The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill utilizes a configuration resolver (scripts/resolve.py) that reads project-specific metadata and markdown profiles (e.g., .agents/skills-config/docker-compose-guardrails/host-policy.md) and incorporates their content into instructions for the agent.
Ingestion points: Project-specific config.yaml and profile .md files located in the repository's .agents/skills-config/ directory.
Boundary markers: While the resolver adds markdown headers to separate sections, it lacks explicit delimitation or instructions to prevent the agent from executing instructions potentially embedded in untrusted project files.
Capability inventory: The skill can execute docker compose and docker inspect via structured subprocess calls and writes generated instruction sets to a local cache directory.
Sanitization: The implementation includes strict path-containment validation to mitigate directory traversal risks, though the text content of configuration files is processed without sanitization.
[COMMAND_EXECUTION]: The skill invokes external binaries using the Python subprocess module to interact with the Docker ecosystem.
Evidence: The scripts/check_compose_guardrails.py script executes the docker compose config command.
Context: These executions are restricted to static command lists and are necessary for validating resource guardrails within the target environment.

docker-compose-guardrails