The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from dependency pull requests (titles, bodies, and changelogs), which presents a potential surface for indirect prompt injection.
Ingestion points: PR metadata and body content are retrieved in SKILL.md using gh pr list and gh pr view.
Boundary markers: Step 0 of SKILL.md contains explicit instructions to treat all PR content as data and never as instructions, specifically warning about compromised dependencies placing payloads in changelogs.
Capability inventory: The skill has the ability to approve and merge pull requests using gh pr review and gh pr merge.
Sanitization: The skill relies on safety gates that extract specific metadata (e.g., version components and commit verification) rather than executing instructions found within the PR text.
[EXTERNAL_DOWNLOADS]: The skill performs network requests to an external registry to verify dependency information.
Evidence: Fetches crate metadata from the official crates.io API in SKILL.md to check the release date of dependencies as part of a security cooldown gate.
[DYNAMIC_EXECUTION]: The skill executes a localized script to process data retrieved from an external service.
Evidence: Employs a hardcoded Python one-liner (python3 -c) in SKILL.md to parse JSON output from the crates.io API to extract a timestamp.

pr-deps-merge