dsh-llm-approver

In the DeepSeek Harness web GUI, every sandbox escalation retry triggers an approval dialog, even for obviously harmless operations like writing a new file or listing a directory. In fast-paced development, these interruptions quickly add up to noise. dsh-llm-approver tackles exactly this: before the dialog appears, an independent-context LLM call judges whether the pending command is safe.

The core mechanism is a fail-closed pre-review gate. The request contains only the pending command and its justification, never the conversation history. When the LLM determines the operation is reversible, non-destructive, and touches no sensitive data, it is allowed immediately. Any failure mode, whether timeout, parse error, unparsable output, or user interruption, degrades to the original approval prompt. The gate can only grant, never reject, and every edge case resolves by falling back to the user.

It is built for developers who use the DeepSeek Harness web GUI daily, trigger sandbox escalations frequently, and want to cut through the noise of trivial approval prompts. Switch the permission preset to Workspace Write, LLM Review, and safe operations pass silently while genuinely risky commands still surface the dialog for your decision. All audit logs remain exactly as before; LLM verdicts are written to the host log under the llm-approver name and never touch the session event vocabulary.