The status line · Claude Code

Only what nothing else can show.

Claude Code lets you replace the status line with your own script. Mine shows five figures, and the choice of which five is the whole design: these reach the status-line command and are written nowhere else, so no other tool on the machine can display them. Everything the HUD can compute for itself was removed from here, because two readouts of one session are only worth having while neither restates the other.

It also spawns nothing. The previous tenant of this line ran roughly forty-seven short-lived helpers per render, and every process launch on this machine is scanned before it starts — so the cost was paid forty-seven times, ten seconds apart, forever. Every field is now read with the shell's own pattern matching: one process per refresh, no children.

The five figures

BILL

Claude Code’s own billed total for the session. The HUD prices the transcript instead and lands a few percent under, so this is the figure that settles it.

EDITS

Lines added and removed — the tool’s own tally of what IT changed, which is not the same question as the git diff and is not answerable from one.

CONTEXT

Percentage of the window used, stated against the REAL window size. Anything reading the transcript has to infer that size; this does not.

CACHE

The cache read on the request in flight, this instant. Every other reader sees the last COMPLETED response, which is a different moment.

MODEL

The model id with its context-variant suffix. The transcript records the bare name and never the variant, so this is the only place the distinction survives.

The dollar figures are an API-equivalent usage gauge — what the same tokens would cost on the metered API — not a real subscription charge. It’s there to keep me honest about how much a given approach actually burns.

Take it

The whole script is public — one file, no dependencies beyond jq. A hook auto-publishes it every time I change it, so the gist is always the current build.

View the gist on GitHub