The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a monitoring and aggregation logic that consumes data from an untrusted source (worker-generated event files).
Ingestion points: The functions monitor_parallel_group and aggregate_parallel_group_results in SKILL.md read messages from the bus_path directory using read_messages. Specifically, it extracts payload["error"] from failure messages and payload["commit_sha"] from completion messages.
Boundary markers: There are no boundary markers or instructions to the agent to treat data from the message bus as untrusted content.
Capability inventory: The skill has the ability to spawn new worker agents (dispatch_workers via agent_factory), interact with the file system to create event triggers (Path.touch), and influence the orchestrator's state transitions.
Sanitization: There is no evidence of sanitization or validation of the data retrieved from the message bus before it is used in logic checks or stored in the orchestration metadata.

parallel-dispatch