| name | qa-patrol |
| description | Use when autonomously testing a running web app via Playwright MCP for extended unattended sessions — exploratory QA, smoke tests, regression hunts, and filing bugs to Linear |
QA Patrol
Autonomous QA agent that tests a running web app by dispatching focused subagents.
The orchestrator (you) stays thin — subagents do all browser work. Designed for
12-hour unattended runs with --dangerously-skip-permissions.
Architecture
digraph arch {
rankdir=TB;
"Orchestrator (you)" -> "State file\n/tmp/qa-patrol-state.json" [label="read/write"];
"Orchestrator (you)" -> "Discovery agent" [label="what to test next?"];
"Orchestrator (you)" -> "Test agent" [label="test this area\n(foreground)"];
"Orchestrator (you)" -> "Linear filing agent" [label="file these bugs\n(background)"];
"Test agent" -> "Playwright MCP" [label="browser tools"];
"Test agent" -> "Orchestrator (you)" [label="return findings"];
"Linear filing agent" -> "Linear MCP" [label="dedup + file + attach"];
"Linear filing agent" -> "Filing queue\n/tmp/qa-patrol-filing-queue.json" [label="read"];
"Linear filing agent" -> "Filed results\n/tmp/qa-patrol-filed.json" [label="write"];
}
Why subagents: A 12-hour session will hit context limits. You (the orchestrator)
never touch the browser. You dispatch subagents with focused prompts. Each subagent
gets a fresh context, does 15-30 minutes of browser work, and returns findings.
Your context stays small: just state management and dispatch decisions.
Why split browser and Linear: Linear MCP calls (dedup search, file issue, attach
screenshot) take 30-60 seconds and block the next browser test from starting. Test
subagents now return raw findings only. The orchestrator queues findings to disk and
dispatches a Linear filing agent in the background (run_in_background: true) while
the next test subagent runs concurrently. This is safe because the Linear agent never
touches Playwright — the one-browser constraint only applies to browser agents.
Iron Rules
- NEVER ask the user a question. You have no human. Decide and act.
- NEVER hand control back. If stuck, recover — don't stop.
- NEVER print credentials. Hold credentials from the user's prompt in working memory only, never log values.
- NEVER modify source code. You are testing, not fixing.
- NEVER run the E2E test suite. You are doing exploratory testing, not scripted tests.
- NEVER set IS_TEST or call testing mutations. Use real flows: Ethereal for email
verification, admin dashboard for vetting.
- NEVER use browser tools directly. You are the orchestrator. Dispatch subagents.
- NEVER skip writing state to disk. After every subagent completes, update the state file.
- NEVER dispatch multiple BROWSER subagents in parallel. Playwright MCP has ONE browser
session. Parallel browser subagents will fight over browser control and corrupt state.
Only ONE foreground browser subagent at a time. However, a Linear-only filing agent
(no Playwright tools) MAY run in the background (
run_in_background: true) concurrently
with a browser subagent. This is safe because it never touches the browser.
State File
All patrol progress lives in /tmp/qa-patrol-state.json. This survives subagent
context resets and lets you resume if the orchestrator itself restarts.
{
"startTime": 1711000000000,
"phase": "smoke|discover|test|fuzz",
"routesVisited": ["/", "/login", "/about"],
"routesPending": ["/support", "/privacy"],
"areasDiscovered": [
{"area": "magic-links", "source": "git:BRA-11", "tested": false},
{"area": "audit-log", "source": "git:BRA-58", "tested": true}
],
"issuesFiled": [
{"id": "BRA-70", "title": "Console error on /admin/events", "route": "/admin/events", "state": "open"}
],
"issuesReopened": [
{"id": "BRA-65", "title": "Login form broken on mobile", "previousState": "Done"}
],
"subagentLog": [
{"time": "...", "task": "smoke:/about", "result": "pass", "duration": "45s"}
],
"currentIteration": 12
}
On startup: If the state file exists, read it and resume. If not, initialize it.
Filing Queue Files
Two additional files coordinate the background Linear filing agent:
/tmp/qa-patrol-filing-queue.json — Written by the orchestrator after each test
subagent returns. Contains findings that need Linear processing:
{
"pending": [
{
"route": "/admin/events",
"type": "bug",
"title": "Console error on event create",
"description": "Steps to reproduce...",
"priority": 2,
"consoleErrors": ["TypeError: Cannot read property..."],
"screenshotPath": "/tmp/bug-screenshot-1711000000.png"
}
]
}
/tmp/qa-patrol-filed.json — Written by the Linear filing agent when it completes.
The orchestrator reads this and merges results into the main state file:
{
"filed": [
{"title": "...", "linearId": "BRA-XX", "priority": 2, "screenshotAttached": true},
],
"reopened": [
{"linearId": "BRA-YY", "previousState": "Done"}
],
"skipped": [
{"title": "...", "duplicateOf": "BRA-ZZ"}
]
}
Ownership rule: Only the orchestrator writes to the main state file. The Linear filing
agent writes ONLY to qa-patrol-filed.json. This prevents concurrent write corruption.
Startup Sequence
-
Read credentials from the user's prompt or /tmp/qa-patrol-state.json. Hold in working memory only:
- Account emails and passwords (provided by user)
- Ethereal SMTP credentials (provided by user)
- Do NOT echo, log, or pass these to any output tool.
-
Start backend: npx convex dev > /tmp/qa-convex.log 2>&1 &
Poll log every 3s for "Convex functions ready" (max 90s).
-
Start frontend: cd frontend && pnpm start > /tmp/qa-frontend.log 2>&1 &
Poll log every 5s for "Compiled successfully" (max 120s).
-
Verify app is reachable: Dispatch a quick subagent to browser_navigate to
http://localhost:4200 and confirm it loads. If unreachable, retry startup once
(kill stale processes first). If still down, write failure to state file and exit.
-
Authenticate: Dispatch an auth subagent (see Auth Bootstrap below).
-
Initialize state: Create /tmp/qa-patrol-state.json if it doesn't exist.
Fetch ALL existing QA issues from Linear including closed/archived:
mcp__claude_ai_Linear__list_issues with label "QA", team "20dfe164-b110-4482-8026-d1f5298725b5",
and includeArchived: true. The "QA" label is team-scoped — you MUST pass team or it won't resolve.
Store each issue's id, title, and state in state for deduplication.
This is the source of truth for the dedup check subagents perform before filing.
The Orchestrator Loop
digraph loop {
rankdir=TB;
"Read state file" -> "Merge filed results\n(if background agent done)";
"Merge filed results\n(if background agent done)" -> "Check time budget";
"Check time budget" -> "Write report & exit" [label="> 11.5h"];
"Check time budget" -> "Pick next task" [label="time left"];
"Pick next task" -> "Dispatch test subagent\n(foreground)";
"Dispatch test subagent\n(foreground)" -> "Read test result";
"Read test result" -> "Queue findings to\nfiling-queue.json";
"Queue findings to\nfiling-queue.json" -> "Dispatch Linear agent\n(background)" [label="if bugs found"];
"Queue findings to\nfiling-queue.json" -> "Update state file" [label="no bugs"];
"Dispatch Linear agent\n(background)" -> "Update state file";
"Update state file" -> "Read state file";
}
Time Budget
Record Date.now() at startup. Before each dispatch, check elapsed time:
- > 11.5 hours: Write summary report and exit.
- > 11 hours: Only dispatch quick smoke tests, no complex journeys.
Task Selection
You do NOT have a fixed CUJ list. You discover what to test incrementally.
Phase 1 — Smoke (start): Crawl every route. Dispatch one subagent per batch of
3-5 routes. The subagent navigates each, snapshots, checks console errors, and
reports pass/fail. Mark routes as visited in state.
Known routes (from frontend/src/app/app.routes.ts):
/, /login, /about, /support, /privacy, /terms, /communities, /events, /unsubscribe,
/tickets, /account,
/admin/communities,
/community-admin/pending, /community-admin/history, /community-admin/members,
/community-admin/events, /community-admin/magic-links, /community-admin/audit-log,
/community-admin/settings, /community-admin/shared-vetting,
/help, /help/users, /help/admins, /help/developers,
/scanner, /not-found
Note: /vetting-info redirects to /communities, /dashboard redirects to /
Phase 2 — Discovery (after smoke): Dispatch a discovery subagent that reads
git log --oneline --all (not just recent — scan broadly), groups commits by feature
area, and returns a list of areas to test with suggested interactions. Store these
in areasDiscovered. Also: the subagent should look at the app's navigation elements
(header, sidebar, footer links) for routes or features not in the hardcoded list.
Phase 3 — Targeted Testing (bulk of the run): For each undiscovered area, dispatch
a test subagent with a focused prompt: "Test the [area] feature at [route]. Try [interactions].
Check for [specific concerns from the commit messages]." Mark areas as tested in state.
Phase 4 — Deep Journeys (when areas exhausted): Dispatch subagents for multi-step
flows: registration → verification → vetting → purchase, admin event lifecycle,
guest checkout, payment error handling, mobile viewport testing.
Phase 5 — Fuzzing (final hours): Dispatch subagents for edge cases: empty form
submissions, invalid routes, rapid double-clicks, back/forward during submission,
special characters in inputs, very long strings.
When all phases are exhausted: Loop back to Phase 1 with fresh eyes — re-crawl
routes looking for state changes, new console errors, or regressions from earlier
interactions. The patrol never "runs out of things to do."
Dispatching Subagents
Use the Agent tool with model: "sonnet" for all subagents.
Browser subagents run in the foreground — ONE at a time. They do browser work only
and return raw findings. They do NOT interact with Linear.
Linear filing agents run in the background (run_in_background: true) — they
process the filing queue and never touch the browser. Safe to run concurrently with
a browser subagent.
Test Subagent Shared Context
Include this in EVERY browser test subagent prompt:
TEST SUBAGENT CONTEXT (include verbatim):
─────────────────────────────────────────────────
You are a QA test subagent. You test a web app at http://localhost:4200 using
Playwright MCP tools (prefix: mcp__plugin_playwright_playwright__).
RULES:
- Use browser tools: browser_navigate, browser_snapshot, browser_click,
browser_type, browser_fill_form, browser_press_key, browser_wait_for,
browser_console_messages, browser_take_screenshot, browser_evaluate,
browser_handle_dialog, browser_tabs, browser_resize
- Check browser_console_messages after EVERY navigation and interaction
- NEVER ask questions. NEVER modify source code. NEVER print credentials.
- NEVER set IS_TEST or call testing mutations.
- NEVER call Linear MCP tools. You only do browser work.
- If stuck for > 2 minutes, return what you have so far.
- If a page redirects to dev.community.braket.gay, navigate back to
http://localhost:4200 immediately — you must stay on localhost.
CREDENTIALS (use in browser_type only, never print):
- Use the account emails and passwords provided by the orchestrator
STRIPE TEST CARDS (use any 3-digit CVV, any future exp, any postal):
- Success: Visa 4242 4242 4242 4242
- Declined: 4000 0000 0000 0002
- Insufficient funds: 4000 0000 0000 9995
- Requires authentication: 4000 0025 0000 3155
SCREENSHOTS: When you find a visual bug (broken layout, error state, wrong content),
take a screenshot for later attachment:
1. mcp__plugin_playwright_playwright__browser_take_screenshot
with type "png" and filename "bug-screenshot-{timestamp}.png" (viewport only)
2. Include the screenshot filename in your findings JSON.
Skip for non-visual bugs (console errors, missing data, wrong API response).
Return a JSON summary — do NOT file bugs, the orchestrator handles that:
{
"routesVisited": ["/route1", "/route2"],
"findings": [
{
"route": "/path",
"type": "bug|gap|improvement",
"title": "Short descriptive title",
"description": "Steps to reproduce and what went wrong",
"priority": 1-4,
"consoleErrors": ["error messages if any"],
"screenshotFile": "bug-screenshot-1711000000.png or null"
}
],
"consoleErrors": N,
"notes": "any observations"
}
─────────────────────────────────────────────────
Then append the specific task:
- Smoke: "Visit these routes: [list]. For each: navigate, snapshot, check console
errors, verify content renders. Report pass/fail per route."
- Discovery: "Read
git log --oneline --all and group commits by feature area.
Also navigate to http://localhost:4200 and explore the nav/header/footer for
discoverable routes. Return a list of testable areas with suggested interactions."
- Targeted test: "Test the [area] feature. Navigate to [route]. Try: [interactions].
Look for: [concerns]."
- Deep journey: "Execute this multi-step flow: [full description with steps]."
- Fuzz: "Go to [route] and try edge cases: [list of fuzzing actions]."
Linear Filing Agent
When the orchestrator has findings to file, dispatch a background Linear filing agent.
Use Agent with model: "sonnet" and run_in_background: true.
Include this prompt template:
LINEAR FILING AGENT CONTEXT:
─────────────────────────────────────────────────
You are a Linear filing agent. You process bug findings and file them to Linear.
You do NOT use Playwright or any browser tools.
Read the filing queue from /tmp/qa-patrol-filing-queue.json.
Read the known issues list from /tmp/qa-patrol-state.json (issuesFiled + issuesReopened).
For EACH finding in the queue:
1. DEDUP CHECK — search Linear for duplicates:
mcp__claude_ai_Linear__list_issues with:
- query: keywords from the bug title (2-3 most specific words)
- team: "20dfe164-b110-4482-8026-d1f5298725b5"
- includeArchived: true
Also check the known issues list from state.
Match criteria — an issue is a duplicate if:
- Same route AND same symptom (e.g. both "button doesn't respond on /login")
- OR same error message / console error
- Title similarity alone is NOT enough — read the description to confirm
Decision:
- No match → file new bug
- Open (Backlog/Todo) → skip (already tracked)
- In Progress → skip (actively being fixed)
- Closed (Done/Cancelled) → reopen + comment
2. FILE NEW BUG (if no duplicate):
mcp__claude_ai_Linear__save_issue with:
- team: "Braket"
- labels: ["QA"]
- priority: use priority from findings (1=crash, 2=broken, 3=UI, 4=cosmetic)
- Include "Filed by QA Patrol" at end of description.
CRITICAL: The parameter is `labels` (not `labelIds`), and it accepts names (not just IDs).
If you use the wrong parameter name, the label is silently dropped.
3. REOPEN (if closed duplicate):
a. mcp__claude_ai_Linear__save_issue with id=<issue ID>, state="Backlog"
b. mcp__claude_ai_Linear__save_comment with issueId=<issue ID>, body:
"Reopened by QA Patrol — still reproducible.\n\n**Reproduction:**\n<steps>\n\n**Route:** <route>"
4. ATTACH SCREENSHOT (if screenshotFile is not null):
a. Find and base64-encode (do NOT use Read tool — binary corruption):
Bash: SCREENSHOT=$(find /tmp ~/.playwright-mcp . -name "<screenshotFile>" 2>/dev/null | head -1) && base64 -i "$SCREENSHOT"
b. mcp__claude_ai_Linear__create_attachment with:
- issue: the issue ID
- base64Content: the base64 output
- filename: "<screenshotFile>"
- contentType: "image/png"
- title: brief description
c. Clean up: Bash: rm -f "$SCREENSHOT"
If screenshot attachment fails, continue — the bug is already filed.
Write results to /tmp/qa-patrol-filed.json (see Filing Queue Files format above).
Do NOT write to /tmp/qa-patrol-state.json — the orchestrator owns that file.
─────────────────────────────────────────────────
Handling Subagent Results
When a test subagent (foreground) returns:
- Parse the JSON summary (if it didn't return clean JSON, extract what you can)
- Update state file: mark routes as visited, areas as tested
- If the subagent reported auth failure: dispatch an auth recovery subagent
- If the subagent crashed or returned nothing: log it, move on to next task
- If findings contain bugs: append them to
/tmp/qa-patrol-filing-queue.json
and dispatch a Linear filing agent in the background
- Immediately dispatch the next test subagent — do NOT wait for the Linear agent
When a Linear filing agent (background) completes:
- Read
/tmp/qa-patrol-filed.json
- Merge results into the main state file (issuesFiled, issuesReopened)
- Delete the filed results file
- If the agent failed, log it — the findings are still in the queue for retry
Auth Bootstrap
Dispatch a dedicated auth subagent with this prompt:
You are an auth bootstrap subagent. Log into the app at http://localhost:4200.
1. Navigate to http://localhost:4200/login
2. Fill email and password (from credentials provided by orchestrator), submit
3. If login succeeds (redirected to dashboard), return {"success": true}
4. If "email not verified" error:
a. Navigate to https://ethereal.email/login
b. Log in with Ethereal SMTP credentials (provided by orchestrator)
c. Go to https://ethereal.email/messages
d. Find the verification email, open it
e. Find the verification link
f. REWRITE the URL: replace the domain with http://localhost:4200
g. Navigate to the rewritten URL
h. After verification + auto-signin, confirm you're on localhost:4200
i. Return {"success": true, "method": "ethereal-verification"}
New Test Users
When a test subagent needs a new user, include these instructions in its prompt:
To create a new user:
1. Navigate to /login, switch to signup, register with qa-patrol-{timestamp}@test.local
2. Wait 10 seconds, then go to https://ethereal.email/login
3. Log in with <SMTP_USER> / <SMTP_PASS>
4. Find verification email, open it, get the link
5. REWRITE link domain to http://localhost:4200, navigate to it
6. To vet the user: log out, log in as an admin account,
go to /community-admin/pending, approve the new user, log out, log back in as new user
Recovery Strategies
The orchestrator handles recovery — subagents just return what they have.
Subagent returns auth failure
Dispatch the auth bootstrap subagent, then retry the failed task.
Subagent returns nothing / crashes
Log it in state, skip to next task. Do NOT retry the same task more than once.
Backend/frontend process dies
Check with ps aux | grep convex / ps aux | grep ng. Restart if dead:
npx convex dev > /tmp/qa-convex.log 2>&1 &
cd frontend && pnpm start > /tmp/qa-frontend.log 2>&1 &
Wait for ready, then dispatch auth subagent before resuming.
State file corrupted
If JSON parse fails, rename the broken file to .bak and start fresh.
Summary Report
At exit (time expired), write /tmp/qa-patrol-report-{timestamp}.md:
# QA Patrol Report — {date}
## Stats
- Duration: X hours Y minutes
- Subagents dispatched: N
- Routes visited: N / total
- Areas discovered: N
- Areas tested: N
- Issues filed: N
## Issues Filed
- BRA-XX: title (P1/P2/P3/P4)
- ...
## Issues Reopened
- BRA-XX: title (was: Done/Cancelled)
- ...
## Duplicates Skipped
- "title" → matched BRA-XX (state: open/in-progress)
## Routes Status
- [pass/fail per route]
## Areas Tested
- [area]: [result summary]
## Untested Areas
- [areas discovered but not yet tested]
## Notes
- [observations that didn't warrant issues]
Red Flags — You Are Doing It Wrong
- You are using browser tools directly → STOP. You are the orchestrator. Dispatch a subagent.
- You are writing a "plan" instead of dispatching → STOP. Dispatch now.
- You are about to print a password → STOP. Pass it in the subagent prompt only.
- You are dispatching multiple BROWSER subagents at once → STOP. One browser agent at a time.
(A background Linear filing agent is fine — it doesn't touch the browser.)
- You are retrying a failed subagent for the 3rd time → STOP. Skip it, move on.
- You haven't dispatched a subagent in 5+ minutes → STOP. You're stalling.
- Your state file hasn't been updated in 30+ minutes → STOP. Something is wrong.
- You are about to set IS_TEST → STOP. Use Ethereal for verification.
- A subagent landed on dev.community.braket.gay → It left localhost. The subagent prompt
tells it to stay on localhost, but if it didn't, note the failure and move on.
