dsh-gpt-compat

When driving a DeepSeek Harness (DSH) profile with GPT or Codex-class models, a subtle but persistent issue often shows up: these models tend to attach sandbox escalation parameters (sandbox_permissions and justification) even during routine write operations. To the DSH sandbox engine each of those fields reads as a privilege-escalation request, so the same operation gets denied over and over and tool calls fall into a retry loop.

dsh-gpt-compat addresses exactly that gap with a precise, fail-closed approach. It strips the two escalation fields only when it can prove the request is redundant, meaning the requested permissions are no wider than the current session policy, so the tool simply executes under the existing policy. Genuine escalation requests that carry a non-empty justification are left untouched and continue through DSH's native approval flow. Isolated fields, empty-justification escalations, unknown patterns, and policy-parse failures are all preserved for the DSH core to validate or reject. Configuration is minimal: declare the provider list and the tool-name list you want covered.

If your DSH profile connects to GPT, Codex, or similar providers that habitually emit escalation parameters, and you are repeatedly hitting sandbox rejections caused by those redundant fields, this plugin removes the "looks-like-escalation-is-actually-redundant" noise while keeping the security boundary intact.