Inbox Triage Cron

Fetch unread emails from Gmail using the Gmail API.

GET https://gmail.googleapis.com/gmail/v1/users/me/messages?q=is:unread&maxResults=20

For each message, GET the full message to extract: sender, subject, date, snippet.
Classify each email:
- 🔴 Urgent: needs reply today (from known contacts, contains "urgent", "ASAP", time-sensitive)
- 🟡 Action Required: needs a response but not time-critical
- 🔵 FYI: newsletters, notifications, CC'd threads
- ⚪ Low Priority: automated receipts, marketing

Return: total unread count, then the top 5 by urgency with sender, subject, and classification.

Analyze the input and classify it into the defined categories.

Classification approach:
1. Extract key signals from the content (keywords, sender, urgency markers, topic)
2. Match against category definitions
3. Assign confidence score (high/medium/low)
4. For ambiguous cases, classify as the more important/urgent category (err on the side of caution)

Output for each item: { category, priority, confidence, reasoning }

If an item could belong to multiple categories, pick the primary one and note the secondary.

Generate a draft response based on the context.

Guidelines:
- Match the tone of the original message (formal → formal, casual → casual)
- Keep it concise — aim for 2-4 sentences for routine responses
- Include specific references to the original message
- End with a clear next step or call to action
- Don't make commitments or promises on behalf of the user

Mark the draft clearly: "[DRAFT — Review before sending]"
Include a confidence score: High (routine/clear intent), Medium (judgment call), Low (complex/sensitive).

For Low confidence drafts, explain what needs human judgment.

Label/tag emails by category in inbox

Steps:
1. Validate all required inputs are available
2. Execute the operation described above
3. Verify the result meets expected output format
4. Handle errors gracefully — retry transient failures, log and alert on persistent ones
5. Return structured output with status and any relevant data

If any required data is missing, request it from the user before proceeding.

Send a notification to the user via the configured messaging channel.

For Pushover:
POST https://api.pushover.net/1/messages.json
Body: { token: {APP_TOKEN}, user: {USER_KEY}, message: "{notification_text}", title: "{title}", priority: 0 }

For Twilio SMS:
POST https://api.twilio.com/2010-04-01/Accounts/{SID}/Messages
Body: To={phone}, From={twilio_number}, Body={message_text}

Keep the message concise — under 160 characters for SMS, under 500 for push.
Include the most actionable information first.
Log: timestamp, channel, recipient, message preview, delivery status.

Analyze the input and classify it into the defined categories.

Classification approach:
1. Extract key signals from the content (keywords, sender, urgency markers, topic)
2. Match against category definitions
3. Assign confidence score (high/medium/low)
4. For ambiguous cases, classify as the more important/urgent category (err on the side of caution)

Output for each item: { category, priority, confidence, reasoning }

If an item could belong to multiple categories, pick the primary one and note the secondary.