When a zstd-compressed request (e.g. from Codex CLI) hits /v1/responses with a cloud model the request failed. Fix by decompressing zstd bodies before model extraction, so cloud models are detected and proxied directly without the writer being wrapped.