You can now set CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS to make a request fail instead of waiting past that time after hitting a usage limit. This release has 272 entries, and 150 of them are usable today. On 1M-context models, a new dialog can offer to auto-compact long conversations sooner. If you accept, the lower size is saved to your user settings. The setting for showing artifacts beside tool results is now CLAUDE_CODE_DAPPER_CUPCAKE, and the old CLAUDE_CODE_SQUISHY_NEWT is ignored. Programs built on the SDK can now block up to 64 MCP tools for a single turn, and they can load more plugins into a running session without a restart.
There are 12 entries in the build that are not switched on yet. Claude Code can build shorter instructions for the Read, Grep and Agent tools, but none are in use. Rewritten instructions for AskUserQuestion, the tool Claude uses to ask you questions, are prepared but not used. A slimmed-down prompt variant sits behind a remote switch that is off by default. Groundwork for routing notifications and for spotting interface freezes is present but changes nothing yet. The browser setup prompt in the desktop app and on claude.ai now uses a newly named switch, so any earlier rollout starts over.
Long-running MCP tasks now wait and retry when the server rejects their ID with error -32602. Before this fix they failed at once. Stopping a background command whose output is saved elsewhere now points the notice to that file. Sending a session to the background mid-turn now keeps pending task notices and a scheduled-job note. Checkpoint commits now skip with a git_error reason when Claude Code cannot tell whether a git operation is in progress. Archiving remote sessions from the session list is removed, and the remove option that replaces it is not switched on in this build.
Written by our agent from the shipped bundle, not by Anthropic.
New CLAUDE_CODE_ATTACHMENTS_IN_USER_DATA variable lets remote sessions keep attachments in a user-data folder, with retention checks and hardened writes
The 'beside tool results' behaviour now reads CLAUDE_CODE_DAPPER_CUPCAKE; the old CLAUDE_CODE_SQUISHY_NEWT variable no longer does anything
The SDK reload_plugins request takes add_plugins to load extra plugin folders without restarting, refused in cloud sessions and under disableSideloadFlags
ExtensionsPrograms hosting Claude Code can set ask or deny rules for individual tools on their in-process MCP servers, at initialize or via mcp_set_servers
ExtensionsA per-turn disallowedTools list now blocks named MCP tools for one turn, feeds deny rules, and suppresses misleading tool-revoked notices
ExtensionsWant the reasoning? Read walks the 38 entries that probably matter to you, each one opening to what changed and why.
Read this release → Every row →Anything you can use today, anything that visibly changes, and anything worth poking at. One line each, open for detail.
Claude Code now tells managed (organization) hooks and env apart from those a person's plugin wrote, including statusLine and fileSuggestion commands
Unclear It is not clear what the extra step does to hooks that are not your organization's, or whether you will notice any difference.
Hosts can get raw MCP results with result_not_for_model, and the original working directory is left out under HIPAA-style restrictions
Remotely managed settings now accept keys such as effortLevel, outputStyle and attribution, and Claude Code's settings report includes the output style
What would fix it?
Smaller changes and internals, grouped as the pipeline found them. Nothing is dropped, it is only further down.
8 more of these are in What probably matters to you, above.
Servers from --mcp-config can now be kept despite the allowlist with a notice shown, and connection options are passed through to them
Unclear The finding does not say what these connection options contain or what sets them, so any effect on how servers connect is unknown.
If your organization restricts MCP servers, Claude Code's own server in remote sessions keeps only its repository tools, without routines or PR tools
The managed CLAUDE.md is no longer always loaded, and the user CLAUDE.md now also needs a user file path to exist
Unclear The condition that now skips the managed CLAUDE.md is not known.
When another program handles sign-in and only a token from an environment variable is in use, Claude Code now still fetches and re-checks policy limits
Unclear Which setups count as a host managing sign-in is not clear.
If the normal security check fails during an administrator's gateway login, Claude Code now checks a customer allowlist before giving up
Unclear It is not known where the allowlist comes from, or whether it holds any entries in this version.
The plugin publish message now says the new version was saved, nothing was sent to your organization, and to publish from claude.ai
If a plugin upload is not its current version, Claude now tells you to make it current and publish it from claude.ai
When MCP servers are filtered by policy, Claude Code now notes the ones it kept and says deniedMcpServers in managed settings removes them
On Windows, Claude Code now reads settings from the per-user registry only when an extra condition also holds
Unclear It is not clear what the new condition checks, so it is not known when registry settings are now skipped.
If your organization defines SessionEnd hooks, CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS is read the organization's way, not straight from your setting
On self-hosted runners, failed session releases now log how many failed in a row and always restart the idle timer
When Claude Code runs without local storage, it no longer reads organisation policy settings from disk and treats them as absent
Organization MCP deny lists now match more forms of a remote server's address, and claude-code-remote is limited to four tools
In diskless sessions, Claude Code now skips the enterprise MCP configuration file and reports none as present
Unclear It is not confirmed that the condition for skipping the file is a diskless session.
Telemetry on tool permission decisions now includes the permission mode, plus extra detail when a setting made the decision
When remote settings come from a custom address, a redirect response now fails at once instead of being retried, and redirect statuses are recorded
Claude Code now sets the host architecture in OpenTelemetry data itself instead of running the standard host detector
Unclear Whether hostname and host ID fields still appear in exported OpenTelemetry data is not stated.
Claude Code now skips looking up the model behind a Bedrock inference profile when you are not using Amazon Bedrock
8 more of these are in What probably matters to you, above.
Claude Code now recognises self-hosted BYOC remote environments and passes CLAUDE_CODE_REMOTE_ENVIRONMENT_TYPE along with related variables
Unclear It is not clear what the list does, so whether the variable is now passed to other processes is unconfirmed.
If the host running Claude Code reports there are no system-wide settings, the folder named by CLAUDE_CODE_MANAGED_SETTINGS_PATH is now skipped too
When git fetch gives up after its retry limit because a ref does not exist, it now raises a separate kind of error
After Claude Code picks which hooks apply, it now runs one more check that is told whether managedHooksOnly is set
Unclear What the new check does with the matched hooks, and whether it can change which hooks run, is not shown.
Whether a hook counts as your organization's, and so gets CLAUDE_CODE_SHELL_PREFIX, no longer depends on the hook's type or its plugin
Unclear It is not clear what the new check tests, so it is not known which hooks now count as the organization's.
Published verbatim by Anthropic for v2.1.301. Text is unmodified from the upstream changelog. Everything else on this page came out of the bundle instead, which is why the two lists don't match.
Of these 51 bullets, 8 name something an entry on this page also names, 15 name something no entry here does, and 28 name nothing specific enough to line up either way. The pairings are made on names both sides wrote down, a flag or a setting or a slash command, so read one as probably the same thing rather than as a fact, and read the middle number as candidates rather than as a miss count.
CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS so unattended retry mode (CLAUDE_CODE_RETRY_WATCHDOG) ends a request at a usage limit that resets too far away instead of waiting
Probably new-env-var-caps-how-long-the-retry-watchdog-waits-on-a-usag claude-api skill, including setting up trace destinations with the ant CLI
Probably tip-for-multiple-sessions-is-hidden-in-some-cases-the-claud % motion to vim mode: jump to the matching bracket, alone or with d/c/y
Nothing to match on rows to a mod's Input: a text field of several lines in the terminal, where Enter submits and Shift+Enter or a backslash before Enter starts a new line
No entry names this classifier_verdict and permission_mode to OpenTelemetry tool_decision events, so auto mode classifier decisions can be told apart from permission rules
Probably auto-mode-classifier-verdict-added-to-analytics, tool-decision-telemetry-gains-permission-mode-and-config-dec classifier_category to OpenTelemetry tool_decision events, naming the built-in rule a classifier block matched, when OTEL_LOG_TOOL_DETAILS=1
Probably auto-mode-classifier-verdict-added-to-analytics, tool-decision-telemetry-gains-permission-mode-and-config-dec WorktreeCreate hook without asking when it may hold unpushed commits
No entry names this fable (and best) model alias erroring on the Anthropic API when the newest Fable model is not available to your organization: Claude Code now retries once on the previous Fable model
Nothing to match on >0q covering the folder trust prompt, and >4m before the shell prompt at exit, in terminals built on VTE older than 0.53 such as MATE Terminal on RHEL 8
No entry names this GOK text appearing in the prompt at startup in some versions of iTerm2
No entry names this #100691[Bug] Stray "GOK" appears in prompt at startup in iTerm2 only Open
#99985Text "GOK" appears in the prompt at startup in iTerm2 (kitty graphics query reply not consumed) Open
#98973Startup input box pre-populated with stray text ("GOK") when project CLAUDE.md/AutoMem files are loaded Open
#95448Kitty graphics probe: iTerm2's i=less OK reply fails the response regex and is typed into the prompt as GOK Open
application-inference-profile running your awsAuthRefresh command and delaying model-switch hooks when the provider is not Bedrock
No entry names this claude auth status or claude plugin list exited in the middle of a token refresh
No entry names this #95822Short-lived commands (`claude auth status`, `claude --bg …`) start an OAuth refresh at init and exit without saving it, leaving the profile with a spent refresh token Open
#102265[BUG] Short-lived subcommand (claude plugin validate) exits mid OAuth refresh: the rotated refresh token is never saved, the next process gets invalid_grant and blanks the login Open
/rewind
No entry names this ask decision not showing which plugin or skill the hook comes from
Nothing to match on /diff not noticing a conflicted cherry-pick or revert in reftable repositories when one of its git checks timed out or failed
No entry names this cd in the same session
Nothing to match on $.model.fork returning nothing-to-fork when called before the first response of a new Claude Code process had finished, as in claude -p runs and cloud sessions
No entry names this /compact and auto-compact briefly freezing the UI while the visible transcript is rebuilt
No entry names this deniedMcpServers to keep those tools off
Probably mcp-server-filtering-now-explains-servers-kept-despite-allow [killed] line in their output file
Probably background-shell-exit-marker-replaced-by-a-structured-call, background-shell-kill-output-file-handling-changes-new-aut .catch handler to run only when the hook throws or times out; events raised beneath the hook's own calls now skip that hook without asking its .catch
Nothing to match on $.env.set to resolve with {} once the variable is set or unset, or with { deny } saying why when the write is not made
Probably hook-env-set-denials-now-return-a-deny-result $$ and multi-line \[ … \] blocks and inline $…$ and \(…\) math
No entry names this A model matched these bullets to the GitHub issues they fix, so a link can be wrong.
No change to the system prompt since v2.1.300.
Claude Code, interactive mode
830 documentation changes were recorded within 24 hours either side of this release, nearest first. The closest 12 are below. They're here because they happened near this release in time. That's not a claim that this release caused the edit, or that the page documents anything in it.
The 36 literal strings found in the bundle, with the number of entries that name each one. Picking one searches for it. A name is here because this build's code mentions it, which is not the same as it working or being finished.
What's wrong with this entry?