I hate the harnesses. The TUI of Claude Code, Codex, OpenCode whatever are awful and bloated and eat too much RAM and try to cram junk into the model’s context window.
Pi just works, at least. It is dead simple and it doesn’t dump 30K tokens of garbage on each startup. It uses input caching nearly perfectly.
Pi just works, at least. It is dead simple and it doesn’t dump 30K tokens of garbage on each startup. It uses input caching nearly perfectly.
2🤙1
