The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill audits external GitHub repositories by cloning them and reading untrusted content (e.g., README.md) which could contain adversarial instructions intended to influence the agent's scoring or subsequent remediation suggestions.
Ingestion points: The scripts/collect_evidence.py script and references/audit-checklist.md instructions prompt the agent to read repository metadata and the README.md file.
Boundary markers: The skill does not provide explicit delimiters or "ignore embedded instructions" warnings to the agent when it is instructed to process the external repository content.
Capability inventory: The skill workflow gives the agent access to execute shell commands (gh, git) and Python scripts that can interact with the network and local filesystem.
Sanitization: No sanitization or filtering of the ingested file content is performed before it is analyzed by the agent.
[COMMAND_EXECUTION]: The workflow relies on the agent executing various shell commands to evaluate repository health, which involves handling user-provided repository identifiers.
Evidence: references/audit-checklist.md provides bash templates using gh (GitHub CLI), git, grep, wc, and ls to gather metrics and check file presence.
[EXTERNAL_DOWNLOADS]: The skill performs repository clones from external sources based on the user-provided target.
Evidence: references/audit-checklist.md uses git clone --depth 1 to download the target repository to a temporary directory for analysis.
[DYNAMIC_EXECUTION]: The internal evidence collection script uses dynamic command invocation to interact with Git and the GitHub CLI.
Evidence: scripts/collect_evidence.py executes gh and git via subprocess.run using arguments derived from the command-line inputs.

repo-audit