Rocket.Chat Code Analyzer (Gemini CLI Extension)

This project is an MCP-based Gemini CLI extension focused on context reduction for code analysis in large repositories.

Goal: help agentic workflows stay within strict token budgets by reducing unnecessary code context before inference.

This extension exposes the following tools:

• analyze_directory
  • Builds a lightweight AST-based structure map of .js and .ts files.
• search_codebase
  • Performs structural keyword matching against functions, classes, and limited JSDoc/file-name signals.
• get_dependencies
  • Extracts shallow import/require dependencies for a target file.
• view_file_skeleton
  • Returns a compressed code skeleton (signatures/types/classes, stripped implementations).
• get_pruned_context
  • End-to-end pruning pipeline: search -> choose anchor file -> compress -> include dependencies until tokenBudget is reached.

The semantic retrieval is currently basic and must be enhanced for production-scale conceptual queries.

• search_codebase is mostly identifier-driven, not full semantic intent retrieval.
• Natural language queries like "message sending flow" may return empty matches if similar identifiers are absent.
• This is a known gap and the next planned improvement area.

• server.js: MCP server entry point and tool registration.
• src/indexer.js: directory indexing (AST structure extraction).
• src/searcher.js: structural keyword search.
• src/dependency.js: dependency extraction.
• src/compressor.js: skeleton/compression logic.
• src/pipeline.js: budgeted pruned-context assembly pipeline.
• gemini-extension.json: Gemini extension config.
• GEMINI.md: operational prompt protocol for using tools efficiently.

• Node.js 18+ (recommended)
• npm
• Gemini CLI with extension support

Clone/copy this folder to your machine, then open a terminal inside:

cd <path-to>/code-analyzer

npm install

Use your Gemini CLI local extension workflow and point it to this folder (the folder that contains gemini-extension.json).

The extension starts server.js over stdio using:

• command: node
• args: ${extensionPath}/server.js

Ask Gemini to run one of the tools, for example:

• "Use analyze_directory on <absolute-project-path> and summarize the architecture."
• "Use get_pruned_context with tokenBudget: 6000 for query buildPrunedContext."

You should see a tool execution line similar to:

• get_pruned_context (code-analyzer MCP Server)

Use get_pruned_context on:
targetDirectory: D:\path\to\repo
query: "authentication flow"
tokenBudget: 6000

Then explain the critical modules and possible bug risks.

• This is a working MVP extension for context reduction in code analysis.
• It is intentionally focused on tooling mechanics, not full semantic retrieval quality yet.
• Semantic ranking and conceptual search quality are the primary next improvement track before large-monorepo validation.