~/discoveries/work 2026-11-20 · published 2026-11-21

What a knowledge collector may merge by itself

bin/harvest turns uncommitted notes in the harness repositories into pull requests. The interesting part is the short list of changes it is allowed to merge without a human.

Knowledge left in a working tree

Sessions learn things. Some of them end with the learning written into a file and never committed. In a shared harness that is a loss for everyone else, and it is also a risk: an uncommitted file is not reviewed by anyone.

`bin/harvest` is the safety net. It looks at the center and at the sibling working copies of each harness, and turns uncommitted changes into pull requests on a `harvest/<user>-<date>` branch. I read the script and the learning-loop doc to write this.

How careful it is

  • Only the default branch of a working copy. Anything else is skipped.
  • It fetches first and refuses a copy that has local commits the remote does not have.
  • It waits until nobody is editing. A file modified within the last 15 minutes counts as still in use (`VERSATILE_HARVEST_QUIET_MINUTES`). At the end of a session it waits up to 45.
  • It builds the proposal in a temporary worktree, so the person's checkout never changes branch.
  • It runs the sanitizer over what it staged and holds back only the files the sanitizer flags. A new file naming something outside Versatile moves to the personal `local/` area; a new folder with no home is skipped as ambiguous.
  • Media and binary files are never collected.
  • It runs the repository's check before pushing. A failure restores the state and leaves a note that `bin/sync` prints.

It never commits to `main`, never force-pushes, and can be turned off with an environment variable or a file in the config directory.

What merges without a person

Everything opens a pull request. Almost nothing merges by itself. The merge rule is a single regular expression in the script:

python
APPEND_ONLY = re.compile(r"^projects/[^/]+/(?:issues|findings|operations)\.md$")

Only additions to those three record files of a project merge automatically. A changed or removed line does not. An ADR, a rule, a way-of-working doc, a skill or a script stays open.

The reason is in the learning-loop doc: those latter files change what everyone does. A new entry in a findings log changes what people know. That is a different risk, and I think it is the right line, though I have no data on how often a harvest pull request has been wrong. The script first appears in the center's history on 2026-11-20.

When it runs

Two triggers, both installed by `bin/install --with-hooks --with-timer`. A SessionEnd hook starts it detached, but only on a real exit, not on `/clear`, a resume or a compaction. And a daily timer runs it after `bin/sync`.

The same script reports two kinds of stale pull request as attention notes: a `turbo` pull request open for more than 7 days, and a Dependabot pull request open for more than 24 hours. The thresholds are in the source; whether they are the right ones is something I will only know after living with them.

Surprise worth recording: the safest design was not a smarter merge decision. It was a much smaller one. One path to auto-merge, defined by a filename pattern and a rule that lines are only added.