[BUG]

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Bug Report: API requests fail with ECONNRESET after update (2.1.222 / 2.1.223); rolling back to 2.1.210 fixes it
Summary

After Claude Code auto-updated to 2.1.222 on 2026-08-05 (and subsequently 2.1.223), every task that makes an API call fails with a connection reset and retries indefinitely. The same machine and network had worked for months on prior versions. Downgrading the CLI to 2.1.210 immediately and fully resolves the problem. The Claude desktop app's "Code" tab remains broken, since it bundles its own (affected) version that cannot be easily rolled back.

The raw network path to api.anthropic.com is healthy — curl (HTTP/1.1), ping, and tracert all succeed to the same endpoint — but Claude Code's own API connections are actively reset (ECONNRESET). This strongly suggests a client-side networking change between 2.1.210 and 2.1.222.

Environment
OS: Windows 11 (win32-x64)
Claude Code install: native installer, C:\Users\<user>\.local\bin\claude.exe
Account / auth: Claude Max account (logged in; not an API key)
Affected versions: 2.1.222, 2.1.223
Last known-good / working version: 2.1.210
Also affected: Claude desktop app "Code" tab (bundled Claude Code version — still failing after CLI rollback)
claude doctor result: healthy — "No installation issues found," Search OK, auto-updates enabled, native install method

What Should Happen?

The new version should work not spin trying to connect.

Error Messages/Logs

Symptom / Observed errors

Every prompt or command that requires the API hangs and retries. Local-only commands (e.g. /status) work fine; anything that reaches the API resets.

Exact error strings observed:

Terminal CLI: ✻ Unable to connect to API (ECONNRESET) · Retrying in 7s · attempt 5/10
Desktop app "Code" tab: Request failed · retrying (6/10) · 42s

Reproduces on:

A trivial prompt (hi)
/doctor (its connectivity check)
Every real task
Timeline
Worked reliably for months on prior versions.
2026-08-05: auto-updated to 2.1.222 → tasks began hanging.
Later updated to 2.1.223 → same failure.
2026-08-06: rolled CLI back to 2.1.210 → fixed; desktop app "Code" tab still fails.
Key diagnostic evidence
Auth is fine — /status shows a valid Claude Max login. The failure is a connection reset, not a 401/403/429.
Raw connectivity to the API works:
ping api.anthropic.com → resolves to 160.79.104.10, 0% loss, ~7 ms.
curl.exe -v https://api.anthropic.com → succeeds over HTTP/1.1; returns the normal 404 + Anthropic API banner, real Cloudflare/Anthropic certificate (no MITM/SSL-inspection).
tracert api.anthropic.com → completes cleanly (~11 hops, ~8 ms), no corporate proxy in path.
Protocol difference: curl here negotiates HTTP/1.1 and works; Claude Code uses HTTP/2 and is reset. HTTP/2 could not be tested from this machine (the bundled Windows/schannel curl lacks --http2), but the pattern (HTTP/1.1 OK, HTTP/2 reset) is consistent throughout.
The reset is active (ECONNRESET / RST), not a silent timeout — indicating something terminates Claude Code's connection specifically.
Everything ruled out (does NOT fix it)

The following were each tested and made no difference to the Claude Code failure:

Working directory — was on a mapped network drive (Z:); moved to local C: — no change.
IPv6 — found genuinely broken on this network (curl -6 to api.anthropic.com times out; IPv4 works). Disabled IPv6 on the adapter — no change to Claude Code. (Real issue, but not the cause.)
MTU black hole — found genuinely present: ping -f -l 1472 (1500 B) to 8.8.8.8 = 100% loss; -l 1400 works; LAN handles 1472. Lowered adapter MTU to 1400, then 1200 (netsh interface ipv4 set subinterface) — no change to Claude Code. (Real issue, but not the cause.)
Antivirus (Bitdefender) — disabled all modules → no change; fully uninstalled + rebooted → no change; netsh winsock reset + reboot (to clear leftover LSP/WFP filter drivers) → no change.
Corrupted install — reinstalled the same version (2.1.223) fresh → no change.
DNS / endpoint — pinned api.anthropic.com to its IPv4 in the hosts file, flushed DNS → no change.
Reboots — multiple, throughout.
What DID fix it
claude install 2.1.210

After downgrading the CLI to 2.1.210, a simple hi prompt returns a normal response immediately. Auto-update must be disabled afterward (otherwise it re-updates back to the broken build).

The Claude desktop app "Code" tab is still broken, because it bundles its own newer Claude Code version and there is no straightforward way to roll that back.

Hypothesis

A networking/HTTP-client change introduced between 2.1.210 and 2.1.222 causes Claude Code's API connections to be reset (ECONNRESET) on this network, while HTTP/1.1 tools (curl) to the same endpoint succeed. The network here includes a double-NAT / double-router setup (192.168.1.1 → 192.168.2.1), so the new client behavior may interact badly with a middlebox or with HTTP/2 handling. Older versions are unaffected.

Requests to Anthropic
Investigate networking / HTTP-client changes between 2.1.210 and 2.1.222 (e.g. HTTP/2 usage, TLS/ALPN, connection reuse, keep-alive, retry behavior).
Provide a way to pin or roll back the desktop app's bundled Claude Code version, or ship a fix so the "Code" tab works again.
Consider more graceful fallback (e.g. HTTP/1.1) when HTTP/2 connections are reset by a middlebox.

Steps to Reproduce

How to reproduce (this environment)
On the affected Windows machine, run claude and send any prompt (e.g. hi).
Observe Unable to connect to API (ECONNRESET) · Retrying … looping.
Confirm curl.exe -v https://api.anthropic.com succeeds from the same shell (rules out raw connectivity).
Run claude install 2.1.210, relaunch, send hi → succeeds.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.210

Claude Code Version

2.1.223

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_