Skip to content

Open the setup screen on recommended models and name the CPU - #103

Merged
oney-erge merged 1 commit into
mainfrom
feat/first-screen-recommendations
Nov 20, 2026
Merged

oney-erge merged 1 commit into
mainfrom
feat/first-screen-recommendations

Conversation

@oney-erge

Copy link
Copy Markdown
Owner

Problem

A new visitor lands on the Setup screen and sees hardware detail first: GPU rows, two memory meters and the budget controls. The answer to "which model should I run?" sits below the fold behind a button. The CPU row also read AMD64 Family 25 Model 80 Stepping 0, AuthenticAMD on Windows.

Change

  • Recommendations load on their own. Once the hardware is read, the three picks (Recommended, Faster, Higher quality) for a general assistant appear with no click. This runs once. After that only the user's own requests change them.
  • One request owns the panel. Each recommendation request takes a sequence number and only the latest may write the panel or show an error. A slow automatic request that finishes after the user pressed Recommend models for me is dropped. The automatic request shows no error toast and tells the user how to retry.
  • The System card is one line plus a disclosure. The line names the GPU and memory, the CPU, the RAM and the fit budget. The hardware inventory, live VRAM meter and budget controls sit under Hardware details and fit budget, closed by default. Every element id is unchanged.
  • CPU name. Windows reads ProcessorNameString from the registry and macOS uses sysctl machdep.cpu.brand_string, both falling back to the old value. Linux is unchanged.
  • Web asset version ui34 to ui35; docs/UI.md and the changelog describe the new first screen.

Verification

  • pytest on the UI, web, hardware, recommendation, starter-pack and documentation tests: 144 passed, 1 skipped (the bakeoff UI test, already marked as disabled). This includes the Playwright file with Chromium installed.
  • New tests: picks appear without a click and the first card starts above y=900 at 1440x900; the summary line text; the folded controls work once opened; a held automatic response released after the user's request does not replace it.
  • Mutation check: with the sequence guard removed, the race test fails because the stale automatic answer replaces the user's choice. The guard is restored in this PR.
  • CPU name: registry success, registry failure fallback and macOS, with mocks. On this Windows machine the real value is AMD Ryzen 9 5900HS with Radeon Graphics.
  • Screenshots of a real run on a fresh profile (Ollama off) at 1440x900 and 390x844, dark and light: no console errors, no horizontal overflow, details open and close.
  • ruff check ., node --test tests/js/frontend-modules.test.mjs, git diff --check.

Not verified: the macOS CPU lookup on a real Mac (mocked only), and the screen with Ollama running and models installed (covered only by the mocked Playwright run).

The first screen led with hardware detail, so the answer to 'which model should
I run?' sat below the fold behind a click. Once the hardware is read, the three
recommendations now load on their own, and the System card folds its detail
under one line naming the GPU, CPU, RAM and fit budget. Each recommendation
request takes a sequence number and only the latest may write the panel, so a
slow automatic request cannot replace what the user asked for.

On Windows the CPU came from platform.processor(), which returns the raw
'AMD64 Family 25 Model 80' identifier. Read the registry name instead, and use
sysctl on macOS. Bump the web asset version to ui35.
@oney-erge
oney-erge merged commit f53064e into main Nov 20, 2026
9 checks passed
@oney-erge
oney-erge deleted the feat/first-screen-recommendations branch November 20, 2026 16:36
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant