baton

Brief

I write code with Claude Code and Codex, and a task is often halfway through on the computer when I have to leave. To check on it from my phone or add a line, there were two options: a remote desktop, or waiting until I was back.

Process

A daemon on the computer reads the session records both tools already write to disk. Through it, the phone continues an existing session or starts a new one, which then appears on the computer. Every turn’s state is saved, because a phone losing its connection is normal.

Before building, I did some research, and the commonest mistake there was taking issue titles, docs and naming as evidence. So I made a rule: any claim about how something behaves needs support from data, source code or a test, and where there is none, I write that I don’t know. The rule overturned several numbers. What looked like 310 Claude sessions was 137; the rest were sub-agent traces. What looked like 148 Codex sessions was 37.

Two more rules came from mistakes. A session’s identity is its file name, with no separate mapping: three tools that kept their own mappings all drifted. And baton uses the agent installed on the system rather than bundling its own; in tests, a bundled older version had a third of the usable context.

Outcome

baton is written for my own few machines, so it has no multi-user permissions and none of an IDE’s file trees or diffs. The repository is private.