dsh-auto-approve

When running an Agent in DeepSeek Harness, every file write or shell command triggers a manual approval dialog. For low-risk, reversible day-to-day operations, this click-by-click confirmation quickly becomes a source of disruption. dsh-auto-approve steps in when the session preset is set to Auto-approve: it takes over the approval pipeline and uses keyword rules together with a configurable judge model to decide whether each tool call is allowed, rejected, or escalated to a human.

The core pipeline has three layers. The first is a set of keyword buckets (reject / human / allow) that match tool names, commands, paths, working directories, and custom-tool argument values—yet never inspect written contents or code bodies, guarding only the catastrophic-with-zero-context red lines. The second layer is the judge model, which receives the same fields as the Web tool card and returns a category id plus an independent risk level (low / medium / high); the plugin then executes the action defined in the corresponding row × level cell. The third layer is the fallback: whenever the model output is unparseable, times out, or the call simply fails, the decision reverts to the original Web approval dialog rather than silently passing through. Extra controls include model-initiated human review, a fully editable judge prompt template, and an adjustable output-token budget.

This plugin is aimed at developers who drive Agents heavily inside DeepSeek Harness and want trusted operations to flow automatically while keeping a hard safety net. If you are tired of ls ./src still demanding a manual tap while rm -rf / is caught by a single keyword, dsh-auto-approve makes the approval chain truly risk-tiered.