active experiment

code-with-pi

A terminal broker where an AI coding agent and its user watch or take over the same live shell. I wrote it in C11 over Windows ConPTY and POSIX PTY with authenticated local IPC, and added CI tests on Windows and Linux.

Repository →

Writing about this project

  1. Known limits of a shared persistent terminal

    Failure boundaries in shell recovery, cancellation, replay, backpressure, terminal negotiation, and interactive programs.

  2. One shell, two observers: the persistent terminal architecture

    How a native broker owns one PTY or ConPTY while an agent and a human share observation and controlled input.

  3. What I tested in the persistent terminal broker

    Observed Windows and Linux checks for PTY lifecycle, authenticated IPC, state persistence, cancellation, replay, and human takeover.