The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the 5dive CLI, often requiring sudo privileges to manage systemd units, Linux users, and tmux sessions associated with background agents.
[REMOTE_CODE_EXECUTION]: Commands such as 5dive crew install and 5dive agent import facilitate the downloading and execution of external code, including git-based projects and agent personas with associated hooks.
[EXTERNAL_DOWNLOADS]: The skill facilitates fetching external resources for agent provisioning and system updates, targeting vendor-specific repositories and user-specified URLs.
[DATA_EXFILTRATION]: Provides built-in commands for transmitting data externally, specifically through 5dive push for GitHub integrations, 5dive deploy for Vercel deployments, and a 5dive bug command that transmits diagnostic telemetry to the vendor's repository.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection due to the ingestion of data from various sources.
Ingestion points: Inter-agent messages via agent ask/agent logs, shared task data in tasks.db, and external inbound messages from Telegram or Discord channels.
Boundary markers: The CLI implements a messaging envelope ([5dive-msg from=<name> id=<id>]) to help the agent distinguish between peer messages and instructions, though these do not prevent adversarial injection.
Capability inventory: The agent can execute administrative commands, write to shared databases, push/deploy code, and interact with other agents via tmux send-keys.
Sanitization: The instructions recommend quoting message bodies to avoid shell interpolation, but there is no evidence of automated filtering or sanitization of ingested content before it enters the agent's context.

5dive-cli