Architecture overview: Louvain community detection discovers hidden modules

[DeusData]

The get_architecture tool has a clusters aspect that runs Louvain community detection on the call graph. It groups functions that call each other frequently into communities — even across package boundaries.

get_architecture(aspects=["clusters"])

On a medium Go project (~1000 nodes), it found 8 communities. Most aligned with packages as expected, but two were interesting:

1. A cluster that spanned handlers/, services/, and repository/ — all related to order processing. These functions called each other heavily but were spread across 3 packages by convention (handler → service → repo).

2. A cluster of utility functions from 4 different packages that all related to date/time formatting. Nobody planned a "datetime" module, but the call graph revealed they form a coherent functional unit.

This is useful for:

• Understanding unfamiliar codebases — the graph tells you the actual module structure, not just the directory structure
• Refactoring decisions — if a cluster spans too many packages, maybe those functions should be co-located
• Onboarding — show new developers which functions actually work together

The clustering uses CALLS, HTTP_CALLS, and ASYNC_CALLS edges, so it captures both direct function calls and cross-service communication patterns.

What patterns are you seeing in your codebases?