The Agent Skills Directory

[COMMAND_EXECUTION]: The Tools/cmux.ts script utilizes Bun.spawn to execute shell commands, including the cmux CLI and ssh. These operations are fundamental to the skill's purpose of terminal management.
Evidence: Subprocess calls are present in functions like runProcess, cmuxExec, and commandMiniFleet.
Mitigations: The script implements validateSshTarget which uses a whitelist regex (/^[A-Za-z0-9._@:%+/\-]+$/) to prevent command injection in SSH host strings.
[PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection by reading output from terminal screens and returning it to the agent context.
Ingestion points: Screen content captured via the read-screen command in Tools/cmux.ts (used by the read and monitor subcommands).
Boundary markers: No explicit delimiters or instructions are used to wrap the screen output when it is ingested by the agent.
Capability inventory: The skill can execute local commands through cmux and ssh, and send notifications to a local API (localhost:31337).
Sanitization: No sanitization is performed on the screen content before it is processed by the agent.

CMUX