Skip to content

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++
5.8 KB
Raw Blame

T10 · Build Engine v3

Phase 2 · 17 findings (2 high, 9 medium, 6 low) · Depends on: T01, T02, T03

Goal

Replace Tuning with designs/Engine v3.dc.html: auto-tune with the live animation, workloads with the two modes (leave room for other apps / full speed), and an Expert mode with every parameter, presets and the raw config editor.

Files to change

These are the likely files. Read them before editing, and change others only if the fix needs it.

  • src/screens/Tuning.tsx (rewrite as Engine)
  • src/components/EngineExperience.tsx
  • src/lib/tuning-controls.ts
  • src/lib/config.ts
  • src/lib/experience.ts
  • server/engine.mjs
  • designs/Engine v3.dc.html (reference: structure, copy, motion timing, colours)

Findings to fix (high and medium)

# Severity State Issue Suggested fix
1 Medium engine-beginner Dial tick labels (4K 8K 16K 32K) sit in a straight row under a 270° dial, not at their ticks labels around the arc, or use a slider
2 Medium engine-beginner Context range tops out at 32K; model supports 1M range from engine capabilities
3 Medium engine-beginner "Reset" link is very pale grey (~2:1) ink-muted #6f6f6f minimum
4 Medium engine-* Sticky action bar is ~140px tall with "More actions", disabled "Apply now…", disabled "Save for next load", plus two lines of notes; covers content and both actions look disabled at rest one-line bar that appears only when there are changes ("2 changes · Discard · Save")
5 Medium engine-resource-controls Beginner mode exposes GPU/layer split, KV mode (INT8), CPU threads move to Expert; Beginner gets workload + headroom only
6 Medium engine-expert Two stacked segmented controls (Beginner/Expert, Parameters/Raw config) look identical mode as a toggle in the header; sub-tabs as underline tabs
7 High engine-config-invalid Invalid JSON shows no visible error; the error line sits under the sticky action bar; Export silently disables inline error banner above the editor with line/column; keep editor bottom clear of the bar
8 Medium engine-expert-advanced Several Expert parameters (RoPE scaling, Expert cache / RAM mode) are "Preserved only · adapter does not apply this value" — shown as live inputs in small grey caption disable the control + "Not used by this engine" tag, or hide behind "Show unsupported"
9 Medium engine-expert-changed-only-empty Status says "Draft" and "Save for next load" is enabled while "Changed only (0)" Draft/Save only when changes > 0
10 High engine-expert-* At 1280×720 Expert mode shows ONE parameter per screen: mode tabs, sub-tabs, two disclosures, filter row and the sticky bar take ~85% of the height compact table rows (name · key · value · revert), collapse header chrome on scroll
11 Medium engine-expert-server "Server access (server.host)" duplicates the Serving page setting, and is "Preserved only" here one home: Serving; link from Expert

Optional polish (low)

Do these only if the task is otherwise done and the budget allows.

# Severity State Issue Suggested fix
1 Low engine-prompt-batch Second dial in blue (context dial is green); "Automatic · recommended 2,048" checkbox unchecked while value is 2,048 one accent; Automatic on by default, dial disabled until unchecked
2 Low engine-raw-config Plain textarea: no line numbers, highlighting or inline validation; bottom clipped by sticky bar code editor with gutter errors
3 Low engine-* Two-column layout leaves right column empty below "Expected experience" sticky summary card on the right
4 Low engine-discard-confirmation "Return all three editors…" — user doesn't know there are three "Discard all unsaved engine changes"
5 Low engine-actions Recommended defaults / Validate / Speed test hidden in "More actions" surface "Reset to recommended" and "Run speed test" as visible buttons
6 Low engine-adapter-details "llama.cpp adapter… native Strata configuration requires a different adapter" move to Expert › About engine

How to check

Run these and fix anything that fails:

  • npm run typecheck
  • node --test tests/ (all existing tests stay green)
  • npm run check (typecheck, UI audit, contrast, measure, luminance)
  • npm run build

Task-specific checks:

  • Beginner mode shows no expert controls.
  • Expert mode shows all parameters on one scrollable page grouped in sections; at 1280×720 more than one section is visible.
  • Invalid raw config shows an inline error next to the editor and disables Apply; the sticky action bar never covers content.
  • Auto-tune motion respects prefers-reduced-motion. Run npm run motion and npm run motion:budget.
  • The header shows the model via formatModelName() and has no "Change model".
  • Each finding above is resolved, or marked "already fixed by Txx" or "not reproducible" with a one-line reason.

Screens to re-capture and compare with the audit screenshots in docs/evidence/audit-2026-10-07/:

  • engine-beginner-light.png
  • engine-resource-controls-light.png
  • engine-prompt-batch-light.png
  • engine-expert-light.png
  • engine-raw-config-light.png
  • engine-config-invalid-light.png
  • engine-discard-confirmation-light.png
  • engine-actions-light.png
  • engine-adapter-details-light.png
  • engine-expert-advanced-light.png
  • engine-expert-changed-only-empty-light.png
  • engine-expert-server-light.png

Done when

  • All checks pass.
  • Commit on branch fix/t10-engine-page with message T10: Build Engine v3.
  • Append a short entry to audit-fixes/PROGRESS.md: what changed, which findings were fixed, skipped or not reproducible, and anything the reviewer should look at.