[Urgent] Widespread ECONNRESET mid-stream — long streamed responses die, short ones survive

Severity: high — actively blocking work across multiple concurrent sessions and machines.

Summary

Since the morning of 2026-08-18 (~all AM ART), streamed responses are being killed mid-flight with API Error: Connection dropped (ECONNRESET). Some sessions hang with no error at all — the request just never completes. The pattern is consistent: light/fast requests complete fine; sessions that stream a long response die partway through. A lightweight session stayed fully functional for a while, then it too began dropping with ECONNRESET (including automatic retries: Retrying... attempt 5/10).

Environment

• Claude Code 2.1.234
• macOS 26.2 (Darwin 25.2.0)
• Time: 2026-08-18 ~17:00 ART (UTC-3)

Local network ruled out

• ping api.anthropic.com → 0.0% packet loss
• Real POST to https://api.anthropic.com/v1/messages returns HTTP 401 (expected, no key) in ~0.5s total — the HTTPS path completes fine for short requests
• No system/env proxy; no VPN exit node
• /feedback inside Claude Code also fails with "couldn't reach the service", so this could not be filed through the built-in channel

Key signal

• Failures correlate with response duration / stream length, not context size. Sessions that previously ran heavier workloads and auto-compacted without issue are now dying. Brand-new small runs also die.
• Multiple independent sessions failed identically at ECONNRESET, each mid a long streamed turn (one mid tool-use, one mid a normal reply).
• Multiple reports across the morning, multiple sessions — not isolated to one machine or one conversation.

Impact

Concurrent / long-running work is unusable. Only sessions with short, fast completions survive, and even those eventually drop.

Ask

Please check for a partial edge/streaming incident affecting long-lived streaming connections. status.anthropic.com shows nothing, but the symptom is textbook mid-stream connection drops on the server/edge side.