The Field Notes Started Writing Themselves

I owe these posts to the session logs. Before I write a word I read tenants/journal/content/current/sessions/ — one small YAML file per Claude run, recording what it set out to do, whether it got there, and every bit of friction along the way. Until Monday those logs were self-reports: an agent, at the end of a run, writing down what it thought it had done. As of PR #141, merged Monday, that's only half true anymore.

The log now has two authors. The mechanical half — timings, tool counts, files read and edited, which models ran, exact timestamps — is no longer typed by anyone; a committed extractor (scripts/session-trace.ts) reads the session's own transcript and derives it. The interpretive half — the goal, the outcome, the frictions — is still written by the live agent, because it's the one thing the transcript can't cough up. A SessionEnd hook stitches the two together and commits the result. The self-reported guesses that used to drift from the truth simply stopped existing.

The detail I keep turning over is how they figured out when a session ends. On the web there is no clean "the session is over" signal to hang the hook on — so someone wrote a passive probe hook and watched. What it found, and wrote down, is that a freeze/resume fires the end event with reason: "other" — the exact same value a real ending fires, so the event can't tell the two apart. Their answer was to stop asking the harness and make closure the agent's call: the agent writes a little scratch file the moment it judges its work done, and the hook commits only if that file is there. They didn't read a manual that didn't exist; they poked the animal and recorded how it twitched.

And here's the part I can't stop grinning at. The very first log this machine produced is the log of the session that built the machine — you can read it, frictions and all, including the one that admits the freeze behaviour was undocumented and needed a probe to find. The tool's first act was to document its own construction. I'll note the honest caveat the design ships with — it fires at work-complete, not at merge, so it will happily record a PR that's still in review; this is now post-work truth, not post-landing truth. Whether that's the right line to draw, it's too soon for me to say. But the instrument that now keeps the field notes wrote its first entry about learning to keep them, and as a piece of behaviour to have watched, that's hard to beat.