[404oops]

Reserve a bounded transcription window before forwarding its configuration or audio. Retain immutable committed-item identities separately from bounded completed-item history. Pause audio while an exhausted window settles, then renew its reservation. Include pending transcriptions in disconnect and response-completion settlement decisions.

Record completion once. Failure, timeout, missing usage and upstream closure use a conservative fixed fallback and release credit after recording. Capacity overflow closes the connection without evicting older unresolved identities. Distinguish missing provider realtime models from explicitly unpriced ones, and refuse unpriced models before forwarding buffered frames.

Billing policy for review: enabling input transcription reserves 64 items at $0.10896 each, or $6.97344 before deployment cost factors. The window uses the largest supported transcription cost factor and retains unused capacity until consumption or connection closure. Failed or unreported items charge the fixed per-item budget. Reservations are budgets, not caps on reported usage. These choices can refuse low-balance accounts and can charge more than actual usage when final usage is unavailable. All limits and fallback behavior are documented.

Validation: the transcription layer passed 108 relay tests; the final combined relay passed 110 tests after the completed-history helper reuse and continuation patch. Mock-provider tests cover delayed transcription-only disconnects, response.done while transcription remains pending, duplicate/late events, failure/timeout/upstream close, shared-budget exclusion, window backpressure and renewal, explicit unknown/missing/dated models, and an expensive oldest identity followed by 64 more pending items. Tests assert exact charges and hold release. No paid provider calls were made; actual provider incidence and financial loss are not measured.

Final combined validation at 7adeb371e177a9532bac5d78a3f83318d1af4757: 625 backend tests, 122 SDK conversation/history/transport tests, 10 built-SDK mock-provider tests and 3 cross-platform API checks passed. Backend build, SDK declaration checks and lint passed; repository typecheck reports no new errors with 23 known baseline errors. No paid/live provider traffic or production rollout test was run.

Stack: follow-up 5/7, overall 11/13. Based on #4454 (fix/realtime-active-eligibility). Review and merge bottom-up. The existing six-PR stack ends at #4441.

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	95.07% ⬇️ -0.03 pp	40849 / 42966
🔵	Statements	93.24% ⬇️ -0.03 pp	44864 / 48116
🔵	Functions	92.78% ⬇️ -0.05 pp	7712 / 8312
🔵	Branches	84.72% ⬇️ -0.10 pp	31354 / 37006

File Coverage

File	Stmts	Branches	Functions	Lines	Uncovered Lines
Changed Files
src/backend/services/realtime/OpenAIRealtimeRelayService.ts	90% ⬆️ +1.13 pp	82.98% ⬆️ +1.12 pp	94.21% ⬆️ +2.46 pp	92.9% ⬆️ +0.93 pp	241, 263, 353, 364, 477, 512, 594, 628, 650-651, 653, 655-656, 673, 687, 693, 705, 904, 906, 976-989, 995, 1007, 1044, 1082, 1118-1119, 1150, 1168-1169, 1205, 1208, 1278-1279, 1294, 1375, 1393, 1485-1486, 1493, 1519, 1562-1563, 1586-1591, 1595, 1602-1606, 1643, 1670, 1717, 1728, 1751-1752, 1773, 1781-1782, 1786-1790, 1793, 1805-1806, 1824, 1831, 1844-1845, 1853-1855, 1861-1863, 1889, 1906-1917, 1922, 1954

Generated in workflow #3278 for commit d5b0970 by the Vitest Coverage Report Action