Distillation

Date: Sep 16, 2026

• 🔴 (10:59) User supplied a test-file excerpt covering approximately lines 137–1108 for Responses recall-aware streaming. The code includes SSE construction for response.completed; assistant message output items; terminal-event/status parameterization; replacement of principal provider diagnostics with a fixed public failure envelope; provider diagnostic aliases "error" and "response.error"; sequence assertions such as [0, 1]; completion callbacks receiving { response, successful }; recall suppression via callbacks that throw "should not run" or "should not be called"; malformed or incomplete tool-call output; and sanitization of a private value "private-incomplete-reference".
• 🔴 (11:00) User supplied the continuation of the recall-aware Responses streaming tests, approximately lines 1089–1670. It covers sparse/atomic tool-call behavior for response ID "resp_atomic_sparse" with recalls expected to remain 0; principal failure using created("resp_principal_failure", "gpt-5.6-terra"); mixed output containing assistant messages and function_call items such as call_id: "call_1"; preservation/inspection of generated message IDs including "msg_resp_mixed_0"; mixed-tool handling where runFollowUp must not execute; a read call identified by call_id: "call_read"; function-call argument events for item_id: "fc_args"; and error-path tests involving streamResponsesRecallAware, retryPrincipal, onComplete, logger warn/error, drain(client), and failures before follow-up.