QVAC-18410 test[skiplog]: add e2e coverage for transcribe metadata flag (PR-1701) by Victor-Rodzko · Pull Request #1883 · tetherto/qvac

Victor-Rodzko · 2026-05-04T14:23:13Z

Note: be concise and prefer bullet points.

🎯 What problem does this PR solve?

PR QVAC-17184 feat[api]: propagate whisper per-segment metadata to SDK users #1701 (QVAC-17184) added an opt-in metadata: true flag to transcribe() / transcribeStream() that returns TranscribeSegment[] instead of the joined transcript. The new public boundary (shape + emission order + engine routing + duplex code path) had no e2e coverage.

📝 How does it solve it?

New transcription tests in transcription-tests.ts:
- transcription-metadata-batch — transcribe({ metadata: true }) returns TranscribeSegment[].
- transcription-metadata-streaming — drives the live bidirectional transcribeStream({ modelId, metadata: true }) session: decodes a 16kHz mono WAV fixture to f32le, pumps it through session.write() in chunks plus a trailing silence pad, then iterates yielded segments. Exercises transcribeStreamDuplexMetadata / processLineMetadata — not the deprecated upfront-audio overload.
New sad-path in parakeet-tests.ts:
- parakeet-tdt-metadata-rejected — metadata: true is whisper-only; parakeet engine must throw. Pins errorContains: "does not support metadata" so unrelated failures (model load, audio I/O) can't masquerade as a successful rejection.
Shared helpers in tests/shared/transcription-segments.ts:
- validateSegments() — asserts shape { text, startMs, endMs, append, id } per segment, plus non-decreasing startMs and id so emission stays in audio-time order (append: true revisions may repeat a startMs). Content-agnostic — transcript correctness is already covered by existing transcription tests.
- runMetadataStreamDuplex() — shared driver for the live duplex metadata flow (WAV decode → f32le → chunked write() → end() → collect → validate). Used by both desktop and mobile executors.
Desktop + mobile (Bare/RN) executor parity for all three flows.

🧪 How was it tested?

Desktop (macOS, Bare): npx qvac-test run:local:desktop --filter "transcription-metadata-|parakeet-tdt-metadata-rejected" → 3/3 ✅.
iOS (physical iPhone): same filter → 3/3 ✅.

github-actions · 2026-05-12T15:12:40Z

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*

opaninakuffo · 2026-05-12T16:48:15Z

clean PR. tight scope: covers the three new code paths (transcribe({ metadata: true }), transcribeStream({ metadata: true }) duplex, parakeet rejection), and the parakeet error pin ("does not support metadata") is anchored to a real, engine-specific substring in parakeet-transcription/plugin.ts so it can't accidentally pass on a different failure. validators stay content-agnostic — nice. desktop+mobile parity maintained.
one minor non blocking thing below.

…t metadata (PR-1701) - transcription-metadata-batch / transcription-metadata-streaming: validate TranscribeSegment[] shape returned by Whisper when metadata: true - parakeet-tdt-metadata-rejected: assert metadata flag is rejected by non-Whisper engines (Parakeet) - shared validateSegments() helper, content-agnostic shape check - desktop + mobile executor parity for transcription and parakeet flows Also folds in a tooling fix surfaced while authoring these tests: tests-qvac/package.json gains clean:sdk-snapshot which wipes node_modules/@qvac/sdk plus the iOS/Android consumer build snapshots before reinstalling, so install:build:full no longer reuses a stale @qvac/sdk copy on any platform. Co-authored-by: Cursor <cursoragent@cursor.com>

…on reason

github-actions · 2026-05-13T14:19:53Z