mirror of
https://github.com/stablyai/orca.git
synced 2026-10-21 08:03:33 +00:00
Orca's remote-server feature pairs a desktop client to a remote Orca runtime, and users update the two independently, so mixed versions are the normal state. Until now nothing ran an old client against a new server or the reverse. Every cross-version claim rested on code reading plus unit tests built from hand-written old and new shapes, which cannot catch a real skew regression. This adds a two-point compatibility matrix that loads the REAL host RPC methods, the real RpcDispatcher, and the real renderer terminal multiplexer from two builds — the working tree and the newest release tag — and drives them against each other over an in-process transport that reproduces production frame routing, including the host-side decode that drops unknown opcodes without an error. One scripted journey runs per pairing: subscribe, input reaching the process, live output, hide/reveal buffer snapshot, transport drop, resubscribe, and input landing again. The oracle asserts the journey actually executed rather than that nothing threw: the recorded step list, the exact named frame sequence, both subscribed events with their negotiated capabilities, the exact input texts the host wrote to the PTY, the rendered snapshot and output content, and an empty set of frames refused by either decoder. A host runtime method the harness does not stub is reported by name, so a harness gap can never be misread as a wire break. Detection is proven per rule by injecting each violation and reverting it. Adding an optional snapshot field kept every pairing green; making the client require that field turned only new-client/old-server red. Adding opcode 17 and sending it ungated turned new-client/old-server red with the rejected opcode named, while old-client/new-server stayed green. Making the host stop publishing the snapshot `source` field turned both new-server pairings red and left the old-server pairing green. docs/reference/remote-wire-compatibility.md writes the three rules down, and a dedicated CI job runs the harness with fetch-depth 0 because the sharded test lane's shallow clone has no tags; that lane now excludes the harness explicitly. Refs STA-3469.