pi-model-thinking

Auto-set and remember thinking levels per model in pi.

No more Ctrl+P → Shift+Tab → Shift+Tab when switching models. This extension remembers what thinking level you set for each model and restores it automatically.

Install

pi install npm:@ogulcancelik/pi-model-thinking

For local development:

pi install ~/Projects/pi-extensions/packages/pi-model-thinking

How it works

The extension uses a single file: ~/.pi/agent/model-thinking.json.

This file is both your initial config and your live state. You can seed it manually, or let it grow organically as you use pi.

If the model is in the file

On model_select (Ctrl+P, /model, session restore), the extension automatically applies the stored thinking level.

If the model is NOT in the file

The extension does nothing. Pi handles thinking natively, exactly as before.

When you manually change thinking (Shift+Tab)

The extension writes the new level into model-thinking.json for that exact model. Next time you use that model — even in a new session — it starts with your chosen level.

If your change matches a provider-level default in the file, the exact model entry is cleaned up automatically.

Config format

~/.pi/agent/model-thinking.json:

{
  "models": {
    "anthropic/claude-sonnet-4-5": "high",
    "openai/gpt-5.2-codex": "medium"
  },
  "providers": {
    "fireworks": "low"
  }
}

Resolution order:

1. models["provider/modelId"] — exact match
2. providers["provider"] — provider-wide fallback
3. Not in file → extension ignores, pi handles natively

You must create the config file. Only models explicitly listed (exact or by provider) are managed by the extension. For unmanaged models, pi handles thinking natively and the extension does not interfere.

Commands

/model-thinking — shows current resolution and whether the active model is managed.

/model-thinking reset — deletes model-thinking.json, clearing all remembered levels.

Example flow

Create ~/.pi/agent/model-thinking.json with your defaults:

{
  "providers": {
    "anthropic": "high",
    "openai-codex": "medium"
  }
}

1. Ctrl+P to Claude → thinking is high (provider default).
2. Shift+Tab change thinking to low. Extension writes "anthropic/claude-sonnet-4-5": "low".
3. Quit pi, start a new session, Ctrl+P back to Claude. Thinking is automatically low.
4. Shift+Tab back to high. Extension removes the exact-model override (matches provider default).
5. Switch to GPT (managed via openai-codex provider). Thinking is medium.
6. Shift+Tab on GPT to high. Extension writes "openai/gpt-5.2-codex": "high".
7. Switch back to Claude → high. Switch to GPT → high. No manual tweaking needed.