FengLingYaaa/dsh-gpt-perm-strip

This is a native plugin built exclusively for DeepSeek Harness, and it only operates on GPT-family models. Its core purpose is to resolve a common error triggered when GPT sends duplicate sandbox permission fields that are not strictly wider than the current session’s permission level. Before the sandbox escalation check runs, the plugin strips unnecessary sandbox_permissions and justification fields while leaving genuine permission escalation requests untouched. It works by wrapping each tool’s execute method to modify a cloned copy of the call arguments, so the original tool call record stored by DSH still shows the full output the model generated.

Any developer using GPT-family models with DeepSeek Harness will benefit from this plugin, as GPT commonly adds permission fields even when they are not needed for a new tool call. Without this plugin, calls that include duplicate permission fields fail immediately with an error about the escalation not being strictly wider. The plugin fixes this failure case automatically without requiring any manual changes to how you write prompts or configure tools. It only activates for matching GPT models, so all other models you use with DSH will not be affected by the plugin’s logic.

Installing the plugin is straightforward, and you can add it directly from GitHub with the single command dsh plugin --profile web add github:FengLingYaaa/dsh-gpt-perm-strip. If you prefer to modify the plugin or install it from a local clone, you can run git clone, install dependencies with pnpm, run tests, build, then add it to your DSH profile from the local directory. The plugin offers several configuration options, including toggling logging of stripped fields, adding custom model patterns, and toggling a GPT-specific context prompt. It is released under the open source MIT license, so you can use, modify, and distribute it freely for any purpose.