mirror of
https://github.com/heygen-com/hyperframes.git
synced 2026-11-16 06:19:21 +00:00
Address review findings on the streaming interruption recovery. Classifier (engine): walk the error cause chain (depth 5, including AggregateError members) so undici's "fetch failed" wrapper and a dual-stack localhost AggregateError classify by the errno they carry. Recognise the loopback spellings Node, Bun and Chrome emit: unbracketed "::1", "[::1]", "127.0.0.1" and "localhost" with ETIMEDOUT, ECONNREFUSED or ECONNRESET, and Chrome's ERR_TIMED_OUT, ERR_CONNECTION_TIMED_OUT, ERR_CONNECTION_REFUSED, ERR_CONNECTION_RESET, ERR_CONNECTION_CLOSED and ERR_EMPTY_RESPONSE. Normalise the IPv6 host to "::1". Export isLoopbackConnectionLoss, LoopbackConnectionLoss and CaptureEndpointDiagnostic. Endpoint provenance: CaptureStageError now forwards the classified endpoint, so the orchestrator's re-classification (which early-returns the stage error) sees the real host and port instead of always logging an unknown owner. Retry gate: replace `transient_browser && workerCount === 1` with resolvePreFrameLoopbackLoss in the new render/preFrameRecovery module. Only a failure that names a loopback endpoint, on exactly one worker, with zero frames written, qualifies. Target closed, Page crashed, a browser that cannot launch or a runtime that never becomes ready fail immediately as before the widening; a SIGTERM-killed Chrome cannot hide behind same-host retry work. recoverPreFrameFileServer owns the health log and the restart-on-unhealthy decision so both are unit-testable. File server: one createRenderFileServer factory for probe discovery, frame capture and the pre-frame restart. A rejected body.cancel() no longer marks a healthy server unhealthy; the probe error carries the errno (ECONNREFUSED on Node, ConnectionRefused on Bun). Audio: share canRetryMixLocally and resolveAutomationRerun between the grouped and ungrouped mixes so retryable failures (external interruption, managed deadline/inactivity, missing FFmpeg) never trigger a compatibility or automation-degradation rerun, an interruption landing on the rerun is reported instead of discarded, and a cancelled group submix uses the canonical cancelled/user classification. Tests pin each guard: dropping endpoint forwarding, deleting the restart block, workerCount >= 1, dropping the zero-frame gate, forcing the identity-header check true, removing the bare ::1 pattern, removing the cause walk, reverting the ungrouped audio guard and discarding the rerun interruption each turn a test red. Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>