Summary

After the upstream WebSocket server sends its terminal event and then closes the connection (e.g. with code 1011 on keepalive timeout), Bifrost would hand out the same connection to the next request on the same client session. The connection appeared live (IsClosed() returned false, age was within limits) because the WS close frame sat buffered in the socket and had never been read. The next caller would succeed on WriteMessage (OS buffer accepts the write) and then immediately fail on the first ReadMessage when the buffered close frame was returned as an error. This caused a fallback to the HTTP bridge and an error frame to the client.

The root cause is that Pool.Get relied only on an atomic bool (set by Bifrost's own Close call) and wall-clock expiry checks to determine connection health. Neither of those detects a server-side close that Bifrost never consumed.

Changes

• transports/bifrost-http/websocket/pool.go: added (*Pool).isLive helper that sets a 1ms read deadline on an idle connection and calls ReadMessage. A timeout error (no queued data) confirms the connection is alive. A close frame, EOF, or any other error reveals it is stale. The probe runs in Pool.Get after the existing IsClosed / isExpired checks, before incrementing inFlight and returning the connection to the caller. Stale connections are closed and the loop continues to the next idle candidate or a fresh dial.
• transports/bifrost-http/websocket/pool_test.go: added TestPoolGetEvictsStaleSessionConn. A mock WS server accepts one connection, then closes it server-side after the test returns the connection to the pool. Pool.Get is called again and must produce a fresh connection (new dial), not the stale one.

Type of change

• Bug fix

Affected areas

• Transports (HTTP)

How to test

go test ./transports/bifrost-http/websocket/... -cover -run TestPoolGetEvictsStaleSessionConn -v
go test ./transports/bifrost-http/websocket/... -cover
go vet ./transports/bifrost-http/...

Expected: TestPoolGetEvictsStaleSessionConn passes, all existing tests pass, no vet warnings.

Integration test (manual): configure a WS upstream that closes connections after the terminal event. Send two sequential requests on the same Bifrost WS session with a pause between them. Before this fix, request 2 would receive an error frame. After this fix, request 2 succeeds with a fresh upstream connection.

Breaking changes

• No

Related issues

Closes #3002

This fix was discovered while working on PR #2775 (openai-oauth-passthrough). The feature branch in that PR contains an earlier version of the liveness probe, which has been extracted here and removed from the feature branch in a separate commit.

Security considerations

None. The probe is a local read with a 1ms deadline on a connection that Bifrost owns exclusively while it is in the idle pool. No external input is involved and no credentials are affected.

Checklist

• I read docs/contributing/README.md and followed the guidelines
• I added/updated tests where appropriate
• I updated documentation where needed
• I verified builds succeed (Go and UI)
• I verified the CI pipeline passes locally if applicable