Composer Predictions

Get it for macOS

Suggests the message you're likely to send next as ghost text in the empty composer, one Tab from an editable draft. Nothing is sent for you: Tab only fills an empty message box, and typing, pasting or Escape hides the suggestion until the conversation moves on. Suggestions are saved across restarts and checked against the conversation before they reappear, so an unchanged thread costs no new request. Every reply in a thread you've opened costs one request in a hidden helper thread, which receives up to 48,000 characters of recent messages and tool output.


Overview

View source

After an agent replies, the next thing you type is often obvious. Composer Predictions reads the conversation and suggests that message as faint ghost text in the empty message box. Press Tab and it becomes an ordinary draft you can edit. Nothing is ever sent for you.

What you get

  • Tab to insert. Plain Tab fills the empty box with the suggestion and focuses it. Shift+Tab keeps its usual behaviour.
  • Out of your way. Typing, pasting or Escape hides the suggestion until the conversation moves on. Nothing shows while the thread is running or the draft has text, mentions or attachments.
  • Your choice of model. By default it uses the conversation's own provider and model at low reasoning effort. In Settings → Plugins → Composer Predictions, pick a separate provider, model, effort and service tier with BB's native picker.
  • Survives restarts. Suggestions, abstentions and dismissals are saved and checked against the current conversation before they are shown again.
  • Silent when unsure. The model can suggest nothing when there is no clear next move.
  • bb composer-predictions show|regenerate|dismiss <thread-id> and a bundled agent skill for checking why a suggestion is or isn't showing.

Safety and cost

Tab only fills an empty draft through BB's public composer API, and the plugin never submits a message. Each prediction is one request in a hidden helper thread, never in your conversation, archived and stopped as soon as it finishes, fails or reaches its 60-second timeout. Normal provider usage applies. Once you open a thread with an empty message box, each later reply there costs one request, even when you are not looking at it, until the thread is archived or deleted or the plugin restarts. At most two helpers run at once, and an unchanged conversation reuses its saved result.

Each request carries up to 48,000 characters of recent messages and tool output by default. A separate prediction model's provider receives that too. The plugin's database stores the suggestion, its status and a fingerprint, not your conversation. BB keeps each helper's history as an archived hidden thread. The helper is told not to use tools and runs in auto permission mode; it is stopped if it asks for approval or input, but it is not a tool-free sandbox.

How it works

The plugin rebuilds up to 400 recent events (your messages, the agent's replies and completed tool records), capped at a character budget of 48,000 by default. A fingerprint of the conversation, prompt, model and budget decides whether a saved result can be reused. BB has no ghost-text slot yet, so the overlay relies on BB 0.46 composer markup; if it cannot find its own composer, it shows nothing and leaves Tab alone.

Requirements

bb 0.46 or later, and a provider that supports BB's auto permission mode, with the prediction model available on the conversation's machine.

More from MacHatter1

More in Thread Content 59

Change what people see or do inside an open thread.

View all