Skip to content

Liner Code 0.2.1 (alpha)

Choose a tag to compare

@lynnarize lynnarize released this 04 Nov 05:39
· 13 commits to main since this release

Update from 0.2.0 if you have it. On a Mac Liner Code does it by itself (Restart to update), or
download the new DMG; on Windows, download the new zip.

New

  • Liner 2.0: what agents make for Liner Code is drawn in its own look. Every add-on is made
    with a design skill of Liner Code's surfaces, buttons, rows and type, and gets the app's colours
    to draw with, so a panel or a page reads as part of the app.
  • A model in an app: an add-on can ask a model and get the answer back as it's written. The
    first ask opens Liner Code's own popup, where you pick one of your models and allow it; the app
    keeps it and shows it on its strip, where a click changes it. It's only a model, with no tools,
    files or agent: an API endpoint's, or Claude Code's with its tools, settings and servers off.
    Each app gets two answers at a time and 300 a day, and its usage is logged under it.
  • Each mode leads with its icon in the picker, on Home's strip and in the chip beside a session's
    title: Standard's sparkle, Workshop's blocks, and your own modes with their agent's glyph.

Changed

  • Creator mode is now Workshop, in name only: your sessions and settings carry over.
  • Workshop only when you ask for it: it starts from Create your add-on, from "Make it an app",
    or when you pick it under the box. Everything else starts in your default mode.
  • Models come from one tier list built on Artificial Analysis (Terminal-Bench 4.0, FrontierCode
    v1.1, cost per task) instead of a maker-by-maker order. Opus and GPT Astra plan and take the hard
    jobs; DeepSeek V4.1 Flash builds wherever it's ready, then Sol, then Sonnet. A model never builds
    for itself: when Sol is your best, Sol plans and Terra builds. Critics come from different makers.
  • OpenCode Go's quota lasts longer: its frontier-class models have a small monthly cap, so none
    of them plans, takes hard jobs or builds in Dual. DeepSeek V4.1 Flash does it all, the newest GLM
    Flash judges, and MiMo V2.6 Pro and the newest GLM critique. Capped models still read.

Fixed

  • Home's strip no longer runs past the box. It holds where the task runs and the branch on the
    left and the mode at the right, each cut short with an ellipsis; Scratch folder, Carry on a
    session and the keys move to a quiet line under it.