dsh-approval-ai-review

In DeepSeek Harness permission mode, every tool call can trigger a human confirmation prompt, even for low-risk operations like reading a config file. Approval fatigue slows the workflow down and tempts teams to loosen policies, which paradoxically weakens the guardrail on the operations that truly need oversight.

dsh-approval-ai-review inserts a single auxiliary model call ahead of any human answerer. The tool name, exact call arguments, and the asker reason are sent to the review prompt: an approve verdict resolves the request immediately; an escalate verdict, a timeout, or a malformed response delegates the request down the existing approval waterfall with the review analysis appended to the reason, so the eventual human sees exactly what the operation would do and what it risks. The plugin is inert by default, and the review call is a standalone stream request that consumes zero conversation-transcript tokens and leaves the KV cache untouched.

It is aimed at teams that keep an approval waterfall in their deployment, want to cut down on repetitive manual confirmations, and still want a human in the loop for anything the model does not clear. The plugin never rejects on its own authority, so safe operations get auto-granted while the review effort is spent only where it matters.