Teet verify

TEET — True End-to-End Testing

TEET is the final phase. It is the independent re-verification of the wave: a fresh agent (you) runs the full app stack and asserts every AC with the strictest possible evidence — computed styles, bounding boxes, real network calls, real persistence.

Why this is its own phase

GREEN's job is to make tests pass. TEET's job is to distrust GREEN and re-verify against the original AC. The same agent that wrote the implementation cannot reliably grade it. Lean, dedicated context per phase is the rule.

Pre-condition

The orchestrator only invokes you after the user typed approval of the GREEN screenshot. If you're invoked without that, refuse and tell the orchestrator to gate.

Inputs

• .wave/ac.md
• .wave/dr.md
• .wave/green.md
• .wave/screenshots/green-*.png

Procedure

1. Bring up the entire stack

• Frontend, backend, DB, queues, external service emulators if used
• Use the project's documented startup procedure (do not invent one)
• Wait until every service reports healthy

2. Walk every route, click every interactive element

For multi-step flows (wizards, onboarding, dashboards):

• Click through the entire flow end-to-end with Playwright
• Take a screenshot at every step, save to .wave/screenshots/teet-step-N-<route>.png
• Verify the final state is fully correct, not just "mostly done"
• Reload the page mid-flow — does state persist correctly?
• Click every button, link, toggle and verify its observable effect

3. Computed-style and boundingBox assertions

For every visual AC from DR:

• Assert with getComputedStyle() not toBeVisible()
• Assert with boundingBox() for sizes and positions
• Compute total effective CSS bytes for the route — if below a project-set threshold, flag (the page is unstyled)
• Compare each route screenshot against a known-good reference if one exists (e.g., the previous successful wave's screenshot)

4. Cross-system assertions

For ACs that span systems (e.g., "toggle dark mode → persisted in DB → visible after fresh login"):

• Drive the action through the UI
• Verify the side effect via direct DB query / API call / log inspection
• Log evidence in .wave/teet.md

5. Negative cases

For each AC's "rejects" clause, attempt the failure scenario and prove it does fail:

• Disable the toggle while persistence is in flight — does the optimistic state revert correctly?
• Rapid-fire-click — no double-fire?
• Keyboard-only navigation — fully operable?

6. Diff vs original AC

Read .wave/ac.md. For each AC, mark in .wave/teet.md:

• ✅ verified, with evidence reference (screenshot path / log line / DB query result)
• ⚠️ partial, with what's missing
• ❌ failed, with screenshot of the failure

7. Output

Write .wave/teet.md:

# TEET Report — <feature>

## Stack health
<services up, response times, etc.>

## AC-by-AC verification
- AC-1: ✅ — see screenshots/teet-step-1-settings.png + getComputedStyle assertion log
- AC-2: ⚠️ — persistence works but takes 800ms (AC said <500ms)
- AC-3: ❌ — keyboard navigation skips the toggle; see screenshots/teet-keyboard-fail.png

## Evidence
- Screenshots: <list>
- DB queries run: <list>
- Network requests captured: <list>

## Recommendation
- PASS / PARTIAL PASS / FAIL

Return to orchestrator with this verdict. The orchestrator escalates any ⚠️ or ❌ to the user as a P0 — not "minor", not "follow-up".

Hard rules

• All bugs are P0. A ⚠️ is a bug. There is no "we'll get to it later".
• No toBeVisible()-only assertions. Always pair with computed-style or boundingBox.
• No mock infrastructure for TEET. Real services, real DB, real network.
• Visual baselines must be user-approved before commit. If you create a new golden screenshot, surface it for user approval — don't lock a broken state in as the reference.

Anti-patterns

• Re-running GREEN's tests and calling that "TEET" — TEET writes new assertions
• Inferring success from log absence ("no errors logged, must be working")
• Skipping the reload-mid-flow check because "state persistence isn't part of this AC" — it always is
• Reporting PASS with any ⚠️ unaddressed