Since 2026-07-04 ~22:00 UTC, all outbound HTTPS calls from my service to api.openai.com fail with "Premature close" — the request reaches OpenAI (confirmed billed on their side), but the response body arrives truncated. We isolated it with a side-by-side probe running inside the same container:

Request over a keep-alive connection (Connection: keep-alive, Node https.Agent keepAlive:true): fails in <1s with "Invalid response body ... Premature close" — even for tiny ~1KB responses;

The identical request over a fresh connection (keepAlive: false): succeeds consistently (~500-900ms).

Reproducible 100% of the time, single or concurrent connections, affects both buffered and SSE responses. The same calls work fine from other networks with the same API key. No dependency/runtime changes on our side (locked lockfile; reproduced across multiple deploys). We are working around it by disabling connection reuse, but this looks like an egress/NAT/proxy regression mangling HTTP keep-alive framing. Was there an egress change on 2026-07-04? Region: [confira a região do seu serviço em Settings]. Happy to run further probes if useful.