Autonomous Builder

Autonomous Builder

Build from idea to working artifact with minimal supervision.

When To Use

• User says "build this in background"
• User says "autonomous", "headless", "overnight"
• User provides idea and wants hands-off execution
• User says "just build it" or "come back with something working"

Core Philosophy

"Tell it an idea, come back with the thing built."

Like a capable underling: takes an idea, uses best judgment at every step, and returns with something tangible (50-99% complete).

Quick Start

# From project directory (runs in resilient tmux session):
oneshot-build "A CLI tool that converts markdown to PDF"

# Monitor progress:
tail -f .agent/STATUS.md

# Attach to running session:
tmux attach -t oneshot-build

# Check tasks:
bd list --json

Resilient by default: All builds run in tmux. Survives terminal disconnect.

How It Works

Phase 1: Planning (Automatic)

1. Uses front-door skill to analyze idea
2. Creates structured plan with create-plan
3. Parses plan into beads tasks with dependencies
4. Commits plan file

Phase 2: Building (Loop)

while tasks remain:
  1. bd ready --json → pick next task
  2. bd update <id> --status in_progress
  3. Implement the task
  4. git commit after each file
  5. bd close <id> --reason "commit: hash"
  6. Check for stuck condition

Phase 3: Completion

• All beads tasks closed
• Final status written to .agent/STATUS.md
• Git history contains all work

.agent/ Directory

Created automatically for each build:

.agent/
├── STATUS.md        # Real-time progress log
├── ITERATIONS.md    # Current iteration count
├── LAST_STATE.md    # Beads state hash (for stuck detection)
├── LAST_ERROR.md    # Last error if any
└── SCRATCHPAD.md    # Working notes

Loop Detection

Prevents infinite loops via:

When stuck, writes to .agent/LAST_ERROR.md and stops gracefully.

Configuration

Environment variables:

MAX_ITERATIONS=100      # Maximum build iterations
STUCK_THRESHOLD=5       # Iterations without progress
CLAUDE_CMD=claude       # Claude CLI command

Manual Mode

If you want more control, run phases manually:

# Phase 1: Plan only
claude -p "Use front-door, then create-plan for: [idea]. Create beads tasks, then stop."

# Phase 2: Build one task at a time
claude -p "Run bd ready, implement one task, commit, close task, stop."

# Repeat Phase 2 until done

Decision Making

The autonomous builder uses these principles:

1. Prefer simple - Choose straightforward solutions
2. Commit often - Every file edit gets committed
3. Use existing - Leverage ONE_SHOT patterns and skills
4. Best effort - 50% working is better than 0% perfect
5. Stop gracefully - When stuck, save state and stop

Resilient Execution (v7.4)

All builds run in tmux by default - survives terminal disconnect, SSH drops, etc.

How It Works

oneshot-build "idea"
  ├─ Creates tmux session: oneshot-build-{timestamp}
  ├─ Starts heartbeat (every 30s)
  ├─ Starts checkpointer (every 5 min)
  ├─ Runs Claude with aggressive sync rules
  └─ Logs everything to .agent/session.log

Aggressive State Persistence

During build, state is saved:

• Before each task: bd update && bd sync
• After each task: bd close && bd sync
• Every commit: bd sync
• Every 5 minutes: checkpoint commit + bd sync
• On any error: bd sync (preserve before crash)

If Terminal Disconnects

The build continues! When you reconnect:

# 1. Check if session running
tmux list-sessions | grep oneshot

# 2. Reattach
tmux attach -t oneshot-build

# 3. If session died, check beads
bd sync && bd ready --json

# 4. Resume from state
claude -p "Resume from beads"

Recovery

If the builder stops unexpectedly:

# Check what happened
cat .agent/STATUS.md
cat .agent/LAST_ERROR.md
cat .agent/HEARTBEAT | tail -5  # When did it die?

# See remaining tasks
bd sync
bd ready --json

# Check session log
less .agent/session.log

# Continue manually or restart
oneshot-build "same idea"  # Will pick up from beads state

Integration with ONE_SHOT

The builder uses these skills internally:

• front-door - Initial interview and spec
• create-plan - Structured planning
• implement-plan - Task execution
• beads - Persistent task tracking
• failure-recovery - Loop detection

Example Session

$ oneshot-build "A Python CLI that fetches weather data and displays it nicely"

=== ONE_SHOT Autonomous Builder ===
Idea: A Python CLI that fetches weather data...
Max iterations: 100

Initialized .agent/ directory
Phase 1: Planning
- Created epic: weather-cli
- Created 8 tasks in 3 groups
- Plan saved: thoughts/shared/plans/2025-01-03-weather-cli.md
Planning complete

Phase 2: Building
Iteration 1: Building
- Implementing: Create project structure
- Committed: feat(setup): create project structure
Iteration 2: Building
- Implementing: Add API client
- Committed: feat(api): add weather API client
...
Iteration 8: Building
- Implementing: Add CLI interface
- Committed: feat(cli): add rich CLI display

All tasks complete!

=== Build Complete ===
Check .agent/STATUS.md for details

Keywords

autonomous, headless, background, overnight, build it, just do it, hands-off, come back with