mirror of
https://github.com/stablyai/orca.git
synced 2026-10-23 00:03:52 +00:00
Brennan Benson
5e1a5bf664
Routing a removal to the confirmed host is useless if that host's row can never be confirmed. `worktreeId` is `repoId::path` with no host component, and every id-keyed projection folded the second host away. Store index: dedup by (host, id) instead of id. The duplicate this dedup was built for — the createWorktree-appends / fetchWorktrees-replaces race — is the SAME workspace twice and still collapses; two hosts are two workspaces and now both survive. Adds getIndexedWorktreesById for callers that need every owner. Sidebar: rowKey, host-context labels, and the lineage emit/child sets are keyed by host-qualified identity. The lineage set mattered most — keyed by bare id, the second host's row was never emitted at all. Repo grouping falls back to per-row host labels when one repo id spans two hosts, since the per-repo label cannot tell those rows apart. Delete lineage: a colliding child id resolves to the parent's host, so "delete all" cannot route a descendant at the other machine's checkout. Reconciles the two id indices, which disagreed (first-wins in worktree-by-id-index, last-wins in worktree-repo-index) and so resolved the same colliding id to different rows. Both are first-wins now, matching the `.find()` they replaced; same-host duplicates are already collapsed by freshness before that point, so first-vs-last only decides genuine two-host collisions, which no bare-id lookup can resolve. Documented on both, and both are marked as not-evidence-of-ownership. The index keeps its one-read-per-row budget (retained selectors assert it) by recovering the id from the identity key instead of re-reading it.