Skip to content

04-test-plan.md

Latest commit

 

History

History
30 lines (21 loc) · 2.33 KB

04-test-plan.md

File metadata and controls

30 lines (21 loc) · 2.33 KB
Raw

04 — Test Plan (WHAT to test, per phase)

The app-specific WHAT lives in the project's own TEST-PLAN.md. This file is HOW to write and run it — any project, app or website. Golden rule: never test on silent failure — a failed step stops the run, gets captured, classified.

Writing rows

Each row = action + expected proof (log line AND visible change) + safety note if real-world side effects:

- [ ] **Row name:** do X → expect `log/line Y` + screen shows Z (screenshot/dump/selector).
      Safety: only test account/safe number/staging; screenshot proof before destructive tap.
  • Group by phase/day in build order. Tick only on verified proof, never on "looks fine".
  • Keep a re-test section: rows once trusted-but-unproven get one line each with exact proof needed. Start capture before the first action, not mid-test.
  • After every run: tick boxes, note evidence names, move FAILs to the problem ledger (05).
  • Name the locked gate once per stack (lockfile → package runner; pyproject/requirements → pytest; go.mod → go test; Cargo.toml → cargo test; else ask) and record it — "run the gate" is unactionable until the command is written down.

Running (any surface)

  1. Prep session (env up, tunnel/dev-server healthy, state clean) → begin top-to-bottom, in order. Metered resources (paid endpoints, metered devices, API credits): request only when the run is fully queued, release the moment live work ends — never hold them on standby.
  2. Per step: act → capture (log tail + current screenshot/dump) → assert BOTH expected line AND visible change → record PASS/FAIL/HANDOFF in RESULT.md.
  3. FAIL → failure protocol (06 § Failure): freeze + capture, classify A/B/C/D, never reclassify silently.
  4. Close: RESULT complete (one line per step + artifacts), evidence pruned (keepers only), toggles/data restored, summary → PROGRESS.

Lessons

  1. Gate discovery. "Run the locked gate" is unactionable when nobody knows which command it is. Cost: skipped verification on new projects. Rule: detect once per stack, record it as the locked gate, never guess twice.
  2. Metered resources. Holding paid endpoints, metered devices, or API credits on standby burns budget for zero work. Cost: meter running between sessions. Rule: request only when the run is fully queued; release the moment live work ends.