Paths: File paths are relative to skills repo root.

Code Knowledge Graph

Type: Standalone Utility Category: 0XX Dev Environment

Indexes codebase into a layered graph (tree-sitter AST → SQLite) and provides dependency analysis, path tracing, references, implementations, and architecture overview via MCP tools.

Inputs

When to Use

• Starting work on an unfamiliar codebase → index + architecture
• Before refactoring a function/class → find_symbols + inspect_symbol + trace_paths
• Understanding call flow → trace_paths
• Finding a symbol quickly → search

Workflow

Phase 1: Index

Check if graph exists (.hex-skills/codegraph/index.db in project root).

If NOT exists:

Call: index_project({ path: "{project_path}" })

If exists (re-index on demand):

Call: index_project({ path: "{project_path}" })

Idempotent — skips unchanged files automatically.

Phase 2: Query

Route based on user intent:

Canonical selector rule: Semantic tools accept exactly one selector:

• symbol_id
• workspace_qualified_name
• qualified_name
• name + file

Preferred flow: use find_symbols only after narrowing path as much as practical, then feed the returned workspace_qualified_name into inspect_symbol, trace_paths, find_references, or find_implementations for exact follow-up queries.

Query boundary rule: find_symbols is name-based discovery only. For code fragments like export function or unresolved member-call patterns like app.get(...), use grep_search instead of treating them as symbols.

Ambiguity rule: if find_symbols returns truncated: true or a large candidate_count, refine with path, then name + file or workspace_qualified_name instead of widening the graph query.

Path rule: path may be the indexed project root or any file/subdirectory inside that indexed project.

Dataflow anchors: trace_dataflow requires source.anchor and optional sink.anchor. Use:

• param for function parameters
• local for local variables
• return for function returns
• property with access_path for bounded property flow

Precision controls: inspect_symbol, trace_paths, and find_references support min_confidence (low, inferred, exact, precise) when the caller wants to suppress weaker parser-only facts.

Phase 3: Present Results

1. Show MCP tool output directly (markdown tables)
2. For code snippets referenced in results, use hex-line read_file with line ranges; add edit_ready=true, verbosity="full" only when you intend to carry revision/checksums into an edit
3. Suggest follow-up queries based on results:

• After find_symbols with a clean top match → suggest inspect_symbol with workspace_qualified_name
• After find_symbols with truncated: true → suggest narrowing path or switching to name + file before any deeper graph tool
• After inspect_symbol → suggest trace_paths if refactoring
• After trace_paths → suggest find_references or find_implementations depending on symbol kind
• After empty trace_paths from a broad or module-level selector → suggest inspect_symbol or analyze_architecture instead of assuming there are no dependencies

Supported Languages

MCP Server Setup

Add to .mcp.json:

{
  "mcpServers": {
    "hex-graph": {
      "command": "node",
      "args": ["{skills_repo}/mcp/hex-graph-mcp/server.mjs"]
    }
  }
}

Definition of Done

• Project indexed (index_project returns success)
• Query results shown to user
• Follow-up suggestions provided

Version: 0.1.0 Last Updated: 2026-03-20