Overview & ArchitectureThe Trust Problem100% Deterministic2 min read
Why AI Agents Need a Deterministic Referee
Why line-based git diffs and LLM-on-LLM reviews fail when autonomous coding agents refactor code.
Executive Overview
Autonomous AI coding agents (Claude Code, Cursor Agent, Antigravity, Devin, Copilot) write code in bursts across dozens of files and self-report success. A standard git diff only sees line additions +1/-1, while using an LLM to review an LLM is a probabilistic hallucination loop. Change Firewall acts as the neutral, compiler-grounded referee between agent claims and your main repository.
⚠️ Critical Behavioral Warning
A 1-line return type change (e.g., returning { user } instead of user) looks innocent in git diff, but breaks every single client component at runtime.
Reference Matrix & Behavior Specs
| Analysis Vector | Traditional Git Diff | Change Firewall AST Engine |
|---|---|---|
| Contract Mutations | Hidden as plain text change | Analyzed against AST function contracts |
| Downstream Blast Radius | 0% caller awareness | Reverse import graph traversal up to 3 hops |
| Merge Readiness | Requires human deciphering | Deterministic 0-100 score + blocking gate |
| AI Self-Healing | Manual review cycles | Stdio JSON-RPC MCP server for autonomous fix |
Key Capabilities & Specifications
🛡️
🔒 100% Offline & Local (zero token costs, zero telemetry)
⚡
Instant TypeScript Compiler API AST traversal in milliseconds
✦
Full reverse dependency graph mapping across your entire repository