Meeting-Transcription Pipeline — EBO + Implementation + Live

🟢 built / wired / live-verified 🟡 residual — needs a real Meet or an owner account action SEAM testable via uploadAudio + sim_call_tts (no live Meet) RESIDUAL one golden-journey per infra change

Provider recommendation

One line, then the deciding factors for THIS codebase.

Recommendation Why (for this repo)

Fireflies HYBRID (not Fathom Business) Repo is 100% Fireflies-wired; native Hungarian AI summaries; client_reference_id deal-id round-trip already passed by ingest_call_audio; CI-able uploadAudio test seam. Fathom Business has no HU summaries (6 non-HU langs), no external-id passthrough, and no upload seam → a full client/handler/HMAC/desktop rewrite for a strictly worse fit. Fathom only wins on bot-free >3 coverage, which the bot relay handles.

Recommendation	Why (for this repo)
Fireflies HYBRID (not Fathom Business)	Repo is 100% Fireflies-wired; native Hungarian AI summaries; client_reference_id deal-id round-trip already passed by `ingest_call_audio`; CI-able `uploadAudio` test seam. Fathom Business has no HU summaries (6 non-HU langs), no external-id passthrough, and no upload seam → a full client/handler/HMAC/desktop rewrite for a strictly worse fit. Fathom only wins on bot-free >3 coverage, which the bot relay handles.

The pipeline (delivered)

Capture → webhook → match → store → AI → surface. The smart relay chooses SDK (≤3) vs bot (>3).

① Capture (smart relay)

≤3 → SDK· >3 → bot·upload ·paste

route_capture() + handle_participant_joined(store=); bot dispatch single-flighted via put_if_absent (NOT set(ttl=)).

② Webhook + match

/fireflies/transcription→ clientReferenceId→cal_id →email

NEW: clientReferenceId read FIRST, then the email→CRM→meeting_id chain.

③ Store + AI (HU)

deal store→ Gemini extract→HU summary

Notes-only — nothing auto-sent / auto-tasked to the lead (human-gate intact).

④ Surface

board pill+ Notion calls-DB row

Idempotent Notion row (dedup on Transcript ID), Lead→CRM relation, markdown transcript body.

EBO — positive scenarios 8

ID	Scenario	Verify layer	Behavior
S1	≤3 call → bot-free Meet SDK capture (common case)	🟡 RESIDUAL (capture) / 🟢 SEAM (rest)	Bot-free SDK records, webhook → match → store → HU summary → board pill → Notion row.
S2	>3 call → Notetaker BOT auto-invited (the relay)	🟢 BUILD/SEAM / 🟡 RESIDUAL (join)	route_capture → 'bot' → addToLiveMeeting once (single-flight). LIVE-verified: success:true.
S3	No-bot audio upload (record-anything → public URL)	🟢 LIVE/SEAM	ingest_call_audio → uploadAudio with client_reference_id = deal_id. Already wired.
S4	Manual paste (bridge for any transcriber)	🟢 LIVE + BUILD (Notion row)	process_pasted_transcript → extract + HU summary + NEW Notion calls-row (Source=Upload).
S5	Idempotent webhook (duplicate delivery)	🟢 SEAM	Second delivery is a no-op; Notion row deduped on Transcript ID (query-before-create).
S6	HU AI summary as CRM NOTES only (human-gate)	🟢 LIVE/BUILD	Summary + items are notes; nothing auto-sent to the lead, no auto-task. Test asserts it.
S7	Board card pill reflects state	🟢 LIVE	📞 Sales call megvolt · N perc + Transcript: betöltés…/megvan/hiányzik. Pre-built.
S8	Notion calls-DB row complete + Lead-linked	🟢 BUILD	Name/Lead→CRM/Recording/Date/Duration/Source/Transcript ID/status/AI summary + markdown body.

Scenario

Verify layer

Behavior

≤3 call → bot-free Meet SDK capture (common case)

🟡 RESIDUAL (capture) / 🟢 SEAM (rest)

Bot-free SDK records, webhook → match → store → HU summary → board pill → Notion row.

>3 call → Notetaker BOT auto-invited (the relay)

🟢 BUILD/SEAM / 🟡 RESIDUAL (join)

route_capture → 'bot' → addToLiveMeeting once (single-flight). LIVE-verified: success:true.

No-bot audio upload (record-anything → public URL)

🟢 LIVE/SEAM

ingest_call_audio → uploadAudio with client_reference_id = deal_id. Already wired.

Manual paste (bridge for any transcriber)

🟢 LIVE + BUILD (Notion row)

process_pasted_transcript → extract + HU summary + NEW Notion calls-row (Source=Upload).

Idempotent webhook (duplicate delivery)

🟢 SEAM

Second delivery is a no-op; Notion row deduped on Transcript ID (query-before-create).

HU AI summary as CRM NOTES only (human-gate)

🟢 LIVE/BUILD

Summary + items are notes; nothing auto-sent to the lead, no auto-task. Test asserts it.

Board card pill reflects state

🟢 LIVE

📞 Sales call megvolt · N perc + Transcript: betöltés…/megvan/hiányzik. Pre-built.

Notion calls-DB row complete + Lead-linked

🟢 BUILD

Name/Lead→CRM/Recording/Date/Duration/Source/Transcript ID/status/AI summary + markdown body.

EBO — negative / anti-fragility 9

ID	Scenario	Verify layer	Behavior
N1	No CRM match, real external attendee	🟢 LIVE/SEAM	ffseen marked + Slack #ai-new-lead flag; no silent drop, no auto-link.
N2	Webhook before transcript ready (early)	🟢 LIVE/SEAM	get_transcript raises → not marked done → 10-min backup poll retries.
N3	Transcription never arrives (missing)	🟢 LIVE/BUILD	Pill stays 'hiányzik'; Notion status not flipped to done; nothing fabricated.
N4	>3 escalation when SDK would truncate	🟢 BUILD/SEAM / 🟡 RESIDUAL	4th joiner re-fires the event → route 'bot', single-flight; ≤3 SDK is the floor.
N5	Fireflies API/key failure	🟢 LIVE/SEAM	Structured {ok:false,error}; no unhandled raise; backup poll + paste remain.
N6	Duplicate bot dispatch / rate limit	🟢 BUILD/SEAM	put_if_absent('meet_bot_sent:{url}') → exactly one dispatch. LIVE single-flight verified.
N7	Consent: NO consent UX (Q5)	🟢 DESIGN	Zero consent code added — explicit non-requirement (spoken consent verbally).
N8	Account-level Fireflies webhook missing	🟢 LIVE	Per-upload webhook + 10-min poll cover it; owner registers account webhook (checklist).
N9	SA lacks Meet/Workspace-Events scopes	🟡 DESIGN/RESIDUAL	SDK + participant.joined impossible until owner adds scopes; build FAILS CLOSED, no crash.

Scenario

Verify layer

Behavior

No CRM match, real external attendee

🟢 LIVE/SEAM

ffseen marked + Slack #ai-new-lead flag; no silent drop, no auto-link.

Webhook before transcript ready (early)

🟢 LIVE/SEAM

get_transcript raises → not marked done → 10-min backup poll retries.

Transcription never arrives (missing)

🟢 LIVE/BUILD

Pill stays 'hiányzik'; Notion status not flipped to done; nothing fabricated.

>3 escalation when SDK would truncate

🟢 BUILD/SEAM / 🟡 RESIDUAL

4th joiner re-fires the event → route 'bot', single-flight; ≤3 SDK is the floor.

Fireflies API/key failure

🟢 LIVE/SEAM

Structured {ok:false,error}; no unhandled raise; backup poll + paste remain.

Duplicate bot dispatch / rate limit

🟢 BUILD/SEAM

put_if_absent('meet_bot_sent:{url}') → exactly one dispatch. LIVE single-flight verified.

Consent: NO consent UX (Q5)

🟢 DESIGN

Zero consent code added — explicit non-requirement (spoken consent verbally).

Account-level Fireflies webhook missing

🟢 LIVE

Per-upload webhook + 10-min poll cover it; owner registers account webhook (checklist).

SA lacks Meet/Workspace-Events scopes

🟡 DESIGN/RESIDUAL

SDK + participant.joined impossible until owner adds scopes; build FAILS CLOSED, no crash.

What was built (additive)

Change File EBO

route_capture() smart relay + handle_participant_joined(store=) single-flight app/transcription.py S1/S2/N4/N6

handle_meet_participant_joined() + POST /meet/participant-joined (dormant) app/flows_calls.py · app/modal_app.py S1/S2/N9

clientReferenceId-first deal match app/flows_inbound.py §4 #1 / S3

write_call_record() idempotent Notion calls-DB writer, wired into both handlers app/flows_calls.py S8/S5/N3

get_transcript requests transcript_url/audio_url app/clients/fireflies.py S8 (Recording)

NOTION_DB['call_recordings'] wired to the live DB app/config.py S8

Landmine fix: store().set(ttl=) → put_if_absent TRANSCRIPTION_WIRING.md S5/N6

Live bug fixed: AddToLiveMeeting selection = success only app/transcription.py S2 (live)

Change	File	EBO
`route_capture()` smart relay + `handle_participant_joined(store=)` single-flight	app/transcription.py	S1/S2/N4/N6
`handle_meet_participant_joined()` + `POST /meet/participant-joined` (dormant)	app/flows_calls.py · app/modal_app.py	S1/S2/N9
`clientReferenceId`-first deal match	app/flows_inbound.py	§4 #1 / S3
`write_call_record()` idempotent Notion calls-DB writer, wired into both handlers	app/flows_calls.py	S8/S5/N3
`get_transcript` requests `transcript_url`/`audio_url`	app/clients/fireflies.py	S8 (Recording)
`NOTION_DB['call_recordings']` wired to the live DB	app/config.py	S8
Landmine fix: `store().set(ttl=)` → `put_if_absent`	TRANSCRIPTION_WIRING.md	S5/N6
Live bug fixed: AddToLiveMeeting selection = `success` only	app/transcription.py	S2 (live)

TDD: +27 tests (test_capture_router, test_call_record_notion, test_meet_participant_route) red→green; legacy test_transcription_s6 updated to the HYBRID contract. Suite 558 green.

Live status — what you can test NOW

Now (live on v102) Blocked on an owner account action

Dash + full funnel render (200, badge c88ec77-dirty+v102) ≤3 bot-free SDK capture — Workspace admin approval + Fireflies SDK plan

Upload route (S3) + manual paste (S4) → Notion calls row Account-level Fireflies webhook (so bot/SDK transcripts fire)

/meet/participant-joined: ≤3→sdk, >3→bot dispatch (success:true) Workspace-Events participant.v2.joined subscription (for live trigger)

Duplicate-webhook idempotency + Slack no-match flag One real ≤3 + one real >3 golden-journey Meet (after SDK approval)

Now (live on v102)	Blocked on an owner account action
Dash + full funnel render (200, badge `c88ec77-dirty+v102`)	≤3 bot-free SDK capture — Workspace admin approval + Fireflies SDK plan
Upload route (S3) + manual paste (S4) → Notion calls row	Account-level Fireflies webhook (so bot/SDK transcripts fire)
`/meet/participant-joined`: ≤3→sdk, >3→bot dispatch (success:true)	Workspace-Events `participant.v2.joined` subscription (for live trigger)
Duplicate-webhook idempotency + Slack no-match flag	One real ≤3 + one real >3 golden-journey Meet (after SDK approval)

Owner checklist (ordered) — the remaining prerequisites

# Action Unlocks

1 Confirm FIREFLIES_API_KEY valid (Pro for the floor; Business for the bot-free SDK) All routes

2 Register the account-level Fireflies webhook → <PUBLIC_BASE_URL>/fireflies/transcription + set the x-hub-signature shared secret Bot/SDK transcripts fire (N8)

3 Confirm PUBLIC_BASE_URL in the Modal env (matyas--clientsflow-pipeline-web.modal.run) Per-upload webhook + form links

4 Google Workspace admin approval for the Fireflies Meet SDK + enable the integration ≤3 bot-free capture (S1)

5 (For the live >3 trigger) Workspace-Events sub for meet.participant.v2.joined → /meet/participant-joined (Meet scopes on DWD + Pub/Sub topic + publisher role + renewal cron). Lighter: Fireflies calendar auto-join with „wait for ≥2". Automatic >3 relay (S2)

6 Sign the Fireflies DPA (GDPR; EU Private Storage = Enterprise). Consent stays verbal — no code. Compliance

#	Action	Unlocks
1	Confirm `FIREFLIES_API_KEY` valid (Pro for the floor; Business for the bot-free SDK)	All routes
2	Register the account-level Fireflies webhook → `<PUBLIC_BASE_URL>/fireflies/transcription` + set the `x-hub-signature` shared secret	Bot/SDK transcripts fire (N8)
3	Confirm `PUBLIC_BASE_URL` in the Modal env (`matyas--clientsflow-pipeline-web.modal.run`)	Per-upload webhook + form links
4	Google Workspace admin approval for the Fireflies Meet SDK + enable the integration	≤3 bot-free capture (S1)
5	(For the live >3 trigger) Workspace-Events sub for `meet.participant.v2.joined` → `/meet/participant-joined` (Meet scopes on DWD + Pub/Sub topic + publisher role + renewal cron). Lighter: Fireflies calendar auto-join with „wait for ≥2".	Automatic >3 relay (S2)
6	Sign the Fireflies DPA (GDPR; EU Private Storage = Enterprise). Consent stays verbal — no code.	Compliance

Artifacts

EBO: plans/fireflies-ebo/EBO.md · Implementation: plans/fireflies-ebo/IMPLEMENTATION.md · Wiring + landmine fix: TRANSCRIPTION_WIRING.md · Notion calls DB: „Call Recordings & Transcripts" (92cae219-a975-4a34-ae2e-a8981982a22c) · Live dash: https://matyas--clientsflow-pipeline-web.modal.run/dash · git: merged to local main, NOT pushed.