The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes a local utility script scripts/ollama-task.sh which uses curl and jq for API interaction.
Evidence: The script uses jq's --arg and --argjson flags to safely inject variables into JSON bodies, preventing injection attacks.
Evidence: The script uses set -euo pipefail to ensure robust error handling and exit status propagation.
[DATA_EXFILTRATION]: The skill communicates with an Ollama API over the network.
Evidence: By default, the skill targets http://localhost:11434, which is a safe local-only endpoint.
Evidence: For remote hosts, the skill instructions explicitly require the agent to ask the user for the host address rather than guessing or hardcoding external URLs.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a bridge to external models.
Evidence: The skill documentation in README.md clarifies that the target Ollama models operate in a restricted environment with no tool use, file access, or shell capabilities, which limits the scope of any potential prompt injection.

ollama-delegate