Get it for macOS

Track context usage as a percentage and safely continue long tasks in fresh threads. A meter above the chat input shows how full the context window is: it uses your theme's main color below 20% (light gray in dark mode), then turns yellow at 20%, orange at 30%, red at 40% and dark red at 65%. Each thread has an Auto handoff switch, off until you turn it on. Turn it on below 20% and the thread hands off once it reaches 30%; turn it on at 20% or above and it hands off at the next safe stopping point. The agent saves its plan and progress as notes, and a fresh thread confirms it read them, then continues with the switch on. Turning the switch off cancels a pending handoff.


Overview

View source

See how much of its context window each thread has used, and move a long task to a fresh thread before the window fills. The new thread picks up from notes the agent writes down first.

What you get

  • A context meter above each thread's chat input, showing the percentage of the context window in use.
  • An Auto handoff switch beside it, set per thread.
  • A bb context-handoff command and a context-handoff skill that agents use to save their notes and confirm them.

Sub-agent threads are left alone.

Meter colors

The bar changes color as the thread fills:

  • Below 20%: your theme's main color (light gray in BB's dark theme, dark gray in its light theme).
  • 20% to 30%: yellow.
  • 30% to 40%: orange.
  • 40% to 65%: red.
  • 65% and up: dark red.

The percentage is the share of the context window that the provider reports. When a provider reports no usage or no window size, the meter says so and automatic checks wait.

Auto handoff switch

A thread you start has the switch off and shows only the meter.

  • Turned on below 20%: the thread keeps working, a tick marks 30% on the bar, and the handoff is requested once usage reaches it.
  • Turned on at 20% or more: the handoff is requested at once, for the next safe stopping point. An idle thread starts a turn to do it.
  • Turned off: a pending handoff is cancelled and the thread carries on. A new thread that already confirmed its notes keeps running.

A request never cuts off a turn or a tool. It waits for the current turn, and the agent finishes the step it is on before it writes its notes.

What a handoff does

The agent saves the original request and plan in PLAN.md, and its progress, decisions and next step in HANDOFF.md, inside a .handoff folder in the workspace. The plugin then starts a new thread with the same workspace and provider. That thread must read the notes and confirm their checkpoint with bb context-handoff resume before it does any work, and the old thread shows a link to it.

The new thread starts with the switch on, so a long task keeps handing off at 30%. Every new part is a direct child of the first thread, so nesting stays one level deep.

Requirements

  • BB with plugin SDK 0.6.132 or newer (tested on BB 0.49), and a provider that reports token usage and context window size.
  • An agent that follows the checkpoint request. The handoff happens at its next safe stopping point, not at an exact percentage.
  • Keep the .handoff notes folder out of git.

bb context-handoff status shows a thread's switch, stage and usage. bb plugin config context-handoff set enabled false pauses handoffs for every thread.

More in Token Usage & Limits 41

Understand or control token, context-window, and provider-quota use.

View all