The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external logs and local project files which could theoretically contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The agent reads container stdout/stderr logs via the GET /api/v1/docker/{owner}/{name}/logs endpoint and processes local project source code/Dockerfiles during the deployment preparation phase.
Boundary markers: There are no explicit instructions for the agent to wrap log output in delimiters or treat the content as data rather than instructions when troubleshooting.
Capability inventory: The skill uses curl for API interactions, tar for creating build contexts, and involves writing sensitive credentials (API keys) to local .env files.
Sanitization: The skill provides a 'Best Practice' to use an inspected, allowlisted staging directory for builds to prevent the accidental upload of sensitive local data.

cloudish