🤡 clownhead

A status board for the Claude Code sessions already running on your machine: which are busy, which are idle, and which one is blocked waiting on you.

Install

curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install clownhead

uv tool install git+https://github.com/rooterkyberian/clownhead gets you whatever is on main, which is ahead of the last release.

Requirements

• Claude Code 2.1.227 or newer on PATH.
• macOS or Linux. Developed on macOS with iTerm2; CI runs the suite on both.
• Sessions are found whatever terminal they run in. Tab colours are iTerm2's alone, kitty gets notifications, and everything else (an IDE's embedded terminal included) falls back to the bell and a tab renamed to ⚠ <session>: <why>. Raising a window is macOS-only.
• Sessions in a JetBrains IDE get their terminal tab selected as well, which takes an Accessibility grant for whichever application clownhead is running in: System Settings, Privacy & Security, Accessibility. Without it the window still comes up and the board says why the tab did not.

Keys

QUIET is time since the session last beat, AGE is time since its process started, and the pane below the table carries the id, path, process and terminal the columns cannot fit.

• → (or a click on the row) opens that session's conversation beside the board, usually the fastest way to tell what it is actually doing. ↑ and ↓ scroll it, ← closes it.
• enter gets you into that session: a live one has its terminal focused, and one that has ended is resumed here, which ends the board.
• f focuses its terminal: attention, then the window brought to the front, and in a JetBrains IDE the session's own tab selected once it is.
• o opens the pull request the session was working on, asking which when it named more than one.
• / filters by name, status, path, or session id — or by pull request or issue, below.
• p opens the pull requests you have open on GitHub, below.
• n starts a new session for the pull request or issue being filtered on.
• c folds in the sessions that have already ended. The count in the top bar is that same switch, and clicking it works too.
• y copies its resume command.
• r renames it.
• t asks whether to send its process SIGTERM, and can close its tab behind it.
• , opens the settings.
• ^p opens the command palette.
• q quits.

Tab colours

iTerm2 only. Every reload tints each session's tab to match its state, so the herd is readable from the tab bar of a terminal the board is nowhere near. Turn it off in the settings and the tabs it tinted are cleared on the way out.

Pull requests and issues

Nothing on a session records which pull request or ticket it belongs to, so no column can show one. Paste a pull request or issue URL into / and clownhead reads the transcripts of whatever the board is showing. Finished work is usually in a session that has ended, so c first, then the URL:

 🤡  2 of 137 sessions · acme/data-platform#309                        ⟳ 5s
 STATUS  NAME                    WHERE
 idle    invoice-parser          data-platform ⇢ invoice-parser
 closed  design-system:87e26be1  ~/dev/acme/design-system

A search of the live herd alone that comes back empty says so, and says that c would widen it, instead of folding the closed ones in uninvited.

owner/repo#309 and repo#309 name the same thing more briefly. Jira is named by its URL.

The pane below the table names the pull requests the selected session worked on, freshest first, which is the same reading done backwards — and o opens one in the browser.

A PRS column puts the freshest one on every row instead. Turn it on in the settings, or ask ls for it:

$ clownhead ls --closed --columns status,name,prs,where
STATUS  NAME                  PRS                       WHERE
busy    settlement-retries    payments-api#309 +2       ~/dev/acme/payments-api
idle    invoice-parser        data-platform#362         data-platform ⇢ invoice-parser
idle    index-rebuild         web-platf…#88 +14         web-platform ⇢ search-index
closed  design-system:87e26be1  -                       ~/dev/acme/design-system

It is the one column that costs a pass over the disk, which is why it is off until asked for — 55 ms for 151 MB of transcripts, read once and kept until ^r. A row says ? until that read lands and - once it has and the session named nothing; the repository gives way before the number does when the column is too narrow, since the number is the only part you read a pull request by.

Your pull requests

p asks GitHub what you have open, and the transcripts which sessions here worked on each:

 49 open · reading status 31/49… · 6 with sessions here
 PR                                TITLE                              CHECKS  REVIEW    SESSIONS  UPDATED
 acme/payments-api#309             feature: settlement retries        ✗ 2     changes          1       3h
 acme/data-platform#362            chore: migrate httpx to httpx2     ✓       approved         2       1d
 acme/web-platform#88              spike: search index rebuild        ⟳ 1     required         0       4d

What wants you leads: something red or objected to first, then what is approved and green and waiting on somebody to press merge, then what is still out for review. Drafts sink below all of it.

enter leaves the list and points the board at the sessions that worked on that pull request, ended ones folded in, which is the same place pasting its URL into / arrives at. o opens it on GitHub and y copies its URL.

This is the one view that asks somebody else, so it needs gh and says which of no gh, no auth and no network it ran into rather than showing an empty table. The three answers arrive separately — the list in one request, the statuses one apiece, the sessions from a pass over every transcript — and the table fills in as each lands.

Starting one

Every ticket starts with the same three steps: find the checkout, make a worktree, tell a fresh session what to work on. clownhead <url> does all three.

$ clownhead https://github.com/acme/data-platform/issues/2

The board opens filtered to that issue, the ended sessions already folded in. enter gets you back into whichever one you pick; n starts a new one in plan mode, in a worktree named after the issue.

Commands

Every view is also a one-shot subcommand, so the same data pipes into a script.

--columns names what ls shows and the order to show it in: status, name, quiet, age, pid, tty, worktree, prs, where, resume. Everything but pid, tty, worktree and prs is on by default, resume included. A listing you are reading in order to get back into something should hand you the command that does it:

$ clownhead ls --pr acme/payments-api#309 --closed --columns name,resume
acme/payments-api#309 · 2 of 74 sessions
NAME             RESUME
payments-api-7c  (cd /Users/you/dev/payments-api && claude --resume 4e020900-df7c-4665-a804-d973b14a1926)
index-rebuild    (cd /Users/you/dev/web-platform && claude --resume 8b1c4f22-0d31-4f0a-9c2e-3a7b1e5d6f08 --worktree search-index)

A resume command is the longest thing on the board and the one truncation ruins, so naming fewer columns is how you get one whole; --columns name,resume is the pair worth remembering.

How it works

How it works: discovery, the attention signals, the control socket behind renaming, and what happens to a session when it is terminated or resumed. Alternatives surveys the other tools for managing a herd, and where this one differs.

Problems

Open an issue. The form asks for clownhead --version and claude --version up front, since a surprising number of surprises are a disagreement between the two.

Development

mise install
mise run check    # lint + typecheck + test
mise run demo     # re-record docs/demo.gif from docs/demo.tape, which needs vhs