pi-auto-router

A pi extension that automatically routes prompts to the right model.

You type a prompt. Haiku classifies it (~300ms, fractions of a cent). The request goes to Sonnet or Opus. Simple tasks stay cheap. Complex tasks get the power they need. You never think about model selection.

flowchart LR
  A[You type a prompt] --> B[Haiku classifies]
  B -->|simple / moderate| C[Sonnet 4.6]
  B -->|complex / deep reasoning| D[Opus 4.6]
  C --> E[Response]
  D --> E

Loading

Install

pi install npm:pi-auto-router

Or try it without installing:

pi -e npm:pi-auto-router

How it works

The extension hooks into before_agent_start. Before each LLM call:

1. Sends your prompt to Haiku with a classification instruction
2. Haiku responds with one word: sonnet or opus
3. Extension calls pi.setModel() to switch
4. Status bar shows → sonnet or → opus

Manual override: Switch models with Ctrl+P or Ctrl+L and the router respects your choice for that turn, then resumes auto-routing.

Routing rules

Haiku routes to Opus for:

• Complex architecture and system design
• Multi-file refactors with tricky interdependencies
• Subtle debugging (race conditions, memory leaks, flaky tests)
• Novel algorithm design
• Nuanced writing or deep analysis
• Long chains of reasoning

Everything else goes to Sonnet: file reads, lookups, commands, standard edits, straightforward questions, writing tests.

When in doubt, Haiku picks Sonnet. Opus is only for tasks that genuinely need deeper reasoning.

Cost

The Haiku classification call adds ~$0.0001 per prompt. Over a full day of coding, maybe $0.01-0.05 extra. The savings from not running Opus on simple tasks far outweigh this.

Configuration

The extension uses these model IDs by default:

Role	Model ID
Router (classifier)	claude-haiku-4-5
Standard target	claude-sonnet-4-6
Complex target	claude-opus-4-6

To change these, edit the constants at the top of extensions/model-tier-coach.ts.

License

MIT