Keep a decision journal

Every agent harness now records what happened. None of them record why.

A session leaves behind hundreds of tool calls — every file read, every command run, every edit applied. Not one of them contains the sentence that matters: we chose Postgres over SQLite because reconnect had to converge, and we ruled out the queue because it needed a broker we do not run. That sentence lives in the model's context for about an hour, and then compaction destroys it.

Six months later somebody asks why the code is shaped this way. The commit history answers what. The journal is where why survives — and, more valuably, where the alternatives you already rejected survive, so nobody spends a week rediscovering that they do not work.

Quickstart