Change Firewall v0.3.1•The AI Agent Trust Layer

Your AI agent wrote the code.We are the referee before it merges.

Autonomous agents rewrite 20 files in 5 seconds and self-report success. Change Firewall provides deterministic, compiler-grounded verification of what the agent actually touched vs. what it claimed, maps downstream blast radius, and proves runtime crashes with zero guesswork.

Add to Cursor / Claude MCP Documentation

Neutral Non-LLM Referee

Compiler AST Grounded (Zero Hallucination)

100% Offline • Claude, Cursor & Devin Ready

AST Behavioral Diff & Blast Radius Simulator

Interactive Radar

Scenario: AI refactored return in user.ts (return user ➔ return { user })

Blast Radius: 6 Broken Downstream Callers

Tap any node to inspect broken callsite & runtime exceptionAST Graph View

src/controllers/user.tsLine 42

-return user;

+return { user };

AI wrapped the return payload in a nested object without updating downstream contract deserializers.

High Risk Mutation Detected

Risk Score: 74 / 100

💥6 Downstream Callers Severed

AST analysis discovered that 6 consumer files depend on the direct properties of user. Without Change Firewall, this innocent 1-line change silently breaks web, auth, and billing.

Automated CI Gate:MERGE BLOCKED (Exit 1)

v0.3.1 THE AGENT TRUST ARCHITECTURE

The 4 Pillars of the AI Code Referee

Autonomous coding agents rewrite 20 files in seconds and grade their own homework. Change Firewall replaces blind trust and probabilistic LLM reviews with deterministic, compiler-grounded verification.

Trust Layer Architecture

THE DETERMINISTIC AI AGENT REFEREE (v0.3.1)

│

┌─────────────────────────┬──────────────────┴──────────────┬─────────────────────────┐

THE AGENT REFEREE CORE (v0.3.1)

Pillar 1: The Brain

The Behavior Graph & 11-D Fingerprint Matrix

Elevates analysis beyond file-level imports into semantic behavioral boundaries.

11 Orthogonal VectorsAST Semantic Graph

Living Semantic Boundaries Evaluated

API Boundaries

REST, Fastify/Express routes, Next.js server actions, tRPC routers

Database Boundaries

Prisma schemas, Drizzle models, TypeORM entities, raw SQL migrations

Event Boundaries

Kafka/RabbitMQ pub-sub, Redis streams, EventEmitter hooks

Auth & Security

JWT validators, session guards, RBAC middlewares, token verifiers

Test Boundaries

Direct test suites and indirect downstream integration specs

The 11-Dimensional Fingerprint Matrix (Click to Inspect)

Vector 1 of 11

DIMENSION: API CONTRACTDeterministic AST Inspection

REST endpoints, Fastify/Express routes, Next.js server actions, and tRPC signatures.

Trigger Condition

Changed return type from User to { user: User } or removed required parameters.

Downstream Impact

Immediate 400 Bad Request or deserialization TypeError in all client consumers.

COMPLETE COMMAND GUIDE

Learn the CLI through real runs

Each example fills editable text in the playground terminal. Press Enter to run the installed CLI and read its actual output.

Analyze

Analyze working changes

Compare saved working files with HEAD.

change-firewall

Analyze

Analyze help

Ask the installed CLI for its current flags.

change-firewall analyze --help

Analyze

Machine-readable report

Print the real JSON result.

change-firewall analyze --json

Analyze

Analyze staged changes

Limit collection to the Git index.

change-firewall analyze --staged

Analyze

Compare an earlier base

Useful after the current edits are committed.

change-firewall analyze --base playground-baseline

The installed command’s --help remains the source of truth. The guide never supplies replacement terminal output.

REAL COMMANDS · REAL GIT STATE

Edit the sample. Run the actual CLI. See what changes.

The lightweight preview loads immediately. Start a live disposable runtime when you are ready; if it is unavailable, the terminal says so and never invents results.

Loading the interactive workspace…

The Trust Layer Primitives

Why AI Agents Need a Non-LLM Referee

When autonomous agents rewrite entire codebases in seconds, probabilistic LLM reviews fail. You need deterministic, compiler-level enforcement before anything merges.

Zero Guesswork

Symbolic Runtime Crash Proofs

Traces backward-slice data flows from widened return types to consumer call sites. Mathematically proves unhandled TypeErrors and null dereferences at exact line numbers with zero false alarms.

Pillar 3• Active Protection

Live Sandbox

Interactive Crash Simulator Sandbox

Simulates real Node.js runtime exceptions in the dashboard with active stack traces, visual call ladder progression, step-by-step symbolic proofs, and 1-click preventative fixes.

Pillar 3• Active Protection

1-Click Auto-Fix

AI Remediation Command Center

Converts detected behavioral shifts and arity breaks into surgical, zero-hallucination prompts tailored for Cursor, Claude Code, Antigravity, and Copilot with Safe Fix, Minimal, and PR modes.

Pillar 4• Active Protection

11-D Multi-Vector

11-Dimensional Fingerprint Matrix

Evaluates every commit across 11 orthogonal vectors: API contracts, auth guards, nullability widening, database queries, event hooks, state stores, and error semantics for complete coverage.

Pillar 1• Active Protection

Temporal Memory

Persistent Firewall Memory Engine

Zero-config local store in .firewall/memory/ that tracks contract invariant history over time. Flags unexpected mutations on functions that have been historically stable for 100+ commits.

Pillar 2• Active Protection

Agent Safety

"AI Agent Intent vs Reality" Guard

Performs bidirectional semantic grounding on stated prompts vs AST diffs. Accurately verifies feature scope, detects missing parameter defaults (= null), and rejects unrelated claims.

Pillar 4• Active Protection

Graph Intelligence

Living Behavior Graph Boundaries

Models the system beyond file imports into semantic API, Database, Event, Auth, and Test boundaries. Traces multi-hop critical paths and downstream consumer blast radius up to 3 hops deep.

Pillar 1• Active Protection

CI/CD Gate

One-Command CI/CD Merge Gate

Enforces deterministic preflight verification in GitHub Actions and Husky pre-commit hooks. Strictly exits with Code 0 (Safe to merge) or Code 1 (Blocked review required).

Merge Gate• Active Protection

Open Standard Protocol (MCP)

Give Your Agent a Neutral Referee

Configure once. Claude Code, Cursor, Google Antigravity, and Windsurf query Change Firewall natively via MCP before presenting code to you — catching breaking mutations autonomously in their own feedback loop.

Claude Desktop Configuration

~/Library/Application Support/Claude/claude_desktop_config.json

mcp_config.json

2 "mcpServers": {

3 "change-firewall": {

4 "command": "npx",

5 "args": ["-y", "change-firewall", "mcp"]

6 }

7 }

How it works: Claude Desktop natively spawns Change Firewall as a local process over stdio. A hammer icon appears in chat with all 4 tools ready to execute.

Simulated Claude Desktop ExperienceMCP Active

Claude 3.7 Sonnet

18ms

User Input

“Refactor user session authentication and check if anything broke.”

Autonomous Tool Invocation

change-firewall.analyze_changes

Deterministic Feedback

AST diff complete: 0 contract breaks. Risk Score: 12/100 (Safe).

Checked 14 export signatures across 3 files
2 downstream consumers auto-updated
All parameter nullability contracts verified

Safety Gate Result:READY TO MERGE

Exposed MCP Tools:Click for details

Developer Updates & AST Superpowers

Never miss a breakthrough update.

Get notified when new superpowers launch: Automated AI Self-Healing Patches (--fix), Python/Go AST engines, and GitHub PR review bots.

Interactive Predictive Arc Field

Compiler-Grounded Invariant Proofs