The Agent Skills Directory

[SAFE]: No malicious patterns detected. The skill follows best practices for security-sensitive operations, requiring human confirmation before modifying any files and providing clear instructions on how to interpret external data safely. The instructions explicitly prohibit the agent from committing, pushing, or opening PRs autonomously.
[INDIRECT_PROMPT_INJECTION]: The skill identifies that reading third-party data (like manifests, changelogs, or security advisories) is an attack surface and provides a dedicated 'Author authority' framework to mitigate risks. It explicitly instructs the agent to treat such text as data, never as instruction, and to flag any embedded instructions as attack signals.
Ingestion points: Project manifests (package.json, Cargo.toml, composer.json, go.mod, pom.xml), container definitions, and forge-resolved metadata (GitHub Actions, GitLab CI).
Boundary markers: The skill enforces a mandatory planning phase where all intended changes are presented for human review before any write operations occur.
Capability inventory: Utilizes subprocess execution via Bash and file editing tools for repository maintenance.
Sanitization: Instructs the agent to specifically look for and ignore instructional patterns in data content, treating them as attack signals rather than valid commands.
[COMMAND_EXECUTION]: The skill performs functional use of repository-local and system-provided package management tools (pnpm, npm, cargo, composer, go) and forge CLIs (gh, glab) for auditing and dependency resolution. The instructions emphasize honoring the repository's own configuration and avoid installing unauthorized toolchains or bypassing safety gates like the release-age gate.

update-deps