Claude Code Re-Sent a Request Your Org Policy Had Already Rejected
2.1.239 fixes a request refused by an organization policy check being re-sent before the refusal was ever displayed. Same release, same root cause: the retry path treats a decision as an outage.
NeuroX AI · August 23, 2026
The Claude Code 2.1.239 changelog fixes a request rejected by an organization policy check being re-sent before the rejection was shown. The control fired. The client retried anyway. The human found out second.
Read the ordering carefully, because it's the whole bug. The policy check returned a verdict, the retry path consumed it as a failed attempt, and the resend went out before anything reached the screen. From the operator's seat there was one request. From the policy layer's seat there were two, one of them after an explicit no.
The same release patches the identical mistake in a second place: persistent retry mode (CLAUDE_CODE_RETRY_WATCHDOG) now fails immediately on organization spend-limit and out-of-credits errors instead of waiting indefinitely for a reset that isn't coming. One bug re-sends what was denied; the other blocks forever on a limit that a wall clock will never clear. Both come from the same missing distinction — a retry loop that reads every non-200 as weather.
This is the failure mode we find in most in-house agent harnesses. Backoff gets written early, against timeouts and 5xx, and it works. Then quota, policy, entitlement, and content refusals start arriving through the same channel, and the loop that was designed for a flaky network is now the thing deciding whether your governance layer means anything.
Transient means try again. Everything else is an answer, and answers don't get retried.