Connect a zero-cost mechanical layer to a duoduo brain job so the LLM only runs when there is something worth acting on. During idle periods the brain job stays dormant — no polling, no token spend.
The three-layer model
Mechanical layer (shell / Python script — zero LLM)
│
│ duoduo session notify ← fires only when new data arrives
▼
Brain job (keepalive — dormant by default)
│
│ Notify ← fires only when signal meets your threshold
▼
Main session (your foreground chat)
- Mechanical layer — a plain script scheduled by launchd, cron, or a
systemd timer. It polls the source, compares against last-seen state, and
calls
duoduo session notifyonly when something new is detected. No LLM involved at any point. - Brain job — a
keepaliveduoduo job. It wakes on notify, processes the incoming data, decides whether the result is worth surfacing, and sends a Notify to your main session only when the signal clears your threshold. Then it goes dormant again. - Main session — your regular foreground chat. It receives a single, pre-filtered brief only when the brain has found something actionable.
Typical use cases
- Feed / account monitoring — check a public feed every few minutes; wake the brain only on new posts.
- Price or metric alerts — poll an API for a number; wake the brain only when it crosses a threshold.
- File or repo change detection — watch a directory or git remote; wake the brain only on new commits or file changes.
- Webhook relay — a lightweight receiver script forwards payloads to the brain via notify, keeping the brain dormant between events.
- Data-source health watch — detect when an upstream source goes stale and surface an alert without running a full LLM cycle just to check.
References
- Pattern: mechanical layer + brain job — how to build the two-layer setup end-to-end.
- Notify routing and signal filtering — how the brain decides what to surface, how to format briefs, and how to route them without flooding your main session.