Spec Workflow

Working posture (ADR-0056). Adversarial review is a named, bounded operation. This workflow invokes review passes (frame-critique / craft), whose skeptical register belongs inside those isolated reviewer subagents. Outside a review, default to collaborative and solution-forward; don't carry the adversarial stance into ordinary conversation.

Spec 003 promoted this skill from stub to active. The deterministic state mutations live in workflow.py; this SKILL.md drives the judgment layer.

What this skill does

• Guides SPIDR-splitting a new spec into vertical slices (Spike last, not first — try Rules / Data / Interface / Path first).
• Flags slices that look like horizontal phasing (no user-facing layer touched).
• Drives the spec lifecycle state transitions via workflow.py.
• Coordinates implementer + reviewer subagent invocations at the right points.
• Enforces the reconciliation checklist before a slice goes DONE.
• Consults docs/memory/glossary.md when drafting ACs to surface unknown domain terms.
• Surfaces skill-routing observability via workflow.py routing-stats [--days N] — a read-only histogram of which skills fired (jig baseline vs. richer/"other" skill per category) from .claude/skill-usage.jsonl (slice 041-02).
• Surfaces use-case coverage via workflow.py coverage [--project-dir DIR] — a read-only, advisory (non-blocking), project-wide bidirectional check (slice 068-03): a deterministic set-difference over the use_cases: trace links that reports use cases with no implementing spec (coverage gap) and specs citing no parent use case (scope creep). No-op when the project has no ## Use cases section.
• Guards the status board via workflow.py check-board <project-dir> — a read-only audit that exits non-zero when docs/specs/README.md no longer matches the spec records, or when two spec directories claim one number. Writes nothing, so it is safe to run in CI against a checkout.
• Surfaces gate-bypass telemetry via workflow.py gate-stats [--days N] — a read-only per-gate histogram of how often each gate honored its env-var override (an override-frequency audit trail, not a gate-value verdict) from .claude/skill-usage.jsonl (slice 078-02).

SPIDR splitting

All non-trivial specs are SPIDR-split into vertical slices before implementation begins. Spike is the last resort — try Path / Interface / Data / Rules first.

• S — Spike: research/learning activity. Only when none of P/I/D/R apply. AI agents default to spiking too eagerly — resist.
• P — Path: split by alternative paths through the story (happy path first, edge paths later).
• I — Interface: split by UI / platform / channel (minimal first, polish later).
• D — Data: split by data subset or format (less data first).
• R — Rules: split by business rules (simple first, edge cases later).

Anti-horizontal-phasing rule: every slice must touch the user-facing layer and deliver end-to-end value. A slice that touches only the DB or only the parser is horizontal phasing — re-split.

See worked-example-spidr-split.md for one applied example per axis plus a jig-native dogfood case (spec 017's three-axis split). The canonical primer for all five axes lives at docs/spec-workflow/spidr-primer.md.

Spike slices

When SPIDR's S axis fires during decomposition (none of P / I / D / R apply because the team doesn't yet know enough to pick), the resulting slice is marked kind: spike in its frontmatter — the typed enum that spec_lint.py validates.

When to introduce a spike during decomposition. Reach for S only after trying R / D / I / P. The bias to resist is "let me research this first" as a prelude to "now let me build it as one big slab" — that is horizontal phasing in a trench coat. If the spike would conclude with "now ship the implementation," the implementation IS the slice, and the research goes inside it.

Body shape (four labelled blocks). A kind: spike slice carries four blocks alongside the standard Goal / DoR / AC / DoD scaffolding. Each label must be written with the trailing colon (**Question:**, etc.) — that is what spec_lint.py matches against.

• Question: — one sentence stating the open question. Set at DRAFT.
• Time-box: — explicit budget (e.g., "1 day", "4 hours"). Set at DRAFT.
• Findings: — bullet evidence collected during the spike. Filled during IN_PROGRESS.
• Outcome: — one of ADR-NNNN created / spec NNN-NN unblocked / abandoned (reason). Multiple outcomes separated by ; (e.g., ADR-0007 created; spec 030-02 unblocked). Set at DONE.

spec_lint.py soft-warns when a kind: spike slice is missing any of the four labels — mid-flight spikes legitimately have empty Findings / Outcome, so this is a warning, not a hard error.

Always nested, never standalone. Spike slices live inside a real spec — never as a standalone docs/spikes/ artifact. The 1-slice-spec case (no clear downstream spec yet, just an investigation) collapses to "spawn a normal spec where the only slice is kind: spike." This forces the investigator to articulate the downstream change up front and keeps jig at two numbered families (specs+slices, ADRs).

Abandoned-spike manual-reshape failure mode. When a spike's Outcome is abandoned (reason), dependents are NOT automatically cascade-flagged. The human (or the next session) audits each dependent slice and decides whether the original design still holds. Automation here over-fires: "approach A abandoned" often means "approach B from the same findings still satisfies the dependents." workflow.py deliberately stays out of the cascade business; the SKILL.md hand-off is the documented gate.

How to use

Creating a new spec

0. Step 0 — confirm the project is scaffolded (spec 063 / ADR-0011). BEFORE reserving a number or drafting ANY docs/specs/ or slice structure, confirm this project is a scaffolded jig project. If it isn't, route — do not hand-roll directories:

   • Greenfield (no jig structure yet) → tell the user to run /jig:scaffold-init. It lays down conventions, templates, hooks, the status board, and a seed reference spec.
   • Existing spec/slices/ layout, but not jig-scaffolded (no scaffold.json) → tell the user to run /jig:migrate. It adopts the existing layout into jig structure.

   You don't have to decide the state yourself: workflow.py new (step 2) classifies and routes for you (spec 063-01) — a scaffold.json-bearing project proceeds; a greenfield project is refused naming /jig:scaffold-init; an adoptable spec-driven project is refused naming /jig:migrate. The deterministic gate and this human-readable precondition agree by construction, so don't restate the detection heuristic here — run the helper and let it route. (Bypass for a deliberate out-of-band flow: JIG_SCAFFOLD_PRECONDITION=0.)

   The anti-pattern this step exists to kill: an auto-triggered spec-workflow run improvising a loose slices/ folder (or any ad-hoc docs/specs/ skeleton) because /jig:scaffold-init was skipped. That produces a non-jig layout that then needs migrating — exactly the reported failure. When in doubt, route to setup first; never invent the structure by hand.

1. Confirm the work needs a spec. Trivial fixes don't.

   • Reference moved? Reframe first. If a load-bearing reference changed from outside the system (a design system, vendor / API contract, test infra, compliance regime, platform, or product-positioning / strategic-vision shift), reach for /jig:reframe before drafting — re-baseline the corpus onto the new reference so you draft against it, not the dead premise (spec 067 / ADR-0024). 1a. Read the vision ## Use cases section as framing — before drafting (spec 068-02 / ADR-0025). If the project's docs/product-vision.md carries a ## Use cases section (the breadth-layer behaviors captured at init), read it first as framing context for this spec — which captured behavior does this work serve? The section is the shared frame specs anchor against; reading it before you draft is what keeps behavior-dense projects from each spec inventing its own slice of the world. (If there is no ## Use cases section, the layer isn't adopted for this project — skip this step; nothing here applies.) You record the answer as a trace link in the spec's use_cases: frontmatter (step 2a). 1b. Cross-check tracked bugs before drafting ACs. For feedback/triage specs, read docs/bugs/README.md and any overlapping docs/bugs/NNN-*.md record before turning reported symptoms into acceptance criteria. If an item is a tracked defect with a bug record, route it through bug-fix (or link to that bug as an explicit dependency) instead of copying the defect into the spec as a second owner. Keep polish/design/new-behaviour items in the spec.

2. Reserve the next free number on origin/main:

   python3 "${CLAUDE_PLUGIN_ROOT}/skills/spec-workflow/workflow.py" new <slug>
   

   The helper computes max(NNN) + 1 across docs/specs/, writes a minimum stub docs/specs/NNN-<slug>/spec.md (frontmatter + Overview

   • SPIDR-analysis headers), commits it as docs(specs): reserve NNN-<slug>, and pushes to origin/main. If the push is refused by branch protection / permissions, the helper automatically falls back to a reserve/NNN-<slug> branch + gh pr create. This locks the number team-wide before any drafting begins, killing the parallel-worktree spec-number-collision failure mode logged across specs 014/015/016/017.

   Works from any branch or worktree (ADR-0015 / spec 051). The helper routes on the current branch: on main it runs the proven in-place flow (clean tree required, since the commit lands on local main); off main — a feature branch or a linked .claude/worktrees/* worktree — it builds the reservation commit in an ephemeral detached worktree checked out at origin/main and pushes it by SHA, never touching your branch, cwd, or working tree. You no longer need to switch to main (and a linked worktree can't, anyway).

   Flags: --no-push for solo machines without a remote, or for an off-main provisional reservation committed on the current branch (the number is local-view and may collide at merge — treat it as provisional); --pr to skip the direct-push attempt on protection-locked main.

2a. Record the use-case trace link — and grow the vision on discovery (spec 068-02 / ADR-0025). The stub seeds an empty use_cases: frontmatter list. Fill it with the UC-N id(s) (from the vision ## Use cases section, step 1a) this spec serves — the dependencies:-style flow-list shape, e.g. use_cases: [UC-1, UC-3]. This is the machine-resolvable trace link the reconcile-phase coverage check (slice 03) reads.

The discipline is soft — an empty/absent use_cases: never blocks a transition (AC4 / ADR-0011). But it is not silent. The trigger is mechanical and deterministic — the classify_spec predicate in skills/_common/use_cases.py computes one of no_section / empty / resolved / unresolvable for this spec — not a voluntary "is this new?" self-report. Whenever classify_spec returns empty or unresolvable (the spec cites nothing, or cites a UC-N with no match in the vision) at draft/framing, surface a three-path prompt — every path is one step and none blocks drafting:

• (a) cite an existing use case — this spec serves a behavior already in the vision: put its UC-N id(s) in use_cases:. Done.
• (b) grow the vision — this spec serves a behavior not yet captured: reuse vision-elicitation's capture loop, seeded with the existing entries (so the author sees the current set), → normalize → confirm, then write additively (append, never discard-and-replace) and assign the next free UC-N (use_cases.next_use_case_id allocates max + 1; retired numbers are never reused). The confirm step guards grow quality so a reachable trigger can't silently bloat the section: (i) enforce goal-level grain — reject spec-shaped / requirements-level phrasing, re-running slice 01's normalize ("[actor] can [goal]"); (ii) run a near-duplicate check against the seeded existing entries (use_cases.is_near_duplicate) — on an apparent match, route back to path (a)-cite rather than minting a duplicate. Then record the new id in use_cases:.
• (c) decline — legitimately untraced (infra / refactor / no user-facing behavior) or defer: leave use_cases: []. The vision is unchanged, and any resulting gap is slice 03's advisory coverage backstop. No-op.

CRITICAL — the no-section no-op. When classify_spec returns no_section (the project has no ## Use cases section — the breadth layer is not adopted, e.g. jig's own repo), the prompt is suppressed entirely: nothing prompts and nothing errors. A project with specs but no use-case layer is wholly unaffected. The trigger fires only on empty / unresolvable, which presuppose the section exists. 3. Create docs/specs/NNN-<slug>/{spec.md,plan.md,tasks.md} with the conventional structure: status frontmatter, overview, SPIDR analysis, ordered slices. 4. SPIDR-split: for each slice, the goal is one vertical piece that delivers end-to-end value. Spike is the last resort, not the first reach. 5. Each slice is a whole-file document with the canonical frontmatter shape (status, dependencies, last_verified) plus DoR / AC / DoD / Close-out sections. workflow.py new (step 2) already emitted a well-formed starter slice-01-tbd.md from the packaged slice template, so you never hand-resolve a template path; add any further slices in that same shape. For the structural reference — what a filled-in slice looks like — mirror the in-project worked example that scaffolding installs at project root: docs/specs/001-adopt-jig/ (spec.md + slice-01-bootstrap.md), the first spec, which your project's CLAUDE.md names as the worked example to imitate. Set status: DRAFT in the frontmatter. Legacy slices that use prose **STATUS: DRAFT** markers still work (lazy migration); no need to rewrite them. 5a. Design-fidelity authoring nudge (spec 104-02 / ADR-0049). When a slice you just wrote in step 5 ships visual design (a mockup, a design system spec, a screen with colours/spacing/sizes/layout rules to hit), don't let "doesn't match the mockup" live only in a picture:

• (a) Extract the design values into checkable ACs. Pull the concrete values — colours, spacing, sizes, layout rules — out of the mockup and write them as this slice's acceptance criteria, the same way any other observable behavior becomes an AC. This is what turns a fuzzy "looks right" into something a reviewer (or an eval) can actually check.
• (b) When fidelity must gate, wire the servo rail. If a screen's visual fidelity needs to be a hard, enforced condition of DONE — not just an eyeballed check — set design_review: true in the slice frontmatter and wire a servo design-eval (screenshots the running app against the reference, scores it with a pinned vision judge) as the done-condition. design_review: true is attested, read-only, at REVIEWED by spec 071's design-review pass (the deriver slice_needs_design_review in workflow.py reads the flag) — jig never re-derives the eval score itself.

Graduated, not mandatory — jig offers, never forces, servo. Not every screen earns a frozen eval:

• Low-stakes visual polish → design-values-in-ACs plus attest-by-eyeball at review time is enough; no servo design-eval required.
• A hard fidelity gate (fidelity must not regress, or is a stated product requirement) → servo design-eval + design_review: true.

Pick the tier that matches the stakes; when in doubt, start with (a) and add (b) only when eyeballing genuinely isn't enough. See spec 071 and ADR-0049 for the full routing rationale; this step adds no new mechanism — teeth stay anchored to the existing design_review flag. 6. Ground your factual claims (spec 064-02 / ADR-0020 §1–§2). Any load-bearing factual claim about a runnable surface — library/API capability, version/perf behavior, behavior of existing code — must be backed by an executed probe (run the command, read the source / node_modules) or a citation. A universal or negative claim ("the only", "never", "always", "one-way", "nothing reads", "otherwise clean") is established by an enumeration — a search you can show returns the complete set — not a single positive citation (ADR-0052): one true example proves nothing about the rest of the set, and these are the highest-value claims a future reader relies on. To claim enumeration, state why the search is exhaustive — what closes the set so nothing escapes. Some sets are closed by syntax and this is easy (imports in a package, call-sites in a repo); many only look grep-bounded — a "nothing reads this" search misses dynamic / reflective / ORM / string-built / config-wired / codegen'd / cross-repo access (illustrative, not a checklist to clear: the burden is to show the search captures every member, not to rule out named escapes). When you cannot show the search is exhaustive, an empty result is absence of evidence, not an enumeration: weaken the claim, tighten the boundary until the search genuinely closes the set, or move it to ## Assumptions. Never dress an empty search as an enumeration — the frame-critique reviewer treats "I searched and it was empty" as ungrounded until you have shown what closes the set. Everything you cannot verify goes in the spec stub's risk-gated ## Assumptions section, marked explicitly — never asserted as fact. This makes mandatory + derived the existing informal "Current state (verified …)" discipline that the 064-01 retro found jig already half-practices by hand: it was grounding-by-probe all along, just reliant on author diligence. The ## Assumptions you surface here has downstream value — slice 064-04 derives the frame_review trigger mechanically from it, so honest framing now is what decides later whether the adversarial frame-critique pass fires. The section is risk-gated: write "None" / omit when there are no unverified load-bearing assumptions; don't pad with boilerplate. (For a worked example of marked assumptions plus probe-grounded claims, see ADR-0020 ## Assumptions A1–A4 + ## Kill criteria, and the spec 064-01 retro, which probe-verified its three most load-bearing claims before recording them.) 7. Let the assumptions decide frame_review (spec 064-04 / ADR-0020). You are not asked "is frame-review needed?" — the ## Assumptions you just surfaced decide it, mechanically. Set the slice's frame_review flag from workflow.py frame-review-needed:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/spec-workflow/workflow.py" \
  frame-review-needed "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>"

The rule is a derivation, not a judgment call: true iff the slice's ## Assumptions section carries ≥1 real (non-placeholder) assumption — so honest framing in step 6 is exactly what fires (or silences) the adversarial frame-critique pass. An inline-mirror / refactor slice with no unverified assumptions (## Assumptions absent or just "None") stays default-off. ADRs are always-on (OQ3): any ADR gets frame_review: true unconditionally — the deriver returns true for any adr-*.md path. When the value is true, set frame_review: true in the slice frontmatter so the gate + session-plan dispatch the pass. 8. Add rows to docs/specs/README.md (or regenerate via workflow.py status-board).

Picking up a slice

1. Read the automatic jig hint: project-orientation headline injected at SessionStart, or refresh it manually before choosing work:

   python3 "${CLAUDE_PLUGIN_ROOT}/skills/spec-workflow/workflow.py" orient \
     --project-dir .
   

   The headline is computed from scaffold.json and lifecycle artifacts. Treat docs/architecture.md, the spec corpus, and the status board as authoritative; a shallow source-tree listing is not evidence that a scaffolded project is greenfield or that recorded stack decisions are absent.

2. Check docs/specs/README.md for the next slice in READY_FOR_IMPLEMENTATION (or DRAFT for a slice you intend to plan now).

3. Run:

   python3 "${CLAUDE_PLUGIN_ROOT}/skills/spec-workflow/workflow.py" transition \
     "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>" IN_PROGRESS
   

   Claim-on-working-state (spec 049-01, amended by ADR-0045). On a frontmatter (file-per-slice) slice, a transition into a working state — READY_FOR_REVIEW / IN_PROGRESS / REVIEWED / RECONCILED — stamps claimed_by: (the current branch name, or JIG_CLAIM_ID), so spec-level work is marked too, not just implementation. Entering a release point clears it: the two pickup-queue states DRAFT / READY_FOR_IMPLEMENTATION (step 2 above tells you to choose work from exactly those, so a leftover owner there would mark a free slice as occupied), plus the terminal DONE / DEFERRED / ABANDONED.

   It refuses only when the slice is already IN_PROGRESS under a different identifier and you are moving it to IN_PROGRESS (naming the holder, pointing at --release); any other foreign claim — on your copy or on origin/main — is a loud non-blocking warning, because two sessions working one spec can be legitimate. The claim is local by default; add --push (direct) or --pr (via PR) to reserve it on origin/main so other worktrees see it, at any working state, though only an IN_PROGRESS reservation also publishes status: there (race / protected-branch handling mirrors workflow.py new). At a working state that reservation is best-effort (for a target other than IN_PROGRESS): if the trunk copy is already status: IN_PROGRESS under someone else's claim or none, it warns and pushes nothing, because that state is what the start-of-build guard hard-blocks on — stamping a claim over it would move a live lock, or manufacture the enforced pair on an unclaimed copy. Your own trunk claim just reports a benign no-op. To force-release a stale claim: transition <spec> <slice> <state> --release --reason "<why>" (clears claimed_by:, logs to ## Release log).

   Class-A cross-ref guard (spec 112 / ADR-0058). A transition into a working state is refused when the slice is already DONE on origin/main — a stale branch re-advancing already-integrated work. The → IN_PROGRESS case is caught by the start-collision guard (spec 051-04); the other working states (READY_FOR_REVIEW / REVIEWED / RECONCILED) by _refuse_integrated_advance, which reads the shared cross_ref_state.identifier_state_on_ref primitive. For a sanctioned re-open / supersession of integrated work, pass --reopen (a first-class, audited bypass distinct from the blanket JIG_CROSSREF_GATE=0); the → IN_PROGRESS path's own escape remains JIG_START_COLLISION_GATE=0 (convergence tracked in docs/refinement-todo.md). Unreachable origin/main → non-blocking warning, transition proceeds.

   Do not read a blank claimed_by: as "free". It means no claim is recorded: claims are local unless pushed, so another worktree's unpushed claim is invisible, and a plain Edit-tool write to a slice takes no claim at all. A claim that IS present names the session that last moved the slice into a working state — a presence hint, not a live lock. When it matters, ask rather than assume — see bug 014.

4. Fill in / refresh plan.md and tasks.md for the slice.

5. Spawn the implementer subagent with the spec path. Prefix the Task prompt with [jig:phase=implementation] [jig:spec=NNN] [jig:slice=NNN-NN] so jig-telemetry.sh can attribute implementation-phase cost. Implementer writes the deliverable to disk (TDD — failing tests first).

After implementation

Slices 031-01 + 031-02 + 060-05 wired a multi-pass review flow into the post-implementation step. Every slice runs through two passes before the IN_PROGRESS → REVIEWED transition; two further passes fire on demand — the arch pass when the slice declares arch_review: true, and the code-health pass when it declares code_health_review: true.

The orchestrator runs the passes in this order:

1. Compliance pass — jig:independent-review (always). Spawn the reviewer subagent against the deliverable using the prompt built by review.py implementation. Reviewer is read-only; it evaluates each acceptance criterion and returns pass | fail | needs-changes.

2. Craft pass — pr-review (always). After the compliance pass returns, build the craft-pass prompt with review.py pr-review and spawn a second reviewer-shaped subagent. The reviewer is read-only (Read/Glob/Grep, no Skill tool), so it cannot route to a skill via Claude's skill router; instead review.py hands it a concrete richer-skill path to read-and-apply, falling back to jig's inlined baseline buckets (scope / blockers / nits / strengths). The pass returns the same VERDICT / REASONING / SPECIFIC ISSUES / RECONCILIATION NOTES envelope as the compliance pass, with SPECIFIC ISSUES entries tagged [blocker] / [nit] / [strength].

   Select the richer skill first (spec 096-03 / ADR-0040 D3). --richer-skill is a required argument on pr-review (and arch-review / code-health). Before building the craft prompt:

   • Run the candidate step: review.py candidates pr_review "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>" --pass craft. It prints a tiered list — [high-confidence] (with descriptions) and [speculative] (names only) — and writes the shown set to a sidecar.
   • Pick the single best HIGH-CONFIDENCE candidate for the category and pass it as --richer-skill <name>. If several are genuine, pick one — do not refuse; the alternatives are recorded as the shown-and-declined set. If none fits, pass --richer-skill none for jig's baseline.
   • The pick is a heuristic, not a guarantee — an explicit review.<category>_skill in scaffold.json (096-01) overrides it, and candidates never picks for you (a [speculative] false positive like a briefing skill must not be selected on lexical grounds alone).
   • CI / no orchestrator: pass --richer-skill none --non-interactive; config remains the reproducible path. Omitting --richer-skill entirely, or invoking a pass with no sidecar + no config + no --non-interactive, fails fast (naming the missing candidates step) rather than silently using the baseline.

3. Arch pass — arch-review (on-demand). Before running this pass, query the slice's arch_review: frontmatter flag via workflow.py arch-review-needed. When the helper prints true, build the arch-pass prompt with review.py arch-review and spawn a third reviewer-shaped subagent. The pass produces the four canonical arch buckets (summary / strengths / concerns / open questions) wrapped in the same verdict envelope, using the same candidate-channel resolution as the craft pass (096-03): run review.py candidates arch_review …, pick the best, pass --richer-skill; config (review.arch_review_skill) overrides, else jig's baseline buckets are inlined. When the helper prints false, skip this pass entirely. Slice authors flip the flag by uncommenting the arch_review: true line in the slice template's frontmatter — set it when the slice changes module boundaries, public contracts, or architecture-shaped concerns.

4. Code-health pass — jig:code-health (on-demand, gated). Before running, query the slice's code_health_review: frontmatter flag via workflow.py code-health-review-needed. When it prints true, run health.py yourself (the orchestrator / CI), capture its tight summary, and feed THAT summary into review.py code-health … --summary-file (--summary-file - to pipe it in). Then spawn a reviewer-shaped subagent. The reviewer is read-only (Read/Glob/Grep, no Bash) — it must NOT run health.py; it judges the summary you provide. The reviewer renders the judgment a tool can't: is duplication within the ADR-0002 inline-mirror budget? is a complex function inherent or fixable? are the lint findings worth blocking on? The pass returns the same verdict envelope, with SPECIFIC ISSUES tagged [blocker] / [nit] / [strength]. Why gated, not always-on: ADR-0017 flags the per-slice review cost (specs 055/057 context-cost discipline) and recommends gating it like arch-review — so it defaults off and slice authors opt in with code_health_review: true. The evidence file is reviews/slice-NN-code-health.md.

When spawning any reviewer Task above, prefix the Task prompt with telemetry tags before the review.py body: [jig:phase=<phase>] [jig:spec=NNN] [jig:slice=NNN-NN]. Use compliance for review.py implementation, craft for pr-review, arch for arch-review, code-health for code-health, and reconciliation for the final reconciliation review.

Block rule for the REVIEWED transition. All required passes (compliance + craft, plus arch when arch_review: true, plus code-health when code_health_review: true) must pass before transition <slice> REVIEWED:

• Any fail verdict from any pass blocks the transition.
• needs-changes from the compliance pass blocks (the implementer addresses findings and re-runs).
• needs-changes from the craft pass does NOT block — the [nit]-tagged entries become reconciliation-log items (the implementer captures them in the deviation log during reconciliation). Only [blocker]-tagged entries from the craft pass block the transition.
• The arch pass follows the same rule as the craft pass: [blocker]-tagged entries block; [nit]-tagged entries and needs-changes become reconciliation-log items.
• The code-health pass follows the same rule: [blocker]-tagged entries block the REVIEWED transition; [nit]-tagged entries become reconciliation-log items.

Measurement tag (non-gating). The craft/arch/code-health passes also self-classify each SPECIFIC ISSUES entry [spec] (an acceptance-criteria / spec-frame issue a smarter clarify pass could have prevented) or [impl] (implementation-discipline — untested edge, brittle fixture, drift). This carries no gating consequence — it exists purely so a future rollup can count [spec]-tagged themes across specs (docs/refinement-todo.md "Instrument the review→learnings→clarify loop before building it").

The gate is mechanical, not advisory (slice 045-03 / ADR-0014 §5). workflow.py transition now refuses the REVIEWED / RECONCILED / DONE moves unless the required review evidence — recorded with review.py record-review as docs/specs/NNN-<slug>/reviews/slice-NN-<pass>.md — exists and clears (verdict: pass). REVIEWED requires compliance + craft (+ arch when the slice declares arch_review: true, + code-health when it declares code_health_review: true); RECONCILED requires the reconciliation verdict and ### Deviation log plus ### Reconciliation sweep subsections; DONE re-validates the post-implementation and reconciliation evidence set (in addition to the existing dependencies: check). A refusal names the missing/invalid artifact and the record-review command to produce it. The gate enforces evidence consistency, not human sign-off (it lives in the agent's trust boundary per ADR-0011). Bypass it for a deliberate out-of-band flow by setting JIG_REVIEW_EVIDENCE_GATE=0 (also false/off/no) — the status still transitions and the DONE dependency check still runs; only the evidence check is skipped.

After all required passes pass:

4. Address any reviewer findings, adding regression tests for any real bugs found.
5. Record each pass's verdict as durable evidence with review.py record-review (writes docs/specs/NNN-<slug>/reviews/slice-NN-<pass>.md — see the independent-review SKILL.md § "Recording and checking review evidence"). The REVIEWED transition is gated on this evidence, so it is not optional.
6. Transition: transition <spec.md> <slice> REVIEWED. The gate re-validates the recorded compliance + craft (+ arch,
   • code-health) verdicts before the status flips (and before the 003-04 auto-tick).

Recovering from a failed review. A fail/needs-changes verdict — or a [blocker]-tagged craft/arch finding, which is recorded as a non-pass verdict — blocks the REVIEWED transition. To recover: address the findings, re-run the pass against the updated deliverable, record-review the new verdict (it overwrites in place the earlier file for that (slice, pass); git history keeps the prior one), then re-run transition … REVIEWED. With every required pass now pass, the gate clears. A non-pass artifact never overwritten by a later pass keeps blocking — the "superseded without a later pass" case (ADR-0014 §4).

When a review retracts a claim, sweep the corpus before re-recording. The recovery above is written for a code-shaped finding, which is local to one file. A finding about content is not: a retracted assertion is usually copied by design into CHANGELOG.md, the slice record, the inbox, and cross-referenced docs. Fixing only the reviewed deliverable leaves the retracted version authoritative in every sibling artifact — and the stale copy is frequently the one the next session reads first (the project's own rules make CHANGELOG.md a read-before-you-fix record), so the pass re-fails round after round on a document you never touched. Before you record-review the new verdict, grep the retracted phrasing across the docs root and CHANGELOG.md, and reconcile every hit — the plain sweep is what reaches the changelog, the inbox, and arbitrary cross-referenced files. Within the spec itself, /jig:analyze is the structured complement: its Duplication and Terminology Drift categories catch a retracted claim surviving across the spec's own slice files and the docs it cross-references (product-vision, accepted ADRs, the glossary, architecture.md) — but it audits one spec's files plus that fixed set, not the whole corpus, so it sharpens the sweep rather than replacing it. Distinguish surviving assertions (the claim still stated as true — must fix) from explicit retractions (the claim named as withdrawn in a changelog or history entry — correct, and worth keeping).

# Compliance pass (always)
PROMPT=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
  implementation "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>" \
  "<deliverable-path-1>" ...)
SUBAGENT=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
  subagent-type implementation)
# … feed "[jig:phase=compliance] [jig:spec=NNN] [jig:slice=NNN-NN]\n\n$PROMPT"
# … to Task with subagent_type: $SUBAGENT, wait for pass …

# Craft pass (always) — spec 096-03: select the richer skill first.
# 1. Show the tiered candidates + write the sidecar:
python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
  candidates pr_review "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>" \
  --pass craft
# 2. Read the [high-confidence] tier, pick the single best (or `none`), then
#    build the prompt with the REQUIRED --richer-skill (config overrides it):
PROMPT=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
  pr-review "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>" \
  "<deliverable-path-1>" ... --richer-skill "<name-or-none>")
SUBAGENT=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
  subagent-type pr-review)
# … feed "[jig:phase=craft] [jig:spec=NNN] [jig:slice=NNN-NN]\n\n$PROMPT"
# … to Task with subagent_type: $SUBAGENT, wait for pass …
# (CI / no orchestrator: --richer-skill none --non-interactive.)

# Arch pass (only when slice frontmatter has `arch_review: true`)
# IMPORTANT: capture the helper exit code — a non-zero exit means the
# slice lookup failed (missing spec / unknown fragment / ambiguous),
# not "no arch pass needed." Surface the error rather than silently
# skipping the pass.
if ! NEED_ARCH=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/spec-workflow/workflow.py" \
    arch-review-needed "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>"); then
  echo "arch-review-needed failed — aborting" >&2
  exit 2
fi
if [ "$NEED_ARCH" = "true" ]; then
  # 096-03: show candidates for arch_review, then pick (config overrides).
  python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
    candidates arch_review "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>" \
    --pass arch
  PROMPT=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
    arch-review "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>" \
    "<deliverable-path-1>" ... --richer-skill "<name-or-none>")
  SUBAGENT=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/independent-review/review.py" \
    subagent-type arch-review)
  # … feed "[jig:phase=arch] [jig:spec=NNN] [jig:slice=NNN-NN]\n\n$PROMPT"
  # … to Task with subagent_type: $SUBAGENT, wait for pass …
fi

# Code-health pass (only when slice frontmatter has `code_health_review: true`)
# The orchestrator runs health.py and feeds its summary IN — the read-only
# reviewer never runs the tool (no Bash).
if ! NEED_CH=$(python3 "${CLAUDE_PLUGIN_ROOT}/skills/spec-workflow/workflow.py" \
    code-health-review-needed "docs/specs/NNN-<slug>/spec.md" "<slice-fragment>"); then
  echo "code-health-review-needed failed — aborting" >&2
  exit 2
fi
if [ "$NEED_CH" = "true" ]; then
  # Ru

…(truncated)