avenoxstatusline

A two-line status line for Claude Code with an ASCII pet.

132 lines of bash. No npm, no node, no build step.

ʕ•ᴥ•ʔ☕  Opus 5 · high · think   5h 41% · 7d 63%
▓▓▓▓░░░░ 62%  ⎇ main*  ● serai:vegrun

---

Why another one

There are plenty of good Claude Code status lines. Two things make this one different.

1. It's a shell script, not a package

Your status line re-runs every few seconds, forever. Every render is a fresh process. Spawning a Node runtime on that loop is a strange thing to do to your laptop.

This is one bash file with two dependencies you already have (git, jq). Install is a curl and one line of JSON. There is nothing to update, nothing to audit, and nothing in node_modules.

2. The pet is a gauge, not an ornament

The bear's mood is derived from session state, so you can read how the session is doing without reading any numbers:

State	Pet	Meaning
context ≥ 90%	ʕ⊙ᴥ⊙ʔ‼ red, flailing	you are about to compact
context ≥ 75%	ʕ•﹏•ʔ💦 yellow, sweating	start wrapping up
pending approvals	ʕ•ᴥ•ʔ❗ yellow, alert	something is waiting on you
effort xhigh / max	ʕ◣_◢ʔ⚡ locked in	the expensive gear is engaged
context ≥ 50%	ʕ◔ᴥ◔ʔ focused	working
otherwise	ʕ•ᴥ•ʔ idle	blinks, glances, leans

Peripheral vision does the monitoring. You only look directly at the line when the bear changes.

And when nothing is wrong, it does things. Roughly every 10 seconds it pulls a ~4-second stunt — nap, dance, wave, table flip (and sets the table back), flower, coffee, sparkle, hug, wander.

---

The trick: animating a stateless script

Claude Code runs the status line command fresh on every render. There is no process to hold a frame counter in, so conventional animation is impossible.

So the frames are indexed by wall-clock time instead of by state:

T=${SL_TICK:-$(date +%s)}
frame() { local -a a=("${@:2}"); PET="${a[$(( $1 % ${#a[@]} ))]}"; }

frame "$T" "ʕ•ᴥ•ʔ" "ʕ•ᴥ•ʔ" "ʕ-ᴥ-ʔ" "ʕ◕ᴥ◕ʔ"   # idle: blinks

Each render independently computes which frame it is right now. The script stays completely stateless and the pet still walks through its frames as time passes. Stunts use the same idea at two scales: T % 10 picks the position within a stunt, (T / 10) % 9 picks which stunt.

SL_TICK pins the clock so you can screenshot or test a specific frame.

---

Install

curl -fsSL https://raw.githubusercontent.com/avenoxai/avenoxstatusline/main/statusline.sh \
  -o ~/.claude/statusline.sh
chmod +x ~/.claude/statusline.sh

Then add to ~/.claude/settings.json:

{
  "statusLine": {
    "type": "command",
    "command": "bash ~/.claude/statusline.sh",
    "refreshInterval": 3
  }
}

Requirements: bash (3.2+, so stock macOS works), git, jq. Tested on macOS and Linux.

---

What's on the line

Line 1 — pet, model, reasoning effort, extended thinking, and your 5-hour / 7-day rate-limit usage.

Line 2 — context bar (green → yellow → red), context %, branch (* = dirty tracked files), worktree, project badge, pending approvals, sync state.

Every segment hides itself when it has nothing to say. In a plain directory with no git, line 2 is just the context bar.

---

Configuration

Env var	Effect
SL_BADGE_CMD	Run any command in the repo root; its first stdout line becomes the project badge. A prefix:value string renders the prefix dimmed.
SL_NO_SERAI=1	Disable the built-in Serai badge.
SL_TICK=<int>	Pin the animation clock. Useful for screenshots and tests.

Custom badge example:

export SL_BADGE_CMD='cat .env 2>/dev/null | grep -m1 ^APP_ENV= | cut -d= -f2 | sed "s/^/env:/"'

The Serai badge

Out of the box, if the repo contains .serai/config.json, the line shows ● serai:<repo> plus pending approval count (⚑2) and sync freshness (●sync live / ○sync stale). That's Serai, a control plane for coding agents.

If you don't use it, nothing renders and nothing is read. SL_BADGE_CMD takes priority over it, and SL_NO_SERAI=1 turns it off entirely.

---

Seeing it without installing it

./demo.sh          # every mood state, side by side
./demo.sh stunts   # all 9 idle stunts, frame by frame

demo.sh pins SL_TICK, so it is also the smoke test — every state, a non-git directory, and malformed input all render or the script is broken.

---

Notes on correctness

Two things this script is deliberate about, because both are easy to get wrong:

Caching. Git calls are cached ~5s, keyed by session_id. Keying a cache off $$ looks reasonable and never hits — every render is a new PID.

Field parsing. The JSON payload is read one field per line, not tab-separated. Tab is IFS whitespace, so IFS=$'\t' read -r a b c silently collapses runs of tabs and every field after an empty one shifts left. Since git_worktree is empty whenever you are not in a worktree — the normal case — the working directory would land in the worktree slot and the session id in the working directory slot, quietly killing every git lookup on line 2. IFS= read -r per line keeps empty fields empty.

---

License

MIT © Avenox