qa — a Claude Code QA plugin

Autonomous, evidence-gated QA for any web app. Parallel browser agents explore, break, and audit your site; every finding is independently reproduced in a fresh browser before it reaches the report. Install once, onboard any project with /qa:init.

Playwright-based QA automation, reusable across projects. Install the plugin once; onboard any project with /qa:init. All three build phases are validated against the fixture pilot — 29 unit/integration tests and a demo site with 8 planted bugs (see Dev below).

The engineering, not the autocomplete

The hard part of agentic QA isn't getting a model to click around — it's stopping it from reporting confident nonsense. This plugin is built around a few deliberate decisions that make agent output trustworthy:

• A single findings contract every producer emits and every consumer reads — so agents, deterministic audit scripts, suite reporters, and visual diffs all flow through one pipeline.
• An evidence rule that discards any finding without a screenshot + repro/codeRef before it can reach a report.
• A verifier gate that re-runs agent-claimed bugs in a clean browser and throws out the ones that don't reproduce.

Those decisions are the point. They were designed and architected by Benjamin Franck and built at speed by driving Anthropic's Fable 5 — frontier models make the output fast; the architecture makes it reliable.

Commands

Command	What it does
/qa:init	Onboard a project: detect stack, crawl routes, interview for flows/auth, scaffold qa/ + qa.config.json; optional persistent-suite setup and headed auth capture
/qa:explore [target] [--video]	Parallel exploratory session: explorer(s) + mobile-tester + breaker agents in isolated browsers → evidence-backed findings → verifier gate → HTML report
/qa:test [generate|run|heal] [flow]	Persistent suites: generate specs from flows (selectors verified live), run them through the findings pipeline, heal stale tests with an honest stale-vs-bug verdict
/qa:audit [a11y|perf|links|seo|design|all]	Deterministic audit scripts (axe-core, web vitals, link/asset sweep, SEO checks) + impeccable design dimension when available
/qa:visual [baseline|check]	Visual regression: per-route × viewport baselines, pixelmatch diffs, regression-vs-intended triage
/qa:full [target] [--no-fix]	Everything above in parallel into one unified, verified report — then the fix loop
/qa:fix [run-id]	Fix loop: reproduce-first → minimal fix → re-verify by replaying the finding's own repro (max 3 attempts, then blocked + diagnosis)
/qa:report [run-id]	Render/open the self-contained HTML report for any run
/qa:status	Read-only digest: last run, open findings, suite health, suggested next action

How it works

• Findings schema (skills/qa-core): every producer — agent, audit script, suite reporter, visual diff — emits the same shape; every consumer (verifier, fix loop, report) reads it. The evidence rule discards any finding without a screenshot + repro/codeRef before it can reach a report.
• Verifier gate: agent-claimed findings are independently reproduced in a fresh browser before reporting; deterministic script/suite findings carry machine evidence (merge --trust-verified).
• Browser layer: @playwright/cli (pinned) for agent sessions; the playwright library + axe-core/pixelmatch for scripts; everything installs into ${CLAUDE_PLUGIN_DATA} — host projects stay clean.

Dev

• npm test — 29 unit/integration tests (findings, reporter, renderer, audits vs fixture site, visual compare)
• npm run demo — fixture site with 8 planted bugs at http://localhost:4173 (PORT= to override); ground truth in fixtures/demo-site/BUGS.md
• claude plugin validate . — plugin structure check
• Try locally: claude --plugin-dir .