xoai/qa

• Announce: "Sage → qa workflow." before starting work

QA Workflow

Browser-based functional testing. Finds integration failures that live in the gap between "code is correct" and "product works."

Report only — /qa does NOT fix bugs. /fix fixes.

Step 1: Prerequisites (Zone 1)

Check Lightpanda MCP availability. Attempt to call a Lightpanda MCP tool (e.g., goto with about:blank).

If available: Ask for URL:

Sage → qa workflow. Lightpanda browser connected.
What URL is the application running on?
(e.g., http://localhost:3000, https://staging.myapp.com)

Verify URL is reachable via goto. If not, report and offer to wait or proceed with code-only.

If not available:

Sage: Lightpanda MCP is not connected. /qa needs a browser to test
your application.

Setup: See sage/core/references/lightpanda-setup.md for install guide.

[B] Set up browser — I'll install Lightpanda
[C] Continue without browser — code-only QA
[X] Cancel

Pick B/C/X, or tell me what you need.

Code-only fallback (no browser): Analyze the diff, identify affected routes, review code for common integration issues:

• Mismatched API contracts (frontend expects shape X, backend returns Y)
• Missing error handlers (happy path only, no error/loading states)
• Untested flows (code paths with no test coverage)
• State management gaps (state written but never read, or vice versa)

Report notes: "Browser testing: not available — code-analysis only."

Step 2: Scope Analysis (Zone 1)

If invoked after /build (cycle context exists): Read the cycle's spec.md and plan.md. Identify affected routes, user flows, and acceptance criteria to verify.

If invoked standalone (no cycle context):

Sage: What should I test?

[1] Everything — full app smoke test (I'll discover routes)
[2] Specific flow — describe what to test
[3] Specific URL — point me at a page

Pick 1-3, type / for commands, or describe what you need.

Produce and present a test scope:

TEST SCOPE:
Routes to test:
- /dashboard (modified: added new widget)
- /settings/notifications (new page)

Flows to test:
- Dashboard loads with new widget visible
- Navigate to notification settings
- Toggle preference → verify API → verify UI update

Acceptance criteria to verify:
- [from spec] Toggle persists across page reload
- [from spec] Widget shows unread count

Ask: "Does this scope look right, or should I add/remove anything?"

Step 3: Route Testing

For each route in scope:

1. Navigate (goto) to the route
2. Check page health:
   • evaluate: check for JS console errors
   • markdown: verify page renders content (not blank/error)
   • semantic_tree: verify expected structural elements
3. Record: status (pass/fail/warning), details if fail

Step 4: Interaction Testing

For each flow in scope:

1. Discover interactive elements (interactiveElements)
2. Execute the flow step by step:
   • click buttons, links, toggles
   • fill form fields
   • After each: check errors (evaluate), verify state (semantic_tree)
3. Test edge cases from acceptance criteria:
   • Reload after state change → persists?
   • Empty form submit → validation fires?
   • Navigate away and back → state survives?
4. Record: status, steps executed, where flow broke

Step 5: QA Report (Zone 2)

Save report to .sage/work/[cycle-id]/qa-report.md (if cycle context) or .sage/docs/qa-report-[topic].md (if standalone).

Use the report template from develop/templates/qa-report-template.md.

🔒 QA REPORT CHECKPOINT:

Sage: QA complete.

Tested: {N} routes, {M} flows
Results: {pass} pass, {fail} fail, {warn} warning
Verdict: {PASS | FAIL | PASS WITH WARNINGS}

Bugs found: {count}
  Critical: {count} — [brief descriptions]
  Major: {count}
  Minor: {count}

Report: .sage/[path]/qa-report.md

[A] Approve report  [R] Revise — retest something
[F] → /fix to address bugs  [N] New session

Pick A/R/F/N, or tell me what to change.

Next steps (Zone 3):

Next steps:
  /fix            — diagnose → scope → fix → verify (reads qa-report)
  /design-review  — design quality audit
  /reflect        — review the cycle, extract learnings

Type a command, or describe what you want to do next.

Quality Criteria

Good QA output:

• Every route in scope was actually tested (not claimed-as-tested)
• Bugs have concrete reproduction steps, not vague descriptions
• Each bug includes suggested fix classification (Surgical/Moderate/Systemic)
• Evidence is included (console output, element state, response content)
• Untested routes are noted as untested, NOT claimed as pass

Enforcement Contracts

No fake browser testing: If Lightpanda is not available, do NOT fabricate browser test results. Code-only analysis is honest about what it can and cannot verify.

Report completeness: Untested routes must be listed as "not tested" in the report. Do NOT claim a route passes if you didn't load it.

Advisory only: /qa reports. It does NOT fix bugs. If the agent catches itself thinking "I'll just fix this real quick" — STOP. Save the finding to the report and let /fix handle it.

No fixing in /qa: The separation between find and fix is deliberate. An agent that finds AND fixes has incentive to minimize findings. /qa's only job is thoroughness.

Rules

• /qa reports, /fix fixes. Never combine.
• Code-only fallback is honest about its limitations.
• Every bug gets a severity AND a fix classification.
• Evidence is mandatory for fail/warning findings.
• Update manifest.md if cycle context exists.