Rocket.Chat: Code Analyzer Agentic Inference Context Reduction Mechanics

Analyzing massive codebases like Rocket.Chat via iterative LLM agents causes context bloat, rapidly exceeding free-tier limits. This project builds a graph-native code navigator exposed as a Gemini MCP server. An offline indexer parses the full Rocket.Chat source via ts-morph, extracts a typed dependency graph (11 edge kinds, including Rocket.Chat-specific patterns like callbacks.run/add and Meteor method dispatch that break standard import analysis), and compresses function bodies into signature-only skeletons for ~4× token reduction. At query time, three composable MCP tools—search, graph, implement—let an LLM navigate by following real edges rather than ranking documents. An evaluation framework with ground-truth test cases (collected by Rocket.Chat experts and Claude) and 5 measured metrics closes the loop.