beadboard-driver（珠板驱动器）

BeadBoard Driver

Overview

This skill is the operator runbook for agent execution in external repos with BeadBoard as control plane.

Core principle: explicit state + explicit assignment + explicit evidence.

What is BeadBoard

BeadBoard is a real-time dashboard for bd-backed agent work — it surfaces agent liveness, state transitions, swarm progress, and inter-agent mail in a live UI. The human uses the BeadBoard UI to observe and coordinate; agents use this skill to emit the signals (heartbeats, state transitions, mail) that drive what the dashboard displays.

Agents almost never work inside the BeadBoard repo. You work in your own project repo. bb and bd are globally available on PATH. Scripts used by this skill (preflight, mail shim, etc.) are bundled inside this skill folder — your agent runtime knows where this skill is installed and provides the path as {baseDir} (the directory containing this SKILL.md).

The Iron Law

No bead claims, handoffs, or completion statements without:
1) assignee set,
2) coordination checked,
3) evidence recorded.

Requirements

• bd must be installed and available on PATH.
• bb or beadboard must be installed globally and available on PATH.
• Work from the target repository root.
• Install bd: npm install -g beads-cli
• Install bb: clone from GitHub and install globally — see Step 0 bootstrap for the exact commands

Session Runbook (Do Not Skip Steps)

Step 0: Read project.md — Cache-First Decision

ls project.md 2>/dev/null && echo "EXISTS" || echo "MISSING"

MISSING → you are the first agent in this project. Run the Bootstrap Checklist below, then return here.

EXISTS → read the Environment Status Cache table at the top of project.md.

• All rows pass → skip Step 1 entirely, go straight to Step 2. The environment is verified.
• Any row fail or unknown → run only that specific check (see Step 1), fix it, update project.md, continue.

project.md is the token budget. Trust it when it's green. Only spend checks on what's actually broken.

Bootstrap Checklist

{baseDir} is the absolute path to the directory containing this SKILL.md, injected automatically by your agent runtime (Claude Code, Codex, etc.). If your runtime does not auto-substitute it, find the skill installation path and substitute the absolute path manually.

Run once, in order, when project.md is missing.

A. Check bd (beads-cli)

which bd 2>/dev/null || where bd 2>/dev/null

If missing:

npm install -g beads-cli

B. Initialize beads database (if not already present)

ls .beads 2>/dev/null && echo "EXISTS — skip" || bd init

Required before any bd config commands. Skip if .beads already exists.

C. Check bb (BeadBoard)

which bb 2>/dev/null || where bb 2>/dev/null

If missing — bb is required for coordination. Ask the user:

"BeadBoard is not installed. It is required for agent coordination. Would you like me to install it now? I'll clone the repo and install it globally."

If the user agrees:

git clone https://github.com/zenchantlive/beadboard.git ~/beadboard
cd ~/beadboard && npm install && npm install -g .

D. Configure mail delegate

node {baseDir}/scripts/setup-mail-delegate.mjs

Self-resolves the shim absolute path and runs bd config set mail.delegate automatically.

E. Run session preflight

node {baseDir}/scripts/session-preflight.mjs

Must pass before continuing. Checks bd, bb, and confirms delegate is set.

F. Create project.md

cp {baseDir}/project.template.md ./project.md

Fill in every field in the Environment Status Cache table and Project Identity section. Set each verified row to pass with today's date. This is what saves the next agent from re-running all of the above.

Append to the Session Log:

| YYYY-MM-DD | `<your-agent-bead-id>` | Initial project.md created, all checks pass |

Step 1: Environment Check (Skip If project.md All Green)

Only run this step if project.md has a fail or unknown row.

Run only the failing check:

{baseDir} is the directory containing this SKILL.md. Your agent runtime substitutes it automatically.

After fixing: update that row in project.md to pass with today's date.

See Platform Notes if running on Windows native or WSL2.

Step 2: Create Agent Bead Identity + Verify Mail

bd create --title="Agent: bb-<role-name>" --description="<scope>" --type=task --priority=0 --label="gt:agent,role:<orchestrator|ui|graph|backend|infra>"

# Register in bb coordination system and set identity
bb agent register --name <role-name> --role <orchestrator|ui|graph|backend|infra>
export BB_AGENT=<role-name>

Naming convention: Name your bd bead bb-<role-name> (e.g. bb-silver-scribe). Register the same name in bb as <role-name> (e.g. silver-scribe). Set BB_AGENT=<role-name>. This bridges both identity systems: bd agent state bb-silver-scribe... uses the bead ID; bd mail send uses BB_AGENT automatically.

Now that BB_AGENT is set, verify the full mail stack:

node {baseDir}/scripts/ensure-bb-mail-configured.mjs

Expected: ok: true with matching delegate and actor. If it fails, see Failure Modes.

Check inbox before proceeding:

bd mail inbox

Read and ack any pending messages before claiming work. Unacked HANDOFF or DECISION messages may affect your tasks downstream.

Then set lifecycle state:

bd agent state <agent-bead-id> spawning
bd agent state <agent-bead-id> running

Step 3: Note Any Environment Changes in project.md

project.md was already read in Step 0. Only update it now if something changed this session — new package installed, delegate reconfigured, new platform quirk discovered. If nothing changed, skip this step entirely.

If you fixed a fail/unknown row in Step 1, update that row to pass with today's date now.

Step 4: Read Hard Memory and Task Context

# Select your primary domain first:
# memory-arch | memory-workflow | memory-agent | memory-ux | memory-reliability
bd query "label=memory AND label=mem-canonical AND label=<domain> AND status=closed" --sort updated --reverse
bd ready
bd show <target-bead-id>

Pick the domain matching your task: memory-arch (architecture decisions), memory-workflow (session protocol), memory-agent (agent setup/identity), memory-ux (UI/UX), memory-reliability (errors/recovery). Domain anchor IDs are in references/memory-system.md.

bd ready lists all unblocked, unassigned tasks ready for pickup. Review the output and select a <target-bead-id> to pass to bd show.

Minimum: read task contract, dependencies, success criteria, and blockers.

Step 5: Claim Work with Assignee + Hook Slot

bd update <target-bead-id> --status in_progress --assignee <agent-bead-id>
bd slot set <agent-bead-id> hook <target-bead-id>
bd agent state <agent-bead-id> working

Never use --claim. Use explicit --assignee.

Step 6: Execute + Heartbeat + Coordinate via Mail

During execution:

bd agent heartbeat <agent-bead-id>

LLM agents (Claude Code): Heartbeat at turn start and immediately before long-running commands (builds, tests). Inter-turn silence is expected — do not treat it as a health failure. The every-5-minutes cadence applies to persistent daemon-backed agents only. Note: The Witness enforcement layer that automatically marks agents dead on missed heartbeats is not yet running. Heartbeats are recorded and visible in the BeadBoard dashboard but not currently enforced. Daemon implementation is a future epic.

Coordinate through delegated mail:

bd mail vs bb agent: bd mail inbox/send/read/ack are thin wrappers that delegate to bb agent inbox/send/read/ack via the configured shim, injecting BB_AGENT as your sender identity automatically. Always use bd mail in this skill. Raw bb agent commands appear in some reference docs as lower-level equivalents — use them only when the shim is not configured or for direct bb-level debugging.

bd mail inbox
bd mail send --to <agent-or-role> --bead <bead-id> --category <HANDOFF|BLOCKED|DECISION|INFO> --subject "<short>" --body "<details>"
bd mail read <message-id>
bd mail ack <message-id>

When blocked:

• send BLOCKED
• set bd agent state <agent-bead-id> stuck
• resume only after intervention/response

See references/coordination-system.md → Inbox Polling Protocol for minimum polling moments and cadence.

Step 7: Verification Gates (Code Changes)

npm run typecheck
npm run lint
npm run test

Do not claim fixed/done without fresh command output from this session.

Step 8: Publish Evidence, Close, and Update Cache

bd update <target-bead-id> --notes "<commands run + key outputs + files changed>"
bd close <target-bead-id> --reason "<what was completed>"
bd slot clear <agent-bead-id> hook
bd agent state <agent-bead-id> done

Update project.md Environment Status Cache:

• If you ran tests: set Tests last run row to pass/fail + today's date
• If you ran preflight: update session-preflight row
• If you ran bb agent register this session: update bb agent registered row to pass | BB_AGENT=<name> + today's date
• Append one line to the Session Log: | YYYY-MM-DD | <agent-bead-id> | <what you verified/changed> |

This update is what saves the next agent from re-running your checks.

Step 9: Memory Review

If reusable lesson exists:

• create/supersede canonical memory decision bead

If no reusable lesson:

• record: Memory review: no new reusable memory.

Red Flags

Stop and correct if you are about to:

• close without --assignee history
• skip bd mail checks at session start/claim/close
• claim completion without gate output
• write project context inside skill folder instead of repo project.md
• use deprecated direct command patterns from old docs

Use-The-Right-Doc Map

• references/session-lifecycle.md: Full end-to-end session choreography.
• references/agent-state-liveness.md: Agent states, heartbeat cadence, liveness interpretation.
• references/coordination-system.md: Canonical bb-mail command surface and category semantics.
• references/coord-events-sessions-ack.md: Trigger map, inbox polling protocol, blocked-to-resume walkthrough.
• references/command-matrix.md: Exact command inventory for day-to-day operation.
• references/failure-modes.md: Deterministic diagnosis and recovery paths.
• references/memory-system.md: Memory anchors, injection protocol, promotion/supersede rules.
• references/archetypes-templates-swarms.md: Swarm composition, molecule operations, worker dispatch patterns.
• references/missions-realtime.md: Real-time/watcher/event troubleshooting.
• references/creating-beads.md: Creating epics, tasks, subtasks with proper naming, dependencies, and workflow.

Bottom Line

If you follow this runbook exactly, any agent can enter cold, coordinate safely, and deliver auditable completion without drift.

Platform Notes

Environment Variable Syntax

Mail Delegate Path Format

When configuring bd config set mail.delegate "node <path>/bb-mail-shim.mjs":

Binary Detection

The preflight script checks bd and bb are on PATH. On native Windows, the system uses where instead of which. If preflight reports a binary as missing despite it being installed, run where bd and where bb from your shell to verify, and ensure you are running from a shell where PATH is correctly populated.