Let a local model triage, escalate to a big model only when needed

A local model on a mid-range GPU is plenty for classification, summarising, and naming things. It is free, private, and never rate-limited. It is also not smart enough to refactor a codebase.

So don't ask it to. The n8n workflow looks like this:

1. Event arrives (email, RSS item, webhook).
2. Local model classifies it and writes a one-line summary.
3. If the class is "needs code" or "needs judgement", the item is forwarded to a Claude Code job with the summary as the prompt.
4. The result comes back as a pull request or a note.

Guardrails

• Every job has a dry-run flag and a dead-man timer.
• Escalations are logged with cost, so you can see what the cloud model is actually doing for you.

Cheap model first, expensive model on request, human on exception.