The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the islo CLI utility to manage the lifecycle of job runs and sandboxes. Specifically, it uses islo use <sandbox> -- <command> to execute arbitrary commands within a managed environment for diagnostic purposes (e.g., git status). While these are vendor-provided tools, the ability to execute arbitrary shell commands is a high-privilege operation within the sandbox context.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface. It instructs the agent to diagnose failures by reading error_message, error_code, and output from islo logs. This data—which could originate from untrusted external sources or previous failed stages—is then incorporated into recovery instructions passed to a new agent session via the islo factory line-run steer --reason flag.
Ingestion points: Data enters the context through islo logs, islo job event, and inspection of sandbox files (e.g., git status output).
Boundary markers: The skill does not define specific delimiters or instructions to ignore embedded commands when processing these diagnostic outputs.
Capability inventory: The agent can modify future agent prompts (steer --reason), execute commands in sandboxes (islo use), and delete infrastructure (islo rm).
Sanitization: No explicit sanitization or validation of the ingested error data is described before it is reused in instructions.

line-manager-recovery