The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs repository synchronization using standard git commands (git fetch, git pull --rebase, git rev-parse) to ensure hardening fixes are applied to an up-to-date local branch.
[COMMAND_EXECUTION]: Utilizes a provided shell script scripts/detect_ecosystems.sh to identify manifest files. The script uses benign system utilities such as find, ls, and sed to search for project metadata.
[COMMAND_EXECUTION]: Instructs the agent to run project-specific sanity checks (e.g., npm ci, uv sync, docker build --check) after applying approved changes. These executions are time-limited and restricted to the scope of the project being audited.
[EXTERNAL_DOWNLOADS]: References official security tools and configurations from well-known sources, including Trivy (Aqua Security), Cosign (Sigstore), and pip-audit (PyPA), to assist the user in hardening their development pipeline.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it reads and analyzes untrusted data from project manifest files and GitHub Action workflows.
Ingestion points: Project manifest files (package.json, pyproject.toml, requirements.txt, Dockerfile, .github/workflows/*.yml) read during Phase 1 and Phase 2.
Boundary markers: The skill uses explicit markdown code blocks and instructions for verbatim quoting to maintain separation between tool output and agent instructions.
Capability inventory: The skill has capabilities for file system read/write (after approval) and shell command execution (git, package managers) as documented in SKILL.md.
Sanitization: Relies on verbatim quoting and human-in-the-loop review cycles during the four-phase workflow to mitigate the risk of malicious instructions in audited files influencing agent behavior.

supply-chain-audit