ADR-0055: Granola Meeting Intelligence Pipeline

Context

Joel uses Granola for meeting transcription and notes. Granola exposes an MCP server (https://mcp.granola.ai/mcp) with tools for listing meetings, querying notes, and pulling transcripts. A granola-cli (v0.1.0) wraps this via mcporter for agent-friendly HATEOAS JSON access.

Today, meeting notes sit passive in Granola — Joel has to manually extract action items, decisions, and follow-ups. The agent system should actively monitor for new meetings, analyze transcripts, and surface actionable items into Todoist (ADR-0045) without Joel lifting a finger.

Why Now

Granola MCP auth is working (OAuth via mcporter, enterprise enabled)
granola-cli is deployed and tested against live data (5 meetings, search works)
Todoist integration exists (ADR-0045, ADR-0047 async conversation channel)
Heartbeat cron runs every 15 minutes — natural polling interval

Requirements

Detect new/updated meetings since last check
Pull transcripts and summaries
Extract action items, decisions, follow-ups via LLM analysis
Create/update Todoist tasks with meeting context and links
Notify Joel via gateway when significant items are found
Don’t re-process meetings already analyzed

Decision

Add a Granola check step to the existing heartbeat cron that polls for new meetings, then fans out to a dedicated meeting/analyze Inngest function for transcript analysis and task creation.

Architecture

heartbeat (every 15min)
  └─ step: granola-check
       ├─ granola meetings --range this_week
       ├─ compare against Redis set `granola:processed`
       └─ for each new meeting → emit meeting/noted event
 
meeting/noted (Inngest function)
  ├─ step: pull-details    → granola meeting <id>
  ├─ step: pull-transcript → granola meeting <id> --transcript
  ├─ step: analyze         → LLM extracts action items, decisions, follow-ups
  ├─ step: create-tasks    → task port (ADR-0045) add (with meeting link, context)
  ├─ step: mark-processed  → Redis SADD granola:processed <id>
  └─ step: notify-gateway  → pushGatewayEvent with summary

State Management

Redis set granola:processed: meeting IDs already analyzed. TTL 90 days.
Redis hash granola:last_check: timestamp of last successful poll. Heartbeat uses this for --updated-after on next check.

Analysis Prompt

The LLM analysis step extracts structured data from the transcript:

{
  "action_items": [
    { "task": "Run Amex statements through AI analysis", "owner": "Joel", "deadline": "today", "context": "..." }
  ],
  "decisions": [
    { "decision": "Merge Egghead into Code TV", "rationale": "...", "meeting_link": "..." }
  ],
  "follow_ups": [
    { "item": "Monthly check-ins for expense monitoring", "frequency": "monthly" }
  ]
}

Task Creation (via ADR-0045 Port)

Tasks are created through the task management port (ADR-0045), not a specific adapter. Currently Todoist is the active adapter, but the pipeline is adapter-agnostic.

Title: [Granola] Run Amex statements through AI analysis
Description:
  From: "Egghead business wind-down" (Feb 5, 2026)
  Owner: Joel
  Link: https://notes.granola.ai/d/<id>
  Context: Accountant requested expense analysis for tax filing
Destination: Inbox (agent captures, Joel triages)

Action items from meetings automatically appear in Todoist within 15 minutes
Decisions are captured as structured records, not buried in transcripts
Meeting intelligence is searchable via Todoist and Qdrant (if observations are emitted)

Negative

mcporter OAuth tokens expire — needs refresh handling or re-auth flow
LLM analysis costs per meeting (mitigated: only new meetings, ~5/week)
Granola MCP may change — adapter boundary absorbs this

Follow-up Tasks

Add granola-check step to heartbeat function
Create meeting/noted Inngest function
Add granola:processed Redis set management
Handle mcporter OAuth token refresh in CLI
Add Granola connection health to system-check skill

Implementation Plan

Affected Paths

packages/system-bus/src/inngest/functions/heartbeat.ts — add granola-check step
packages/system-bus/src/inngest/functions/meeting-analyze.ts — new function
packages/system-bus/src/inngest/functions/index.ts — register new function
packages/system-bus/src/inngest/serve.ts — add to function array

Pattern

Shell out to granola CLI (compiled binary at ~/.local/bin/granola)
Parse HATEOAS JSON response
Fan-out via Inngest events (one per new meeting)
Task creation via task management port (ADR-0045) — adapter-agnostic
Gateway notifications via pushGatewayEvent (ADR-0018)

Verification

granola meetings returns data in heartbeat step
New meetings emit meeting/noted events
Previously processed meetings are skipped (Redis SISMEMBER)
Tasks created via port with meeting link and context
Gateway receives notification with meeting summary

Granola Meeting Intelligence Pipeline