Badger

AI agents that won't break your code.

How It Works • Quick Start • Architecture • Safety • Status

---

Badger is a safety-first agentic coding system. It wraps AI coding agents in a protective layer of checkpoints, verification, and automatic rollback — so your codebase stays intact even when an agent makes a mistake.

Every file modification is checkpointed before it happens, verified after it lands, and auto-reverted if anything breaks. No exceptions. No "the change is too small."

How It Works

You: $badger "Add user authentication with JWT"

Badger:
  1. GUARDRAIL CHECK  → Is the target file protected?
  2. BLAST RADIUS     → How many files depend on it?
  3. CHECKPOINT       → Snapshot the repo state
  4. EXECUTE          → Make the change
  5. VERIFY           → Build · Test · Lint · Typecheck
  6. RESULT           → Pass? Commit. Fail? Rollback.

If verification fails, Badger never tries to "fix forward." It rolls back to the last known-good state, logs what went wrong, and tells you exactly what happened.

Quick Start

# Clone
git clone https://github.com/AniketPatel148/badger.git
cd badger

# Install dependencies
bun install

# Run
bun run dev

Commands

Command	Description
$badger "<task>"	Start a safety-first agentic coding task
$badger-status	View checkpoints, rollback history, and safety state
$badger-rollback	Manually rollback to any checkpoint

Architecture

Badger uses a 3-tier prompt architecture with strict role separation:

┌─────────────────────────────────────────────┐
│              Tier 1: System Rules            │
│   Constitution · Safety Protocol · Comms     │
├─────────────────────────────────────────────┤
│              Tier 2: Agent Roles             │
│   Manager · Architect · Executor · Reviewer  │
├─────────────────────────────────────────────┤
│              Tier 3: Skills                  │
│   $badger · $badger-status · $badger-rollback│
└─────────────────────────────────────────────┘

Agents

Agent	Role	Boundary
Manager	Orchestrates tasks, communicates with user	Cannot write code
Architect	Designs systems, plans file structures	Cannot write code
Executor	Implements features, writes code	Cannot skip checkpoints or change architecture
Reviewer	Reviews code quality and security	Cannot write code

Agents communicate only through the Manager — no direct agent-to-agent calls. This ensures the safety protocol is enforced at every handoff.

Safety Layer

🛡️ Checkpoints

A git checkpoint is created before every file modification. These form a LIFO stack that can be unwound at any point.

✅ Verification

After every change, Badger runs: Build → Test → Lint → Typecheck. Commands are auto-detected from your project config.

⏪ Automatic Rollback

If any verification check fails: revert → log → notify. Always.

📊 Blast Radius Analysis

Before touching a file, Badger scans its dependency graph and assigns a risk level:

Risk	Dependents	Action
Low	0–1	Proceed
Medium	2–5	Proceed with caution
High	6–15	Warn and proceed carefully
Critical	16+ or config/entry	Stop and ask user

🔒 Sensitive File Guardrails

.env, credentials, certificates, API keys — auto-detected and protected by default. Agents cannot read, modify, or reference them.

Status

🚧 Badger is under active development. Core safety primitives are being built out. Not yet ready for production use.

Built with

• TypeScript + Bun
• Targets: Claude Code and OpenAI Codex

License

MIT

---

_{Built by Aniket Patel}