Agentic Code Analyzer

GSoC 2026 — graph-native code navigation for Rocket.Chat's 4M+ line monorepo.

GSoC Proposal

Problem

LLM agents analyzing large codebases accumulate context query after query, quickly exhausting free-tier token budgets. The root cause: code exploration is treated as a retrieval problem (rank documents by similarity), but it is actually a navigation problem (follow dependency edges from an entry point).

Rocket.Chat compounds this with five patterns invisible to standard import analysis:

Pattern	Why standard analysis fails
Meteor string-keyed method dispatch	sdk.call('sendMessage') target is a string literal
Event-driven callbacks	callbacks.run/add('afterSaveMessage') — no import between emit and handler
Symbol name collisions	sendMessage has 6+ definitions across client / server / packages
Blaze-to-React migration gaps	.html template names don't appear in TypeScript imports
Hook-based EE extensions	EE modules extend core via callbacks.add, not subclassing

Solution

An offline indexer builds a typed dependency graph (11 edge kinds). Three MCP tools expose it to any MCP-compatible client. A Constitution (AGENTS.md) encodes architecture knowledge as navigation rules. An evaluator closes the loop.

Source (.ts/.tsx)
  → hasher.ts      incremental MD5, skip unchanged
  → skeleton.ts    AST parse: signatures + 11 typed edges
  → embedder.ts    Gemini API: symbol → float32[768]
  → GLOBAL_INDEX   symbols · callGraph · fileDependents · embeddings
        ↓
  AGENTS.md (navigation rules) + MCP tools → LLM
        ↓
  Evaluator (5 metrics) → eval report → targeted fix → repeat

Setup

# Clone both repos side by side
git clone https://github.com/RocketChat/Agentic.Code.Analyzer.git
git clone https://github.com/RocketChat/Rocket.Chat.git

cd Agentic.Code.Analyzer
npm install
export GEMINI_API_KEY=your_key
npm start

The analyzer expects Rocket.Chat as a sibling directory by default. To use a different path:

export ROCKET_CHAT_SRC=/path/to/Rocket.Chat
npm start

Usage with Antigravity CLI (agy)

Install agy:

curl -fsSL https://antigravity.google/cli/install.sh | bash

The project includes .agents/mcp_config.json — agy picks it up automatically:

cd Agentic.Code.Analyzer
agy

Then ask questions directly:

> How does message sending work end-to-end?
> Where is the rate limiter configured?
> What breaks if I change sendMessage?

Quick non-interactive test:

agy -p "Use the search tool to find sendMessage in the server layer"

Other MCP Clients (Claude Desktop, Cursor, etc.)

This is a standard MCP server. Add to your client's MCP config:

{
  "mcpServers": {
    "rocket-ast-analyzer": {
      "command": "npx",
      "args": ["tsx", "/path/to/Agentic.Code.Analyzer/src/server/index.ts"]
    }
  }
}

MCP Tools

Tool	Description
search(query, layer?, question?)	Fuzzy symbol search reranked by embedding similarity (0.4 × fuzzy + 0.6 × cosine). Supports client/server layer filter.
graph(symbol, direction, depth?, edgeTypes?, question?)	BFS downstream or upstream. When question is provided, applies semantic pruning — edges with cosine similarity < 0.1 are dropped.
implement(symbol, filename)	Full source + up to 5 callee skeletons. Capped at 3 calls per question.

Question types → tool strategy

Type	Example	Strategy
Architecture	"How does message sending work end-to-end?"	search(entry) → graph(down)
Locate	"Where is the rate limiter configured?"	search(keyword) → implement
Pattern	"How do I register a new REST endpoint?"	search existing instance → implement
Routing	"How does a DDP method call reach its handler?"	search(dispatcher) → graph(down, edgeTypes=[...])
Impact	"What breaks if I change sendMessage?"	search(target) → graph(up) → implement top callers

Evaluator metrics

Metric	Threshold	What it catches
File hit rate	≥ 95%	Wrong files retrieved
Symbol coverage	100%	Key symbol missing from answer
Retrieval order	≥ 80%	Entry point found too late
Tool call count	≤ 10	Agent taking too many steps
Implement share	≤ 30%	Over-relying on full source reads

Project structure

src/
  server/           MCP server layer
    index.ts          entry point — builds index, starts MCP server
    registry.ts       tool definitions + handlers
    retriever.ts      search / getContext / getImplementation
  indexer/          offline indexer
    index.ts          scan, prewarm, build GLOBAL_INDEX
    skeleton.ts       AST parse → signatures + 11 edge types
    hasher.ts         incremental MD5 change detection
    embedder.ts       Gemini embedding API
    state.ts          GLOBAL_INDEX type definitions
    local-db.ts       index persistence to disk
  eval/             evaluation framework
    session-recorder.ts   record agy sessions (tool calls + AI output)
    evaluator.ts          score sessions against 5 metrics
    testcases.json        ground truth: questions + expected files/symbols
    claude_answers.md     Claude baseline (for comparison)
  config.ts         paths and constants
AGENTS.md           navigation rules (auto-loaded by agy)
.agents/            workspace MCP config (auto-loaded by agy)