The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides templates and instructions for users to manually execute management commands such as ajet-swarm start to coordinate reinforcement learning clusters.
[EXTERNAL_DOWNLOADS]: Generated agent code is designed to reference model weights and datasets hosted on ModelScope's official infrastructure.
[DATA_EXFILTRATION]: The skill describes network communication where clients connect to a swarm server (default port 10086) to synchronize training rewards and metadata, which is the core functionality of the AgentJet framework.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it ingests user requirements to generate executable agent code and reward logic.
Ingestion points: User-provided specifications for agent tasks and scoring metrics.
Boundary markers: The provided templates do not utilize explicit delimiters for user input interpolation.
Capability inventory: Generated scripts have access to the openai SDK and requests library for network-based task completion.
Sanitization: No explicit sanitization or validation of user-provided logic is implemented in the templates, as they are intended for development environments.

write-swarm-client