Radish 🌱

Autonomous coding loops with safety guardrails.

Radish wraps autonomous AI coding sessions (Claude Code, Cursor, Aider, etc.) with blast radius control, auto-checkpoints, kill switches, and structured audit trails. Ship features overnight without the 4am disasters.

Why Radish?

Autonomous coding agents are powerful but dangerous. They can:

• Delete critical files
• Overwrite production configs
• Spiral into infinite loops burning API credits
• Make changes that are hard to reverse

Radish provides the guardrails that let you sleep while agents work.

Features

• Blast Radius Control — Restrict which files/directories agents can modify
• Auto-Checkpoints — Git commits at configurable intervals
• Kill Switches — Automatic termination on violation or timeout
• Structured Audit Trails — Every action logged for review
• Cost Limits — Stop before you burn through your API budget
• Violation Detection — Real-time monitoring for forbidden operations

Quick Start

# Clone the repo
git clone https://github.com/longarcstudios/radish.git
cd radish

# Make executable
chmod +x radish.sh scripts/check_violations.sh

# Configure your session
cp radish.yaml.example radish.yaml
# Edit radish.yaml with your settings

# Run an autonomous session
./radish.sh "Implement user authentication feature"

Configuration

Edit radish.yaml to configure your session:

# radish.yaml
session:
  name: "feature-auth"
  timeout: 3600              # Max session duration (seconds)
  checkpoint_interval: 300   # Auto-commit every 5 minutes

guardrails:
  allowed_paths:
    - "src/**"
    - "tests/**"
  forbidden_paths:
    - ".env*"
    - "*.pem"
    - "secrets/**"
  forbidden_commands:
    - "rm -rf"
    - "DROP TABLE"
    - "format"

limits:
  max_files_changed: 50
  max_lines_changed: 2000
  max_cost_usd: 10.00

on_violation: "stop"         # stop | warn | checkpoint-and-continue

How It Works

1. Pre-flight checks — Validates config, creates initial checkpoint
2. Session wrapper — Launches your agent with monitored stdin/stdout
3. Real-time monitoring — Watches file changes and command execution
4. Violation detection — Compares actions against guardrails
5. Auto-checkpoints — Periodic git commits with session metadata
6. Post-session audit — Generates structured log of all changes

Commands

# Start a guarded session
./radish.sh "Your task description"

# Check for violations in current state
./scripts/check_violations.sh

# View session logs
ls -la logs/

# Rollback to last checkpoint
git reset --hard HEAD~1

Telemetry

Radish v0.1.1+ includes optional cloud telemetry. Every checkpoint sends anonymous session data to help improve Radish.

Data collected:

• Session ID (random, not linked to you)
• Task description
• Agent type
• Checkpoint results (pass/fail)
• Violation types (if any)

No personal data, code, or file contents are ever transmitted.

Opt-out

To disable telemetry, set the environment variable:

export RADISH_TELEMETRY=false

Or run with:

RADISH_TELEMETRY=false ./radish.sh "your task"

Integration

Radish works with any autonomous coding tool:

# With Claude Code
./radish.sh --agent="claude" "Build the API endpoints"

# With Cursor
./radish.sh --agent="cursor" "Refactor the database layer"

# With Aider
./radish.sh --agent="aider" "Add unit tests"

Audit Trail

Every session generates structured logs:

logs/
├── session-2024-01-15-143022/
│   ├── config.yaml          # Session configuration
│   ├── changes.json         # All file modifications
│   ├── commands.log         # Commands executed
│   ├── violations.json      # Any violations detected
│   └── summary.md           # Human-readable summary

Philosophy

Radish follows the Long Arc Studios principle: Permission is not proof.

Just because an agent can do something doesn't mean it should. Radish provides the accountability layer that makes autonomous coding sessions auditable and reversible.

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT License - see LICENSE

---

Built by Long Arc Studios — Infrastructure for the AI Era.