SDD Autopilot

A Claude Code plugin that takes a one-sentence feature description and autonomously produces a reviewed pull request. It breaks the feature into a spec, plan, tasks, implementation, verification, and code review — each handled by a dedicated AI agent — so you get structured, repeatable feature development without managing the process yourself.

Why

Claude Code is great for ad-hoc coding tasks, but building a complete feature — with a spec, architecture decisions, task breakdown, implementation, tests, and review — requires you to drive every step. SDD Autopilot automates that entire ceremony. You describe what you want, it handles the how, and you review the PR.

Quick Start

# Install from Claude Code marketplace
# Search for "sdd-autopilot" and install

# Or from source:
git clone https://github.com/rubenzarroca/sdd-autopilot.git ~/.claude/plugins/local/sdd-autopilot
cd ~/.claude/plugins/local/sdd-autopilot/engine && npm install && npm run build

# Run a feature:
/sdd-auto:run health-check "Add a health check endpoint that returns server status"

How It Works

• Triage classifies your feature by complexity and picks the right execution mode (Express, Light, Standard, or Full)
• Spec/Plan/Tasks phases generate structured artifacts that each subsequent phase builds on
• Implementation runs one agent per task, with automatic parallelization for independent tasks
• Verification + Review catch issues before the PR — with automated fix loops if something fails
• Adaptive learning records metrics from every run and uses them to improve future runs

For the full architecture, see docs/architecture.md.

Commands

Configuration

The pipeline selects an execution mode automatically based on feature complexity:

Headless Mode

Headless mode lets an external orchestrator (CI script, autonomous agent, cron job) run the SDD pipeline without human interaction, PRs, or pushes.

How to activate: pass --headless flag and set the SDD_MODE=headless environment variable.

Output contract:

• Exit code 0 = success. A single atomic commit is ready on the worktree branch.
• Exit code 1 = failure (escalated or awaiting_input — cannot proceed without human).
• Last line of stdout: TLDR: {one sentence summary of what was done}

The external orchestrator is responsible for merge, push, and branch cleanup.

Example:

SDD_MODE=headless claude -p "/sdd-auto:run my-feature 'Add rate limiting to API'" --dangerously-skip-permissions

--headless is compatible with all execution modes (Express, Light, Standard, Full) and with --skip-worktree. It implies --skip-pr — do not pass both.

Requirements

• Claude Code CLI with plugin support
• Node.js 18+
• git and gh (GitHub CLI) for PR creation

Links

• Architecture — pipeline phases, state machine, file structure
• Tools Reference — all 38 MCP tools across 4 categories
• Observability & Metacognition — learning loop, analytics, anomaly detection
• Memory Intelligence — two-layer memory model
• Changelog

License

MIT