Skip to main content

Session collaboration

If you run several Agent sessions at once, you can let them cooperate: one on the frontend, one on the backend, without you relaying messages by hand.

Stable session names​

Collaboration requires being able to name your targets. Give each Agent session a stable name (like frontend, tests), fixed across the whole project and preserved when sessions reopen.

Naming your sessions in the Belfry sidebar (reviewer, frontend, tests, and so on) is itself the signal that gives collaboration clear targets. "Hand it to frontend" refers to that session.

The built-in belfry CLI​

Collaboration happens through the built-in belfry command line, across sessions in the same project:

  • send — dispatch a task to a session
  • wait — wait for the other side to finish
  • done / fail — report completion or failure

Tasks are queued per target Agent. Delivery is held while the other side is busy or waiting for confirmation, and resumes in order once it's idle — messages never get injected into a conversation already in progress.

Approval gates​

Collaboration tasks can have approval gates: sensitive operations need your sign-off before executing. There are also loop and hierarchy limits to prevent sessions endlessly delegating to each other.

Collaboration status panel​

One panel shows it all: who is waiting on whom, how many tasks are queued, and what has been delivered.

Environment diagnostics​

Collaboration depends on a runtime environment (the Belfry skill, CLI login state, collaboration channels). The settings page provides diagnostics for:

  • Whether the Belfry skill is present and up to date
  • Codex / Claude Code login and feature status
  • doctor results and collaboration channel connectivity

Diagnostics can sync the built-in skill automatically or manually. Each client reports its status separately and partial successes are preserved — if Claude is fine but Codex isn't logged in, one failure doesn't bury the other.