| name | full-qa |
| description | Generic autonomous QA suite for any project — reads a test plan (markdown file or inline spec), runs all tests using Chrome CDP + iterate cycles, fixes bugs found, and produces a final QA report. Works with any web stack. Trigger with /full-qa [path-to-test-plan.md]. Use before releases, after significant changes, or whenever you want a full system health check. |
| author | DevOtts |
| author_url | https://github.com/DevOtts |
/full-qa — Autonomous Full QA Suite
You are running an autonomous, project-agnostic QA pipeline. You read a test plan, execute every test, fix bugs as you find them, and deliver a final pass/fail report — without stopping between steps unless a destructive or truly ambiguous action requires confirmation.
This skill incorporates Chrome CDP browser control and iterative bug-fix cycles natively. You do not need to invoke /chrome-cdp-control or /iterate separately.
CDP mechanics live in one shared reference — read ../references/cdp-core.md (relative to this skill's base directory) before any browser work. It canonically owns endpoint resolution (CDP_URL/APP_URL env → grounding/test plan → defaults), the action template, tab selection, the selector ladder, waits, and the failure protocol. Never restate them here; never hardcode a CDP endpoint or app port — parallel runs collide on hardcoded values.
Route guard (cdp-core.md §9): autonomous mode is for TEST ENVIRONMENTS ONLY. If any test case targets the user's authenticated real-Chrome session (their logged-in accounts — posting, sending, buying), REFUSE to run that case autonomously and re-route it to /chrome-cdp-control, whose per-write gate requires explicit user confirmation for each write. Unattended (user asleep) that case ends as BLOCKED/deferred with the reason. There is no autonomous write path on an authenticated session.
Step 0 — INGEST THE TEST PLAN
First action: identify the test plan.
- If the user passed a file path (e.g.,
/full-qa ./E2E-test-plan.md), read that file immediately.
- If the user pasted tests inline, extract them from the conversation.
- If neither: ask the user ONE question: "Where is the test plan? Paste it or give me a file path."
From the test plan, extract:
- Service URLs — all
localhost:PORT or external URLs mentioned
- Auth credentials — any test user emails/passwords
- Test cases — each scenario with its ID, steps, and pass criteria
- Setup steps — any DB resets, seed commands, or imports required before tests
- Stack info — language, framework, DB type (for tailoring fix strategies)
If any of these are missing from the plan, infer reasonable defaults and state your assumptions before starting Phase 1.
Phase 1 — PREFLIGHT
Verify every service the test plan references is alive. Do this silently and fix anything that is down before proceeding.
1.1 Service health checks
For each URL in the test plan:
curl -s -o /dev/null -w "%{http_code}" <SERVICE_URL>/health
curl -s -o /dev/null -w "%{http_code}" <SERVICE_URL>
Expected: 200 (or the status code the plan specifies). Any non-200 is a failure to investigate.
1.2 Chrome CDP check
Resolve the CDP endpoint per cdp-core.md §1 (env CDP_URL → grounding/test plan → the core's default), then:
curl -s "$CDP_URL/json/version" | python3 -c "import sys,json; d=json.load(sys.stdin); print('Chrome:', d.get('Browser','?'))"
If CDP fails, stop and give the user the relaunch instruction from cdp-core.md §2, then say: "Then run /full-qa again."
1.3 Playwright check
python3 -c "import playwright; print('playwright ok')" 2>&1 || pip3 install playwright && python3 -m playwright install chromium
1.4 Preflight report
Print a compact table before moving on:
PREFLIGHT
✓ Service A http://localhost:XXXX 200
✓ Service B http://localhost:YYYY 200
✓ Chrome CDP $CDP_URL Chrome/XXX
✗ Service C http://localhost:ZZZZ ECONNREFUSED ← fixing...
Do not proceed to Phase 2 if any service is down after your fix attempt.
Phase 2 — SETUP / CLEAN SLATE
Run any setup steps from the test plan: DB resets, seed imports, migrations, fixture loading.
If the test plan has no explicit setup steps, check for common patterns:
docker exec ... psql (Postgres via Docker/Supabase)
npm run seed / yarn seed
python manage.py migrate && loaddata
- REST API seed endpoints (POST /seed, POST /admin/reset)
Important: If setup involves destructive operations (DROP, DELETE ALL, reset), print exactly what you're about to run and wait for a "yes" from the user before executing. Exception: if the test plan explicitly says "clean slate required", proceed without confirming.
After setup, verify the expected baseline state (row counts, seed user logins, etc.) using whatever the test plan specifies. If not specified, verify at minimum that auth works for the first test credential listed.
Phase 3 — TEST EXECUTION
Run every test case from the plan, in order. For each test:
- Print
▶ Running [TEST_ID] — [Test Name]
- Execute all steps (API calls, DB queries, or browser actions — see section below)
- Evaluate pass/fail criteria
- Print result:
✓ PASS [TEST_ID] or ✗ FAIL [TEST_ID] — [reason]
- On FAIL: immediately enter the Bug Fix Cycle (Phase 4) before moving to the next test — unless the plan says to collect all failures first
API-only tests (curl + DB)
Use curl for REST/GraphQL endpoints and psql/sqlite3/mysql for DB verification. Collect concrete evidence — response bodies and row counts — not assumptions.
Browser tests (UI interactions)
Every browser action follows the 5-step sequence — no exceptions:
Step 1 — Screenshot (see current state) · Step 2 — Decide the single next action · Step 3 — Execute one action · Step 4 — Screenshot again to verify the expected state · Step 5 — Repeat until the scenario completes.
For every step's mechanics use cdp-core.md verbatim: the canonical action template (§4, with CDP_URL resolved, never hardcoded), tab selection (§5), the selector ladder (§6), the wait strategy (§7). One QA-specific delta on the core failure protocol (§8): after the 2-attempt limit, record FAIL for the test case, document seen-vs-expected, and continue to the next test rather than stopping the run.
Destructive action gate
Before any test step that writes, deletes, posts, sends, submits, or modifies shared state beyond the test environment, STOP and confirm with the user. Tests within a local dev environment (localhost) do not require confirmation for write operations. Authenticated real-Chrome cases never reach this gate: the route guard (top of this file / cdp-core.md §9) already re-routed them to /chrome-cdp-control.
Phase 4 — BUG FIX CYCLE
Triggered immediately on each FAIL. Follows the diagnosis → fix → test → evaluate loop.
DIAGNOSE
DIAGNOSIS: <one sentence root cause>
HYPOTHESIS: <what I believe is happening>
EVIDENCE: <logs / DB query / screenshot / response body>
FIX PLAN: <minimal code change — file:line>
Rules:
- Read logs, check DB state, inspect API responses before touching code
- Form a hypothesis. Verify it with one targeted check before acting.
- Spawn an
Explore subagent for broad codebase research (tracing data flows across >3 files). Keep yourself for reasoning.
FIX
- Change only what the diagnosis identified — no refactoring, no "while I'm here" cleanups
- If fix spans multiple files, apply all before re-testing
- If fix requires a service restart, wait for the startup confirmation message before re-testing
TEST
Re-run the specific test that failed. Collect concrete evidence:
| Task type | Verification method |
|---|
| API route | curl the route, check response code + body |
| DB state | Query the specific table/row |
| UI behavior | Screenshot + check URL + check text content |
| Compilation | Run tsc --noEmit or the project's build command |
EVALUATE
RESULT: PASS | FAIL | PARTIAL
EVIDENCE: <observed output>
REMAINING ISSUES: <if partial>
NEXT ACTION: continue tests | new diagnosis cycle | escalate to user
Escalation rules
- After 3 distinct diagnosis attempts on the same bug → escalate to user with full diagnosis history
- If fix requires a product decision (e.g., "should this return 400 or silently default?") → escalate
- If fix requires a DB migration with DROP/ALTER → escalate before running
Regression check
After fixing any bug: re-run any previously-PASS test that touches the same service/component. A fix that breaks something else is worse than a known failure.
Phase 5 — EXPLORATORY TESTS (loop until dry)
After all plan-specified tests complete, run exploratory testing as a loop that stops only after 2 consecutive dry rounds — never a fixed "top N". Each round:
- Spawn an
Explore subagent to find untested paths:
Explore subagent prompt:
"In [PROJECT_ROOT], identify likely bug-prone areas not covered by the tests run so
far (list them). Focus on:
- Unguarded edge cases in service/controller layers
- Missing error handling in critical flows
- Data display issues in UI components
- Race conditions or state management issues
For each suspicious area, propose a concrete test (what to click or API to call),
the expected behavior, and the likely failure mode.
Return a ranked list of up to 5 NEW tests (not variations of ones already run)
with: ID, description, steps, expected result."
- Implement and run the proposed tests, same PASS/FAIL pattern (failures enter the Phase 4 bug-fix cycle).
- Count the round dry if it surfaced no new issue and no genuinely new test. New issue found → the dry counter resets to zero.
- Two consecutive dry rounds → stop, and record the stop reason as "2 dry rounds". Anything else (round cap, time, token budget) that stops the loop early is a cap and MUST be listed in the report's No-silent-caps section.
Phase 6 — FINAL REPORT
This report is a feeder. Running under /build-it, every result maps onto the conductor's unified report and the Go-Live verdict maps onto the DoD table — it never stands as a second, competing verdict. Standalone runs use the format directly. Either way, append each test's evidence (command · quoted output · verdict) to .taskstate/evidence.md as it happens — the conductor's claim gate and verifier read that ledger.
When all tests complete (or you've exhausted fix cycles), output:
## QA Report — [Project Name] — [Date]
### Summary
| Metric | Count |
|--------|-------|
| Tests run | N |
| Passed | N |
| Failed | N |
| Bugs fixed | N |
| Exploratory tests | N |
| Deferred | N |
### Test Results
| ID | Test Name | Result | Notes |
|-----|-----------|--------|-------|
| T01 | [Name] | ✓ PASS | |
| T02 | [Name] | ✗ FAIL | [what failed] |
| T03 | [Name] | ⚠ SKIP | [reason] |
### Bugs Fixed
| # | File:Line | Description | Fix Applied |
|---|-----------|-------------|-------------|
| 1 | src/foo.ts:42 | [what broke] | [what changed] |
### Known Issues / Deferred
- [Item]: [why deferred — needs product decision / out of scope / requires manual action]
### No silent caps
- [every test skipped, sampled, bounded, or loop stopped early, and why — or "nothing was capped"]
### Go-Live Readiness
**READY** / **NOT READY** — [one sentence verdict; under /build-it this maps onto the conductor's DoD table rather than standing alone]
Autonomy Rules
Proceed without asking:
- Running any test (curl, DB query, browser action) against localhost services
- Reading logs, source code, DB state
- Taking screenshots
- Applying code fixes for clearly-diagnosed bugs
- Restarting local services after a fix
- Creating test data (users, records, sessions) within the test environment
Stop and confirm before:
- Any write to a non-localhost / production service
- Running
DROP TABLE, DELETE FROM without a WHERE, or other destructive DB commands that aren't part of an explicit seed/reset step in the test plan
- Pushing code changes (git push)
- Running migrations that alter shared schema structure (ADD COLUMN with NOT NULL default is ok, DROP COLUMN needs confirmation)
- Killing OS processes with
pkill (exception: killing a service you just started for a resilience test)
- Taking any action on the user's logged-in accounts in external services
Maximum autonomy: Fix bugs without asking. Re-run tests without asking. Restart services without asking. Only escalate when you've genuinely tried 3 different approaches and are stuck, or when the fix requires a product decision.
Quick Reference: Common Stack Commands
Supabase / Postgres (Docker)
docker exec supabase_db_supabase psql -U postgres -c "SELECT ..."
docker exec -i supabase_db_supabase psql -U postgres < migration.sql
curl -s -X POST "http://127.0.0.1:54321/auth/v1/token?grant_type=password" \
-H "apikey: <ANON_KEY>" -H "Content-Type: application/json" \
-d '{"email":"user@example.com","password":"password"}'
Next.js
cd <project> && npm run build 2>&1 | tail -20
npx tsc --noEmit
pm2 logs <app-name> --lines 50
NestJS
curl -s http://localhost:PORT/health
tail -50 <output-dir>/engine.log
FastAPI / Python
curl -s http://127.0.0.1:PORT/health
cd <project> && python3 -m uvicorn main:app --host 127.0.0.1 --port PORT &
Django / Rails / Laravel
Chrome CDP
The canonical action template lives in ../references/cdp-core.md §4 — use it as-is.
Authored by DevOtts.