Keeps this fork current with upstream (stablyai/orca) without a human watching the upstream commit feed.

1. Fetches origin/main (upstream) and fork/main. If upstream's SHA matches the cursor from the last run, it exits immediately — no agent, no tokens.
2. Computes the fork footprint (files the fork changed since its base) and intersects it with the files upstream touched. That overlap is the only place upstream churn can collide with us, and it scopes everything below.
3. Runs a trial merge in a persistent worktree. git answers the textual-conflict question by itself — no agent involved in deciding whether there's a conflict.

Clean merge → pnpm install / typecheck / test / lint. If those pass and upstream touched at least one fork file, an agent reviews the overlap for semantic breakage a clean merge hides: a renamed symbol the fork calls, a changed signature or type contract, a moved file, an i18n key collision. Green on all of that → branch is pushed, PR opened and merged automatically. Anything short of green → PR is left open with the report attached.

Conflict → an agent resolves what it is confident about, leaves markers where it is not, and writes a per-file report: what upstream changed, what the fork changed, why they collided, what a human must still decide. Branch is pushed and the PR is labelled needs-human. Never auto-merged.

• sync.mjs — the pipeline. Lives here, not in config/scripts/, because upstream never touches this path, so the tooling itself is never a conflict.
• ~/.cache/orca-fork-sync/worktree — persistent worktree, so node_modules survives between runs and pnpm install stays incremental.
• ~/.cache/orca-fork-sync/state.json — last-processed upstream SHA.
• ~/.cache/orca-fork-sync/reports/ — every report, also posted to the PR body.

Mutable state deliberately lives outside the repo so nightly runs never dirty it.

node .fork-sync/sync.mjs --dry-run   # everything except push/PR/merge
node .fork-sync/sync.mjs --force     # ignore the cursor and re-run
node .fork-sync/sync.mjs

The agent reads CLAUDE.md — its "Fork Context" section is what tells it what the fork deliberately adds and why. Keep that section current: it is the only thing standing between an agent and a plausible-looking resolution that quietly discards a fork behaviour.

The gates run under node 24 (package.json engines), pinned explicitly in the systemd unit via mise x node@24. This is not cosmetic: on node 26 about 40 unit tests fail with window.localStorage undefined — unrelated to any fork change (they fail identically on the pristine fork base), but enough to keep the test gate permanently red and block every auto-merge.

If your own shell defaults to node 26, pnpm test will fail locally for the same reason. Run it as mise x node@24 -- pnpm test, or pin the repo.

Daily. Upstream moves fast, and small diffs conflict far less than large ones. Requires the machine to be awake; a missed night is harmless — the next run picks up every commit since the cursor.