beady-eye

bdi draws each beads tracker as a tree of its work, each piece of work above the beads that make it up, and redraws the tree as beads change. Run it inside a repository and you see that project. Name several projects in a config file and you see them all on one screen.

Each claimed bead shows the herdr pane its agent is working in. Select the bead and the last lines of that pane appear underneath. You read what the agent is doing on the same screen as the work, and a claim with no pane behind it is flagged.

A bead can wait on a bead in another project. bdi finds that blocker in its own tracker and draws it beneath the bead that waits. bdi --beads lists the work that is ready, and unlike bd ready it counts a blocker in another project.

An agent asks you a question on a bead and moves on to something else. Answer, put it off or leave a comment, and the agent has it.

Agents claim, comment on, close and file beads as they work. To wait on another agent, an agent needs only that agent's bead.

beads itself ships no feed of changes. The watcher's socket is one. A script connects, names the beads it watches, and gets one line of JSON for each bead as it stands, then another each time the bead changes, from its status to its labels and comments. This is how a bar widget, a mirror into another system or a relay to a chat follows beads without reading the tracker, and a text or an email to you when a bead moves is a listener you write on this socket. bdi --json and bdi --beads give a one-off snapshot instead.

Any job that can reach the tracker can write to a bead. A nightly build or a deploy writes once when it finishes, and the agent that waited on it carries on.

Your agent can wait on GitHub without polling it. Whoever starts the wait makes a gate, and bdi gates closes it when the pull request merges, the issue closes, the release ships, the workflow run succeeds, the timer runs out or another bead closes in any configured tracker. A failing check, a conflict with the base or a review leaves the gate open and puts one comment on the waiting bead. It looks again every minute, and with --listen it takes GitHub's webhook too. It ships with beady-eye, and docs/configuration.md has every trigger.

bdi watch reads every configured tracker, including one on a shared Dolt server with no checkout on the machine, and holds what it read. It sees a change within one poll of the tracker, or within a fraction of a second when something reports that the tracker changed. A heartbeat every 20 seconds lets a listener tell a quiet watcher from a dead one.

A session calls watch on a bead and carries on. When the bead changes, the plugin puts a message in front of the session saying what changed, busy or idle. A new status, a comment or the bead becoming ready all count. So to reach an agent, comment on its bead or change its status. If the watcher stops answering, the plugin says that too.

This is the screen above. While the watcher runs, bdi redraws it from the watcher as beads change, and runs no bd of its own.

When a listener has put a question to a person, bdi bd <project> carries the answer, deferral or comment back to that project's tracker with that project's own bd. Whatever relays the word names the project, and bdi finds the tracker.