What it does

Set a budget any time, inside Claude Code:

/redline 10m $5

You get a live burn-down in your statusline, and an agent that paces itself to finish inside it:

redline  ████████████  ⏱ 1:00  ·  💰 $1.85 left

The promise (why it holds):

• 🔴 Never over. The budget is a hard ceiling.
• 🏁 Under is success. It lands around 90% with the task done.
• 🪂 Never abrupt. It always finishes with a usable result. No mid-sentence kill.

One budget covers everything the session spawns. Fanned subagents share the same umbrella, counted and enforced. You never budget a subagent separately. How it works.

Works with Claude Code today. Codex and other agents are on the roadmap.

Install

Needs the Node that ships with Claude Code. Pick one, then restart Claude Code:

Homebrew

brew install sankalpgunturi/redline/redline && redline install

npm (package is redline-cc - redline was already taken on npm; the installed command is still redline)

npm install -g redline-cc && redline install

curl

curl -fsSL https://raw.githubusercontent.com/sankalpgunturi/redline/main/install.sh | bash

git

git clone https://github.com/sankalpgunturi/redline && cd redline && ./install.sh

It backs up your settings, never clobbers an existing statusline, and is safe to re-run. Remove anytime with redline uninstall.

Use

| Command | Budget | |---|---| | /redline 10m | 10 minutes | | /redline 2.5h | 2.5 hours | | /redline $5 | $5 | | /redline 30m $5 | 30 min and $5 | | /redline 1h 200k | 1 hour and 200k tokens | | /redline 80% | plan ceiling: stop when your 5-hour window hits 80% used | | /redline +10% | plan allowance: spend 10 more points of the window from now | | /redline 80% 7d | ceiling on the 7-day window instead | | /redline off | clear |

Invalid or zero/negative values (/redline abc, /redline -100) are rejected with an error and a usage example - no budget is set.

Whichever dimension is closest to its limit drives the bar (tagged with its gauge). Each one counts down, like ⏱ 1:00 left and 💰 $1.85 left.

Plan % budgets are anchored to your plan's current state: /redline 80% when the window is already 52% used starts the bar at 52/80, shows the absolute window level plus its reset countdown (plan 52% ↺1:23), and tells you your headroom when you set it. Even with no budget set, the statusline shows where your window sits. As a bonus, redline observes what 1% of your plan actually costs in tokens (1% ≈ 85k tok in redline pulse) - the number the usage page doesn't tell you.

Fleet

Run ten sessions at once? The shared plan window is the resource that actually runs out. Give the machine one ceiling:

redline global 60%

Every session on the machine - budgeted or not - watches the shared window and starts its landing sequence as it approaches 60% used. Each session lands softly with what it has; nothing is killed. redline pulse shows the cap, the window, and each session's estimated share of it. Optional landing-zone pings: redline notify on (a desktop notification when any session crosses 70% or 90%).

Commands

| Command | What | |---|---| | redline pulse | all your active budgets, live, plus the landed-rate (local, no network) | | redline global 60% | fleet budget: machine-wide ceiling on the shared plan window | | redline notify on | desktop ping when a session enters the landing zone (off to disable) | | redline doctor | check the wiring: hooks, statusline sensor, plan feed | | redline uninstall | remove redline from Claude Code | | redline version | |

Pulse

Every active budget across your sessions, live, plus the one number that matters: did the session land within budget? All local, no network. When you run several sessions at once, pulse also estimates each session's share of the shared plan window (~3.1% of plan), so you can see which one is eating it.

Landings are graded, not just counted: a clean landing means the model wrapped up on its own before the 90% lock ever fired, and each landing records the model's own manifest (LANDING: delivered X; cut Y), so you can see what a budget actually cost you in scope.

redline pulse

  redline · did it land within budget?
  ██████████████░░░░░░░░░░  60% landed  (3/5 sessions)
  landed ≤100%: 3     over budget: 2 · avg overshoot 21%