qa-do

Discovery Questions

Check .agents/qa-project-context.md first. Inspect the project manifest, tests, current changes, configured target, and available tools before asking anything. Ask only for a missing target or requirement that changes the requested work.

Core Principles

1. Route to the most precise skill, not the nearest neighbor. A vague match to a broad skill is worse than a direct hit on a narrow one. "Reproduce this bug" goes to bug-reproduction, not the broader ai-bug-triage; "test our Stripe checkout" goes to payment-testing, not generic api-testing. When in doubt, prefer the skill whose trigger phrases name the exact artifact in the request.

2. One or two skills, in order — never a pile. Output at most two skills. A second skill earns its place only when the first leaves a clear gap (diagnose → measure, risk → checklist). If one skill covers the request, say "Direct" and execute it. Routing to three skills means the situation is ambiguous — ask a clarifying question instead.

3. The routing table is the source of truth, and it goes stale. This router is uniquely prone to drift: every skill added to skills/ is a row that may be missing here. Regenerate the table and the Skill Categories reference against skills/ whenever a skill is added or removed. If a request has no row, fall back to the closest category in the Quick Reference rather than forcing a wrong direct match.

How to Use

Describe what you're trying to do or what problem you're facing — a sentence or two is enough. The router outputs the recommended skill(s) (1-2, in order) and one line per skill explaining the role it plays. Example: "Our E2E tests keep failing in CI but pass locally." → test-reliability (diagnose flaky/environment-sensitive tests), then ci-cd-integration (align the pipeline environment with local behavior).

Execute an Application Check

When the full repository is available, use its bounded runner. Resolve <qa-skills> to this checkout and <project> to the application, even when the working directory differs.

node <qa-skills>/scripts/qa.mjs plan --project <project> --goal "<user goal>"
node <qa-skills>/scripts/qa.mjs run --project <project> --goal "<user goal>"

Read the primary skill and next relevant skill in the plan, then perform their workflows. The runner discovers available test commands and writes report.html, result.json, logs, and a host handoff. A plan is not an executed test. Where only individual skills are installed, execute their recipes directly and retain equivalent evidence.

For URL checks, use a concrete target and application-owned readiness selector. Run the configured checks without treating optional missing integrations as a reason to stop unrelated local work. A missing browser, unavailable target, no observed tests, or an unreadable test result remains blocked or incomplete. Preserve skipped cases.

For a failure, reproduce it, preserve its assertion, and apply a repair only when it is within the user's task. Execute again after changing the code. Stop after two attempts if the failure persists and report the exact blocker; do not weaken an assertion or retry indefinitely. The report covers executed checks, not universal correctness. See the repository's docs/QA-RUNNER.md for configuration and host integration.

Common Situations and Their Skills

When the Situation is Ambiguous

If a description maps to three or more skills with equal weight, one clarifying question narrows it down. Answer it and the router reduces to 1-2 skills.

• "Are you fixing something broken, or building new coverage from scratch?"
• "Is this a process problem (how the team works) or a tooling problem (what's running)?"
• "Is the priority speed of delivery, or confidence in correctness?"
• "Are you the only QA, or is this a team-wide change?"

Disambiguation pairs (the four overlaps to resolve, not guess)

• test-reliability vs selector-drift-recovery — one flaky test healed at runtime → test-reliability; many selectors broken by a planned UI refactor → selector-drift-recovery.
• cross-browser-testing vs mobile-testing — browsers/CSS engines → cross-browser-testing; real devices, Appium/Detox/Maestro, gestures/deep links → mobile-testing.
• ai-bug-triage vs bug-reproduction — classify/dedupe a batch of CI failures → ai-bug-triage; reproduce and understand one specific bug → bug-reproduction.
• qa-start vs qa-project-bootstrap — brand-new project, no QA yet → qa-start; onboarding a QA engineer to an existing codebase → qa-project-bootstrap.

Skill Categories Quick Reference

Regenerate this from the category: field of every skills/*/SKILL.md whenever skills change.

Anti-Patterns

• Routing to a broad skill when a precise one exists. "Reproduce this bug" → routing to ai-bug-triage (batch classification) instead of bug-reproduction (single verified repro). Fix: match the exact artifact in the request to the skill that names it.
• Stacking three or more skills on one situation. That signals ambiguity, not thoroughness. Fix: ask one clarifying question and reduce to 1-2.
• Trusting a stale table. If a recently added skill has no row, the router silently mis-routes to a weaker neighbor. Fix: regenerate the table from skills/ (Core Principle 3) and fall back to the closest Quick Reference category rather than forcing a wrong match.
• Grabbing a request that belongs to a named sibling. "Set up QA on a new project" is qa-start; "capture project context" is qa-project-context. qa-do is last resort only.

Done When

• The next 1-2 skills are named with reasons tied to the current task.
• An action request has executed checks or a specific recorded blocker.
• The report identifies the target, source revision, command, outcome, and remaining gaps.
• Authored, executed, failed, skipped, blocked, and unrun checks are distinguishable.

Related Skills

• qa-start — the sibling most confused with this one. Use it (not qa-do) when starting QA on a brand-new project with no QA in place; it chains context → strategy → planning.
• qa-project-context — capture project setup before using most skills; every skill checks for it first. Route here, don't reimplement it.
• test-strategy — when the situation is "we need a QA strategy" rather than a specific problem to route.