Constant 'API Error / Disconnected / Retry x of 10' on 2.1.268 (macOS 26.2) — auto-mode classifier also timing out

Environment

• Claude Code 2.1.268, terminal CLI
• macOS 26.2 (Darwin 25.2.0), zsh
• Model: Opus 5 (1M context), auto mode enabled
• Session: https://claude.ai/code/session_015y3YMoa29JTPbynYdepzsT
• Time: 2026-09-11 ~11:00–11:30 AEST (~01:00–01:30 UTC)

What happens

For the last ~30 minutes nearly every turn hits API Error … Disconnected … Retry x of 10. Responses drop mid-stream and the retry loop runs repeatedly before a turn completes; several turns needed most of the 10 retries. The session is long-running (large context) but nothing else on the machine has connectivity problems.

The auto-mode permission classifier is failing in the same way. Verbatim tool error returned to the model on a plain read-only find/ls command:

claude-sonnet-5[1m] is temporarily unavailable (timed out), so auto mode cannot determine the safety of Bash right now. Wait a moment and then try this action again.

So both the main model requests (Opus 5) and the classifier requests (Sonnet 5) are timing out server-side in the same window — this looks like an API/edge issue rather than a client network fault.

Expected

Requests complete without a 10-retry loop; if the API is degraded, a clearer signal than repeated disconnects.

Notes

• No ~/.claude/debug logs present to attach; happy to run claude doctor or enable debug logging if told which flag.
• status.anthropic.com not yet checked at time of filing — the user asked for this to be logged immediately.

Filed from within the affected session by Claude Code at the user's request.