The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a workflow where the agent analyzes external application code and metadata to generate configuration files (package.yml, lzc-manifest.yml). This creates a surface where malicious or misleading data in the processed application could influence the agent to generate over-privileged configurations. * Ingestion points: Analysis of user-provided Dockerfiles, source code, and application descriptions mentioned in SKILL.md. * Boundary markers: The instructions lack specific delimiters or ignore directives when processing user-provided application content. * Capability inventory: The skill enables the generation of configurations for shell execution in setup_script and buildscript, network permissions, and high-privilege Docker overrides as documented in references/manifest-spec.md and references/build-spec.md. * Sanitization: No specific validation or sanitization rules are provided for the data processed during configuration generation.
[COMMAND_EXECUTION]: The skill provides guidance on using platform-specific tooling (lzc-cli) and specifies fields for executing shell scripts during the build and initialization phases. These features are documented as standard operations for application packaging and deployment.
[PRIVILEGE_ESCALATION]: The technical specifications include support for requesting high-level system permissions, including privileged: true and compose.override. The documentation explicitly identifies compose.override as a high-risk capability and notes that certain storage access requires administrator authorization.

lazycat-lpk-builder