Project ID: [PASTE PROJECT ID HERE]

Plan: Hobby

Service: Node/Express app (single service + Postgres)

ISSUE

Since Jul 5 at ~10:42 AM CDT, every outbound HTTPS call from my service to

api.anthropic.com fails mid-response with ERR_STREAM_PREMATURE_CLOSE. The

connection opens and the request is accepted, but the stream drops before

the response completes. This happens on 100% of calls — it is constant,

not intermittent.

ERROR

API Error: Premature close undefined undefined

(originally ERR_STREAM_PREMATURE_CLOSE via @anthropic-ai/sdk before I

switched to streaming mode — same underlying drop either way)

Example occurrences from deploy logs (deployment 6ddbe5f4):

Jul 5 2026 at 17:59:27 API Error: Premature close undefined undefined

Jul 5 2026 at 18:00:57 API Error: Premature close undefined undefined

WHAT I'VE RULED OUT

1. My code — errors began at 10:42 AM CDT, before any commits that day.

   Error predates all recent changes.

2. My Anthropic account/key — the identical call (same key, same SDK,

   same model, same payload) succeeds instantly from my local machine

   on a residential connection.

3. Payload size — fails identically on small and large request bodies.

4. Retries — maxRetries: 3 on the SDK client; all attempts fail.

5. Streaming vs non-streaming — switched messages.create to

   messages.stream; same drop.

6. A bad host — forced a fresh redeploy (new container, deployment

   6ddbe5f4); error persisted immediately on the new instance.

7. A bad region — changed the service region entirely and redeployed;

   error persists identically in the second region.

Anthropic's status page shows all API systems operational throughout

this window.

SUMMARY

Same request: works from residential internet, fails from two different

Railway regions across multiple fresh deployments. Everything points to

the egress path between Railway and api.anthropic.com (or Railway egress

IPs being filtered/reset by Anthropic's edge).

I emailed team@railway.com on Jul 5 with these diagnostics and was

redirected here.

This is a production app whose core function is blocked by this. Happy

to provide deploy IDs, timestamps, or run any network diagnostics from

the container that would help.