
+- MUST NOT reorder or drop a step. A step with nothing to do MUST be reported as such, not skipped silently.
+- A ticket returned by the panel MUST have its acceptance criteria amended and its status reset before it is
+  picked up again — see Step 3.
+
+### What every builder brief MUST carry
+
+Three rules this corpus adds, and `/implement` knows none of them. All three MUST reach whoever writes code in
+this spec — in the dispatch when a step is dispatched, and in the session's own working instructions when it
+is not.
+
+- **Debugging is conditional, never a phase.** When a test or build fails and the cause is not known, the
+  builder MUST run `wdi-systematic-debugging` before proposing any fix. A third failed fix attempt is the signal
+  to escalate, not to try a fourth.
+- **The corpus is not the builder's to change.** A builder MUST NOT edit `.what/`, `.how/`, or an `applied`
+  `DEC-`. A deviation from the SDD or an `AD-N` is **reported**, and it becomes a `DEC-` through
+  `wdi-decision` — never absorbed as a code patch.
+- **Verification is run, not assumed.** The commands are this product's, and they live in
+  `.constitution/project/codebase-stack-guide.md` — build, test, and whatever the front end needs, each with
+  the directory it runs from. A skill MUST NOT carry one product's build line. A green registry workflow
+  MUST NOT be reported as proof the code compiles; they answer different questions.
+
+### Step 1 — encode the acceptance criteria as failing tests
+
+- The tests MUST be written **at the seams the spec agreed** — existing seams preferred, highest possible,
+  fewest possible. Inventing a new seam here, after the agreement, is a finding: it means Step 1 is redesigning
+  what Phase 2 settled.
+- Every acceptance criterion MUST have at least one test. A criterion no test can express is not an acceptance
+  criterion — it goes back to the ticket, or to the owner as an intent gap.
+- The tests MUST be **seen red, at the right assertions.** A test that passes before the code exists is testing
+  nothing, and a suite that fails for the wrong reason — an import error, a missing fixture — has not encoded
+  anything yet.
+- A test MUST NOT assert a literal where the behaviour is what matters. The panel treats that as a test that
+  cannot fail, and returns it.
+
+### Step 2 — build
+
+- The owner runs `/implement`, and it MUST be given the ticket and the three brief rules above.
+- It commits to the current branch and **never pushes**. That is its own behaviour and it is what we want; the
+  coordinator is the hand that pushes.
+- **`/implement` calls `/code-review` itself, and that call does NOT satisfy Step 3.** It is the builder
+  reviewing its own work — self-review by construction. Step 3 stands as a separate dispatch regardless of what
+  ran inside Step 2.
+- The full suite MUST be run green once at the end, not only the tests this ticket touched. A ticket that
+  passes its own tests and breaks a neighbour's has not finished.
+- Where the work turns out to need something the ticket does not authorise, it stops and reaches the owner with
+  the question verbatim. The builder MUST NOT widen its own scope.
+
+### Step 3 — panel, then judge
+
+Panel composition follows `risk_accepted`: at `low` a two-reviewer panel is **required** on the code; at
+`medium` and `high` it is available and SHOULD be used when the diff touches money, personal data, or a third
+party. The local Agent Rules govern which CLIs and models staff the panel. **A reviewer MUST be a different
+agent from the builder** — the builder's own review layers, `/implement`'s internal `/code-review` included,
+are self-review by construction and never satisfy the panel. This is the one separation in the pipeline that
+MUST NOT be collapsed: where the session cannot provide it and `risk_accepted` is `low`, the ticket is blocked
+and the owner MUST be told, because the field they set is what makes the panel required.
+
+`code-review` reviews along two axes — **Standards**, against this repo's documented conventions, and
+**Spec**, against what the ticket asked for. Both MUST run. Reporting one axis as the panel is reporting half
+a review.
+
+- MUST adjudicate every contested finding by reading the cited lines. Votes MUST NOT settle a finding. A finding
+  neither reviewer can locate in the diff is dismissed with that reason stated.
+
+**MUST return to Step 2:**
+
+- Breaks a ticket's acceptance criterion, or contradicts the contract, the SDD, an `AD-N`, or an `applied` `DEC-`
+- Wrong behaviour, crash, or data loss reachable from the running app
+- Corpus drift: an `LC` touched but not registered, a contract changed in code but not in `02-contracts/`, a
+  screen added without its `01-ux/` entry
+- An enum value rendered straight to the screen instead of read off its label map
+- Payment or private data reaching a tracked file
+- A weakened guard, or a test that cannot fail — including one asserting a literal instead of the behaviour it
+  claims to cover
+
+**MUST record as follow-up and MUST NOT return to Step 2:** style or naming with no behaviour delta · a
+refactor outside this ticket's scope · a pre-existing defect this ticket did not touch · a speculative risk
+with no reachable path.
+
+A must-fix MUST return as a **ticket amendment**, never a chat instruction: amend what the ticket asks for,
+note the finding that caused it, reset the ticket's status to `ready-for-agent`, then pick it up again. **What
+a ticket `satisfies` is not amendable here** — that is the promise, and a must-fix rooted in the promise is an
+intent gap that goes to the owner through `wdi-product`. Changing the `FR` a ticket serves in order to make the
+code pass is how a corpus starts agreeing with whatever was built.
+
+Cap: **2 return trips.** MUST re-run the whole panel after each fix round — a fix introduces defects. On hitting
+the cap MUST escalate and MUST NOT open a PR carrying an unresolved must-fix.
+
+### When the code turns out to be right and the document wrong
+
+This is normal during G5 and it is **not drift**. Building a thing is how you find out what the thing is.
+
+The builder still MUST NOT edit `.what/`, `.how/`, or an `applied` `DEC-` to make its own code fit — that is
+the corpus learning to agree with whatever was built, and it is a different failure. What happens instead:
+
+1. **Say it once.** Name what the code does, which promise it contradicts, and what that costs. One place,
+   one time. An `AD-N` is the one contradiction that **stops** — `decision-guide.md` owns it.
+2. **The owner decides.** If they adopt the code, that survey is spent.
+3. **The owning skill edits**, in the present tense, as if the design had always said this — `wdi-product`
+   for an `FR`, `wdi-component` for behaviour or design, `wdi-blueprint` for a cross-component rule.
+
+What MUST NOT happen after step 2: raising the same conflict again in a later pass, opening an `OQ-` for it,
+dispatching a review over it, or writing anywhere in the corpus that the change arrived late. The commit is
+that record. `corpus-guide.md` § The corpus is written in the present tense is the binding rule.
+
+### Step 4 — ticket-closing checklist, then push and PR
+
+The checklist is **three items**, and it MUST be answered before the PR opens:
+
+1. Something the next person needs to know? → **into the document that carries it**, and that is almost
+   always where it ends. It reaches `wdi-decision` only when no design document has a home for it —
+   `decision-guide.md` § A decision's first home. A ticket contradicting an `AD-N` **stops** rather than
+   closing; that is the one case where recording is mandatory.
+2. A trap for the next agent? → recorded where the next agent will read it.
+3. Test names matching what `specs.yaml` records?
+
+The five items that left this list moved to Phase 4, where the information actually exists.
+
+- MUST run the repository's commit/push audit before `git push`: refuse the forbidden paths, run the guard test,
+  fix content on failure. A failing guard is a finding about the content — MUST NOT weaken the guard or the test.
+- MUST NOT push directly to protected branches (`primary_branch` or `development_branch` from `policy:`, default `main`), MUST NOT force-push, MUST NOT merge.
+- The coordinator MUST be the hand that pushes and opens the PR.
+
+### Step 5 — watch CI, then judge
+
+- MUST wait for every check to conclude, then confirm the checks belong to the **pushed head SHA**. A green
+  report from a stale run is a false report.
+- **A cloud runner is metered, and this step is where the meter runs.** One unit of work earns **one** cloud
+  run: the workflow MUST NOT be shaped so that every intermediate push starts one — `ci-guide.md` owns the
+  trigger shape, and under a mandate `wdi-autopilot` § Cycle-end CI moves this step's run to § Finish.
+- `korpus.yml` validates the corpus, not the code. Build and test evidence comes from Step 2's own runs.
+- Classify each failure before acting: a defect from this change → Step 1 if the test was missing, Step 2 if
+  the code was wrong, with `wdi-systematic-debugging` when
+  the cause is unknown · infrastructure or flake → re-run **once**, and MUST NOT patch code to mask it; a second
+  identical failure is a defect · a guard failure → fix the content.
+- Cap: 2 return trips. On hitting the cap MUST report red honestly rather than keep pushing.
+
+### Parallel tickets
+
+Tickets with no blocking edge between them MAY run at once — that is what the frontier is — but four
+conditions MUST hold: each concurrent builder in its own worktree; `parallel-tickets-blocked` green for every pair released together;
+the first ticket that establishes a component's shape already closed, so later tickets inherit its code map;
+and no shared registry write in flight.
+
+**A wide refactor is the exception, and it inverts the rule.** Its batches MUST run in sequence, not in
+parallel, because each keeps CI green only while the expand still stands. Where even a batch cannot stay green
+alone, they share an integration branch and green is promised only at the final integrate-and-verify ticket.
+
+The pattern that MUST be preferred: run the biggest blocker alone first, let its shape decisions land, then fan
+out.
+
+## Phase 4 — Close the spec
+
+**Six steps.** Run in this order and stop at the first failure.
+
+1. **Registry catch-up.** Every `LC` the spec's design named is registered in `components.yaml`, and every
+   `touches` value resolves — `lc-registered`. This is the moment those questions have answers.
+2. **Inventories refreshed from code.** Run `.constitution/method/scripts/inventory.py`. The plan-versus-reality
+   difference is reported as a finding; it MUST NOT be patched into agreement by hand.
+3. **Structure maps refreshed** through `wdi-init` intent `structure`, if a base folder was born or removed or a
+   key file moved.
+4. **Distillation.** Every applicable row of the ownership table in `corpus-guide.md` has been landed by its
+   owner. Anything durable in the spec folder leaves it now, or dies with it — **the ticket files included.**
+   Their prose is working output; what survives is the index in `specs.yaml` and whatever the checklist routed.
+5. **RTM green.** Every traceability row for this spec is closed. New risks are in the risk register with an
+   owner.
+6. Mark the spec `status: closed` in `specs.yaml`.
+7. **Housekeeping (Archive or Prune).** Once marked closed, offer the maintainer the choice to clean up
+   the completed spec directory from `.scratch/<spec-id>-<slug>/` using `/wdi-prune-or-archive`:
+   - **Archive:** `/wdi-prune-or-archive --spec <spec-id> --archive` (moves the directory to `.archive/specs/<spec-id>-<slug>/` and updates `spec_folder` in `specs.yaml`).
+   - **Prune:** `/wdi-prune-or-archive --spec <spec-id> --prune` (removes the completed spec directory from git and disk while preserving RTM history in `specs.yaml`).
+   If the maintainer declines or defers, leave the directory in `.scratch/`.
+
+The retrospective step is **repealed**, and `RTR-` with it. It was the only thing size `L` decided, and the
+only thing `V19` checked.
+
+- You MUST NOT close a spec with an open distillation row. The whole point of an ephemeral working layer is that
+  durable truth leaves it first.
+- You MUST NOT reopen a closed spec to add scope. Scope arriving late opens a new spec, or goes through
+  `wdi-decision` if it invalidates what is already planned.
+- A change invalidating more than 30% of a spec's tickets MUST go through `wdi-decision`, not a patch to
+  `specs.yaml`.
+- Fast Path work is recorded as `fastpath` with no gates. If an `FR` turns out to be touched, the work MUST stop
+  and be raised to a spec `S`.
+
+## Red Flags — STOP
+
+- "One reviewer approved, good enough"
+- Judging Step 1 or Step 2 from a report instead of from the test suite
+- **Counting `/implement`'s own `/code-review` as the panel** — that is the builder reviewing itself
+- Writing `SPEC.md` at size `S`, or skipping it at `M` and up
+- Landing a ticket with no `satisfies` — the RTM chain breaks silently and nothing else notices
+- Amending what a ticket `satisfies` to make a must-fix go away
+- A ticket that slices one layer instead of cutting through all of them, outside a wide refactor
+- Running a wide refactor's batches in parallel
+- Claiming this skill invoked `to-spec`, `to-tickets`, or `implement` — it cannot; the owner runs them, or under
+  a mandate a builder reads and follows them, and the ledger says so
+- A builder editing `.what/`, `.how/`, or an `applied` `DEC-` to make its code fit
+- Fixing a failing test without knowing why it failed
+- Opening a PR with an unresolved must-fix, or before the ticket-closing checklist is answered
+- Editing a guard, a test, or an assertion to turn something green
+- Reporting green without checking the head SHA, or reading green `korpus.yml` as a passing build
+- Leaving a dispatched step able to stall on a question nobody is there to answer
+- Naming an orchestration tool as this skill's requirement, or restating a CLI/model mapping the Agent Rules own
+- Closing the spec without the registry catch-up in Phase 4 — that is where five checklist items now live
+- Letting the contract state something `.what/` and `.how/` do not, or restate it without a reference
+
+**Each of these means: stop, return to the step or phase that owns it, or escalate to the owner.**
+
+## Output
+
+MUST follow the Agent Rules `Answer Closing` block, carrying these additions:
+
+- **what was done** — spec, `FR` satisfied, every ticket with its branch and PR URL, what the code now does
+- **what blocked it** — every step that looped, every dismissed finding and why, anything left red, and every
+  Phase 4 item that did not pass
+- **what comes next** — every finding recorded as follow-up, every entry in a contract's `deferred` list, every
+  ticket-closing item routed to another skill, and the plan-versus-code inventory differences. A follow-up
+  absent here is a lost finding.