Skip to content

Development

m4bwav edited this page Nov 21, 2026 · 1 revision

Layout

Path What
.claude-plugin/plugin.json, marketplace.json plugin and marketplace manifests (the version lives in plugin.json)
skills/trustwright/SKILL.md the skill: log, register a stop, compare, verify
skills/trustwright/ the evergreen unit: evergreen.json, RESEARCH.md, LEARNINGS.md, CHANGELOG.md, TESTS.md, MAINTENANCE.md, evals/
references/ taxonomy.md (stop sources, categories, kinds), verify.md (the method behind check), sources.md (vendor sources)
scripts/trust.py the script, standard library only
tests/test_trust.py, tests/fixtures/session.jsonl unit tests and a sample transcript

Tests

python -m unittest discover -s tests

or python -m pytest. Each test sets TRUSTWRIGHT_HOME to a fresh temporary folder, so your own record is never touched. 0.1.1 has 8 tests, two of them regression tests for the 0.1.0 bugs. There is no CI; run them before a pull request.

Evals

skills/trustwright/evals/evals.json holds the skill's eval cases, run with the skill and without it. Results are in skills/trustwright/TESTS.md; scratch homes go under a temporary folder through TRUSTWRIGHT_HOME.

Adding to the lists

KINDS, SOURCES and CATEGORIES at the top of trust.py are the allowed values, and references/taxonomy.md explains them; change both together. A new stop signature for scan goes in STOP_PATTERNS as a (source, regex) pair, with a line in tests/fixtures/session.jsonl and an assertion in the scan test. The first pattern that matches a text wins.

Releasing

Bump version in .claude-plugin/plugin.json and in the metadata.version of SKILL.md, add a CHANGELOG.md entry, merge, tag vX.Y.Z on main and create the GitHub release. Then update Versions and upgrading and the footer of this wiki.

Clone this wiki locally