`5x/REPORT.md`

31 lines
# /5x REPORT — qwen38.agentabrams.com

Target: https://qwen38.agentabrams.com (auth admin/DW2024!)
Date: 2026-08-19

| Sweep | Result | Caught | Fixed | Commit |
|-------|--------|--------|-------|--------|
| 1 | 7/7 passed | 0 | 0 | (no changes needed) |
| 2 | 7/7 passed | 0 | 0 | (no changes needed) |

**Stop condition:** clean twice (two consecutive 0-defect sweeps). DONE at sweep 2 (cap 5).

## Six-way state (both sweeps)
- M1 HTTP contract — 200 text/html ✓
- M2 headless render — screenshot OK (54KB) ✓
- M3 automation E2E — `#send` present, **0 JS errors on load** ✓ (three.js scene inits clean — the setState-fallback + WebGL try/catch fixes hold)
- B4 Chrome / B5 Safari / B6 Firefox — render OK across all three engines ✓
- Click-through — Send button OK; Clear correctly skipped (destructive); 11 SKIPs are controls from another app's default list (not defects) ✓

## Honest coverage boundary (what /5x did NOT exercise)
/3x proves load + cross-browser render + no-JS-errors, but does **not** drive a full
model-backed chat (type → stream → 3D Thinking/Generating/STUCK → final render), because
that needs a ~20s live model generation. That runtime path is covered instead by:
- `npm test` — 18 deterministic regressions against a stub Ollama (auth matrix incl.
  no-colon→401, streaming proxy, think/keep_alive passthrough, upstream-500 no-hang).
- Manual browser use for the visual 3D/streaming behavior.

Net: static + cross-browser + proxy logic are automatically verified clean; the live 3D
chat animation is verified by construction (0 load errors) + the stubbed streaming test,
not by an automated in-browser chat E2E.