feature-qc

name	feature-qc
description	Phase 3 of the feature flow, standalone: the full QC battery over the current feature branch. Use when the user says /feature-qc, "run QC", "quality pass", or wants the branch proven buildable+bootable mid-flight. For just one pass, use /simplify, /code-review, or /feature-lint directly instead.
allowed-tools	Bash(git *) Bash(gh *) Bash(pnpm *)
model	inherit

The QC battery — autonomous

Over the whole branch diff, in order (skip nothing silently — report each step):

1. Convergence: all commits on the feature branch; no stray flow worktrees under .claude/worktrees/; no stray branches.
2. Review fan-out — one Workflow({ scriptPath: ".claude/workflows/qc-review.mjs", args }) call runs ALL finders against the frozen branch diff in a single parallel barrier: the two cleanup-finder angles (reuse+simplification, altitude+efficiency)
   • conventions-finder on sonnet; the bug-finder adversarial + cross-file angles on opus, plus a line-by-line angle (sonnet) that runs only on large diffs (models pinned in the workflow, not prose). Address it by scriptPath, never name — Workflow({ name }) resolves only built-in/registered workflows and does NOT scan the repo's .claude/workflows/, so { name: "qc-review" } silently 404s and degrades to the unbounded /code-review path; the path form runs the repo workflow directly. Measure the diff first (git diff --shortstat <range>) and pass args: { range: "origin/dev...ft/<N>", generated: "<globs>", vetoes: "<plan-frozen decisions>", criteria: "<the ft issue's 'Stack & design acceptance criteria' section>", large: <bool>, effort: "medium" } — criteria is what conventions-finder verifies the built diff against; set large: true on a big diff (roughly >8 files or >200 changed lines) to add the line-by-line bug angle (it returns zero on small diffs), and effort: "high" when the slice adds a table/migration, a new trust boundary (auth, server action, agent tool surface), or touches posting/money paths. It returns a consolidated findings list.
3. Adjudicate + apply (this session). The workflow only reports — the session decides. Plan-frozen decisions in the ft issue are vetoes, not findings; drop them. Apply the survivors. A finding that is real but not-this-slice (a bigger refactor, a scale concern that can't bite yet) → append it to the single living backlog via .claude/skills/feature/scripts/backlog-add.sh "<item; origin #<issue> QC; · agent>" (never a new per-item issue). The applied fix diff stays gated by the tsc + lint pass (step 4) and boot smoke (step 5) — no separate delta-verify pass. Large/risky diff → offer the user /code-review ultra before proceeding.
4. feature-lint (scoped to the feature's changed files — LAST because the review pass mutates code; lint formats the final shape) — biome format + safe fixes + residual fixer agents, gating on a clean pnpm build — the authority on compile correctness.
5. Boot smoke — builds can't see boot failures: background pnpm dev, wait for readiness, assert Next.js reports "Ready" and NO failure signatures (ERROR, "failed", unmet peer, unhandled rejection). Collect WARNINGs for triage; kill the process. Startup output only.
6. Doc sync — subtractive first (the revise-agents-md philosophy at slice scope; ships in the same diff). Default outcome is no change — say so plainly rather than invent additions. A fact earns a doc line only if it is durable, action-affecting, and NOT recoverable from the code a fresh session reads; adding is guilty until proven load-bearing. In order:
   • Subtract what the slice made stale — any AGENTS.md / .claude/rules/ / skill line the diff falsified or made code-recoverable → delete it.
   • Add only a genuine non-recoverable keeper (a new guard, a retired pattern, a new trust boundary): AGENTS.md if always-on, the area's .claude/rules/<area>.md if scoped — create a new nested rule file for a brand-new path-area.
   • Skills: if the slice changed what a skill's body documents (a command, a wiring contract), fix that skill; deeper skill bloat → surface it for /meta-dev:improve-skill, never inline-rewrite it here. Single-source every fact (one home; cross-reference, never restate).

Hard rules: the review fan-out is one barrier of ≤6 finders (5 on a small diff, 6 on a large one) with no delta-verify — well under the ≤10-agents-per-fan-out cap; the qc-review workflow (invoked by scriptPath, see step 2) owns finder parallelism and model pins. Never fall back to /code-review for the fan-out — its per-candidate verify phase is unbounded and defeats the cap the workflow exists to enforce. If any step reveals a dependency MAJOR upgrade, framework migration, or schema/data migration is required — STOP and present options; never fix those autonomously. End by stating: builds ✓ boots ✓ findings fixed ✓ (or what remains).