The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses local project documentation (docs/docker-build-.md) as the primary source of truth for constructing and executing shell commands. This creates a surface where a malicious repository could provide crafted build rules to influence the agent's behavior. * Ingestion points: Reads docs/docker-build-.md and modules.json from the project root. * Boundary markers: Absent; the agent treats file content as authoritative instructions. * Capability inventory: Shell script execution (.cmd, .ps1), Maven (mvn), and Git commands. * Sanitization: No explicit sanitization of extracted strings (paths, parameters) before command interpolation.
[COMMAND_EXECUTION]: The skill's core function is to execute local scripts and CLI tools. While it requires confirmation for the first build and for sensitive actions (e.g., publishing), it bypasses confirmation for 'ordinary follow-up' requests. This design relies on the user thoroughly vetting the initial command 'shape'.
[DYNAMIC_EXECUTION]: The agent dynamically generates shell commands by parsing markdown files and filling templates with extracted values. Without strict validation of the source documentation, this could lead to unintended command execution if the documentation is controlled by an adversary.

devops-package-copilot