Shop Kanban

What you get

• A board docked at the bottom of bb with Queue, WIP, Blocked and Done columns. Every non-archived top-level thread is on it, and child threads roll up into their parent thread.
• Columns follow live thread state, so you never drag a thread. Queue means not started. WIP means the thread or one of its children is running, background work included. Blocked means idle and waiting on you: a pending interaction, an error, a failed result, or a last reply that asks a question. Done means idle with a last reply that reports completion; unread Done threads come first.
• Each thread shows a [category] intent title, a summary of the latest exchange (two lines, three on hover), the update time and status, what the thread has spent together with its child threads and subagents (archived ones included; ~ when part of it is estimated), a ring with the share of the context window in use (amber at 70%, red at 90%), and the icon of the harness running the thread. Rest the pointer on a card for a tooltip beside it with the harness, model, thinking level, tokens, cost, access mode and service tier.
• Clicking a thread opens it at the latest message. The Queue + button starts a new thread from a form at the top of Queue that keeps your draft across reloads and crashes. The thread open in bb is highlighted on the board and kept in view.
• The search bar in the board's title bar finds any thread by what was said in it, archived threads included. Picking a result opens the thread in bb and leaves the board as it is.
• A Reviews section near the top of Queue lists open pull requests waiting on your review on Delta and GitHub. A Me | Team toggle switches between requests of you by name and of your teams, starting on Me; the reviewSearchesMe and reviewSearchesTeam settings pick the searches. Each is a two-line card: a risk chip and title, which opens a drafted new thread asking what changed, why and how risky it is; then age, author, and links to open it on Delta or GitHub, approve (click twice), or request changes with a reason. Archive hides one for good. Ones you already reviewed come back marked re-review only when new commits land.

For agents

Agents get four tools to track their own work: move their thread, spawn a child thread, spawn one independent thread per theory for an investigation, and mark a result as success (Done) or failure (Blocked). A bundled skill tells agents when to use them. The same actions are available from the terminal as bb shop-kanban.

Model-written labels

Titles, summaries, and the Blocked-or-Done call for finished replies come from a small model. The plugin runs a pi-compatible CLI on the thread's host, by default pi with anthropic/claude-haiku-4-5, so labels are billed to that CLI's account. Board titles are written back to the bb thread title unless you rename the thread yourself. Set labelerCommand to empty to turn labels off; threads then use heuristic titles and an excerpt of the latest message.

Requirements

• bb 0.47 or later.
• For model-written labels: a pi-compatible CLI on each host that runs threads.
• For thread search: qmd (npm install -g @tobilu/qmd) on a connected persistent host.
• For the review queue: dt (Delta) and/or gh (GitHub), signed in, on that same host.

Each is optional. The board works without any of them: without a labeler, threads keep heuristic titles; search shows how to install qmd when you use it; and review searches that can't run are skipped quietly, with the reason on hover over the Reviews header.

If you replace bb's omp agent with a custom ACP agent entry, bb draws it with a generic toolbox glyph. The plugin draws bb's omp mark for it instead, in omp's pink (#f84fcc), in bb's sidebar and pickers as well as on the board.