qa-web

Project-scoped skill for QA of web/ — a Next.js app with Vitest unit tests and Playwright E2E. Invoked by the qa-web-agent subagent (for sandboxed black-box exploration) or directly by the main session (for inline QA work).

Overview

You are testing a Next.js dashboard. The app has an existing test suite (Vitest + Playwright projects fast/full/demo/qa-explore), a behavior ledger at web/tests/LEDGER.md, and an evidence convention at output/qa-agent/<date>/<slug>/.

Your four phases:

Scope — inspect what changed on this branch; produce a ranked list of user-visible surfaces to probe. Skip if on main with no caller summary.
Explore — black-box heuristic testing (SFDIPOT, FEW HICCUPPS, a11y) scoped to the P0/P1 surfaces from Phase 0. Every finding gets a screenshot + axe snapshot + repro steps.
Author — convert findings and uncovered behaviors into spec-first tests. Human approval gate on the spec, then delegate code generation to the playwright-test-generator subagent (installed by mise run web:qa:init-agents).
Heal — triage failing Playwright tests via the playwright-test-healer subagent; patch selector drift, escalate behavior changes.

Normal flow

Follow this order every time the skill is invoked: