pi-next-cue

Predicts your next prompt after each agent turn — Tab to fill, Enter to send.

Why

After the agent finishes, you usually know what to type next. Pi-next-cue predicts it and shows a hint above your editor — saving keystrokes and keeping you in flow.

Install

pi install npm:pi-next-cue

Requires Pi v0.37.0+.

How It Works

The hint widget shows one of two states, depending on context:

Predicted prompt (after agent responds):

→ run the tests

Your last message (for reference):

↩ fix the login redirect bug

• → Next prompt you're likely to send — generated by a lightweight LLM after each agent turn
• ↩ The last message you sent — so you never lose context of what you asked

When the editor is empty:

• Tab fills the hint into the editor for editing
• Enter sends the hint directly

When you start typing, the hint stays visible but won't interfere — Tab and Enter only trigger when the editor is empty.

Features

• Adaptive tone — matches your language and communication style from recent messages
• Correction learning — remembers when you ignore suggestions, avoids repeating rejected directions
• Tool-aware — considers whether the last tool succeeded or failed when predicting
• Lightweight — single LLM call with minimal context, no background indexing

Configuration

Pi-next-cue uses your session's current model by default. To use a specific fast/cheap model or customize keybindings, create ~/.pi/agent/pi-next-cue.json:

{
  "provider": "deepseek",
  "model": "deepseek-v4-flash",
  "keys": {
    "fill": "tab",
    "send": "enter"
  }
}

All fields are optional. Defaults: tab to fill, enter to send. Key names follow Pi's keybinding format.

Events

Other extensions can control hint visibility:

// Hide the hint (e.g. when showing a dialog)
pi.events.emit("pi-next-cue:pause");

// Restore the hint
pi.events.emit("pi-next-cue:resume");

License

MIT