[davior]

Summary

GAM's half of speech to text (gecko-suite#6 is the suite's): SA-4 in the suite's plan. Depends on that PR being merged and deployed first. Merging this changes nothing until AI_BACKEND=suite, which is already how enrichment, generation and the assistant are switched.

With AI_BACKEND=suite, the audio GAM extracts goes to the suite's speech to text, which holds the Deepgram key.

• The transcript comes back in the shape Deepgram's client returns, so storing it (edited-segment protection, indexing, headers) is the code that already exists.
• GAM writes no usage row for it: the suite recorded the seconds under GAM and the asset, and a second row would count the transcription twice.
• Whether a key exists is the suite's answer. The upload chain and the URL import both asked "is there a Deepgram key?" of this app's own settings; can_auto_transcribe now asks the suite in suite mode, as GAM for the uploader, with a short timeout since it sits on the upload path. A suite that is down counts as "no key", so an upload never leaves a red row behind and a URL import falls back to the site's captions, exactly as it does today with no key. The library keeps working, and Transcribe can be pressed once the suite is back.
• A job's failures reach the activity feed as the suite's own sentences ("Add a Deepgram provider in Gecko Suite", "That audio is over 300 MB. Send it in parts", "can't be reached right now").
• Settings points at the suite's providers instead of asking for a Deepgram key that nothing would read. The model choice stays: it is sent with each transcription.
• requirements.txt pins gecko-suite-client to the suite PR's head commit (the one with SuiteSpeech). I'll re-pin to main once that merges.

Test plan

• cd backend && pytest -q: 1536 passed (8 new for suite mode, run with real ffmpeg: a transcription end to end with no Deepgram key here and no usage row, the suite's sentence reaching the job, a suite that is down, a recording too long, an upload queued or not by the suite's answer, an upload still succeeding when the suite is down, and local mode unchanged).
• cd frontend && npm run format:check && npm run lint && npm test && npm run build: 474 tests passed.
• CI green, including the Docker image build.
• A real transcription against a real suite and Deepgram. Not run here: everything is against fakes.

Deployment

Merging and redeploying changes nothing for users until AI_BACKEND=suite.

1. Before pulling: gecko-suite#6 must be merged and deployed, and the suite needs a Deepgram provider with a key (check Providers; the importer brought across any Deepgram key from GAM or Notes settings).
2. Configuration: nothing new. AI_BACKEND=suite and SUITE_APP_SECRET are already set if enrichment is on the suite.
3. Deploy: rebuilds the backend and the frontend; the server needs outbound access to github.com for the client:

   git pull origin main

   docker compose up --build -d

   docker compose ps

   curl -f http://localhost:${APP_PORT}/api/health

4. Migrations: none.
5. Other apps: nothing. (Video Creator's narration is text to speech, which is the next PR.)
6. Verify: with AI_BACKEND=suite, Settings shows "uses the Deepgram key kept in Gecko Suite". Upload a video: a transcription is queued by itself, finishes, and the transcript is searchable. The suite's Usage has an stt row under GAM with the asset as its subject (seconds, cost blank). GAM's own Usage panel will not show it, because the suite counts it.
7. Roll back: set AI_BACKEND=local and docker compose up -d; the Deepgram key in Settings is still there. A transcription running at that moment fails and can be re-run.

---

---

Generated by Claude Code