The Agent Skills Directory

Command Execution: The skill executes shell commands using tools such as git, gh, and uv run to retrieve governance data and perform repository audits.
Evidence: SKILL.md defines specific procedures for using git log, gh label list, and uv run to query repository state.
Context: The skill implements robust safeguards against command injection, including strict regex validation for policy keys and repo paths, and the use of placeholders for resolved values.
Indirect Prompt Injection: The skill possesses an attack surface for indirect prompt injection because it ingests and processes untrusted data from various files in the repository.
Ingestion points: Governance files, skill definitions (SKILL.md), and recipe manifests (manifest.yaml) throughout the repository.
Boundary markers: Responses are explicitly prefixed with "Generated by Oracle" and subject to a strict 200-word limit to maintain clear boundaries.
Capability inventory: The skill can read file contents and execute several shell-based auditing tools (git, gh, grep).
Sanitization: It uses input validation to ensure that interpolated values like policy keys conform to safe character sets ([a-z0-9_.]).
External Downloads: Auditing procedures utilize the uv tool to fetch external dependencies required for script execution.
Evidence: Scripts in reference/drift-checks.md use uv run --with pyyaml to process YAML configuration files.
Context: The skill relies on well-known, standard packages such as pyyaml for its core auditing tasks, which is standard operational practice.

repo-oracle