[BUG] Retries after "Connection error" fail on fresh sockets for ~3 min (→ ECONNRESET) while new requests succeed instantly — ESC + "continue" always fixes it (macOS 27, 2.1.301)

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?

Since 2026-10-06, sessions regularly end a turn with API Error: Connection dropped (ECONNRESET). The --debug log shows the original request fails within 25–85 ms with Connection error. and no HTTP status. All of the automatic retries then fail the same way, each on a fresh socket ("Stale connection — disabling keep-alive for retry"). The error appears once the 11 attempts are used up, about 3 minutes later.

What I typed next always worked. Across 25 drops, the next message I sent ("continue", "try again", …) succeeded every time, usually within 2–5 s. Five of those messages went out 0 s after the error, right after ~3 minutes of failed retries. Pressing ESC during the retries and typing "continue" also fixes it reliably.

The network is healthy while this happens. During the 00:13:49–00:15:00 incident below, another Claude Code session on the same Mac, network and API IP (160.79.104.10) got 11 successful responses.

Debug log, one incident (2.1.301, native, --debug). It recovered on attempt 8 this time:

21:13:49.680Z [Stall] tool_dispatch_end tool=Bash outcome=ok durationMs=1181
21:13:49.703Z [dispatch] sent anthropic-dispatch-id=v2d
21:13:49.772Z [ERROR] API error (attempt 1/11): undefined Connection error.
21:13:49.773Z [WARN] [dispatch] connection error with anthropic-dispatch-id=v2d; retrying once with anthropic-dispatch-id=v2p
21:13:49.773Z Stale connection — disabling keep-alive for retry
21:13:49.847Z [ERROR] API error (attempt 1/11): undefined Connection error.
21:13:50.531Z [ERROR] API error (attempt 2/11): undefined Connection error.
21:13:51.626Z [ERROR] API error (attempt 3/11): undefined Connection error.
21:13:53.806Z [ERROR] API error (attempt 4/11): undefined Connection error.
21:13:58.726Z [ERROR] API error (attempt 5/11): undefined Connection error.
21:14:06.978Z [ERROR] API error (attempt 6/11): undefined Connection error.
21:14:23.314Z [ERROR] API error (attempt 7/11): undefined Connection error.
21:15:00.456Z [API REQUEST] /v1/messages ...
21:15:01.196Z [API:timing] first byte after 743ms

The earlier incident on the same day (21:06:54Z) had the same pattern and recovered on attempt 6.

Kernel view of the failed attempts (log show, tcp_connection_summary for the Claude Code process during a full 11/11 failure). Each attempt is one short connection with the same shape:

Duration: 0.07 sec  Conn_Time: 0.022 sec  bytes in/out: 1780/5306  pkt rxmit: 0
SYN in/out: 1/1  FIN in/out: 0-1/0  RST in/out: 0/0

TCP connects, there are no retransmits, and no RST is seen in either direction. The connection is closed about 70 ms in, sometimes by the peer's FIN (LAST_ACK) and sometimes by the client (FIN_WAIT_1). Meanwhile, existing long-lived connections from the same process stay up, including one open for 42 minutes.

Pattern from 26 drops (session transcripts, 2026-10-06 → 2026-10-15):

• Every final error came 2.8–3.2 min after the original request, which is the retry budget running out.
• The failing request almost always comes right after a tool result (17×) or after my reply to a question (9×).
• Only one drop happened before 2026-10-06 ("Connection lost mid-response", 2026-09-21).

What Should Happen?

When a fresh request succeeds on the first try, the automatic retry of a failed request should succeed too, without needing ESC + "continue". At least the retries should not keep failing on fresh sockets for 3 minutes while new requests from the same and other processes work.

Error Messages/Logs

API Error: Connection dropped (ECONNRESET)
[ERROR] API error (attempt N/11): undefined Connection error.

Client request IDs of failed attempts (2026-10-15):

• 26a9b192-c368-4bd1-a4de-87993366e71d: first failure, 21:06:54.873Z
• cb04ea7d-bc08-4606-94b2-a88cfee158db: first retry (v2p), 21:06:54.899Z
• c99cf49d-ad84-4778-b178-62c4d65bddf2: attempt 7, 21:14:23.241Z

Steps to Reproduce

It's intermittent, happening several times a day in long sessions with many tool calls:

1. Run claude --debug on macOS 27 (native build) and work normally with frequent Bash tool calls.
2. At some point the request after a tool result fails within ~70 ms with Connection error.
3. Every retry fails on a fresh socket. After 11 attempts (~3 min) you get Connection dropped (ECONNRESET).
4. Press ESC (or wait for the error), type "continue", and it succeeds immediately.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown. There were no drops before 2026-10-06; 2.1.29x was current then.

Claude Code Version

2.1.301 (Claude Code), native install

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

• macOS 27 (Darwin 27.0.0), Apple Silicon, Wi-Fi (en0, MTU 1500), home router, no proxy, no VPN active.
• Ruled out locally: Wi-Fi was stable at the drop times (RSSI −57 to −61 dBm, no disassociation or roaming), DNS was clean, and there was no system proxy. A previously installed Tailscale was removed; drops started before it was installed.
• Possibly related: #94225 (fresh TLS handshakes to 160.79.104.10 reset on some paths, linked to ClientHello size/MTU), #100706 (ECONNRESET at MTU 1500, OK at ≤1492), #100353 (keep-alive disabled after one connection error). This report differs in that the kernel shows no RST, and new requests keep working while the retried request keeps failing.