The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts on external data including 'rendered prompts' and 'referenced files' to perform code changes and validation (SKILL.md). This represents a standard surface for indirect prompt injection where the ingested content could attempt to influence the agent's behavior. However, the skill includes a 'Quality Gate' that instructs the agent to stop and return 'NEEDS_COORDINATOR' if the prompt is contradictory, diverged, or invalid, which acts as a safety check.
[COMMAND_EXECUTION]: The instructions direct the agent to 'Run the assigned or directly relevant validation' (SKILL.md). No specific shell commands are provided in the skill; it relies on validation logic defined in the external task packet. No evidence of unsafe command construction or injection of user-supplied arguments into shell execution was found in the analyzed file.

plan-runner