Post by Nurdzhan K.

AI coding agents work blind. Here is how we give them a deterministic flight radar. 🛰️🦅 When you let an active agent (like Claude Code or Antigravity) loose on a complex repository, it faces a massive constraint: it lacks structural spatial memory. It reads files and directories flatly, burning tokens, and modifies code without mathematically measuring downstream regressions across the system. If you let an LLM guess your architecture or calculate dependency drift, you get unpredictable, non-deterministic failures. To solve this, I built BlastRadius — an open-source, local-first telemetry and dependency graph tool that acts as a real-time command center for agentic workflows. With the release of version rc9.38, the platform now handles advanced architectural governance: Topological Blast-Radius Mapping: Built-in zero-dependency AST resolvers (JS/TS, Python, Go, Rust, Java) map the exact import graph of your workspace. Edits turn red, reads turn green, and affected consumers turn yellow in real-time. Transitive Risk Inspection (New in rc9.38): The system automatically detects "cold hubs"—critical core files that haven't been edited but possess a massive Fan-In ratio. If an agent's change transitively reaches one of these boundaries, BlastRadius flags it immediately. Native MCP Governance: It injects this telemetry directly into the agent's loop via the Model Context Protocol (check_architecture_compliance). The agent reads the generated structural risk matrix before closing an iteration, adapting its code to your architecture constraints automatically. 100% Offline & Sovereign: No remote cloud telemetry, no data leakage, and a rigorous test suite of 1023 tests in Vitest ensuring local execution stability. The dashboard tracks live agent iterations, isolates active dependency trees, and computes multi-hop impact analysis on localhost. 💬 Engineering Meta-Rule for this Thread: This post is intended strictly for peer-to-peer architectural debate, system telemetry insights, and open-source infrastructure governance. To maintain a high-signal environment for engineers, no self-promotion or external product links are allowed in the comments. Unrelated marketing links or automated spam will be actively removed to keep the focus purely on technical execution. Let's talk systems.