qa-fix

Run instruction-driven QA: test and fix autonomously based on the user's input. Accepts either a module name or free text describing what to test.

Input parsing:

• If args match a known module name → run that module's predefined scenarios
• If args are free text (e.g., "the persona chips on generate page aren't highlighting on click") → run custom QA targeting the described issue
• If no args → ask the user what to test

Customize the module list in .claude/agents/qa/system-qa.md and this skill's Step 0 for your project. The default list is an example.

Step 0 — Parse input

Read the user's arguments from the skill invocation.

Module mode — args match a module name:

• Set MODE = module
• Set MODULE = {matched name}
• Set INSTRUCTION = null

Instruction mode — args are free text:

• Set MODE = instruction
• Set MODULE = null (infer from context during testing)
• Set INSTRUCTION = {the user's text}

Step 1 — Prepare

1. Ensure qa-reports/ directory exists:

   mkdir -p qa-reports

2. Determine the branch to test on. Check current branch:

   git branch --show-current

   Use whatever branch is currently checked out (the user likely has the code they want tested already running).

Step 2 — Run QA test agent

Spawn a sub-agent (subagent_type: "general-purpose", mode: "bypassPermissions") with the system-qa prompt adapted for the current mode.

Read .claude/agents/qa/system-qa.md for the base prompt.

Module mode: Use the base prompt as-is, providing:

• MODULE: {MODULE}
• BRANCH: current branch
• CYCLE: 1

Instruction mode: Use the base prompt but REPLACE the module test scenarios section with the user's instruction. Build a custom test plan:

CUSTOM TEST INSTRUCTION:
{INSTRUCTION}

Based on this instruction, determine:
1. Which page(s) to navigate to
2. What actions to perform
3. What to verify (expected behavior)
4. Take screenshots of any issues found

Write results to qa-reports/custom-bugs.md using the same bug report
format. Use "custom" as the module name.

Wait for the agent to complete.

Step 3 — Check results

Read the bug report:

• Module mode: qa-reports/{MODULE}-bugs.md
• Instruction mode: qa-reports/custom-bugs.md

Count OPEN bugs (ignore OBSERVATIONS).

• 0 open bugs → output "QA PASS — no bugs found" and STOP
• Open bugs exist → proceed to Step 4

Step 4 — Run fix agent (autonomous)

Read .claude/agents/qa/fullstack.md for the base prompt.

Create a fix branch:

git checkout -b fix/{MODULE_OR_CUSTOM}-qa-fix 2>/dev/null || git checkout fix/{MODULE_OR_CUSTOM}-qa-fix

Spawn a fix sub-agent (subagent_type: "general-purpose", mode: "bypassPermissions") with the fullstack prompt, providing:

• MODULE: {MODULE} or custom
• BRANCH: fix/{MODULE_OR_CUSTOM}-qa-fix
• CYCLE: current cycle number

The fix agent will:

1. Read the bug report
2. Fix all OPEN bugs
3. Run tests to verify no regressions
4. Commit, push, and create/update a PR

Wait for the agent to complete.

Step 5 — Ask user to restart servers

After the fix agent completes, ask the user to restart the servers on the fix branch before re-testing.

Wait for user confirmation before proceeding.

Step 6 — Re-test (max 3 cycles)

Track cycle count (started at 1 in Step 2).

After user confirms servers are restarted:

1. Increment cycle count
2. Go back to Step 2, but now:
   • BRANCH = the fix branch
   • CYCLE = current cycle number
3. After Step 3:
   • If 0 open bugs → output "QA PASS after {N} cycles" and STOP
   • If open bugs AND cycle < 3 → go to Step 4 again
   • If open bugs AND cycle = 3 → run one final fix (Step 4), then one verification test. Mark done regardless of result.

Step 7 — Final summary

Output:

QA-FIX COMPLETE
Mode: {module | instruction}
Target: {MODULE or first 60 chars of instruction}
Cycles: {N}
Result: {PASS | PARTIAL — N bugs remain}
PR: {PR_URL or "none (clean pass)"}
Bugs found: {total}
Bugs fixed: {count}
Bugs remaining: {count}

Rules

• ALL testing via Playwright MCP browser tools — no curl/httpx/API testing
• ALL fixes go to a PR — never push to main directly
• Fix and re-test autonomously — no approval gate between cycles
• Only pause for server restarts (Step 5)
• Max 3 fix cycles — don't loop forever
• Sub-agents must use subagent_type: "general-purpose" for Playwright access
• Do NOT modify test files unless the test itself is wrong
• Do NOT add features or refactor — only fix reported bugs
• OBSERVATIONS are informational — do NOT trigger fix cycles for them