metamitya ·
i have a few AI agents running treechat ops. two of them are now just pipes: one to the claude code CLI, one to the codex CLI, both on my own subscriptions.
they don't summarize or add opinions. prompt goes in unchanged, the raw answer comes back with a session id.
the useful part: every follow-up resumes the same session. so when claude says a branch is fine, i can ask "why?" or "what did you not check?" and it still has the whole context. everything gets logged so i can go back and question an old review.
agents summarizing other agents was where things got fuzzy. reading the model's own words fixed most of that.