The Agent Skills Directory

[SAFE]: The skill instructions define a comprehensive security framework for task routing. It emphasizes the principle of least privilege by instructing the agent to pass only the minimum task-local context to downstream providers and to preserve host-native approvals for sensitive operations (e.g., credentials, egress, deployments).
[SAFE]: The instructions explicitly address prompt injection risks by mandating that the agent match structured routing rules only and never execute rule text as instructions. It also implements a 'Lane' system (Fast, Standard, Deep, Operational) to separate low-risk read-only tasks from high-risk mutations.
[DATA_EXPOSURE]: No hardcoded credentials, sensitive file paths, or unauthorized network operations were identified. The skill includes instructions to derive 'effective sensitivity' for data inputs and requires explicit user approval for any sensitive egress or credential usage.
[REMOTE_CODE_EXECUTION]: There are no patterns of remote code execution, package installations, or external script downloads from unknown sources. While the skill mentions a 'Pragman CLI', it does not provide commands to download or install it, instead focusing on the instructional logic for routing within the agent ecosystem.
[INDIRECT_PROMPT_INJECTION]: As a routing component, the skill naturally processes untrusted user requests. However, it mitigates this attack surface by requiring the construction of a 'normalized route input' and 'task contract', effectively sanitizing and structuring data before it is passed to downstream skills.

pragman-router