1
0
  • v0.3.2 1bea2dbc60

    School 0.3.2 Stable

    8ball released this 2026-11-12 21:31:03 +00:00 | 0 commits to main since this release

    School 0.3.2: a harness that fits each model

    Model profiles

    The app now adapts to the model a chat runs on, from what the router reports for it (window, effort levels, default
    effort). Small-window models work exactly as before; large-window models get more room.

    • More goes back to a large model: command output, file reads and search results are cut later on a 500K or 1M
      model than on a 256K one.
    • Fewer tools up front for the Qwen family (at most 20; the rest are found with tool search), as its makers advise.
    • Effort: a new chat starts at the model's own default effort (the setting "Default"); every level in the menu is
      a different thing. When a model is stuck, the next step runs one level higher and drops back after progress.
    • Replies that are only thinking (cut before any answer or tool call) are dropped and asked for again one level
      lower, instead of ending the turn.
    • Qwen working rules: a short prompt section for the Qwen family (batch independent calls, send an edit together
      with the command that checks it, read whole files on a large window). Other models' prompts are unchanged.
    • A second edit format (edit / write_file) exists and can be chosen per model in config.toml
      ([profiles."<model>"]); measured against patches it did not win, so patches stay the default.
    • Details and the measurements: docs/MODEL-PROFILES.md.

    Smaller things

    • The remote machines are called Devices (Settings > Workforce); the old "CC hosts" label is gone.
    • The v2 core's output limit leaves room for the model's thinking.
    Downloads