Skip to content

Releases

kabane 1.0.0-beta.0

kabane 1.0.0-beta.0 Pre-release
Pre-release

Choose a tag to compare

@davidpp davidpp released this 04 Nov 20:14
Immutable release. Only release title and notes can be modified.

kabane 1.0.0-beta.0

The first beta of kabane 1.0. It brings kabane onto current OpenTUI (0.8) and the current Agent Client Protocol stack (SDK 2.x, claude-agent-acp 0.99, codex-acp 2.5). It uses the new protocol features as they are meant to be used, so that kabane stays a thin wrapper that relays everything a harness asks to you.

Install it next to nothing else you care about first:

bun add -g kabane@beta

latest stays on 0.3.0. A 0.3.0 install never updates onto a beta by itself. A beta install follows the beta channel with kabane update, up to the stable 1.0.0 when it ships.

Before you try it, back up your tracker. Opening your tracker with 1.0.0-beta.0 runs forward-only migrations (6: drops old permission-answer decisions; 7: agent_activities.author). Reinstalling 0.3.0 afterwards is not a rollback. See docs/getting-started.md, "Updating", for a consistent backup of the configured database.

Your harness can ask you things now

  • AskUserQuestion and harness forms reach you. kabane now advertises ACP form elicitation. Before this, claude-agent-acp disabled AskUserQuestion under kabane and declined MCP elicitations, and codex's request_user_input forms were blocked the same way.
    • In the loop, a form becomes tracker questions: choices, multi-select, text, numbers and yes/no.
    • In the board copilot, a form is filled one field at a time: digits for choices, toggles plus enter for multi-select, y/n, a text input; Esc declines.
    • kabane never answers on your behalf.
  • Withdrawn requests close. When the harness withdraws a permission request or a form, the loop's question closes as "withdrawn" and the copilot drops the prompt. Before, it stayed open, answering nobody. Stopping a loop claim closes its open questions the same way.
  • Harness notices are shown. For example, "Auto mode unavailable for this model, using Accept edits". Before, kabane mislabelled that fallback as your own setting.
  • Your terminal knows when kabane is waiting. The board reports OSC 7501 program status: blocked on a permission or question, working, or idle. Terminals that support it (Rex today) can show it; others ignore it.
  • Multi-select and free-text answers from the board for loop questions, consistent with kabane loop answer (1, 3 or labels). The board and the loop judge an answer by the same rules.
  • Questions left by a runner that died close as withdrawn, so nothing stays pinned on the task.

The copilot knows your harness

  • The harness's own slash commands (and Claude's skills) appear in the / picker after kabane's shortcuts, labelled with the harness. A prompt starting with one, like /compact, goes to the harness exactly as typed.
  • Context usage ends every footer: ctx 42% · $0.31.
  • Notices show their severity: warnings in the accent colour, errors in the failed colour, on the copilot and the loop alike.

The event log reads like a conversation

  • Each tool call is one collapsed row: status, tool, first line of its arguments, duration. Click it, or press t to open every row. The copilot's calls carry their real tool names, inputs and diffs too.
  • Runs of reads and searches fold into "explored · N reads, M searches".
  • Edits open as diffs: unified, or split on wide terminals.
  • The agent's words render as markdown.
  • A toggled row keeps its place on screen.

The board on OpenTUI 0.8

  • Copy over ssh lands on your machine. y copies through OpenTUI's clipboard: OSC 52 to your own terminal when the session is remote, the system pasteboard locally. xclip and wl-copy are no longer needed.
  • The theme follows your terminal live. Switch light or dark and the open board recolours, markdown included.
  • One animation clock drives every spinner, so an idle board stops waking up.
  • Platform fixes: retained paint, synchronized output over ssh, a fix for the Dvorak Ctrl+J binding, and a clean terminal restore after quitting over a slow link.
  • Fixed rows stay visible beside long lists. Renderer teardown is awaited, so setup hands over to the board cleanly.

ACP

  • SDK 2.1.0 on the stable v1 entrypoint.
  • Adapters: claude-agent-acp 0.99.1 and codex-acp 2.5.0.
    • codex 2.x makes read-only a real read-only sandbox, and renames agent to "Auto review".
    • Both adapters now honour promptRequired for an idle steer.
  • Modes go through ACP session config options, which the spec says supersede session modes. session/set_mode stays as the fallback. kabane still only selects the harness's own auto mode, from its manual mode, and never overrides a mode your settings chose.

The loop

  • An interrupted claim resumes in its harness session (session/resume) instead of restarting blank.
  • A run moves the parent to in_progress and flags it for review when it completes.
  • One fixer session takes every lens finding.
  • Runs default to a worktree.
  • A dead runner can be restarted, and runner liveness is safe against PID reuse.
  • Tool-permission answers stay out of the decision log.

Also

  • Config is discoverable and manageable from inside the app (kabane config); sync and db keys are protected, and secrets are redacted on read.
  • Copilot permission requests queue, and a digit answers only a question that has been on screen long enough to read.
  • kabane-launched sessions get an MCP server that codex does not replace (kabane-session).
  • Release tooling: beta releases (X.Y.Z-beta.N on npm beta, as a GitHub prerelease); waiting for npm to serve a publish before calling it partial.

Known limits in this beta

  • OpenTUI 0.8.1/0.8.2, SDK 2.4 and claude-agent-acp 0.100 were held back by a 5-day release-age policy. They follow in a later beta.
  • Gemini was not exercised live; its paths are covered by tests.
  • The live board checks for copilot permission status, multi-select AskUserQuestion and harness-initiated withdrawal were blocked by a harness usage limit during verification. They are covered by in-process protocol tests.

kabane 0.3.0

Choose a tag to compare

@davidpp davidpp released this 27 Oct 05:20
Immutable release. Only release title and notes can be modified.
b54a53b

kabane 0.3.0

The first release since 0.1.0. It covers everything prepared for the unpublished 0.2.0
and what came after.

Loop extension (new, opt-in)

kabane loop drives a parent issue's subtasks to completion through coding harnesses
(Claude Code, Codex, Gemini) over ACP, from a detached runner. Turn it on with
"extensions": ["loop"] in config.json, or ask the copilot with /setup.

  • Start a run from the board (a on an issue) or with kabane loop start. A form shows
    the run's settings before it fires; per-project defaults and gates go in
    .kabane/loop.json.
  • Ready subtasks are claimed continuously up to a concurrency cap, in dependency order
    (blocks, blocked_by, follows). Human tasks stay in the plan but are never dispatched.
  • Gates (named shell commands) run after each subtask and reopen it on red. An opt-in git
    worktree with checkpoints you can roll back to. Verifier lenses review finished work and
    file findings as child issues.
  • Each run is a board tab with one lane per agent session. Steer it from there: message an
    agent, pause, skip, next, stop.
  • Harness permission requests are relayed to you, never answered by the loop. Answer them on
    the board or with kabane loop answer.
  • Budgets for wall time, concurrency, continuations and spend (where the harness reports
    it), plus stall and deadlock detection. Agents park blocked work with waiting, and parked
    work is never nudged or counted as a stall.

Board

  • Tabs and attention over agent questions, replacing the sidebar.
  • The copilot gets /setup: it sets the copilot's harness and model and turns extensions
    on, without editing JSON by hand.
  • Many fixes to tab, lane, banner and key handling in narrow panes.

CLI, MCP and core

  • Concise agent reads (responseFormat: concise) with bounded receipts and context chunks.
  • MCP accepts familiar comment and work-log input shapes.
  • Official extensions are opt-in by name; an unknown name warns once.
  • kabane --version names the commit when run from a checkout.
  • A session waiting on a human is never swept as stale.
  • Harness launches keep your own CLAUDE_CODE_EXECUTABLE and CODEX_PATH.

Releases

  • Verified, resumable releases and an explicit self-update.