The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the hivemind CLI for managing organization workspaces, authentication, and status checks, and includes functionality to serve a web-based dashboard on a local port.
[EXTERNAL_DOWNLOADS]: It fetches machine learning model weights (~600MB) from Hugging Face for semantic search and updates its own CLI via npm, both of which are well-known technology services.
[REMOTE_CODE_EXECUTION]: The skillify pull command synchronizes instruction sets (referred to as "skills") from the Deeplake cloud to local agent directories, enabling dynamic behavior updates from organization-wide contributions.
[PROMPT_INJECTION]: The agent is instructed to read past session summaries and index files from ~/.deeplake/memory/ to establish context, creating a surface for indirect prompt injection.
Ingestion points: Reads files from the local memory directory using shell tools like cat and grep.
Boundary markers: No specific delimiters or safety instructions are provided in the guide to isolate recalled context from active instructions.
Capability inventory: The agent can execute CLI commands, modify organization state, and read/write to project and global skill directories.
Sanitization: Historical session content is processed and loaded into the prompt without explicit validation or filtering.

hivemind-weapon