Claude Agent SDK
Fixture: transcription-sentinel-01.mp3
Expected sentinel: violet harbor quartz
Assistant output: Violet Harbor Quartz
Endpoint hit · exact usage +$0.000072
Tegy now exposes a narrow, read-only TranscribeAudio capability to both supported agent runtimes. It accepts only audio already mounted in the active chat and reports the provider generation, duration, and exact billed cost.
transcription-sentinel-02.mp3/audio/transcriptions returned HTTP 200 with generation gen-stt-1784290136-jhbLUstXR1DgPofbD2P6.2.784 / 60 × $0.0015.openrouter_exact_response row for the same generation. The shared usage ledger stores money to six decimal places, so the persisted value is $0.000070.Fixture: transcription-sentinel-01.mp3
Expected sentinel: violet harbor quartz
Assistant output: Violet Harbor Quartz
Endpoint hit · exact usage +$0.000072
Fixture: transcription-sentinel-01.mp3
Expected sentinel: violet harbor quartz
Assistant output contained: Violet Harbor Quartz
Endpoint hit · exact usage +$0.000072
Each runtime received a randomly selected preset speech fixture, was asked to decode its sentinel, invoked the hosted /audio/transcriptions path, and returned the correct words. The endpoint event and usage event identify the same OpenRouter generation. For this 2.88-second fixture, both recorded costs equal 2.88 / 60 × $0.0015 = $0.000072.
/mnt/user-data/uploads root.provider_final event records the provider’s exact cost, audio duration, model, format, and generation ID.npm run test:audio-transcription — protocol, authorization, filesystem isolation, endpoint telemetry, and exact-usage tests.npm run verify:audio-transcription-live — real provider request and complete Tegy turns in Claude and Codex.npm run lint, npx tsc -b, and git diff --check.Production verification used the authenticated Chromium Profile 3 / tegy session. D1 was queried read-only for only the test chat’s endpoint and usage records; no user identity or secret data is included here.