project-understanding

To map a project's code directory structure and generate a source understanding file, the skill walks the configured code_dir to inventory entry points like train.py and eval.py, outputting a detailed source.md map.

Tracing the training loop and evaluation harness involves analyzing the code tree at code_dir to identify data interfaces and key modules, which this skill compiles into a coherent source.md document for the project.

Formalization code correspondence compares the formalization.md objectives with the actual code implementation to surface gaps and guide debugging, particularly useful during the DIAGNOSE entry phase after major code changes.

Yes, you need a state.json file with a configured code_dir, along with project.md and formalization.md, to extract entry points, the method module, the training loop, and the evaluation harness into source.md.

You should regenerate the source map when source.md is missing or outdated, particularly when entering a DIAGNOSE phase or after making significant code changes that affect the project's method module or training loop.

Yes, you can automatically identify key entry points like train.py and eval.py by running this skill on a project with a configured code_dir, which inventories the structure and maps the method module into source.md.