Development¶

Validation¶

Validate all canonical skill packages, then check that every loop packages the shared loop contract. --repo . also checks that every skill has a row in the README skills table and that .claude-plugin/plugin.json and the marketplace entry carry the same version. The parallel skills are not loops, so the contract check covers only the *-loop directories:

python3 scripts/validate_skills.py skills --repo .
python3 scripts/validate_skills.py skills/*-loop \
  --contract shared/loop-contract.md

codex/skills/ is generated from skills/. Rebuild it after changing a skill, and commit both:

python3 scripts/build_codex.py

Each Codex copy is the canonical skill plus a "Running in Codex" section and an agents/openai.yaml that turns off implicit invocation. The tests fail when codex/skills/ is stale, so the validator runs on skills/ only.

Lint the installer, and check the plugin manifests when Claude Code is installed:

bash -n scripts/install.sh && shellcheck scripts/install.sh
claude plugin validate .

Releases¶

The version lives in two places, .claude-plugin/plugin.json and the sstack entry in .claude-plugin/marketplace.json; the validator fails when they differ. Bump both, add a CHANGELOG.md entry, and tag v<version> after the merge.

Tests¶

Run the validator tests:

python3 -m unittest discover -s tests

The suite includes the repository-agnostic global Stop runtime and cross-repository/session isolation tests.