SkillEvaluator: a map of its code, package by package.

SkillEvaluator is NVIDIA's framework for evaluating the skills of AI agents. Wingbox built the graph of the repository, with no model, and the map stacks its code by package: 335 files and 1,394 dependencies on 16 planes, the packages the others depend on lower. Click a file to see what a change to it reaches.

SkillEvaluator on 16 package planes, one file selected with its blast radius
Open the live mapRead the reportNVIDIA/SkillEvaluatorDrag to turn the planes, scroll to zoom, click a file for its blast radius.
The report

What Wingbox wrote on the whole repository

REPORT.md as the tool wrote it, without its own guesses of architecture layers.

Wingbox report: SkillEvaluator

Generated on 2026-11-16 by wingbox 0.2.6 from the code graph of this project. Open map.html in this folder for the interactive map.

Key findings

  • A change to src/skillevaluator/constants.py reaches 150 of 854 modules (18%) within 2 hops; 79 modules depend on it directly.

Summary

Modules854
Dependencies between modules1,557
Functions, constants, classes and types17,992
Languagespy 854
Boundary violations0

Most depended-on modules

A change in one of these files reaches the most code (test files are left out). "Within 2 hops" counts the modules that depend on it directly or through one other module.

FileDirect dependentsWithin 2 hops
src/skillevaluator/constants.py79150
src/skillevaluator/models/result.py52115
src/skillevaluator/utils/secure_fs.py3982
src/skillevaluator/utils/redaction.py2889
src/skillevaluator/validators/base.py2545
src/skillevaluator/logging_config.py2455
src/skillevaluator/utils/structured_data.py2366
src/skillevaluator/provider_config.py22115
src/skillevaluator/reporting/base.py1931
src/skillevaluator/tier3/harbor/metrics.py1844
src/skillevaluator/utils/rich_markup.py1794
src/skillevaluator/validators/frontmatter_parser.py1749
src/skillevaluator/utils/path_security.py16115
src/skillevaluator/__init__.py1545
src/skillevaluator/tier3/harbor/adapter.py1528

Where to start reading

Files that nothing else depends on and that pull in the most modules:

  • src/skillevaluator/validators/__init__.py, depends on 20 modules
  • src/skillevaluator/tier1/plugin_checks.py, depends on 19 modules
  • src/skillevaluator/deduplication/plugin/commands.py, depends on 11 modules
  • src/skillevaluator/reporting/retained.py, depends on 10 modules
  • src/skillevaluator/tier3/harbor/local_environment.py, depends on 8 modules
  • src/skillevaluator/tier3/harbor/templates/eval.py, depends on 8 modules
  • src/skillevaluator/reporting/plugin_reports.py, depends on 7 modules
  • ci/benchmark_publish.py, depends on 4 modules
  • src/skillevaluator/tier3/harbor/loose_agents.py, depends on 3 modules
  • ci/prepare_security_baselines.py, depends on 2 modules

Next steps

Check what a change will reach before you make it:

wingbox impact src/skillevaluator/constants.py

Rebuild after the code changes with wingbox build. This report and the map are rewritten on every build.

Limits

The graph comes from static parsing. Dependencies created at run time (dependency injection, reflection, generated code) are not visible here.

Method

How this was made

  • Wingbox 0.2.6 from PyPI on the whole repository at commit 186134b, 2026-11-16: one command, wingbox build .
  • The map leaves out the 519 test files and puts every other file on the plane of its package, the folder of its nearest package manifest. The planes are stacked by the dependencies between packages.
  • Checked against the code. Python dependencies compared with Python's own parser: 1,557 in the graph, 1,524 imports in the code, 0 missing, 4 edges in the graph the parser does not confirm.

Run it on any repository: pipx install wingbox && wingbox build . · Wingbox on GitHub (MIT)