The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a Bash script (scripts/ollama-task.sh) to perform operations via curl and jq. While it uses safe practices such as jq --arg to prevent shell injection, it still performs active command execution on the host system to interact with the Ollama API.
[DATA_EXFILTRATION]: The skill is designed to send text content (prompts, code, or data) to an HTTP endpoint. While the default is localhost, the skill allows and provides instructions for connecting to remote Ollama servers. This could lead to sensitive project data being transmitted to external infrastructure if a user provides a remote URL.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a bridge between the primary agent and a sub-agent (Ollama). The output from the Ollama model is brought directly back into the primary agent's context. If the data being processed contains adversarial instructions that influence the Ollama model's output, it could lead to indirect injection into the primary agent. The skill lacks explicit technical boundary markers (like XML tags) to wrap the sub-agent's output.
Ingestion points: The output of scripts/ollama-task.sh run/start/send is returned to the agent context.
Boundary markers: None explicitly implemented in the retrieval script.
Capability inventory: The skill can read/write state files, execute shell commands (curl, jq, rm), and perform network requests.
Sanitization: The script extracts raw JSON fields from the API response without further filtering or escaping before returning them to the LLM.

ollama-delegate