dsh-skill-gate

In DeepSeek Harness, agents routinely auto-trigger skills based on keyword matches, yet the system-picked skill is not always what the user actually needs. dsh-skill-gate inserts a lightweight human checkpoint at exactly that moment, handing the final decision back to the operator without blocking the automation pipeline.

The flow is simple: when the model emits a skill call, the tools/execute waterfall intercepts the dispatch and opens the native question UI, presenting the system-recommended skill alongside up to four additional candidates, each with a short description. If the user selects an alternative, the original call is short-circuited and a synthetic instruction is injected via agent.inject() so the model loads the chosen skill instead. If no selection is made within the 10-second window, the default skill proceeds untouched. Only one gate exists at a time to prevent stacked pop-ups, and in headless environments every call passes through with zero overhead.

This plugin suits teams running a rich skill catalog in DeepSeek Harness where a wrong auto-trigger carries real cost—think multilingual doc generation, database queries, or code review pipelines—and where preserving human final-say without sacrificing throughput matters.