You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
fix(api): Requesty ignores caller cancellation and the configured request timeout #2510
Requesty does not honour the caller's metadata.abortSignal and the configured request timeout the way the merged providers do. Cancelling a request mid-flight does not stop the streaming half of createMessage, and a completion that lands after the cancel is still delivered to the caller.
Evidence on main
The provider does not race its SDK promise through rejectOnAbort (src/api/providers/utils/abort-signal.ts:226), so a cancellation during SDK retry backoff or a request that never settles waits for the transport.
The caller signal is handed to the SDK instead of a per-request bridge. The pinned openai SDK does signal.addEventListener('abort', () => controller.abort()) in client.mjs and never removes the listener, so a long-lived caller signal accumulates one listener and one request controller per request.
Nothing re-checks the signal after the stream ends, so an iterator that closes without throwing - which the SDK does when it converts an aborted read into a normal end - reports success.
The provider follows the shared contract exactly as requesty.ts (#1537) and zoo-gateway.ts (#2430) do: bridge the caller signal and dispose it on every exit path, race the SDK promise and the model lookup through rejectOnAbort / resolveModelWithAbort, guard the emit loop and the post-loop end, and normalise an SDK-side abort with isRequestAborted so a cancellation surfaces as AbortError with the message The <provider> request was aborted.
Scope
Provider file plus its spec. No change to utils/abort-signal.ts or utils/request-timeout.ts; the timeout-versus-body-consumption question is tracked separately in #2486.
Activity