Skip to content

FAQ

Fonn makes no such promise. Fewer characters in an observation is not the same as fewer tokens or dollars, and real tasks vary. Use --filter observe to see what would be removed without changing anything.

By default, Fonn removes only lines that match strict progress patterns or exact duplicates, and it states the count. At the similar level, it also drops the middle of runs that differ only in numbers. It never does that for lines that mention errors, failures or warnings. The footer tells the agent how to read the full original if it needs it.

No. It rewrites only a new tool observation as it arrives. History, reasoning, instructions and source files are left alone.

Under ~/.local/state/fonn by default, or a --state-dir you choose. They can contain private data. Retention is manual, and they should stay out of Git.

Yes. Native workers inherit the session hooks, so their supported Bash observations go through the same filter for every model. --filter off turns it off for parent and workers.

Can Claude call Codex, or the other way round?

Section titled “Can Claude call Codex, or the other way round?”

Yes, when you ask for it. Fonn’s guidance asks agents to prefer their built-in subagents and not to start other agent CLIs on their own initiative. When you explicitly ask, for example “hand this review to Codex”, the agent may run codex exec "<task>" as an ordinary command through the normal approval prompt. That CLI keeps its own login and billing. fonn codex and fonn claude open interactive terminals, so start those yourself when you want Fonn’s profiles and filtering for the second agent. See working with another agent CLI.

A desktop app is planned but not built. The workspace packages are not published to npm as a supported install route.