The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes instructions from external files such as plan.md, metadata.json, and acceptance criteria to perform file writes and code changes.
Ingestion points: Reads task definitions and criteria from plan.md and related documentation in the workspace.
Boundary markers: The skill lacks explicit instructions or delimiters to isolate the content of plan.md from the agent's instructions, though it includes a check to ensure the plan was previously evaluated by a separate process.
Capability inventory: The agent has capabilities for reading and writing files (read_file, write_file), as well as committing changes to the source control system.
Sanitization: There is no evidence of sanitization, validation, or escaping of the instructions found within the external plan file before they are acted upon.

loop-executor