pi-duplex

pi-duplex gives Pi two cooperating agents:

• A fast foreground answers simple questions and stays responsive.
• A persistent reasoner handles substantive work with Pi's normal tools and speaks directly in the transcript.

You keep using Pi normally. While the reasoner works, the foreground can answer a side question, update the active task, queue later work, or stop the task.

Install

Requires Pi 0.84.3 or newer.

These examples use Codex because I have a ChatGPT subscription, but any Pi provider works.

pi install npm:pi-duplex

Log into Codex through Pi:

/login openai-codex

Then start Pi with a fast foreground model and a smart reasoner model:

PI_DUPLEX_REASONER_MODEL=openai-codex/gpt-5.6-sol \
PI_DUPLEX_REASONER_THINKING=max \
pi --model openai-codex/gpt-5.6-luna:medium

The reasoner appears in a different color. Use /reasoner-model to switch its model and Ctrl+O to expand its tool calls. Pi keeps the first Escape for its foreground work; press Escape twice to stop an active reasoner.

Models

The foreground uses whatever model is selected in Pi. The reasoner must be set with PI_DUPLEX_REASONER_MODEL before Pi starts. Its optional PI_DUPLEX_REASONER_THINKING setting defaults to max.

For an open-weight setup, try DeepSeek V4 Flash as the fast foreground and Kimi K3 as the reasoner. After setting DEEPSEEK_API_KEY and MOONSHOT_API_KEY:

PI_DUPLEX_REASONER_MODEL=moonshotai/kimi-k3 \
PI_DUPLEX_REASONER_THINKING=max \
pi --model deepseek/deepseek-v4-flash:low

Of course, you can use whichever model combination you'd like!

MIT licensed.