Current Agent Loop

server.mjs wires everything together at startup. No decisions — pure plumbing.

• Launches Messages.app with the injected dylib so it can fire typing dots / read receipts / sends that AppleScript can't reach.
• Starts the Drafts UI on :7878 — the cockpit Andrew works out of.
• Starts both pollers chat‑poller (inbound iMessage) and issue‑poller (new WOs), plus the scheduled sender and the AppFolio Playwright runner on :9773.

triggers/issue-poller.mjs + triggers/enrich-issue.mjs. Polls Supabase issues_v2 every 5s. Sets the WO up before the LLM ever sees it.

• Detects the new WO across mapped workspaces (prod/test); dedups on a cursor so a crash never double‑fires.
• Enriches deterministically AppFolio reports API → real unit + clean description; a mini‑LLM (gpt‑5.4‑mini) writes a 3–7 word title and makes the urgency call (PMS flag is just one input).
• Pulls the vendor roster ~36 vendors as the inline candidate list (also the name→id map).

core/orchestrator.mjs running the process_work_order skill, new_issue phase. One LLM turn (gpt‑5.4) on the enriched issue.

• read_memory surfaces vendor beliefs + per‑property quirks for the trade.
• set_vendor writes the pick back to issues_v2 (powers the dashboard + downstream vendor draft).
• send_text produces the 2‑ or 4‑line summary — "Unit 1 829 Ocean Park / Has a leaky faucet. / Should I send Yonic?" — but in sendMode:'draft', so it lands as a draft, not a text.
• Off‑hours hold if outside Mon–Fri 7am–7pm, the draft gets hold_until = next window open (even urgent ones — Andrew can override).

triggers/chat-poller.mjs polls chat.db every 1s; runs the incoming_user_message phase. Key fact: this whole path is sendMode:'draft' — even a one‑word "got it" ack is staged, never auto‑texted.

• Read receipt fires instantly the moment the message is observed, before any LLM work.
• Decides intent dispatch approval / vendor swap / clarifying question / status‑update (silent) / learning‑only — gated by how many open WOs are in the recent‑sends list.
• Drafts the dispatch on an approval: ack + draft_tenant + draft_vendor → three draft rows, all for human review.
• Learns write_memory records the decision (see stage 7). Independent of dispatch.

Runs fire‑and‑forget off the same PM‑reply turn. Closes the loop back into stage 3's read_memory.

• write_memory → observation every PM decision (confirm, override, preference, reason) is logged as cheap evidence.
• belief‑former consolidates async per observation — promotes repeated signals into durable beliefs with confidence.
• sessionizer persists the transcript and groups it into conversational sessions via an LLM boundary judge.
• Optional override Andrew can hand‑edit or delete beliefs in the Memory tab — not required for the loop to run.