Distillation

Date: Sep 16, 2026

• 🔴 (19:38) User specified that accumulated text should be replaced with the final version because the final version is more reliable.
• 🟡 (19:39) Inspection of packages/gateway/src/stream/openai-responses.ts:676-683 confirmed that final output-text handling replaces previously accumulated delta text with the final text, documented by the comment Replace accumulated text with the final version (more reliable).
• 🟡 (19:39) packages/gateway/src/stream/openai-responses.ts:686-727 contains handling for output-text deltas and response.function_call_arguments.delta/final argument events, updating accumulated message or tool_use state by output_index.
• 🟡 (19:39) Terminal-event processing in packages/gateway/src/stream/openai-responses.ts:731-811 uses response.completed, response.incomplete, and response.failed; response status drives the stop reason, including content_filter, and usage processing includes prompt-detail fields such as cache_write_tokens.
• 🟡 (19:39) Terminal reconstruction in packages/gateway/src/stream/openai-responses.ts:820-900 reconstructs assistant text and tool-use output, sets completed statuses, derives tool input, changes the stop reason when tool_use was observed, and returns rawOutputItems from state.rawItems.
• 🟡 (19:39) Responses event validation in packages/gateway/src/stream/openai-responses.ts:911-1123 checks sparse-index bounds, sequence_number, terminal response shape, indexed text/function-call events, response.output_item.added, response.output_item.done, and content-part events; malformed terminal input throws Error("malformed Responses terminal event").
• 🟡 (19:40) Identity and semantic-content helpers identified in packages/gateway/src/stream/openai-responses.ts include reconcileCodexIdentity() at line 1147, responsesEventHasSemanticContent() at line 1184, identity-index binding around lines 1246-1258, effective tool identity handling around lines 1283-1296, Codex item binding around lines 1500-1524, and sparse candidate resolution around lines 1549-1614.
• 🟡 (19:40) Projection/normalization logic in packages/gateway/src/stream/openai-responses.ts:1309-1398 validates supported output-item types and statuses and recognizes message content parts such as output_text with string text.
• 🟡 (19:40) Deferred/continuation parsing in packages/gateway/src/stream/openai-responses.ts:1626-1675 validates optional output_index against maxSparseIndex, checks optional sequence_number, and handles principal lifecycle events including response.in_progress.