Done — Verification Gate

A “done” claim without a verification command and exit code is a guess. This skill makes the gate mechanical.

Protocol

Detect entrypoint. From the repo root:
- If jq -e '.scripts."ci:local"' package.json succeeds → run npm run ci:local.
- Else discover available scripts: typecheck, lint, test, build. Run them in that order. Skip silently if a script does not exist.
- For monorepos with frontend/ and backend/ workspaces (Wraith pattern): also run the frontend variants (typecheck:frontend, lint:frontend, test:frontend).

Capture exit code per stage. Print one line per stage:

[done] typecheck: exit=0
[done] lint:      exit=0
[done] test:      exit=1   <- failure

Refuse green summary on any failure. If any stage exits non-zero:
- Print the failing stage’s last 30 lines of output.
- Print STATUS: FAIL — do not claim done.
- Do NOT commit, push, or mark the task complete.
On full pass: print STATUS: PASS — all gates green. Then proceed with whatever the user asked next (commit, PR, summary, etc.).

Reference — Wraith Repo Specifics

/path/to/your-project exposes npm run ci:local which expands to:

npm run lint && npm run typecheck && npm run test:backend \
  && npm run lint:frontend && npm run typecheck:frontend \
  && npm --prefix frontend run test:run

This is the preferred entrypoint there. frontend/src/generated/ is excluded by lint-staged via --ignore-pattern; the skill inherits that — it only runs project scripts and does not file-walk.

Hard Rules

Run from the repo root. If you’re in a subdirectory, cd to root first.
Never silence non-zero exits with || true or 2>/dev/null.
Long stages (build, full test) → use run_in_background and Monitor. Don’t poll with sleep loops.
If npm install is needed first, that’s a different skill (/check-env). Do not auto-install dependencies.

When Not To Run

The user asked a research-only or exploratory question.
You haven’t actually made any changes this turn.
A specific narrower check was requested (e.g. “just typecheck”).

Why This Skill Exists

The /insights report (2026-04-27) flagged “premature completion claims” as the #1 recurring friction across 82 sessions. Existing prose rules in CLAUDE.md (“verify before claiming done”) were not enough — Claude declared done with failing tsc/vitest 3–4 times per cycle. This skill replaces the prose with a mechanical gate.