The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external data from git history, GitHub Pull Requests, and documentation files, creating a surface for indirect prompt injection.
Ingestion points: The agents/changelog-generator.md subagent reads git commit logs (git log), PR metadata (gh pr list), and issue titles (gh issue list). The agents/docs-updater.md subagent reads project documentation files (*.md, *.rst).
Boundary markers: The subagent instructions do not include explicit delimiters or safety instructions to prevent the agent from obeying instructions that might be embedded within the git history or PR descriptions.
Capability inventory: The orchestrator has significant capabilities, including executing shell commands (git, gh, npm, twine), modifying project files, and publishing to public registries (PyPI, npm).
Sanitization: There is no evidence of data sanitization before the ingested text is provided to the agent for processing.
[COMMAND_EXECUTION]: The skill dynamically detects and executes build commands and scripts defined in the project's configuration files.
Evidence: SKILL.md identifies and executes commands from package.json, Makefile, Cargo.toml, and pyproject.toml (e.g., npm run build, make build, cargo build --release).
[EXTERNAL_DOWNLOADS]: The skill ensures that necessary packaging tools are installed by downloading them from the official Python package registry at runtime.
Evidence: references/publishing.md executes pip install --upgrade build twine to prepare the environment for PyPI publishing.

release-manager