Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
Review checklist — completed 7 October 2026
The final review covers all three phases together. See PROGRESS.md for each task and FINDING-STATUS.md for every finding. Checkboxes are intentionally left open where evidence is incomplete or the requested literal check was adapted.
Every phase
-
npm run checkpasses (typecheck, UI audit, contrast, measure, luminance). Final log:check-final.log. -
node --test tests/*.test.mjspasses: 39 tests, none skipped or removed. The package's directory form is incompatible with the installed Node version. -
npm run buildsucceeds. - Every high/medium finding is implemented or closed with a documented reason in PROGRESS.md / FINDING-STATUS.md.
- Every listed state re-captured in both themes. Partial: 87 targeted screenshots; obsolete screens removed; the read-only synthetic fixture endpoint was blocked by the browser. No bypass attempted.
- Dialog/menu/response checks show no unintended content movement; response menu geometry was measured unchanged.
- Source compared with the baseline archive; changes belong to the audit fixes or supporting tests. Design-kit token files unchanged.
Phase 1
- One formatter supplies loaded-model identity throughout the UI.
- Literal no-Qwen2.5 source grep. Intentional exception:
server/real-model-catalog.mjsretains the verified real-engine model/download source. The design preview uses the supplied Qwen3.8 family. Do not rename a real model into a fictional one. - No multi-model pickers remain; Model owns version selection and Setup reuses the same ladder.
- Repeat listbox works above the task dialog in both themes.
- Enabled/disabled primary states and on/off toggles are visually distinct.
- Shared ErrorNotice strips Fixture prefixes; error fields have clear state and recovery paths.
- Composer shows live estimated speed during streaming; Monitor clamps/flags impossible memory values.
- Engine and Model are primary navigation destinations.
Phase 2
- Model follows the reference structure: running hero, version ladder, details/actions and installed versions.
- Fit/recommendation derives from detected hardware and context, not reference placeholder numbers.
- Model lifecycle action mapping is implemented. Some synthetic lifecycle states were source-reviewed rather than captured.
- Engine follows the reference structure with compact Expert parameters and inline JSON diagnostics.
- Motion/reduced-motion checks pass (
motion.log). - Quantified motion budget passes. Not measured: harness self-check could not detect a known-expensive effect (
motion-budget.log). No performance claim made.
Phase 3
- All required screen findings have a disposition.
- Dark mode passes contrast and luminance checks.
- Time-skip completion opens the selected example result; A/B names stay stable.
- Optional low-priority polish remains explicitly unscheduled rather than silently claimed complete.
Delivery
- Package extracted; requested agent file copied unchanged.
- Original evidence preserved; targeted review evidence in
docs/evidence/audit-fixes-2026-10-07/. - No Git metadata exists here; baseline archive and changed-file inventory supplied instead of branch commits.
- No external publishing, API-key save, model-weight download or network exposure performed for the review.