Code Intelligence — Structural Understanding for AI Agents

TL;DR

• Use to index a codebase for fast structural understanding
• Layers: skeleton, code graph, architecture diagram, smart context
• Default: start from the lightest useful layer
• Next: cm-planning, cm-debugging, or cm-execution

Stop scanning. Start querying. Load only the layer the task actually needs.

When to Use

• Understanding a medium or large codebase
• Tracing callers, callees, or impact before edits
• Building focused context for planning, debugging, or execution
• Auto-triggering structural understanding during project setup

Choose a Layer

Need instant codebase orientation with zero extra setup?
└─ YES → Layer 0: Skeleton Index

Need symbol search, callers/callees, or impact analysis?
└─ YES → Layer 1: Code Graph

Need a visual system / module map?
└─ YES → Layer 2: Architecture Diagram

Need a focused context packet for another skill or agent?
└─ YES → Layer 3: Smart Context Builder

Layer	Summary	Load
0	Grep-based skeleton index for instant orientation	references/layer-0-skeleton.md
1	AST graph and symbol-query workflow	references/layer-1-codegraph.md
2	Mermaid architecture generation	references/layer-2-architecture.md
3	Synthesized focused context for downstream work	references/layer-3-context-builder.md

Conditional References

• Load only the target layer reference for the current task.
• Load references/integration-workflows.md only when wiring cm-codeintell into cm-start, cm-planning, cm-debugging, or cm-execution.

Integration

Skill	How cm-codeintell helps
cm-start	bootstrap structural context early
cm-planning	impact analysis and component boundaries
cm-debugging	trace callers, callees, and failure paths
cm-execution	pre-flight context for focused implementation

Rules

• Start from the lightest layer that answers the question.
• Prefer Layer 0 when simple orientation is enough.
• Regenerate or refresh indexes after major structural changes.
• Do not pay for deep graph or context synthesis unless the task needs it.

The Bottom Line

Pick the cheapest layer that answers the question, then load deeper only when needed.