The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes commit messages and changelog text from the repository history to generate documentation. \n
Ingestion points: The scripts/collect-releases.js script reads data from local Git tags, commit subjects, and CHANGELOG.md files.\n
Boundary markers: Instructions in SKILL.md provide a specific Markdown template and require the agent to "State only claims supported by that evidence," though it lacks strict data/instruction delimiters.\n
Capability inventory: The skill can execute shell commands (git, vale, prettier, pnpm) and write documentation files to the local repository.\n
Sanitization: Content from the Git history is summarized without explicit filtering for adversarial instructions, which is a common surface for indirect injection in text-processing tasks.\n- [COMMAND_EXECUTION]: The script scripts/collect-releases.js executes local git commands using execFile, which is a secure method for subprocess invocation that prevents shell injection by passing arguments as an array. The skill also involves running standard development tools such as vale for linting and pnpm for builds.\n- [EXTERNAL_DOWNLOADS]: The instructions describe using git fetch --tags origin to synchronize the local environment with the remote repository. This is a standard and expected operation for a release management tool.

summarize-releases