Review

OUTPUT STYLE: terse-technical per /_shared/terse-output.md. Drop articles, fillers, pleasantries, hedging. Preserve verbatim: code fences, inline code, URLs, file paths, commands, grep patterns, YAML/JSON, headings, table rows, error codes, dates, version numbers. No preamble. No trailing summary of work already evident in the diff or tool output. Format: fragments OK.

Review — Consolidated Precision Gate

The per-change quality gate. Precision-biased: it runs constantly, so a false alarm is expensive — low-confidence advisory findings are suppressed by default. Selects checks from the shared registry /_shared/check-registry.json and runs both detection lanes; delegates the full 8-invariant run to the sprint-review engine. Companion: /blitz:audit is the recall-biased pre-release deep audit that re-surfaces what review suppresses.

Session registration: follow session-lifecycle.md §Session Registration before any other work.

Verbose progress is mandatory. Follow terse-output.md. Print [review] status at every phase + dispatch. Log skill_start/skill_complete to the activity feed.

Flag Parsing

• --sprint NNN — review the specified sprint (defaults to current / uncommitted changes).
• --auto-fix — apply safe fixes (types, lint, imports) automatically.
• --only {completeness|wiring|framework|design|full} — run one folded concern instead of the full gate (default full).
• --min-confidence {high|low} — advisory-finding gate band (default high ≥0.8). reject-authority findings always surface (bypass the gate). See quality-engine.md.
• --dual — set BLITZ_DUAL_CRITIC=1 (cross-model critic for semantic/security findings).

Two detection lanes (both run; registry-driven)

Per quality-engine.md, select consolidated_target ∈ {review, both}:

1. Deterministic lane (run first, fast, zero-FP): lane == deterministic checks — tsc/lint/test/build (det-11/12), shortcut detectors (det-01..20), anti-mock O2 (o2-, was completeness-gate), wiring O3 (o3-, was integration-check, conditional on new modules), framework rules (fw-*, conditional on Vue/Firestore/Pinia). reject-authority findings may flip the verdict and bypass the confidence gate.
2. Semantic lane (single-pass — diff is small, bias is precision): the 4 parallel reviewer agents (security/backend/frontend/patterns). Single-pass findings start at base_confidence ≈0.5 and MUST pass FP-verification to survive; un-reproduced findings are dropped. FP-verification cannot raise confidence (only aggregation does, and review is single-pass) — so a default-review semantic finding stays a sub-threshold advisory, never a blocker. Aggregation is opt-in (--aggregate); audit is where it re-surfaces.

FP-verification (mandatory before any blocker on a base_confidence < 1.0 finding): re-read the cited code, confirm the flaw reproduces against actual behavior, attach the excerpt. No evidence → downgrade to advisory.

--only folded concerns

--only	Registry checks	Replaces
completeness	o2-anti-mock, o2-artifact-l1l2, det-09/10	completeness-gate (deleted; folded here)
wiring	o3-wiring, o3-orphan-route, det-16	integration-check (deleted; folded here)
framework	fw-firestore-vue-pinia	/blitz:code-doctor rule scan (code-doctor keeps --fix standalone)
design	pillar == design — Layer 0 (universal slop) always; Layer 1/2 gated by the scripts/detect-stack.sh adapter; reconciliation suppresses per stack	impeccable design pillar (new)
full (default)	all review-targeted checks + the sprint-review 8-invariant gate	—

--only runs that concern's registry checks read-only and reports; full delegates to the sprint-review engine. sprint-dev Phase 3.5.0 calls /blitz:review --only wiring.

Pre-Flight Validation (full runs)

1. Sprint exists: sprint-registry.json shows target sprint status: review|in-progress. Else inform + stop.
2. Stories exist: sprints/sprint-${N}/stories/ has ≥1 status: done.
3. No conflicting sessions: no active sprint-review session on the same sprint.

Execution

• --only X: run X's registry checks (read detection.command per row), FP-verify, report ranked by effective_confidence. No engine dispatch.
• --only design: first run the preflight scripts/design/preflight.sh <target-repo> and render its DESIGN_LANE_STATUS banner. If semantic != OK, surface the DESIGN_LANE_UNAVAILABLE line in the summary, run only the lanes that don't need impeccable (the deterministic regex rows), and do not return green for the design pillar — coverage is reduced, not passed. Then resolve the adapter by parsing the DESIGN_ADAPTER primary=… variant=… secondary=… incompat=… confidence=… token line from scripts/detect-stack.sh (the machine line, not the prose); select pillar == design rows where adapter ∈ inclusion(primary) ∪ secondary and not suppressed by reconciliation.relaxFor (where inclusion: none→{universal}, tailwind→{universal,tailwind}, tailwind-md3→{universal,tailwind,tailwind-md3}, vuetify→{universal,vuetify}, quasar→{universal,quasar}). Vendored (semantic) rows share ONE impeccable run and match hits by detection.filter; deterministic L1/L2 rows run their regex command (always available, key-free), then apply the registry's design.exclude set — a two-step scoped filter (drop files matching files OR containing any contentGuard; drop lines matching any lineGuard) so token-definition surfaces (@theme/tailwind.config/quasar.variables/Vuetify theme), comments, and SVG paint don't false-positive. FP-verify each survivor (re-read the cited line; reject token-def/comment/SVG) before reporting as a blocker. The consolidated design-raw-color-literal selects its remediation message from perAdapter by the detected stack. Layer 0 always on; Layer 1/2 gated to the stack. impeccable is a dependency of the target project, not the plugin (npm i -D [email protected]). See detector-rebuild.md and improvements/dependency-resolution.md.
• full: invoke the sprint-review skill (passing --sprint, --auto-fix, --dual). It runs the type-check/lint/test/build gates, parallel reviewer agents, auto-fix, and Phase 3.6 the 8 carry-forward invariants (incl. Invariant 7 = critic LGTM via agents/critic.md).

Output

• Findings by severity (Critical/Warning/Suggestion), ranked by effective_confidence; advisory findings below --min-confidence suppressed (logged, not surfaced).
• --auto-fix: report auto-fixed vs manual.
• Pass/fail per gate; for full, the 8-invariant PASS/FAIL summary.

Gotchas

• Only high-confidence-band findings are actionable; lower bands are advisory — don't escalate them to blockers.
• Lane authority differs: the critic's reject is binding; reviewer findings are advisory. Don't conflate them.
• --only {wiring|completeness|framework} scopes the run; the full 8-invariant PASS/FAIL gate is /blitz:sprint-review's job, not this skill's.
• This is per-change/pre-commit precision review of a diff — not the end-of-sprint gate.