Claude Code v2.1.250

A disconnect classified as a load shed, meaning a Cloudflare-mitigated close or a 5xx/429/408 upgrade rejection, now both counts as a stall for backoff purposes and deletes the cached token lease instead of keeping it through up to three failed dials.

When a stall is judged long-running, the retry delay is drawn from 120 to 240 seconds rather than 30 to 60, so a live artifact watch that keeps failing stops hammering the server every half minute.

A stalled watch used to be abandoned at exactly 30 minutes. Each stall run now picks its own deadline of 30 minutes plus a random amount up to 15 more, so sessions caught in one outage do not all give up together.

When reconnection is abandoned, the notice the model receives now says the watch will not reconnect on its own and is delivered passively rather than as an interrupting task notification.

An error handler in the live-subscription arming path reported an outer captured value instead of the exception it had just caught, so the real error was lost. It now passes the caught error.

When a live artifact watch fails to start, Claude Code now decides whether the failure looks like a transient server problem rather than a permanent one, and routes it to the slow stall-style retry instead of the fast consecutive-failure backoff. A failure counts as a temporary outage if the server replied with a retryable status, or if the request produced no reply at all and hung for at least 10 seconds. Neither of these markers existed on this error path in 2.1.248.

Two end-of-watch notices, one after repeated connection failures and one when live updates get switched off, are now sent passively and carry their own advice text rather than the default "watch it again if you still need that". The give-up notice tells the agent the watch will not reconnect on its own; the flag-off notice says there is nothing to re-watch. Ordinary stop and other-org notices are unchanged.

Connection close reasons of the form cf_mitigated_<status> or upgrade_rejected_<status> are now decoded back into an HTTP status and a Cloudflare flag. The previous build only wrote those strings. A close counts as a load shed when the status is 5xx, 429 or 408, or when Cloudflare mitigated it.

One predicate now decides what counts as a temporary outage: any status 500 or above, plus 429 and 408. It marks watch-token startup responses as unavailable and, with the close-code decoder, decides whether a failed live subscribe is a retryable stall.

The build stamp advances to 2.1.250 with a new build time and commit hash. It feeds the user agent and metric keys; nothing else changed alongside it.

The two hard-coded 15 second timeouts on the artifact read and the watch-token renew request became one named constant, which also feeds the stall-run window calculation. Same values, so nothing changes for users.

The renewal-miss event still reports the HTTP status or "no_response" and clears the renewable marker; only local variable names moved around it.

The same build-constants object is inlined at roughly 66 places, and all of them updated together. The two builds are stamped under two hours apart on the same day.