A transparent discovery signal based on current public GitHub metadata.
Recent activity35% weight
0
Community adoption25% weight
10
Maintenance state20% weight
0
License clarity10% weight
100
Project information10% weight
100
This score does not audit code, security, maintainers, documentation quality, or suitability. Verify the repository and its current documentation before adoption.
README preview
Autonomous Coding Toolkit
Requires Claude Code. Built specifically for the Claude Code + Claude API ecosystem.
A collection of scripts, skills, and infrastructure for agent-driven software development. It implements a structured pipeline — from brainstorming through verified delivery — with quality gates between every batch, fresh AI context per execution unit, and a compounding lesson system that turns production bugs into automated checks.
Works as a Claude Code plugin (interactive sessions) or as standalone Bash scripts for headless, unattended CI/CD execution. The standalone scripts (scripts/) have no Claude Code dependency and integrate directly with any CI system that respects exit codes.
Who This Is For
Claude Code users who want structured, quality-gated autonomous coding workflows — the full Code Factory pipeline (brainstorm → PRD → plan → execute → verify) without building the scaffolding from scratch
Teams using AI-assisted development who find agents going off-track or producing inconsistent output — quality gates, fresh-context execution, and machine-verifiable acceptance criteria address the root causes
Developers who want CI/CD integration without Claude Code — the standalone Bash scripts (quality-gate.sh, lesson-check.sh, run-plan.sh) work independently and exit 0/1 for any CI system
Core insight: Plan quality dominates execution quality at roughly a 3:1 ratio. The pipeline enforces rigor at the stages where agent failures actually originate — not just at code review.
The Code Factory Pipeline
Brainstorm → Research → PRD → Plan → Worktree → Execute → Verify → AAR → Finish
Each stage exists because a specific failure mode demanded it:
Stage
What It Does
Problem It Prevents
Brainstorm
Explore intent, surface edge cases, get approval before code
# Bootstrap a project
act init --quickstart
# Full pipeline — interactive, with brainstorming gate
/autocode "Add paginated list endpoint with cursor-based navigation"
# Execute a plan headless — fresh context per batch, quality gates throughout
act plan docs/plans/my-feature.md --on-failure retry --notify
# Run the quality gate standalone
act gate --project-root .
# See all commands
act help
All scripts live in scripts/. They can be invoked directly (standalone) or through the act CLI.
run-plan.sh — Headless batch executor
Parses a markdown plan file and executes each batch via claude -p, with a quality gate between batches and persistent state across context resets.
# Execute a plan from the beginning
scripts/run-plan.sh docs/plans/2026-02-20-feature.md
# Resume after an interruption
scripts/run-plan.sh --resume --worktree /path/to/worktree
# Retry failures, start from batch 3, send Telegram notifications
scripts/run-plan.sh docs/plans/feature.md --start-batch 3 --on-failure retry --notify
# Multi-Armed Bandit mode: two competing strategies, LLM judge picks winner
scripts/run-plan.sh docs/plans/feature.md --mab
# Parallel patch sampling: generate N candidates per batch, score, take winner
scripts/run-plan.sh docs/plans/feature.md --sample 3
Key options:
Flag
Description
--mode headless|team|competitive
Execution strategy (default: headless)
--on-failure stop|skip|retry
Batch failure handling
--max-retries N
Retry limit per batch (default: 2)
--start-batch N / --end-batch N
Execute a range of batches
--resume
Continue from saved .run-plan-state.json
--mab
Thompson Sampling routing between competing agent strategies
--sample N
Spawn N candidate patches per batch, score and pick winner
--max-budget <dollars>
Abort if cumulative cost exceeds limit
--verify
Run verification pass after all batches complete
--notify
Send Telegram notifications on completion/failure
State survives interruption via .run-plan-state.json. Execution context is assembled per-batch into a 6,000-character budget injected into CLAUDE.md before each Claude invocation.
quality-gate.sh — Composite quality gate
Runs all checks in sequence, fails fast on the first failure. Designed to run between every batch in the Ralph loop or as a standalone pre-commit gate.
# Full gate (lesson check + lint + tests + memory warning)
scripts/quality-gate.sh --project-root .
# Fast inner-loop mode (skip lint and license audit)
scripts/quality-gate.sh --project-root . --quick
# Include dependency license audit
scripts/quality-gate.sh --project-root . --with-license
Checks in order:
Step
What It Does
Fails Gate?
Toolkit validation
Runs validate-all.sh if present (toolkit self-check)
Yes
Lesson check
Scans changed files for known anti-patterns
Yes
Lint
ruff (Python) or eslint (Node), if configured
Yes
Structural analysis
ast-grep patterns for bare-except, empty-catch, async-no-await, retry-no-backoff, hardcoded-localhost
Advisory
Module size
Flags files over 300 lines
Advisory
Test suite
Auto-detects pytest / npm test / make test / bash test runner
Yes
License audit
Flags GPL/AGPL dependencies (--with-license only)
Yes
Memory check
Warns if available RAM < 4 GB
Warning only
Exit 0 if all blocking checks pass. Telemetry is recorded automatically when called from run-plan.sh context.
lesson-check.sh — Anti-pattern detector
Scans files for syntactic patterns extracted from the lessons database. Dynamic: adding a lesson automatically adds its check, no code changes needed.
# Check git-changed files in current directory (default)
scripts/lesson-check.sh
# Check specific files
scripts/lesson-check.sh src/api.py src/db.py
# Show detected project scope (scope-aware filtering)
scripts/lesson-check.sh --show-scope
# Override scope manually
scripts/lesson-check.sh --scope "language:python,domain:myproject" src/
# Bypass scope filtering — run all lessons
scripts/lesson-check.sh --all-scopes
Output format: file:line: [lesson-N] title — same as compiler errors, composable with other tools.
When lessons-db is installed (recommended), checks are sourced from the canonical SQLite database. Falls back to reading detection patterns directly from docs/lessons/*.md if the database is unavailable.
Fully automated end-to-end pipeline: reads a report, uses an LLM to identify the top priority, generates a PRD, creates a branch, runs the Ralph loop to completion, and opens a PR.
# Analyze latest report and execute the full pipeline
scripts/auto-compound.sh /path/to/project
# Point at a specific report
scripts/auto-compound.sh /path/to/project --report reports/2026-03-08-audit.md
# Preview what would happen without executing
scripts/auto-compound.sh /path/to/project --dry-run
# Limit iterations in case of runaway loop
scripts/auto-compound.sh /path/to/project --max-iterations 15
Pipeline stages: analyze-report.sh → /create-prd → git checkout -b compound/<slug> → Ralph loop with quality gates → gh pr create.
Reads any markdown report (test failures, error logs, user feedback, metrics) and uses a local LLM to identify the single highest-impact fix. Outputs structured JSON consumed by auto-compound.sh.
# Analyze a report, output analysis.json
scripts/analyze-report.sh reports/weekly-audit.md
# Use a specific
