claude-busy-monitor 0.1.0.post1

List active Claude sessions for the current user with their state.

claude-busy-monitor

Live view of every Claude Code session on your machine — which one is busy, which one is asking for your input, which one is idle — with cumulative token totals.

Ships as a claude-busy-monitor command and a Python library (get_sessions(), get_state_counts()).

Why

When you run many Claude Code sessions in parallel, keeping an overview of their states is a job in itself. You want every session working — every minute it isn't is time you don't get back:

• A busy session is the productive one — Claude is doing the work you asked for.
• An idle session is waiting for your next prompt — feed it.
• An asking session is fully stalled on a menu (most often a permission prompt) — answer it to unstuck Claude.

claude-busy-monitor surfaces the asking sessions instantly so they don't sit there burning your wall-clock.

Scope

What it sees and does:

• Local Claude Code sessions of the running user (~/.claude/sessions/ is per-user — the tool sees only your own sessions).
• Their live state (busy / asking / idle) and cumulative token totals.

What it deliberately does not do:

• No API calls — works entirely from local files; no quota burn, no auth needed.
• No web or desktop sessions — those don't write the ~/.claude/sessions/ files this tool reads.
• No mutation, no daemon, no IPC — read-only, ephemeral, nothing to start or supervise.

Features

Feature	What it gives you
No daemon, no IPC	Reads the files Claude Code itself writes. Nothing to start, nothing to keep running.
Authoritative state	status comes straight from Claude Code's own probe files — no inference, no heuristics, no race-prone proxies.
Cache-aware token totals	Sums all four usage categories (fresh, cache-create, cache-read, output). Naïvely counting only input_tokens underreports by ~10× on cached prompts.
Summary at a glance	Coloured pill counts on the first line — scannable from across the room.
watch-friendly	Stable, line-oriented output; no spinners, no cursor-move escapes.
Zero configuration	Reads ~/.claude/sessions/ and ~/.claude/projects/. Nothing to set up.

Install

PyPI publish is on the roadmap; until then, the install path is git + make:

git clone https://github.com/pbauermeister/claude-busy-monitor.git
cd claude-busy-monitor
make require   # installs uv if missing (idempotent; Linux/macOS)
make install   # installs the command globally via `uv tool install .`

Only prerequisite is Python 3.11+; make require bootstraps uv for you.

To work on the project itself rather than just use it, run make venv-activate instead — it creates .venv, syncs dev deps, and drops you into an activated shell. make help lists every Makefile target (lint, format, test-unit / smoke / e2e, build).

Usage

Command

claude-busy-monitor                       # one-shot listing
watch -n 1 -c claude-busy-monitor         # live refresh (the -c keeps colours)

The output is one summary line followed by one line per session. Each session line reads:

Column	Meaning
1	session id (first 12 hex chars)
2	state pill — busy (red) / asking (yellow) / idle (green)
3	working directory (basename)
4	output tokens, cumulative (out:)
5	input tokens, cumulative (in: — fresh + cached, summed)

Library

Build your own: a tmux notifier, a status-bar widget, a custom dashboard. The library exposes the same data the command uses.

from claude_busy_monitor import get_sessions, get_state_counts, ClaudeState

# Find sessions stalled on a permission prompt and act on them.
asking = [s for s in get_sessions() if s.state == ClaudeState.ASKING]
for s in asking:
    print(f"waiting: {s.cwd} ({s.session_id[:12]})")

# Or just summarise the states.
counts = get_state_counts()
print(f"{counts[ClaudeState.BUSY]} busy, {counts[ClaudeState.ASKING]} asking")

Public API: ClaudeSession, ClaudeState, TokenStats, get_sessions(), get_state_counts().

How it works

claude-busy-monitor reads two on-disk sources that Claude Code itself writes:

1. ~/.claude/sessions/<pid>.json — one probe file per live session, with the authoritative status field (busy / idle / waiting).
2. ~/.claude/projects/<encoded-cwd>/<sid>.jsonl — the per-session transcript, used only to total token usage.

State classification is a one-row table — no inference, no heuristics. Token usage sums the four usage categories per assistant entry.

For the full design — assumptions the classifier depends on, diagnostic recipes for when something looks wrong, and the repair playbook — see README-STATE-DETECTION.md.

Compatibility

• Operating system: Linux (relies on /proc/<pid>/comm). macOS is not supported yet — collaboration is very welcome. Please start by opening a Discussion so we can align on the approach before any issue or PR.
• Claude Code: v2.1.119 introduced the status field this tool relies on; older versions are silently dropped. /exit and claude --resume <sessionId> will migrate them.

License

MIT — see CHANGES.md for the version history.