Skip to content

Releases

v0.4.0

Choose a tag to compare

@bekir1184 bekir1184 released this 16 Oct 23:51
1a3342e

Thanks to @VlcekM for the side pane, minimise,
dismiss and sliding details (#6).

Added

  • ⇥ moves the crew into a narrow side pane beside the conversation, one card
    per agent; ⇤ (or closing the pane) brings it back above the prompt. The
    crew starts above the prompt, and the choice is kept across sessions.
  • – shrinks the band to one plain line, kept across sessions until ▸.
  • A finished agent's ✕ sends it off the stage.
  • Terminal keys: z to minimise, y to move to the side pane.

Changed

  • Cancelled agents leave the stage right away.
  • Details slide open and shut; the chrome buttons stay dim until hovered.

Agent Crew 0.3.0

Choose a tag to compare

@bekir1184 bekir1184 released this 09 Oct 15:39
cb00010

Added

  • Each row names the model its agent runs on, under the type ("Opus 5.5",
    "Haiku 4.5"); Claude's own model sits beside the title, and the terminal
    shows a short name when there is room.

Changed

  • /crew demo is a 30-second show timed for a screen recording: Claude starts
    alone, then three agents arrive one by one and a helper joins, so the band
    fits on screen; the crew works, thinks, one agent asks for approval, and
    everyone ends with a flag.
  • While the demo is on stage, the prompt box shows the demo's own English hint
    (/crew clear) instead of Claude Code's guess at your next message, so a
    recording shows nothing personal; suggestions are untouched otherwise.
  • The type and model under it are drawn small on the desktop, and a row's task
    text uses the room the row really has instead of being cut early.
  • The README is shorter and visual, led by a recording of the demo, with an
    animated image per feature; a small site (site/) shows the same.
  • Token figures on rows, the title and the thin line count only the work
    itself, input and output. Cache writes and re-reads have their own line in
    the details panel, so a long conversation picked up after its cache expired
    no longer shows a sudden jump of hundreds of thousands of tokens.

Agent Crew 0.2.0

Choose a tag to compare

@bekir1184 bekir1184 released this 09 Oct 11:26
cd63778

Added

  • Claude's own line: while the main conversation works, the framed title shows
    its activity, time and tokens. Alone it reads AGENT, and when the turn ends it
    shrinks to a thin line with the turn's final numbers and an arrow for details;
    subagents join below it as AGENT CREW. Setting "Show Claude's own work", on
    by default.
  • Approval state: an agent stopped at a permission prompt shows "Needs
    approval" with the tool, a critter with a question bubble, and the title
    counts the agents waiting on you.
  • Optional cost estimate (setting "Show estimated cost", off by default): a ≈
    dollar figure per row, for the crew in the title, and broken down in the
    details panel, from each model's list prices.
  • The title row has its own arrow that folds the subagent rows away and back; each
    new crew starts with its rows open.
  • Status line summary while the crew is hidden with /crew.
  • Health signals: slow tool calls (over 30 s) show their time, and agents with
    no activity for 2 minutes are marked "Quiet".
  • The details panel lists an agent's last five tool calls, failed ones marked,
    and the files it changed and read.
  • Secrets in commands, URLs and task text are masked as •••.

Changed

  • Thinking is animated: a crew member keeps taking turns between the
    hourglass, a thought cloud with a question mark and a lightbulb that lights
    up, out of step with the rest of the crew.
  • The terminal leads each row with a dot colored by the state instead of a
    text face, and its arrows take letter hotkeys (h for the title, a, b,
    … for the rows) once ctrl+x tab gives the crew the focus.
  • /clear also clears a finished crew.

Agent Crew 0.1.1

Choose a tag to compare

@bekir1184 bekir1184 released this 09 Oct 09:39
10e975a

Changed

  • A row's details now open from an arrow button at its end (▸ / ▾) instead of
    a click anywhere on the row. The desktop delivers raw pointer events to
    plugin layers unreliably (presses dropped, releases doubled), so clicks were
    missed or undone; a button works every time, and with the keyboard too.
  • The title row is framed in Claude's color. The Clear and Hide buttons are
    gone: /crew hides and /crew clear clears.

Fixed

  • The "doing" column no longer wraps onto a second line on the desktop.
  • Rows use more of the available width, so task names are cut less.

Agent Crew 0.1.0

Choose a tag to compare

@bekir1184 bekir1184 released this 09 Oct 08:52

First release of Agent Crew.

Install

/plugin install agent-crew --marketplace bekir1184/agent-crew

Pick the user scope to have it in every session. Requires Claude Code 2.1.293 or newer.

What's in it

  • One aligned row per subagent above the prompt: a pixel-art crew member holding
    the tool it is using, its task, what it is doing now, a progress bar, the time
    left and the tokens it used.
  • Helpers (subagents started by a subagent) shown indented under their parent.
  • Real step progress when an agent keeps a to-do list (TodoWrite, TaskCreate,
    TaskUpdate).
  • Time estimates learned per subagent type and model (median of the last 20
    runs, with the usual range after three runs). Agents that run long are
    re-estimated from past runs that lasted as long, instead of showing overtime.
    Estimates are local arithmetic and cost no tokens.
  • A details panel per row: full task, steps, token breakdown, model, tool and
    request counts. Click anywhere on a row to open it (▸ turns to ▾); one panel
    at a time, framed with a dashed yellow border.
  • Done, failed and cancelled states: checkered flag, dizzy eyes, asleep.
  • Terminal and desktop rendering; text follows your theme.
  • /crew, /crew demo and /crew clear commands.
  • Built to stay out of the agents' way: tool calls and model requests are only
    noted in memory; a one-second ticker writes them in a single update and
    redraws at most once. Engine-internal loops are ignored, SVGs carry only the
    animations they use, and sprites and bars are cached per state.
  • Agents whose end is never reported are closed from the engine's own agent
    list, with the outcome it reports; a quiet agent it still lists as running is
    never retired. Resumed subagents run again. Every hook falls back to Claude
    Code's own behavior on error.