dsh-gpt-tool-compat

If you route GPT-family models (e.g. gpt-5.6-terra) through a third-party provider, you may hit a frustrating wall: the model attaches privilege-escalation parameters to every single tool call, and DSH parameter validation rejects them outright. The call never executes, you see errors like invalid justification or not strictly wider, and retrying changes nothing.

dsh-gpt-tool-compat exists to remove that friction. It registers a tools/pre-execute hook at the process-level root scope and does exactly one thing: before dispatch, it strips only the privilege parameters that are guaranteed to fail validation under any sandbox mode, restoring the call to the plain form the model actually intended. Calls that would have passed validation are left untouched; genuinely broader privilege requests (e.g. read-only → workspace-write with a valid justification) flow through and still go through real user approval. Any unexpected exception is caught and the original call is forwarded as-is—this layer never makes a situation worse.

If you run a DSH web profile, rely on pwsh, bash, or filesystem tool families daily, and keep getting blocked by third-party-routed model calls, dropping in this compatibility layer restores normal tool-call flow without compromising any real security approval mechanism.