ollama-task-orchestrator

Manage and execute Ollama tasks via SSH on a remote worker, providing queue status, exclusive task locking, and code generation commands.

This skill appears to do what it says: let an agent run Ollama tasks on a remote worker via SSH. Before installing, verify the worker is fully trusted and configured for this use: 1) review runner/run_task.sh and runner/queue_status.sh (they can kill/restart Ollama, eval arbitrary commands, and write files inside PROJECTS_DIR); 2) ensure SSH keys in ~/.ssh/config grant access only to the intended, isolated worker account (prefer a least‑privilege account or VM); 3) set DEFAULT_PROJECT carefully (write/test operations will operate inside that project) and keep ALLOW_NL_EXEC=false unless you trust natural-language→command parsing; 4) do not enable this skill for untrusted or broadly autonomous agents unless you accept the ability to run arbitrary commands on the worker; 5) run the installer and scripts in a sandbox or non-production environment first and audit file permissions. If you want lower risk, use the runner only interactively from a controlled host (avoid giving agents unrestricted access).

Type: OpenClaw Skill Name: ollama-task-orchestrator Version: 1.0.0 The skill provides a remote task orchestration framework that includes an explicit 'exec' command using 'eval' to run arbitrary shell commands on a worker machine (run_task.sh). While this is a documented feature, it creates a significant Remote Code Execution (RCE) surface, especially when combined with the natural language routing logic in run_task.sh which uses an LLM to interpret user requests. If the 'ALLOW_NL_EXEC' environment variable is enabled, the system becomes highly vulnerable to prompt injection attacks that could execute unauthorized commands. The skill also facilitates remote execution via SSH (skill.py), extending the risk to the configured worker host.