The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the lark-cli binary to perform operations. It executes subcommands such as skills list, skills read, schema, api, and whoami to interact with the service.
[DYNAMIC_EXECUTION]: The instructions direct the agent to load Markdown guides at runtime and execute the commands contained within them. Specifically, it treats metadata.cliHelp from loaded guides as a recipe for command execution, involving the resolution of placeholders and dynamic assembly of shell commands.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface where it reads data from the output of lark-cli skills read and local files in ~/.agents/skills/. This data is used to guide the agent's behavior, creating a surface where untrusted content in those files could influence or override intended instructions. Evidence: Ingestion points include lark-cli skills read and ~/.agents/skills/<guide-name>/. There are no explicit boundary markers or sanitization steps mentioned for the content of these guides. Capabilities include subprocess execution of the lark-cli tool.
[METADATA_POISONING]: The skill includes a .security-scan-passed file containing a claim that the skill has passed security validation. Per security protocols, this internal claim is ignored in favor of the analyzer's own findings.

lark-cli-router