autocoder/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

This is an autonomous coding agent system with a React-based UI. It uses the Claude Agent SDK to build complete applications over multiple sessions using a two-agent pattern:

1. Initializer Agent - First session reads an app spec and creates features in a SQLite database
2. Coding Agent - Subsequent sessions implement features one by one, marking them as passing

Commands

Quick Start (Recommended)

# Windows - launches CLI menu
start.bat

# macOS/Linux
./start.sh

# Launch Web UI (serves pre-built React app)
start_ui.bat      # Windows
./start_ui.sh     # macOS/Linux

Python Backend (Manual)

# Create and activate virtual environment
python -m venv venv
venv\Scripts\activate  # Windows
source venv/bin/activate  # macOS/Linux

# Install dependencies
pip install -r requirements.txt

# Run the main CLI launcher
python start.py

# Run agent directly for a project (use absolute path or registered name)
python autonomous_agent_demo.py --project-dir C:/Projects/my-app
python autonomous_agent_demo.py --project-dir my-app  # if registered

# YOLO mode: rapid prototyping without browser testing
python autonomous_agent_demo.py --project-dir my-app --yolo

# Parallel mode: run multiple agents concurrently (1-5 agents)
python autonomous_agent_demo.py --project-dir my-app --parallel --max-concurrency 3

YOLO Mode (Rapid Prototyping)

YOLO mode skips all testing for faster feature iteration:

# CLI
python autonomous_agent_demo.py --project-dir my-app --yolo

# UI: Toggle the lightning bolt button before starting the agent

What's different in YOLO mode:

• No regression testing (skips feature_get_for_regression)
• No Playwright MCP server (browser automation disabled)
• Features marked passing after lint/type-check succeeds
• Faster iteration for prototyping

What's the same:

• Lint and type-check still run to verify code compiles
• Feature MCP server for tracking progress
• All other development tools available

When to use: Early prototyping when you want to quickly scaffold features without verification overhead. Switch back to standard mode for production-quality development.

React UI (in ui/ directory)

cd ui
npm install
npm run dev      # Development server (hot reload)
npm run build    # Production build (required for start_ui.bat)
npm run lint     # Run ESLint

Note: The start_ui.bat script serves the pre-built UI from ui/dist/. After making UI changes, run npm run build in the ui/ directory.

Architecture

Core Python Modules

• start.py - CLI launcher with project creation/selection menu
• autonomous_agent_demo.py - Entry point for running the agent
• agent.py - Agent session loop using Claude Agent SDK
• client.py - ClaudeSDKClient configuration with security hooks and MCP servers
• security.py - Bash command allowlist validation (ALLOWED_COMMANDS whitelist)
• prompts.py - Prompt template loading with project-specific fallback
• progress.py - Progress tracking, database queries, webhook notifications
• registry.py - Project registry for mapping names to paths (cross-platform)
• parallel_orchestrator.py - Concurrent agent execution with dependency-aware scheduling
• api/dependency_resolver.py - Cycle detection (Kahn's algorithm + DFS) and dependency validation

Project Registry

Projects can be stored in any directory. The registry maps project names to paths using SQLite:

• All platforms: ~/.autocoder/registry.db

The registry uses:

• SQLite database with SQLAlchemy ORM
• POSIX path format (forward slashes) for cross-platform compatibility
• SQLite's built-in transaction handling for concurrency safety

Server API (server/)

The FastAPI server provides REST endpoints for the UI:

• server/routers/projects.py - Project CRUD with registry integration
• server/routers/features.py - Feature management
• server/routers/agent.py - Agent control (start/stop/pause/resume)
• server/routers/filesystem.py - Filesystem browser API with security controls
• server/routers/spec_creation.py - WebSocket for interactive spec creation

Feature Management

Features are stored in SQLite (features.db) via SQLAlchemy. The agent interacts with features through an MCP server:

• mcp_server/feature_mcp.py - MCP server exposing feature management tools
• api/database.py - SQLAlchemy models (Feature table with priority, category, name, description, steps, passes, dependencies)

MCP tools available to the agent:

• feature_get_stats - Progress statistics
• feature_get_next - Get highest-priority pending feature (respects dependencies)
• feature_claim_next - Atomically claim next available feature (for parallel mode)
• feature_get_for_regression - Random passing features for regression testing
• feature_mark_passing - Mark feature complete
• feature_skip - Move feature to end of queue
• feature_create_bulk - Initialize all features (used by initializer)
• feature_add_dependency - Add dependency between features (with cycle detection)
• feature_remove_dependency - Remove a dependency

React UI (ui/)

• Tech stack: React 18, TypeScript, TanStack Query, Tailwind CSS v4, Radix UI, dagre (graph layout)
• src/App.tsx - Main app with project selection, kanban board, agent controls
• src/hooks/useWebSocket.ts - Real-time updates via WebSocket (progress, agent status, logs, agent updates)
• src/hooks/useProjects.ts - React Query hooks for API calls
• src/lib/api.ts - REST API client
• src/lib/types.ts - TypeScript type definitions

Key components:

• AgentMissionControl.tsx - Dashboard showing active agents with mascots (Spark, Fizz, Octo, Hoot, Buzz)
• DependencyGraph.tsx - Interactive node graph visualization with dagre layout
• CelebrationOverlay.tsx - Confetti animation on feature completion
• FolderBrowser.tsx - Server-side filesystem browser for project folder selection

Keyboard shortcuts (press ? for help):

• D - Toggle debug panel
• G - Toggle Kanban/Graph view
• N - Add new feature
• A - Toggle AI assistant
• , - Open settings

Project Structure for Generated Apps

Projects can be stored in any directory (registered in ~/.autocoder/registry.db). Each project contains:

• prompts/app_spec.txt - Application specification (XML format)
• prompts/initializer_prompt.md - First session prompt
• prompts/coding_prompt.md - Continuation session prompt
• features.db - SQLite database with feature test cases
• .agent.lock - Lock file to prevent multiple agent instances

Security Model

Defense-in-depth approach configured in client.py:

1. OS-level sandbox for bash commands
2. Filesystem restricted to project directory only
3. Bash commands validated against ALLOWED_COMMANDS in security.py

Claude Code Integration

• .claude/commands/create-spec.md - /create-spec slash command for interactive spec creation
• .claude/skills/frontend-design/SKILL.md - Skill for distinctive UI design
• .claude/templates/ - Prompt templates copied to new projects

Key Patterns

Prompt Loading Fallback Chain

1. Project-specific: {project_dir}/prompts/{name}.md
2. Base template: .claude/templates/{name}.template.md

Agent Session Flow

1. Check if features.db has features (determines initializer vs coding agent)
2. Create ClaudeSDKClient with security settings
3. Send prompt and stream response
4. Auto-continue with 3-second delay between sessions

Real-time UI Updates

The UI receives updates via WebSocket (/ws/projects/{project_name}):

• progress - Test pass counts (passing, in_progress, total)
• agent_status - Running/paused/stopped/crashed
• log - Agent output lines with optional featureId/agentIndex for attribution
• feature_update - Feature status changes
• agent_update - Multi-agent state updates (thinking/working/testing/success/error) with mascot names

Parallel Mode

When running with --parallel, the orchestrator:

1. Spawns multiple Claude agents as subprocesses (up to --max-concurrency)
2. Each agent claims features atomically via feature_claim_next
3. Features blocked by unmet dependencies are skipped
4. Browser contexts are isolated per agent using --isolated flag
5. AgentTracker parses output and emits agent_update messages for UI

Process Limits (Parallel Mode)

The orchestrator enforces strict bounds on concurrent processes:

• MAX_PARALLEL_AGENTS = 5 - Maximum concurrent coding agents
• MAX_TOTAL_AGENTS = 10 - Hard limit on total agents (coding + testing)
• Testing agents are capped at max_concurrency (same as coding agents)

Expected process count during normal operation:

• 1 orchestrator process
• Up to 5 coding agents
• Up to 5 testing agents
• Total: never exceeds 11 Python processes

Stress Test Verification:

# Windows - verify process bounds
# 1. Note baseline count
tasklist | findstr python | find /c /v ""

# 2. Start parallel agent (max concurrency)
python autonomous_agent_demo.py --project-dir test --parallel --max-concurrency 5

# 3. During run - should NEVER exceed baseline + 11
tasklist | findstr python | find /c /v ""

# 4. After stop via UI - should return to baseline
tasklist | findstr python | find /c /v ""

# macOS/Linux - verify process bounds
# 1. Note baseline count
pgrep -c python

# 2. Start parallel agent
python autonomous_agent_demo.py --project-dir test --parallel --max-concurrency 5

# 3. During run - should NEVER exceed baseline + 11
pgrep -c python

# 4. After stop - should return to baseline
pgrep -c python

Log Verification:

# Check spawn vs completion balance
grep "Started testing agent" orchestrator_debug.log | wc -l
grep "Testing agent.*completed\|failed" orchestrator_debug.log | wc -l

# Watch for cap enforcement messages
grep "at max testing agents\|At max total agents" orchestrator_debug.log

Design System

The UI uses a neobrutalism design with Tailwind CSS v4:

• CSS variables defined in ui/src/styles/globals.css via @theme directive
• Custom animations: animate-slide-in, animate-pulse-neo, animate-shimmer
• Color tokens: --color-neo-pending (yellow), --color-neo-progress (cyan), --color-neo-done (green)