dsh-strip-sandbox-permissions

Some post-trained models automatically attach sandbox_permissions and its companion justification field to tool-call arguments. When the session already runs with sufficient permission, for example full access, carrying a value that is not strictly wider triggers a false escalation rejection from the DSH sandbox. The tool call is perfectly valid, yet the user sees a confusing sandbox error that blocks the workflow.

dsh-strip-sandbox-permissions is a zero-dependency DSH plugin that intercepts every model tool call at the llm/stream waterfall, the single choke point shared by all adapters. It parses the arguments object in each block-end chunk, removes only the top-level sandbox_permissions and justification keys, and re-serializes the result for the assembler. Every other argument is preserved verbatim. If neither key is present, the original block passes through unchanged at zero cost. Invalid JSON, non-object arguments, and non-tool-call blocks are left untouched.

Built for DSH developers running multi-step tool-call pipelines who occasionally hit spurious sandbox permission rejections. Drop it in, restart DSH, and the noisy fields are silently gone with no changes to adapters or session configuration.