dsh-apply-patch

When working with GPT-family models inside DSH, one problem almost always surfaces: these models were trained heavily on the OpenAI Codex apply_patch tool, so they instinctively emit apply_patch calls during a conversation. DSH, however, does not provide that tool by default, and the result is a hard error. dsh-apply-patch exists to close that gap. It registers a fully Codex-compatible apply_patch tool through the DSH plugin system, letting GPT models create, update, delete, and move files as naturally as they would in the native Codex environment.

Three capabilities stand out. First, the patch grammar is a drop-in match for the Codex Begin Patch, Update File, Add File, Delete File, Move to, and End Patch syntax, with lenient parsing that also accepts the gpt-4.1 heredoc style. Second, every file write and deletion is routed through the DSH built-in sandbox file system and sandboxed bash; in read-only mode any write is rejected with a clear error, and the sandbox is never bypassed. Third, the plugin supports conditional injection by model: enable it only for gpt- models, for all models, or disable it entirely. The toggle lives in a settings card and applies in real time.

This plugin is best suited for developers who frequently use GPT-family models (gpt-4, gpt-4.1, etc.) in DSH and want to keep the patch-based file-editing workflow uninterrupted. If you have hit the apply_patch not registered error, or you want a unified Codex-style file editing experience across all models, this is the plug-in for you.