wangyong1972/dsh-auto-approval

README 摘要

dsh-auto-approval A DeepSeek Harness Cordis plugin that automatically approves safe permission requests using a local LLM (via LM Studio), so you are only prompted for genuinely ambiguous or risky operations. It listens on the approval/request waterfall event ahead of the built-in UI answerer (via prepend ), decides each request through a layered defense, defers uncertain or denied-list operations to the human, and learns operations you approve so they are auto-approved in the future. How it works Key behaviors: - The deny-list never silently rejects. A deny-list hit (e.g. .ssh/ , /etc/ ) is deferred to the human prompt, so you always keep the final say. The operation only runs if you approve it. - Learning is opt-in. Set patternFile to enable it. Every operation you approve after the plugin deferred is remembered; a future identical or pattern-matching operation is auto-approved without prompting or an LLM call. - Only human approvals are learned. Operations the plugin auto-approved (allow-list or LLM-safe) are not learned — there is nothing to learn, and the pattern file only grows from your explicit decisions. - Safe by default on errors. If LM Studio is unreachable or returns a…

分类