dsh-tool-policy

In the DeepSeek Harness ecosystem, tool execution already has robust primitives like sandbox policies and one-shot approvals, but lacks a deployment-owned, unified policy layer to govern every tool call, especially for third-party and MCP tools. The dsh-tool-policy plugin addresses this gap by acting as a declarative per-call policy layer that applies allow, ask, or deny rules before tool execution. Its core capability lies in flexible matching based on tool names and optional argument patterns, with a deny-by-default baseline for enhanced security. For instance, it can reject known destructive command patterns, require human approval for MCP tools, or run allowlists for unattended jobs, while offering policy decision tracing to help operators understand rule triggers. This plugin is ideal for developers, security ops personnel, or enterprise deployers who need fine-grained control over tool calls, especially in scenarios demanding unified policy management across diverse tools and protection against sensitive data leakage in DeepSeek Harness deployments.