The Agent Skills Directory

Command Execution: The verify-lint-mirror.sh script employs eval to run a sequence of hardcoded linting commands. While the use of eval is generally noted for review, here it is restricted to a static list of well-known local tools such as ruff and pylint, as well as local scripts like lint-auth-signals.sh.
Indirect Prompt Injection Surface: The skill ingests data from external sources, specifically merged Pull Request titles and descriptions via the GitHub CLI (gh pr view), to generate changelog entries.
Ingestion points: Untrusted data enters the skill context through the scripts/list-changes-since-tag.sh script, which fetches PR metadata.
Boundary markers: To mitigate the risk of embedded instructions in PR data influencing the agent, the skill utilizes a structured sanitizer (assets/entry-sanitizer.md). This requires the agent to rephrase and summarize PR content into a strict, two-sentence format before inclusion in the changelog.
Capability inventory: The skill has permissions to modify local repository files (CHANGELOG.md, pyproject.toml) and perform network operations such as git push and gh pr create to the project's official repository.
Sanitization: The workflow enforces three mandatory human-review checkpoints: confirming the version bump rationale, reviewing the changelog diff, and addressing any linting failures before a PR is opened.

cut-release