Auto-Coder Multi-Agent

Auto-Coder Multi-Agent is a DAG-driven orchestration tool that reads a project spec (app_spec.txt), generates executable tasks, and runs multiple coding agents in parallel waves with testing, review, retries, and progress tracking.

Key Features

• Automatic planning from spec to task_list.json
• Parallel execution based on dependency DAG
• Task-level context sharing via .context/task_<id>_summary.json
• Review loop (single reviewer or functional+security cross-review)
• Permission sandboxing (enabled by default)
• Reliability controls (timeouts, retries, DAG validation, stale-task recovery)
• Cost guardrails (--max-cost, --max-task-cost)
• Optional git worktree isolation for conflict reduction

Repository Layout

dancodingagent/
├── run.py                # Main orchestrator entrypoint
├── planner.py            # DAG generation + plan review
├── scheduler.py          # Task state, DAG ops, dynamic proposals
├── worker.py             # Claude/Droid session runners
├── git_ops.py            # Git init/commit/worktree/conflict utilities
├── progress.py           # CLI progress + .progress/state.json writer
├── prompts/              # Coder/reviewer prompt templates
├── app_spec.txt          # Project specification input
├── task_list.json        # Generated DAG + runtime task states
├── test_changes.py       # Regression tests
└── logs/                 # Per-task session logs

Requirements

• Python 3.10+
• git
• claude CLI available in PATH
• Optional: droid CLI (if using --review-backend droid)

Droid Backend via CLIProxyAPI

If you need to route Droid through a server built from a subscription source, see:

• https://github.com/router-for-me/CLIProxyAPI

CLIProxyAPI can convert subscription configs into a server endpoint that can then be used by Droid.

Quick Start

Run on a target project directory (must contain app_spec.txt on first run):

python3 run.py --project-dir ./your_project

Example with review and dashboard:

python3 run.py \
  --project-dir ./your_project \
  --max-workers 4 \
  --review \
  --review-mode cross \
  --review-backend droid \
  --sandbox \
  --dashboard

Important CLI Options

• --project-dir: target project root
• --max-workers: max parallel agents (default: DAG-derived)
• --max-waves: cap execution waves
• --review: enable task review/fix loop
• --review-mode {single,cross}: one reviewer or dual reviewer mode
• --review-backend {claude,droid}: review execution backend
• --test-timeout: timeout for task test command
• --session-timeout: timeout for agent session
• --max-cost: total cost circuit breaker
• --max-task-cost: per-task cost circuit breaker
• --sandbox/--no-sandbox: permission sandbox toggle
• --worktree-isolation: run tasks in separate git worktrees
• --review-plan: review generated task DAG before execution
• --replan: regenerate task DAG from updated spec

Execution Flow

1. Generate task DAG from app_spec.txt (or load existing task_list.json)
2. Select ready tasks (dependencies satisfied)
3. Run a wave of tasks in parallel
4. Persist summaries, update statuses, optionally run reviews/fixes
5. Repeat until all tasks are terminal (done/failed/skipped)

Runtime Artifacts

• task_list.json: source of truth for task statuses
• .context/: dependency summaries, structured task results, proposals
• .review/: reviewer verdicts and fix logs
• logs/: per-task session logs
• .progress/state.json: live dashboard state when --dashboard is enabled

Testing

python3 test_changes.py

Security Notes

• The default sandbox blocks high-risk operations (for example sudo, destructive commands, and git push).
• A repository scan for common secret patterns (private keys, API key/token signatures, and raw IPv4 literals) found no hardcoded private keys or API tokens in tracked project files.