Development

Rules for contributors and agents are in AGENTS.md. The scripts stay standard-library Python 3.9+, and every change gets a test.

python -m unittest discover -s tests
python scripts/budget.py

The test run for this wiki:

Ran 53 tests OK

CI runs the tests on Ubuntu, macOS and Windows, on Python 3.14 everywhere and on 3.9 on macOS and Windows, plus the size budget.

Evals

claude plugin eval . --no-publish --trust-plugin --case "trigger-*"

Trigger and decoy cases work there. Trap cases need real writes, and plugin eval mounts the workspace read-only, so they run through subagents. Results and limits are in TESTS.md.

Releases

A release bumps .claude-plugin/plugin.json, plugin.json, VERSION in grw.py, metadata.version in SKILL.md and the CHANGELOG together, then gets a tag.

The examples on this wiki come from ai-docs/notes/2026-11-03-wiki-verify.py, which runs against a fresh clone. Its saved output is beside it.