Skip to content

v0.0.4

Choose a tag to compare

@github-actions github-actions released this 12 Nov 18:22
· 5 commits to main since this release
028c50e

Added

  • Jobs leave an outcome (JobHandle::finish: what they found, or why they failed), kept after they
    end; a command that starts a job returns {"job": id} (jobs::job_reply). The MCP host then
    answers the agent when the job ends, with that outcome, instead of at once with null.
  • A command can return an image ({"image": {"data": base64, "mime_type": …}}): the MCP host
    sends it as image content, so an agent can look at it.
  • kitchen-shell is published on crates.io too (0.0.3 published by hand on 2026-11-12, from the
    v0.0.3 tag with this packaging: README, licence texts, keywords); the release workflow
    publishes it from now on (it needs its trusted publisher on crates.io: this repository,
    release.yml, environment crates-io). Apps that publish their own crates take kitchen-ui and
    kitchen-shell from crates.io.
  • Agents see each command under a tool name MCP clients accept (hist.find_cells →
    hist_find_cells; tool_name); a call by either name reaches the command. Antigravity refused
    every dotted name, which left the chat's agent without lumen's tools.
  • AgentWorkspace::set_preamble: what the agent is told before the first prompt of each session
    (where it works, what is open, which tools act in the app).
  • kitchen-ui timeline (timeline): an episodic memory of the work (after Scott Jenson's
    prototype): what the person and the agent did, by lane, as bars coloured by kind, as wide as
    they lasted and as tall as they were busy, on an hour axis; a card on hover; a click shows the
    episode's context cards: each action, where it came from and what it produced. Many actions
    are summarised into a few episodes (episodes: same-kind neighbours merge, the shortest fold
    into their neighbours), or shown one by one. A gallery page and a headless test.
  • kitchen-shell activity log (activity): every command run (who: the person or an agent; how:
    menu, key, search, title bar, agent; its arguments in short and its result), the time in each
    workspace, every job from start to end, and what a step produced; activity::view is a
    timeline that follows it. Registry::run_as records; the frame, the jobs and the MCP host feed
    it by themselves.
  • kitchen-agent ACP client (acp): the chat's agent is any program speaking the Agent Client
    Protocol (agy acp, Claude Code's adapter), started by the app with its MCP servers; prompts,
    cancel, new sessions; streamed answers, its own tool calls and its permission questions come
    back as AcpUpdates. The agent keeps its own login: no key in the app.
  • AgentWorkspace::connect: the conversation runs on a connected ACP agent; the agent's questions
    show above the prompt with its options; the model chip names it.
  • kitchen-agent bridge as a library function (bridge::bridge_stdio), so an app's own program
    can be its MCP server for an agent (kitchen-mcp-bridge uses it).
  • kitchen-shell project environments (env): the Python venv of a project's scripts (its own,
    the active one or the app's default; version and packages read without running it), made and
    synced by uv from pyproject.toml and uv.lock; the Rust workspace and rust-toolchain.toml.

Install

cargo add kitchen-ui (crates.io). The other crates are used from this tag:
kitchen-shell = { git = "https://github.com/neuro-kitchen/kitchen", tag = "<this tag>" }.