# Claude Code v2.1.250

> Claude Code v2.1.250, released 27 Aug 2026 (2026-08-27). 15 entries read out of the shipped bundle. Unofficial, and not affiliated with Anthropic.

Web version: https://changelogs.core-directive.com/v/2.1.250

Five entries land the reconnection rework for live artifact watches without switching it on. Lease mode, the flag `tengu_slate_lantern_ember`, defaults to off, so a stock build always takes the long 120 to 240 second backoff tier once a stall repeats rather than the two-tier behaviour the code describes. Failed watch startups can now be classed as temporary outages, meaning a retryable server status or a request that hung at least 10 seconds with no reply, and routed to the slow retry schedule. Close reasons written as `cf_mitigated_<status>` or `upgrade_rejected_<status>` are decoded back into an HTTP status and a Cloudflare flag, where the previous build only wrote those strings. Two end-of-watch notices, one for giving up after repeated failures and one for live updates being turned off, are sent passively with their own advice text.

A disconnect classified as a load shed, meaning a Cloudflare-mitigated close or a 5xx, 429 or 408 upgrade rejection, now deletes the cached token lease instead of carrying it through up to three failed dials, and counts as a stall for backoff purposes. A watch judged to be stalling for a long time waits 120 to 240 seconds between attempts instead of 30 to 60, so a persistently broken watch stops hitting the server every half minute. Giving up is no longer fixed at 30 minutes: each stall run picks its own deadline of 30 minutes plus up to 15 more, so sessions caught in a single outage do not all abandon at once. The notice sent when reconnection is abandoned now says the watch will not reconnect on its own, and arrives passively rather than as an interrupting task notification.

An error handler in the path that arms a live subscription reported an outer captured value rather than the exception it had just caught, which lost the real error. It now passes the caught error through.

## What probably matters to you

Anything you can use today, anything that visibly changes, and anything worth poking at. One line each, open for detail.

### Long-stall backoff band and randomised give-up window for live watches

Stalled watches now wait 2-4 minutes between retries and give up at a random point between 30 and 45 minutes.

**What**

Repeated reconnect stalls now back off 2 to 4 minutes instead of 30 to 60 seconds, and a stalling watch gives up somewhere between 30 and 45 minutes rather than at exactly 30, so clients in one outage do not all abandon at the same moment. The long band is picked when a stall repeats and either the connection was closed by Cloudflare mitigation or lease mode is off, and lease mode (`tengu_slate_lantern_ember`) defaults to off, so a stock build takes the long band.

**Details**

- New timing constants: `longStallMinMs: 120000`, `longStallMaxMs: 240000`, `stallOutJitterMs: 900000`.
- Give-up deadline is now per run: `maxStallMs` plus a random amount up to the jitter.
- A counter event fires once per stall run carrying a startup flag and, when known, `http_status`.
- The numbers are compiled in and copied into state; nothing in this build overrides them remotely.

**Evidence**

`stallOutJitterMs`, `stall_run_started`

- Flag `tengu_slate_lantern_ember`: On for this account, and not off by default (read for one account on one subscription tier against v2.1.250; this account: on, anonymous baseline: on, compiled default: on)
- Area: Artifacts
- Names: `tengu_slate_lantern_ember`
- Tier: You'll notice
- Useful: 3/5
- Signal: 4/5

### Which stall backoff tier is used depends on the lease-mode flag

Repeat stalls take the long 2-4 minute retry tier because lease mode is off in a stock build.

**What**

The reconnect delay picker now has two tiers. The long one, 120 to 240 seconds, is used on the second and later attempts of a stall run when the close was Cloudflare-mitigated, lease mode is off, there is no lease, or the lease will not outlive the next attempt. Lease mode is `tengu_slate_lantern_ember` with fallback false, so without a remote value every repeat stall takes the long tier.

**Details**

- Short tier remains 30 to 60 seconds.
- The timing table is a plain in-bundle constant, never reassigned, so the numbers are not server-tunable.

**Evidence**

`longStallMaxMs: 240000,`

- Flag `tengu_slate_lantern_ember`: On for this account, and not off by default (read for one account on one subscription tier against v2.1.250; this account: on, anonymous baseline: on, compiled default: on)
- Area: Artifacts
- Names: `tengu_slate_lantern_ember`
- Tier: You'll notice
- Useful: 2/5
- Signal: 4/5
- Present in the build but not switched on

## Improvements

### Load-shed disconnects drop the reused token lease

Overload-related disconnects now back off and fetch a fresh token instead of reusing a stale one.

**What**

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.

**Details**

- The shed classification is computed before the stall check in the socket lifecycle handler.

**Evidence**

`...(Ne && { shed: !0 }),`

- Area: Artifacts
- Tier: You'll notice
- Useful: 3/5
- Signal: 2/5

### Persistently broken live watches back off to minutes

A repeatedly failing live watch now retries every few minutes instead of every half minute.

**What**

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.

**Details**

- The long class is chosen on a Cloudflare-mitigated close, when lease mode is off, or when the lease is absent or about to expire.

**Evidence**

`longStallMinMs: 120000,`

- Area: Artifacts
- Tier: You'll notice
- Useful: 2/5
- Signal: 2/5

### Giving up on a stalled watch is spread across a 15-minute window

Stalled watches now give up somewhere between 30 and 45 minutes, not all at once.

**What**

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.

**Details**

- The new jitter constant is `stallOutJitterMs: 900000`.
- The window deciding whether a fresh stall continues the previous run also widened, and now accounts for the long backoff band and twice the 15 second request timeout.

**Evidence**

`stallOutJitterMs: 900000,`

- Area: Artifacts
- Tier: You'll notice
- Useful: 2/5
- Signal: 2/5

### "Connection lost" notice reworded and made passive

When a live watch gives up reconnecting, you get a quieter note instead of an interrupting alert.

**What**

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.

**Details**

- The notifier helper takes an options object with advice text and a passive flag instead of a trailing string.
- What the passive flag does further down the notification queue is not visible from this change.

**Evidence**

`the live connection kept failing and reconnecting has stopped`, `it will not reconnect on its own`

- Area: Artifacts
- Tier: You'll notice
- Useful: 2/5
- Signal: 1/5

## Bug Fixes

### Wrong value reported when arming a live subscription failed

When a live subscription fails to start, the error you see is now the actual failure.

**What**

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.

**Evidence**

`armLiveSubscription: late-park fall-through lost its park`

- Area: Elsewhere
- Tier: You'll notice
- Useful: 2/5
- Signal: 1/5

## In Development

### Failed live-watch startups are treated as temporary outages and retried on the slow schedule

Live artifact watches that fail to start from a server hiccup now retry on the patient schedule instead of burning attempts.

**What**

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.

**Details**

- The retryable-status test returns true for HTTP status 500 or above, 429 and 408; such a status attaches an `unavailable` marker to the startup response.
- A startup error code of `boot_request_error` attaches a `noAnswer` marker, meaning the request errored out with no reply.
- The startup builder records the time just before the request is issued, so a hung startup can be measured; a `noAnswer` result plus elapsed time of at least `stallThresholdMs`, the reconnect timing default compiled at 10000 ms, counts as unavailable.
- The failure helper gained `unavailable` and `httpStatus` options and passes a `stalled` flag plus the HTTP status into the reconnect scheduler.
- All of this sits inside the live-watch feature, which is off unless the `tengu_slate_lantern` flag is switched on for your account.

**Evidence**

`boot_request_error`, `...(i.errorCode === "boot_request_error" && { noAnswer: !0 }),`

- Flag `tengu_slate_lantern`: On for this account, and not off by default (read for one account on one subscription tier against v2.1.250; this account: on, anonymous baseline: on, compiled default: on)
- Area: Artifacts
- Tier: You'll notice
- Useful: 3/5
- Signal: 3/5

### Watch-ended messages can be delivered without interrupting the model

When a watch gives up or is switched off, you get a quieter notice that explains it instead of interrupting.

**What**

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.

**Details**

- The announcer's final argument changed from an advice string to an options object with `advice` and `passive`.
- Both advice strings are new in this build.

**Evidence**

`{ advice: Un, passive: !0 },`, `there is nothing to re-watch while live updates are switched off`

- Area: Artifacts
- Tier: You'll notice
- Useful: 3/5
- Signal: 3/5

### Cloudflare and load-shed disconnects are recognised and reported

Claude Code can now tell when a watch was cut off by traffic protection or server overload rather than a real error.

**What**

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.

**Details**

- Failure telemetry now carries `shed`, `http_status` and `cf_mitigated` fields; before it could only report a fixed stalled marker.
- A shed close is folded into the stalled decision and no longer clears the cached token lease.

**Evidence**

`cf_mitigated`, `/^(cf_mitigated|upgrade_rejected)_(\d{1,3})$/`

- Area: Artifacts
- Tier: Under the hood
- Useful: 2/5
- Signal: 3/5

## Internal Changes

### Shared test for transient HTTP statuses

One rule now decides which server errors count as temporary, so retries behave consistently.

**What**

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.

**Evidence**

`return e !== void 0 && (e >= 500 || e === 429 || e === 408);`

- Area: Elsewhere
- Tier: Under the hood
- Useful: 2/5
- Signal: 2/5

### Version bump to 2.1.250

The version stamp moves to 2.1.250.

**What**

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.

**Evidence**

`VERSION: "2.1.250"`, `GIT_SHA: "2f71b9f41af67bfcea88aef7a9f8a0d406825b99"`, `2f71b9f41af67bfcea88aef7a9f8a0d406825b99`

- Area: Elsewhere
- Tier: Under the hood
- Useful: 2/5
- Signal: 1/5

### Request timeout hoisted into a shared constant

Two 15-second timeouts became one shared value; timing is unchanged.

**What**

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.

**Evidence**

`handshakeDeadlineMs`, `Wxt = 15000;`

- Area: Artifacts
- Tier: Under the hood
- Useful: 1/5
- Signal: 0/5

### Lease-renewal miss telemetry unchanged

Lease-renewal failure reporting works the same; only variable names moved.

**What**

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

**Evidence**

`` `renew_miss_${ee.status ?? "no_response"}` ``

- Area: Artifacts
- Tier: Under the hood
- Useful: 1/5
- Signal: 0/5

### Build stamp repeated across the bundle

The build stamp is copied throughout the bundle and all copies updated together.

**What**

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.

**Evidence**

`BUILD_TIME: "2026-08-27T21:19:23Z"`

- Area: Elsewhere
- Tier: Under the hood
- Useful: 1/5
- Signal: 0/5