Diagram Intelligence
Turn source evidence into a typed graph, then render diagrams for GitHub, slides, and interactive inspection from the same structure.
The problem
One source can be drawn as a slide, a repository diagram, or a detailed architecture map. Drawing each version independently loses the relationship between the picture and the evidence behind it. It also makes it hard to tell a sourced connection from an inference.
What I built
Diagram Intelligence is a Python plugin for Codex and Claude Code. It discovers candidates and evidence from an authorized source, builds a typed graph, analyzes the relationships, selects a view, and renders the result. Mermaid, DOT, HTML, JSON, and editable PowerPoint use the same graph and layout plan. The core uses the Python standard library; PowerPoint support is optional.
flowchart LR
A[Authorized source] --> B[Evidence and candidates]
B --> C[Typed graph]
C --> D[Relationship and view policy]
D --> E[Layout plan]
E --> F[Mermaid / DOT / HTML / JSON / PPTX]The CLI can validate a source, run the pipeline, and evaluate a six-domain fixture corpus. A NavGator adapter can read a component and connection export without changing the graph contract. Output bundles are local and hashed, with different public, internal, and private detail levels.
Evidence and limits
The repository includes unit tests and an evaluation corpus. Its plugin archive can be packaged for both hosts, but the README says it is not globally installed or published. Public bundle mode strips source locators and private graph fields; the operator must still review any labels they choose to expose. Diagram Intelligence does not call a network rendering service.