craigvandotcom

ac-board

'Read-only pipeline board — the whole factory in one glance: human gates split into decisions and actions, plans by stage, beads by stage, WIP plus CI health, and the agents now active. Observes only; never writes, closes, promotes, or prompts. Triggers: ''board'', ''dashboard'', ''show the board'', ''state of the pipeline'', ''pipeline status'', "what''s the factory doing", ''WIP status'', ''board overview'', ''full board''. To ACT on human gates use ac-human; to reconcile or re-prioritize use ac-align.'

craigvandotcom 0 Actualitzat fa 4 setm.

Recursos

1
GitHub

Instal·la

npx skillscat add craigvandotcom/agent-compounds/ac-board

Instal·la des del registre de SkillsCat.

SKILL.md

You are the factory window. Render the whole pipeline — human gates, plans, beads, WIP, CI, active agents — in one glance. You observe; you never act.

ac-human invokes you to open a session; standalone, this render is the whole response. The shared read is ac-pipeline/references/board-scan.md — consume it, never reimplement a scan. The lens here is counts only, one stacked block per section, every line within 40 columns so a phone never wraps it; bead ids appear only in 🎯 NEXT.

I/O Contract

Input None (reads project state directly). Optional: "org-wide".
Output One board render: verdict first, the ranked 🎯 NEXT moves last. No prompts.
Artifacts NONE — never writes, closes, promotes, archives, labels, or asks.
Verification Every count traces to a scan; an unreadable read renders ?, never a guessed number.

Prerequisites

  • br installed — verify with which br
  • _plans/ (optional — empty renders as —)

Phase 1 — render (one call)

BOARD="$(git rev-parse --show-toplevel)/.claude/skills/ac-board/scripts/board.sh"
[ -f "$BOARD" ] || BOARD="$(git rev-parse --show-toplevel)/skills/ac-board/scripts/board.sh"
"$BOARD"

Print its stdout verbatim — it is the board. board.sh runs every read of
ac-pipeline/references/board-scan.md (Scans A · B · E · F + docket-health, no loop-boundary
filter) plus waves, PRs and the agent roster in parallel; render.py derives the verdict
(RUNNING · IDLE · STUCK · EMPTY), reduces every section to counts, and renders ? plus the failing command
for any read that cannot answer. Never re-derive a count it printed; never run the scans by hand.
Its 🎯 NEXT block (top three moves in the pull order — ac-pipeline/references/stage-table.md § Pull order) is the routing — add nothing after it.

Asked "org-wide" → run board.sh --compact inside each .beads/ repo (in parallel) and print
its one verdict line per repo. Asked for a live dashboard → point to board.sh --watch [secs].

Principles

  1. Read-only is sacred — no writes, no br mutations, no label changes, no prompts. If you want to fix what you see, you are the wrong skill: route it.
  2. A glance, not an audit — cheap reads only, counts over prose, one screen if possible. Depth belongs to ac-human (drive the docket), ac-tidy (reconcile) and ac-align (strategy).
  3. No loop-boundary filter — the loop side is the answer to "is the factory running". Show it.
  4. Never guess a count — ? plus the failing command beats a plausible number.
  5. Exit silently — render and stop. The 🎯 NEXT block is the hand-off.