BD Swarm Auto - Autonomous Backlog Completion

YOU are the conductor. Execute this workflow autonomously. Do NOT ask the user for input.

Architecture: Parallel Terminal Workers

Spawn 3-4 terminal workers max, each handling 1-2 issues with skill-aware prompts.

BAD:  10 terminals x 1 issue each    -> statusline chaos
GOOD: 3-4 terminals with focused prompts -> smooth execution

Quick Reference

Step	Action	Reference
1	Get ready issues	`bd ready --json`
2	Calculate workers	Max 4, distribute issues
3	Create worktrees	Parallel, wait for all
4	Spawn workers	TabzChrome API
5	Send prompts	Skill-aware prompts with issue context
6	Poll status	Every 2 min, check context
7	Merge & cleanup	Kill sessions first
8	Visual QA	tabz-manager for UI waves
9	Next wave	Loop until empty

EXECUTE NOW - Wave Loop

# Get ready issues
READY=$(bd ready --json | jq -r '.[].id')
[ -z "$READY" ] && echo "Backlog complete!" && exit 0

# Count and distribute to max 4 workers
ISSUES_COUNT=$(echo "$READY" | wc -l)
# 1-4: 1-2 workers, 5-8: 2-3 workers, 9+: 3-4 workers

# Create worktrees (parallel)
for ISSUE_ID in $READY; do
  ${CLAUDE_PLUGIN_ROOT}/scripts/setup-worktree.sh "$ISSUE_ID" &
done
wait

# Spawn workers, send prompts, monitor
# ... see references/wave-execution.md

# FULL CLOSEOUT: Use wave-done skill (recommended)
/conductor:wave-done $READY
# Runs: verify workers → kill sessions → merge → build → review → cleanup → push → summary

# QUICK CLEANUP (alternative, skip review):
# ${CLAUDE_PLUGIN_ROOT}/scripts/completion-pipeline.sh "$READY"

# Check for next wave
NEXT=$(bd ready --json | jq 'length')
[ "$NEXT" -gt 0 ] && echo "Starting next wave..." # LOOP

Key Rules

NO USER INPUT - Fully autonomous, no AskUserQuestion
MAX 4 TERMINALS - Never spawn more than 4 workers
SKILL-AWARE PROMPTS - Include skill hints in worker prompts
YOU MUST POLL - Check issue status every 2 minutes
LOOP UNTIL EMPTY - Keep running waves until bd ready is empty
VISUAL QA - Spawn tabz-manager after UI waves
MONITOR CONTEXT - At 70%+, trigger /wipe:wipe

Worker Prompt Template

## Task: ISSUE-ID - Title

[Explicit, actionable description - what exactly to do, not just "fix the bug"]

## Context
[WHY this matters - helps Claude generalize and make good decisions]

## Key Files
- path/to/file.ts (focus on lines X-Y)
- path/to/other.ts

## Guidance
Use the `/skill-name` skill for [specific aspect].
Follow the pattern in [existing-file.ts] for consistency.

## When Done
Run `/conductor:worker-done ISSUE-ID`

Prompt Guidelines:

Be explicit - "Fix null reference on line 45" not "fix the bug"
Add context - Explain WHY to help Claude make good decisions
Reference patterns - Point to existing code for consistency
Avoid ALL CAPS - Claude 4.x overtriggers on aggressive language
File paths as text - Workers read files on-demand, avoids bloat

Full guidelines: references/worker-architecture.md

Visual QA with tabz-manager

After completing a wave with UI changes, spawn tabz-manager in a separate terminal for visual QA:

# Via TabzChrome API
curl -X POST http://localhost:8129/api/spawn \
  -H "X-Auth-Token: $(cat /tmp/tabz-auth-token)" \
  -d '{"name": "Visual QA", "command": "claude --agent conductor:tabz-manager --dangerously-skip-permissions"}'

tabz-manager can take screenshots, click elements, and verify UI changes work correctly.

Context Recovery

At 70% context, run /wipe:wipe with handoff:

## BD Swarm Auto In Progress
**Active Issues:** [list in_progress IDs]
**Action:** Run `/conductor:bd-swarm-auto` to continue

Auto vs Interactive

Aspect	Auto	Interactive
Worker count	All ready	Ask user
Waves	Loop until empty	One wave
Questions	Make defaults	AskUserQuestion ok
Context	Auto /wipe	Manual

Troubleshooting

Problem	Solution
Worker not responding	`tmux capture-pane -t SESSION -p -S -50`
Merge conflicts	Resolve manually, continue
Worker stuck	Nudge via tmux send-keys
Worker failed	Check logs, re-spawn or close with 'needs-review'

Execute this workflow NOW. Start with getting ready issues.

bd-swarm-auto