Distillation

Date: Sep 16, 2026

• 🔴 [enforced-stream-behavior] (17:20) User stated that a request already beyond the chain token budget never dispatches recall, as documented near packages/gateway/test/openai-responses-recall-aware-stream.test.ts:4910 in test rejects an over-budget principal response before its first recall.
• 🔴 [enforced-validation-invariant] (17:21) User stated that sparse numeric identities can never need more address space than the configured production ceiling, and that the frame override is capped at that ceiling so arithmetic remains practical.
• 🔴 [enforced-validation-invariant] (17:21) User stated that index + 1 always changes the value.
• 🟡 (17:21) packages/gateway/src/stream/openai-responses.ts:1883 documents that a callback is invoked only after an event has passed strict validation and mutation; nearby stream state includes activeContentIndexByItem = new Map<number, number>(), production-ceiling frame capping, event parsing and validation, applyResponsesEvent(state, event, parsed), terminal handling for response.failed, and special handling for terminal snapshots whose type === "item_reference".
• 🟡 (17:22) Search of packages/gateway/src/stream/openai-responses.ts found two identity-to-index mappings: line 1037 sets state.itemIndexById.set(addedItem.id, outputIndex as number), and line 1040 sets state.callIndexById.set(addedCallID, outputIndex as number).
• 🟡 (17:22) Search of packages/gateway/test/openai-responses-recall-aware-stream.test.ts found 19 item_reference occurrences at lines 1792, 1986, 1993, 2002, 2028, 2419, 2807, 3800, 3804, 4208, 4212, 4609, 4613, 4647, 4654, 4977, 4981, 5956, and 5960. Referenced IDs include msg_server_only, msg_reference_incomplete, fc_data_only_followup, msg_server_followup, call_shared_reference, call_principal_recall, and ref_retry_baseline.