Auto-triage Salesforge Primebox replies with labels and Slack alerts

Build me an agent that triages every new reply landing in my Salesforge Primebox, not just the positive ones, and keeps my inbox self-organized.

Trigger: run on a cron every 2 hours during business hours (default 9am to 6pm my local timezone, weekdays only). The schedule and timezone should both be easy to change.

On each run:

1. Call Salesforge List Threads to get threads updated since the previous run (use a rolling watermark, default lookback 2 hours plus a small overlap). For each thread call Salesforge Get Thread to read the latest inbound message and any prior context in that thread.

2. Classify each reply into exactly one of these buckets: interested or meeting requested, out of office (parse a return date if one is present), referral to a colleague (capture the referred person's name and email if visible), not interested, opt-out or spam complaint, or genuine question needing a human. If the model is not confident, leave the reply in an "uncertain" state instead of guessing.

3. Apply the matching Primebox label using Salesforge Update Thread Label. The label ID for each bucket should be configured up front (ask me for the label names or IDs during setup and store them in the workflow config). Uncertain replies stay unlabeled.

4. For opt-out and spam-complaint replies, push the sender email and its domain to my workspace suppression list via Salesforge Bulk Add Do-Not-Contact Entries so they can never be emailed again from any sequence in the workspace.

5. For referrals, append a row to a Google Sheets "Follow-up queue" tab with the referred person's name and email, the original prospect, the referring thread link, and the date. For out-of-office replies with a parsed return date, append a row with the contact, the parsed return date, and the thread link. Use Google Sheets Append Values. The spreadsheet ID and tab name are configuration.

6. Post a compact Slack digest to a channel of my choosing using the Slack Bot Send a Message action. The digest should surface only the actionable stuff, in this order: any opt-out or spam complaints called out loudly at the top of the message, referrals to chase today, out-of-office return dates hitting within the next 7 days, any uncertain replies that need a human look, and finally a one-line summary count of each bucket so I can see the reply mix at a glance. Skip the digest entirely if nothing actionable happened this run.

Idempotency: before labeling or logging a thread, check whether it already carries one of my agent's labels from a previous run and skip it if so. Never label a thread twice and never double-log the same thread to the sheet or the do-not-contact list.

Config I should be able to edit without touching code: cron schedule and timezone, business-hours window, the mapping from bucket name to Primebox label ID, the Google Sheets spreadsheet ID and tab name, the Slack channel ID, and the do-not-contact list scope (workspace by default).