DeepSeek Harness Plugin: dsh-dual-model-eval

DeepSeek Harness introduces the `dsh-dual-model-eval` plugin, leveraging its “everything is a plugin” architecture to meet the need for concurrent code-task evaluation across multiple models. The plugin supports enabling comparison test mode, running 2 to 4 models concurrently, and independently generating code for each model within an isolated Git worktree. It provides real-time tool-call traces, line-change statistics, and detailed metrics, and allows users to adopt a satisfactory model result as the baseline for the next round. Running with local user permissions, the tool creates local commits without pushing to remote, helping ensure safe validation of model performance in real code environments.