Redline

Context fill, session cost and pace on one line, with a warning before the red line.

ctx 82% ████████░░ RED · $4.12 · 3.9k tok/min · full in ~7m · 5h 46% · 7d 13%

Use it

• Status line (under the prompt), updated on every model request:
  • context fill as a percentage with a 10-cell bar, and the level as a word: nothing under 60%, AMBER from 60%, RED from 80% (the status line has no colour, so the word carries it);
  • the session cost so far: Claude Code's own figure (what /cost shows) where it keeps one, shown plain; otherwise an estimate from the price table, marked ≈;
  • new tokens per minute over the last 10 minutes;
  • the pace: when the context hits 100% at the rate it grew over the last 10 minutes (full in ~25m);
  • your plan's 5-hour and weekly windows (5h 46%, 7d 13%), with out in ~40m added when a window will run dry before it resets.
• Toast at 80%: one toast suggesting /compact or a fresh session (/clear), once per crossing. It re-arms only after the context drops under 75% (a compaction, a fresh session), so it never nags on every redraw.
• /redline opens the detail view in a pane: context used against the window, the pace and its rate, input / output / cache read / cache write tokens, the cost (Claude Code's own ledger, plus a per-category estimate from the price table), time in session, each plan window with its reset time and run-out estimate, and the five busiest turns. Buttons: show or hide the band, close. Where no surface places the pane, the same view is printed as the command's output.
• /redline band toggles a band above the prompt with the same gauge, in colour (green, amber, red), for people who hide the status line. Its Hide button turns it off. The choice is remembered across sessions. It steps aside while a survey holds that space.

Settings

None. The thresholds (60%, 80%, re-arm under 75%) and the 10-minute window are fixed.

What it can touch

• Reads: what Claude Code reports about the session ($.session.usage(), the session.measure event), each model request's token counts (turn.step), each turn's prompt text (turn.start, shown only in the busiest-turns list), and compactions and /clear.
• Writes: one remembered value in the plugin's own store (whether the band is on). No files.
• No processes, no network, no model calls. It costs nothing to run.

Where the numbers come from

• Context window: Claude Code reports it for the model in use, so a 1M-token model reads correctly. No table is involved.
• Context fill: Claude Code's own figures (context.tokens and context.percent, the status line's used percentage) whenever it has measured the latest request. Between its measurements the gauge counts the main conversation's last request itself, input side (uncached + cache read + cache write), the same way, so it moves on every request. Subagent requests count toward tokens and cost, never toward the session's context.
• Session cost: Claude Code's own total (cost.usd, what /cost shows), shown without ≈ since it is not redline's estimate. Where Claude Code keeps no ledger, redline's built-in price table prices the requests it counted (marked ≈), and a model the table does not know shows tokens only, never a guessed price.
• Price table (hooks/prices.ts): list prices per model family as of 2026-10-08, used for the per-category and per-turn estimates. It leaves out long-context premiums, fast mode, 1-hour cache writes, batch and regional pricing, so every figure from it is marked ≈.
• Plan run-out: the average pace a window has been used at since it opened (its reset time less its length). Shown for the 5-hour and weekly windows; a gateway's spend limit is shown with no estimate.

Limits

• Plan windows appear only on a Claude subscription, after the first reply of a session; with an API key there are none, and the detail view says so.
• Token counts (and the price-table estimate) start when redline loads; the ledger cost and context fill cover the whole session.
• The pace needs a minute of growing context inside the last 10 minutes. Flat or shrinking context gives no estimate. Right after a compaction the fill is Claude Code's own estimate of the compacted size until the next reply measures it.
• The status line and band update on model requests and measurements, not on a timer, so the session time in the detail view refreshes with them.
• Terminal and desktop are supported and tested. VS Code and mobile draw the pane through the same hooks but are not tested.

Built from a prompt

This mod was built from its prompt.