Skip to content

v0.73.3

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 12 Nov 22:32
· 157 commits to main since this release

Added

  • A provider can be switched off in the model settings of every client: it keeps its key
    but is not asked, its models are not offered and discovery skips it. The config says
    enabled: false on it. The provider of the chosen model cannot be switched off.

  • Google AI as a kind of provider (kind: google): the Gemini API with its own model
    catalogue, streamed answers, thought summaries, function calls and picture input.

  • The model settings of every client show a provider's rate limits when its answers
    report them (OpenAI-compatible x-ratelimit-* and Anthropic anthropic-ratelimit-*
    headers): requests and tokens left of the limit, and when they are full again. Read off
    the answers that come anyway and kept in memory only.

  • GUI: slash commands in the message box. Typing / lists them, Tab or a click completes
    one; commands that only show something open the window's own place for it.

  • GUI: a new chat opens on a line in a warm serif that changes with the time of day.

  • domain-structure skill: laying out and auditing any codebase as domains with front
    doors and one reason to change per file; docs/structure-audit.md lists what here does
    not fit yet.

  • GUI: a Sources module in the side panel lists what is attached to the chat; each
    source goes to the model once, and a switch on it sends it with every request.

  • GUI: a falling star crosses the header's sky now and then, and mist rises at the foot
    of a window left alone for a few minutes; a click or a key clears it. Appearance has a
    switch for each.

  • GUI: a long unfolded block (a turn's activity, thinking, a step, a run of steps) ends
    with its heading again, so it folds from where the reading ended.

  • GUI: a firefly glows over the message box for each subagent at work; a turn that runs
    past sunset sees the header's glow go on into a deep blue night; a turn that ends
    while the mist stands lifts it with a warm light along the foot of the window.

  • GUI: Memory and context has the context window and the sizes of the memory and of the
    work history.

  • GUI: a step's icon says what it is: the terminal, a file, a search, an edit, the web,
    another agent, a connector's tool; thinking and reports have their own. Its color says
    how it stands.

  • GUI: the context ring breathes once the context is more than 85% full.

  • GUI: a subagent's chat shows its task and reports as markdown, its thinking folded, its
    errors apart, and who it is at the top; one agent can be stopped from its row or its
    chat, and finished ones cleared from the list, one by one or all at once.

Changed

  • The context report and a thread's Tokens events count with the same tools and model,
    so Statistics and the ring by the message box agree; Statistics follows the turn.
  • GUI: the quiet mode's moon is filled while it is on; Connectors stands above Artifacts
    in the rail; the sky's stars are soft points with short rays and a rare tinted one;
    More models names each provider in bold and folds it.
  • GUI, Dusk: a chosen value in a row of choices is tinted with the accent instead of
    filled with it; the settings page is wider and stands in the middle; the frame is a
    step lighter and the quiet text a step brighter; the message box lights up a little
    while it is written in.

Fixed

  • GUI: the wheel over the font list in Appearance scrolls the list, not the page.
  • GUI: a long last answer can be scrolled down again after scrolling up in it.
  • GUI: background jobs show in the Jobs module as they start and end.