test-debuggerlisted
Install: claude install-skill YevheniiaQA/qa-agent-toolkit
# Test Debugger
Generated tests are a draft until they have run. This skill turns a draft into a
suite worth trusting, in two passes that answer two different questions: *does the
test run correctly* and *does the test check anything*.
## Pass 1 — functional debugging
Run each test in isolation first (`--workers=1`, one file at a time). Isolation
separates genuine failures from interference, and interference is itself a finding.
Classify every failure before changing anything. Fixing without classifying is how
real defects get "fixed" out of the suite.
| Class | Signals | Action |
|---|---|---|
| **Test bug** | Wrong endpoint, wrong payload shape, bad locator, missing await, wrong expected value | Fix the test |
| **Product defect** | Test matches the contract and the requirement; system disagrees | Do not fix. Log to `DISCREPANCIES.md`, mark `test.fail()` |
| **Environment** | Connection refused, 502, expired cert, missing env var, wrong base URL | Fix config, do not weaken the test |
| **Data** | Missing user, stale record, conflicting leftover state | Move provisioning into a fixture |
| **Flakiness** | Passes on retry; timing, animation, or race dependent | Fix the wait strategy — never add a retry to hide it |
### Scenarios that were always meant to fail
Before classifying anything, check the scenario's handoff fields. A test whose
scenario carries `status: known-defect` and `doNotRelax: true` was written to
fail. Its failure is not a finding to investigate — it i