dsh-approval-flow-poc

Running AI tools inside DeepSeek Harness often leaves teams in an awkward middle ground: the official tools/pre-execute hook is available, but there is no out-of-the-box enterprise approval policy. dsh-approval-flow-poc closes that gap by classifying every tool call into allow, deny, or requireApproval based on a risk level (low / medium / high / critical), so you do not need to build a full security gateway from scratch.

The design follows a minimal-control philosophy. Critical calls are rejected outright; the developer, production, and strict presets map to local development, controlled pre-production, and high-sensitivity demos respectively, and switching is as simple as overriding one config block in the profile. When a human decision is needed the plugin returns DSH-native ask decisions, so the approval interaction stays entirely inside the built-in service and Web UI. Every call also writes a redacted JSONL audit row that records only the parameter summary and policy outcome, never the raw arguments.

It suits small and mid-size teams that already run agent workflows on DSH and want one extra human-confirmation gate on tool invocations. The project is explicit about its limits: no SSO, RBAC, multi-level co-signing, or compliance certification. It is a minimal PoC that can be validated quickly and later extended with Feishu or WeCom approval transports.