The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes external data from GitHub releases (gh release view) and project changelogs to identify breaking changes. This presents an indirect prompt injection surface where a malicious package author could embed instructions in release notes to influence the agent's behavior. While instructions explicitly tell the agent to extract only specific technical details and treat other content as noise, there are no formal sanitization or strict boundary markers (like XML tags with 'ignore instructions' warnings) for this external content. Ingestion occurs via CLI tool output which is then interpolated into the agent's context for summarization.
[COMMAND_EXECUTION]: The skill makes extensive use of local shell commands to audit the dependency tree and verify project integrity. It uses bun, npm, pnpm, yarn, gh, jq, rg, and find for operations ranging from lockfile diffing to running test suites and build gates. This is a primary function of the skill and requires significant environment permissions.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of external and dynamically generated code. It instructs the agent to create and run a local TypeScript file (peer-check.ts) for peer dependency validation and encourages the use of tool runners (bunx, npx, pnpm dlx, yarn dlx) to execute remote packages like semver, taze, and various ecosystem-specific codemods (e.g., @next/codemod). These tools download and execute code from public registries at runtime.
[EXTERNAL_DOWNLOADS]: The skill fetches package metadata and release information from well-known services including GitHub and official package registries (npm/pypi). These network operations are documented as standard requirements for dependency auditing and target trusted infrastructure.

deps-upgrade