[Bug] ECONNRESET on streaming requests with bundled Node v26.3.0

Bug Description

## Summary Claude Code CLI (native binary, bundled Node v26.3.0) on Windows 11 fails with ECONNRESET on nearly every streaming request. The same machine, network, and account work fine in the claude.ai browser UI. Reproduces in both the terminal CLI and the VS Code extension. ## Environment - Claude Code: 2.1.245 (native install, C:\Users\<user>\.local\bin\claude.exe) - Node: v26.3.0 (bundled with the binary; system Node is v24.15.0 and is not used) - OS: Windows 11 - Model: claude-opus-5[1m] - Auth: claude.ai subscription - Corporate environment with Zscaler installed ## Symptoms Nearly every request fails through all 10 retries with: Stream connection error (ECONNRESET) — retrying streaming (N/10) API Error: Connection lost mid-response. The response above may be incomplete. ## Key finding 1: the reset arrives at response start, not after a timeout Elapsed time from "Stream started - received first chunk" to ECONNRESET: attempt 1: 789ms attempt 2: 0ms attempt 3: 0ms attempt 4: 1ms attempt 5: 1401ms attempt 6: 709ms attempt 7: 1ms Multiple 0ms cases mean the RST was already in the socket buffer when the first chunk was read. This is not an idle timeout and not volume-dependent. Time-to-first-byte is consistently healthy (~1.1s). ## Key finding 2: an unrelated host fails identically In the same session, the Atlassian MCP server (mcp.atlassian.com, no relation to Anthropic infrastructure) drops the same way: MCP server "atlassian": HTTP connection dropped after 31s / 62s / 0s uptime Connection error: SSE stream disconnected: TypeError: The socket connection was closed unexpectedly So this is not specific to api.anthropic.com. Long-lived HTTPS/SSE connections from this process are being killed across multiple unrelated destinations, while Chromium on the same machine is unaffected. ## Ruled out - Network path in general — browser on the same machine/network is fine - Zscaler — identical behavior with it ON and OFF - claude.ai MCP connectors — disableClaudeAiConnectors: true, no change - Local config/cache — clean CLAUDE_CONFIG_DIR, no change - Claude Code version — up and down, no change. Note that every version ships the same bundled Node v26.3.0, so the runtime never actually varied across those tests - Node runtime — cannot be tested. Both the npm package and the native install resolve to the same bundled binary; no cli.js exists to run under a different Node ## History - Roughly one month ago the same symptoms appeared and then resolved spontaneously with no local changes - Another member of the same organization hit this once; they believe a downgrade fixed it, but given that the bundled runtime does not change across versions, this was likely coincidental timing - No other member is currently affected ## Possibly related Issue #86473 reports the same shape (Windows 11, raw HTTPS to api.anthropic.com healthy, model requests reset, reproduces in a clean room, browser surfaces unaffected). Anthropic support confirmed two server-side incidents in that thread. ## Request Given that the reset arrives at response start rather than after a timeout, and that it reproduces against a non-Anthropic host from the same process, I would like help determining whether this originates client-side (bundled Node v26 HTTP stack) or elsewhere. Happy to provide full debug logs.

Environment Info

• Platform: win32
• Terminal: vscode
• Version: 2.1.245
• Feedback ID: 404f1410-f034-4251-9840-120df0870a1b

Errors

[]