Get it for macOS

Shows up to three prompts you are likely to type next after each finished turn, one key from your draft. Click a card, or press Tab or → in the empty composer, to put it in your draft; nothing is sent unless you set clicks to send. Show them as numbered cards above the composer, as grey text inside it, or both. Choose the model with bb's provider and model picker: Claude Code and Codex models run as one CLI call with no tools, and any other provider runs as a hidden worker thread. Only the thread's user and assistant messages are sent, never tool output, and hidden and child threads are skipped. Each suggestion request runs on a provider you already use; the default, Claude Sonnet 5.5, needs the claude CLI signed in on the bb server.


Overview

View source

See the instruction you were about to type before you type it.

What you get

  • Up to three suggestions after each finished turn: numbered cards above the composer, the best one as grey text inside the empty composer, or both.
  • Click a card to put it in the draft, or set clicks to send at once. Press Tab or → in the empty composer, or swipe right on a touch screen, to take the best one. Nothing is sent without your action.
  • Silence when there is nothing to suggest: no "push it" after a commit, no "check the build" while the agent says it is running.
  • bb next-prompt to show, generate, and debug suggestions.

How it works

When a visible top-level thread goes idle, the plugin sends the first request and the newest user and assistant messages (no tool output) to a model, Sonnet by default. It keeps the short suggestions the model returns as JSON until the thread moves on, across reloads and restarts. Nothing is written to the thread or the agent's context.

Requirements

Pick the model with bb's provider/model picker. Claude Code and Codex models run as a lean CLI call (claude or codex logged in on the bb server) and use your existing subscription; other providers run as a hidden worker thread. An Anthropic- or OpenAI-compatible HTTP API can replace either.

More from Serge Baranov

More in Thread Content 99

Change what people see or do inside an open thread.

View all
Generative UITurns agent replies into charts, diagrams, maps, plans and forms you can use, right in the chat. Agents choose from 86 components on their own, with no setup: sliders and filters recompute in the page, and buttons, choices and forms put your answers in your message, which reaches the agent when you send it. Test reports, logs, lint errors and coverage files render as their own components instead of pasted text, and a thread's changes, commits, pull request and workers are read live from BB. Adds about 6,400 tokens of instructions to sessions whose replies you read, and looks up photos, places and map tiles on Wikimedia Commons, Wikipedia, Photon and OpenStreetMap unless you turn web lookups off. Relies on BB internals, so a BB update can show the blocks as code until the plugin catches up.ComicBit1.1K
Agent TracesSee where your agents' time goes, from every thread down to each command, tool call, and subagent of a turn. A turn opens as a timeline that splits the agent's time into waiting on you, tools, subagents, thinking, writing, and model time, and it grows live while the agent works. Select a call to see its exit code, output, and arguments, or jump to the reply that followed it in the chat. A Traces page in the sidebar compares your recent threads by active time, calls, and errors, and two turns can be compared phase by phase. Turns export as OpenTelemetry or an HTML snapshot, and the bb agent-traces command gives agents the same answers. Traces come from the thread history on your bb server; an OpenTelemetry collector is optional.Nicolas Charpentier390