feat: support live transcription replay

[jpr5]

Summary

Aligns OpenAI transcription HTTP, Realtime, recorder, and replay behavior with the current nested GA wire contract.

• Dedicated intent=transcription sockets ignore regular realtime query models.
• GA session events serialize audio.input.format and audio.input.transcription; Beta keeps its flat shim.
• Current input-transcription configuration emits progressive transcript events, retains languages, preserves empty transcript fixtures, honors stream timing and interruption controls, and ends at typed transcript.text.done without a [DONE] sentinel.

Baseline vs Candidate Local Proof

The exact probe ran against a clean detached 8d425a4 worktree and the isolated candidate worktree. It used real local HTTP, WebSocket, recorder proxy, persisted fixture, and replay paths.

Red: clean 8d425a4

INTENT_MODELS gpt-realtime-2 wrong-model
CURRENT_GA {"voice":null,"input_audio_format":"audio/pcm","output_audio_format":null,"input_audio_noise_reduction":null,"input_audio_transcription":{"model":"gpt-transcribe"}} session.created,session.updated,input_audio_buffer.committed
INVALID_MODEL Model gpt-realtime-2 does not support session type transcription
HTTP interrupt 200 "data: {\"type\":\"transcript.text.delta\",\"delta\":\"ab\"} ... data: {\"type\":\"transcript.text.done\",\"text\":\"abcdefgh\"}\n\nevent: done\ndata: [DONE]\n\n"
HTTP missing 404 "{\"error\":{\"message\":\"No fixture matched\",\"type\":\"invalid_request_error\",\"code\":\"no_fixture_match\"}}"
HTTP error 422 "{\"error\":{\"message\":\"bad transcription\",\"type\":\"invalid_request_error\",\"param\":null,\"code\":null}}"
RECORDED_FIXTURES ... "response":{"transcription":{"text":"hello"}}

Green: candidate 6ae6ec4

INTENT_MODELS gpt-transcribe gpt-transcribe
CURRENT_GA {"input":{"format":{"type":"audio/pcm"},"noise_reduction":null,"transcription":{"model":"gpt-transcribe"}},"output":{"format":null,"voice":null}} session.created,session.updated,input_audio_buffer.committed,conversation.item.added,conversation.item.input_audio_transcription.delta...,conversation.item.input_audio_transcription.completed
INVALID_MODEL Model wrong-model does not support session type transcription
HTTP interrupt FETCH_ERROR TypeError: terminated
HTTP missing 404 "{\"error\":{\"message\":\"No fixture matched\",\"type\":\"invalid_request_error\",\"code\":\"no_fixture_match\"}}"
HTTP error 422 "{\"error\":{\"message\":\"bad transcription\",\"type\":\"invalid_request_error\",\"param\":null,\"code\":null}}"
RECORDED_FIXTURES ... "response":{"transcription":{"text":"hello","languages":[{"code":"en"}]}}

Validation

prettier --check  PASS
eslint            PASS
tsc --noEmit      PASS
Focused protocol suites: 312 tests PASS
Full pnpm test and pnpm build completed after the GA conformance update with no failure output.

Fourth-Pass Red-Green

Red: terminal reported four stale GA expectations (legacy audio transcription field and missing ).

Green:
RUN v3.2.4 /private/tmp/aimock-live-transcribe-pr.0mGErr

✓ src/tests/ws-realtime.test.ts (83 tests) 7614ms

Test Files 1 passed (1)
Tests 83 passed (83)
Start at 14:39:20
Duration 8.13s (transform 235ms, setup 0ms, collect 300ms, tests 7.61s, environment 0ms, prepare 44ms) passed 83/83. Direct clean-baseline vs candidate VAD probe changed into nested , progressive live deltas, and journal interruption before completion. Full format, lint, typecheck, 145-file test suite, and build passed.

Fourth-pass proof: red terminal ws-realtime test had four stale GA audio expectation failures. Green exact file: 83 of 83 passed. Clean baseline VAD emitted an error; candidate returned nested audio.input.turn_detection server_vad, progressive deltas, and truncateAfterChunks journal interruption. Full format, lint, typecheck, 145-file tests, and build passed.

Fifth-pass red-green: baseline emitted GA session type conversation, top-level turn_detection, and accepted model mutation. Candidate emits realtime for normal GA sessions, keeps VAD under audio.input, rejects established model mutation, and preserves explicit transcription/translation session types. Red conformance test at ws-api-conformance line 618 is now green (66/66); full format, lint, tsc, tests, and build passed.

Sixth-Pass Red-Green

Red: clean baseline 2871866 accepted a real WebSocket mixed model/type update and persisted gpt-4o-transcribe transcription must-not-mutate.

Green: candidate returned error invalid_session_config; its next valid update reported gpt-realtime-mini realtime safe, proving rejection occurs before any session mutation. Live-transcription interruption now uses socket destruction and journals the interruption.

Validation: focused ws-realtime suite 83/83 passed. Full format check, lint, typecheck, test suite, and build passed locally.

P2 Coverage Hardening

Commit f2ffc7e adds tests only; it does not change runtime behavior. The clean 71452a7 baseline already demonstrated the documented intent=transcription strict no-match and ErrorResponse paths, versioned model matching, abrupt interrupted socket termination with journal details, and scheduled inter-frame delivery.

The new direct tests cover those established paths, HTTP stream=true JSON errors, and JSON/SSE recorder-to-local-replay preservation of languages and usage. Applied as a test-only diff to detached 71452a7, the focused suite passed 132/132; the candidate focused suite also passed 132/132. Full format, lint, typecheck, test, and build passed on the candidate.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@345

commit: 23aca19