mirror of
https://github.com/stablyai/orca.git
synced 2026-09-28 16:02:45 +00:00
Jinwoo-H
e442e4fa55
#19689 shipped because no lane paired two builds over `orchestration.*`. The cross-version harness only covered the terminal stream and `agentSession.*`, and every federation "old peer" test wrote a protocol_version integer into a row and read it back with the same build, so a coordinator that omits a field and a host that starts requiring it were both invisible. Pairing against the newest tag alone would not have caught it either: the newest tag already contains the fix for anything broken and repaired inside the last release cycle. A worker host and the desktop that dispatches to it update on their own schedules, so N-1 is an ordinary peer. The suite therefore extracts the two newest stable tags and runs both skew directions against each. Each pairing drives one journey through the real coordinator modules — `startFederatedWorker` composes the attach params, `syncFederatedDispatch` drives the relay — against the peer build's real dispatcher over its real store, with a JSON round trip so an `undefined` field drops exactly as on the wire. What the old side has is read from its checkout; nothing about a release is written down. `pr-code-change-scope.mjs` did not route orchestration source at this job, so the lane would not have run on the PR that broke it; the federation source prefixes are added with cases in its test. The job's file list is now derived from the directory, so a new suite that is never named in the workflow fails instead of covering nothing.