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++
6.3 KB
Raw Blame

32 annotation review — 9 October 2026

Completed in /Users/nikoveit/Documents/Strata-Design/backups/strata.

App: http://localhost:5160 · Host API: http://127.0.0.1:5161

Checklist

Comment Result
1 Chat model rows show name – quant – GB on one line.
2 All seven model API states have explicit labels and actions; transitions disable repeat actions, progress is shown, and errors retry download or loading according to whether files are installed.
3 Removed the shared-model heading from the chat model menu.
4 Running version sorts first in the chat menu.
5 Removed Manage model versions from the chat menu footer.
6 Removed the example-agents footer from the subagent popover.
7 Removed the Design example badge from chat responses.
8 Time-skip card arrow points right toward the side view.
9 Empty chats show a usage dashboard with Overview / Models, All / 30d / 7d, six statistics, activity heatmap, stacked model usage chart, token breakdown and percentages. Generic heading contains no personal name.
10 Every skill uses the same glyph from the existing icon set.
11 GPU inventory lists all detected devices. Temperature and utilization are matched per device, with unavailable sensor readings shown explicitly.
12 Physical processors are listed separately. Aggregate core counts remain available to Engine tuning. The local M4 now correctly reports 10 physical cores and 10 logical threads.
13 Reset to recommended and Run speed test are in the Engine header after the mode button.
14 Removed the specified working-memory warning from the visible validation list. Validation still runs.
15 Removed the specified fast-memory warning from the visible validation list. Validation still runs.
16 Conversation context is linear from 4K to the model maximum, in exact 1K increments. The host validates, saves and translates every supported integer K value.
17 Removed the performance Details disclosure.
18 Removed the Speeds · Demo estimate caption. Provenance remains available to assistive technology and on hover.
19 Removed the Sample workload selector from the performance preview.
20 Removed Shared by all chats from the model hero.
21 Removed Ready from the model hero.
22 Removed Version context selection; the Model page uses the saved Engine context.
23 On this computer rows show quant and GB without Smallest / Balanced names.
24 Removed View running version.
25 Removed View version.
26 The precision explanation sits inline beside Choose a version.
27 Model hero includes a family selector populated from the actual catalog, with versions filtered by the selected family.
28 Removed Not set beside API key.
29 API key label remains on one line.
30 Same badge removal as comment 7.
31 Removed the Mode heading and corrected option heights. All five choices fit without scrolling at the current desktop viewport.
32 User message hover/focus shows relative sent time, copy and Rewind to here. Rewind preserves the original chat and opens a branch with the earlier prompt ready to edit. It is disabled during generation.

State mapping

API state Picker presentation
available Download
downloading Downloading with percentage and progress
installed Downloaded; Switch version
loading Loading; repeat action disabled
loaded Running when it is the active model, otherwise Downloaded
unloading Unloading; repeat action disabled
error Retry; installed files retry loading, otherwise retry download

Data and hardware limits

  • Usage derives from saved browser history. Message and token counts deduplicate copied branch turns. Showcase and Time-skip example turns are excluded. Token totals include only reported engine usage; demo, estimated and unavailable usage do not become measured totals.
  • The current build runs in simulated mode and its catalog contains one model family. No extra production families were invented. A test covers multiple families in a fixture.
  • NVIDIA sensors use stable GPU UUIDs. Linux DRM sensors are read from each device's own sysfs path. Processor topology uses macOS physical/logical/package counts, Linux socket/core topology, and Windows processor instances.
  • This Mac does not expose GPU temperature/utilization through these probes, so its card shows Unavailable. Two-processor and multiple-GPU cases were verified with fixtures, including distinct temperatures and a missing sensor.
  • Hardware inventory does not assume that different GPUs' VRAM can be pooled by the inference adapter.

Primary probe references: NVIDIA SMI, Linux CPU topology, Windows processors, Linux AMD GPU sensors.

Verification

  • npm run typecheck — passed.
  • npm run build — passed; 117 modules.
  • node --test tests/annotations-32.test.mjs tests/dictation.test.mjs tests/agent-updates.test.mjs tests/tasks-and-sharing.test.mjs — 20 passed.
  • Live API accepted 33K; rejected 33.5K and 1025K. Engine launch argument tests cover 4K through 1024K, including non-power-of-two values.
  • Browser: context ArrowRight moved 32K → 33K; test draft discarded, saved context remains 32K.
  • Browser: Mode menu client height and scroll height both 297px. Model menu client height and scroll height both 306px; all model rows 42px.
  • Browser: available version opens Download; installed version opens Switch; running version remains first.
  • Browser: dashboard tabs and periods, model family menu, removed labels, hardware cards, API key layout, skill glyphs, message hover and Time-skip arrow checked.
  • Rewind test branch s_l1uci3r3idc9 was archived. The original give me timeskip transcript and its Show me multi draft were preserved.

Screenshots

Evidence is in docs/evidence/annotations-32-2026-10-09/:

  • usage-overview.jpg, usage-models.jpg
  • model-picker.jpg, mode-menu.jpg
  • models.jpg, installed-versions.jpg
  • engine.jpg, monitor.jpg, skills.jpg, serving.jpg
  • subagents.jpg, message-hover-timeskip.jpg, rewind-branch.jpg