Autonomous Coding Toolkit

An autonomous AI coding agent that executes implementation plans with quality gates, fresh context per batch, and 79 community-contributed lessons that prevent the same bug twice. Built for Claude Code — works as a Claude Code plugin (interactive) or npm CLI for headless CI/CD execution.

Goal: Code better than a human on large projects — not by being smarter on any single batch, but by compounding learning across thousands of batches across hundreds of users.

Install

npm (recommended)

npm install -g autonomous-coding-toolkit

This puts act on your PATH.

Claude Code Plugin

# Add the marketplace source
/plugin marketplace add parthalon025/autonomous-coding-toolkit

# Install the plugin
/plugin install autonomous-coding-toolkit@autonomous-coding-toolkit

From Source

git clone https://github.com/parthalon025/autonomous-coding-toolkit.git
cd autonomous-coding-toolkit
npm link  # puts 'act' on PATH

Platform Notes

| Platform | Status | Notes | |----------|--------|-------| | Linux | Works out of the box | bash 4+, jq, git required | | macOS | Works with Homebrew bash | macOS ships bash 3.2 — install bash 4+ via brew install bash. Also install coreutils for GNU readlink: brew install coreutils | | Windows | WSL only | Run wsl --install, then use the toolkit inside WSL. Native Windows is not supported |

macOS ships bash 3.2 (2007) due to licensing. The toolkit requires bash 4+ for associative arrays and other features.

# Install modern bash and GNU coreutils
brew install bash coreutils jq

# Verify
bash --version  # Should show 5.x

Homebrew bash installs to /opt/homebrew/bin/bash (Apple Silicon) or /usr/local/bin/bash (Intel). The act CLI invokes scripts via bash — as long as Homebrew's bin is on your PATH (which brew sets up automatically), scripts will use the correct version.

Quick Start

# Bootstrap your project
act init --quickstart

# Full pipeline — brainstorm → plan → execute → verify → finish
/autocode "Add user authentication with JWT"

# Run a plan headless (fully autonomous, fresh context per batch)
act plan docs/plans/my-feature.md --on-failure retry --notify

# Quality check
act gate --project-root .

# See all commands
act help

See examples/quickstart-plan.md for a minimal plan you can run in 3 commands.

The Pipeline

Idea → [Roadmap] → Brainstorm → [Research] → PRD → Plan → Execute → Verify → Finish

Each stage exists because a specific failure mode demanded it:

| Stage | Problem It Solves | Evidence | |-------|------------------|----------| | Brainstorm | Agents build the wrong thing correctly | SWE-bench Pro: removing specs = 3x degradation | | Research | Building on assumptions wastes hours | Stage-Gate: stable definitions = 3x success rate | | Plan | Plan quality dominates execution quality ~3:1 | SWE-bench Pro: spec removal = 3x degradation | | Execute | Context degradation is the #1 quality killer | 11/12 models < 50% at 32K tokens | | Verify | Static review misses behavioral bugs | Property-based testing finds ~50x more mutations |

Full evidence with 25+ papers across 16 research reports: docs/RESEARCH.md

How It Compares

| Tool | Approach | This Toolkit's Difference | |------|----------|--------------------------| | Claude Code /plan | Built-in planning | No quality gates, no fresh context per batch, no lesson system | | Aider | Interactive pair programming | Aider is conversational; this is batch-autonomous with gates | | Cursor Agent | IDE-integrated agent | No headless mode, no batch isolation | | SWE-agent | Autonomous GitHub issue solver | Single-issue scope; this handles multi-batch plans with state |

Core differentiators: (1) fresh context per batch, (2) machine-verifiable quality gates, (3) compounding lesson system, (4) headless unattended execution.

Quality Gates

Mandatory between every batch:

1. Lesson check (<2s, grep-based anti-pattern scan)
2. ast-grep patterns (5 structural checks)
3. Test suite (auto-detected: pytest / npm test / make test)
4. Memory check (warns if < 4GB available)
5. Test count regression (tests only go up)
6. Git clean (all changes committed)

Community Lessons

79 lessons across 6 failure clusters, learned from production bugs. Adding a lesson file to docs/lessons/ automatically adds a check — no code changes needed.

Submit new lessons via /submit-lesson or open an issue.

Requirements

• Claude Code v1.0.33+ (claude CLI)
• Node.js 18+ (for the act CLI router)
• bash 4+, jq, git
• Optional: gh (PR creation), curl (Telegram notifications), ast-grep (structural checks)

Learn More

| Topic | Doc | |-------|-----| | Architecture and internals | docs/ARCHITECTURE.md | | Research (25+ papers, 16 reports) | docs/RESEARCH.md | | Contributing lessons | docs/CONTRIBUTING.md | | Plan file format | examples/example-plan.md | | Execution modes (5 options) | docs/ARCHITECTURE.md#system-overview |

Attribution

Core skill chain forked from superpowers by Jesse Vincent / Anthropic. Extended with quality gate pipeline, headless execution, lesson system, MAB routing, and research/roadmap stages.

Research Sources

The toolkit's design is grounded in peer-reviewed research. Key papers:

• SWE-bench Pro (Xia et al., 2025) — 1,865 programming problems; removing specifications degraded agent success from 25.9% to 8.4%
• Context Rot (Hong et al., Chroma 2025) — 11 of 12 models scored below 50% of short-context performance at 32K tokens
• Lost in the Middle (Liu et al., Stanford TACL 2024) — Information placed mid-context suffers up to 20 percentage point accuracy loss
• Agentic Property-Based Testing (OOPSLA 2025) — Property-based testing finds ~50x more mutations per test than traditional unit tests
• Bugs in LLM-Generated Code (Tambon et al., 2024) — Empirical taxonomy of AI code generation failures
• Cooper Stage-Gate — Projects with stable, upfront definitions are 3x more likely to succeed

16 research reports synthesizing 25+ papers: docs/RESEARCH.md

License

MIT