Commands cancelled before they start are cleaned up right away#
A shell command that is already cancelled when it is created is now cleaned up immediately
You can now use claude-sonnet-5 with high effort and automatic thinking switched on by default. The high effort option is now labelled Recommended instead of Default. Admins can use the new managed setting allowedProviders to limit which API providers Claude Code may run on. Setting CLAUDE_CODE_DISABLE_WEB_FETCH turns off WebFetch, the tool Claude uses to download web pages. Claude Code can now load relevant memories saved to your account before a turn. The new claude plugin configure command shows a plugin's options and can save values piped in as JSON.
Seven changes are in the build but not switched on yet. A new plugin test command runs a mod's tests and is marked as early access. Waking and retrying background MCP tasks now sits behind a remote switch that is off by default. A report-delivering tool can end the agent's turn, but only when a server switch allows it. The precomputeCompactionEnabled setting still stays hidden unless a server switch turns it on.
Directory sync for cloud sessions is now fully removed. The CLAUDE_CODE_DIR_SYNC_* variables and CLAUDE_CODE_DISABLE_DIR_SYNC no longer do anything. Launching with --cloud no longer offers to sync your folder, and bundle upload still works. On macOS and Linux, CLAUDE_CODE_LEGACY_BUNDLE is now ignored with a warning. Passing unrestrictedPaths to the agent toolset now throws an error. An @-mentioned audio file no longer gets turned into a transcript for Claude.
Written by our agent from the shipped bundle, not by Anthropic.
New ui_* messages let desktop, mobile and VS Code surfaces attach to a session and draw plugin UI, and a gated list_directory request is added
ExtensionsNew claude plugin configure, plugin install --config reaching bundled MCP server settings, and plugin list --data-size for scripts
claude --desktop opens a session in the Claude Desktop appThe new --desktop flag opens Claude Desktop instead of the terminal, and can pick a session with --continue or --resume <id>
ElsewhereThe new managed setting allowedProviders lists the API providers allowed, and Claude Code refuses to run on any other
Claude Code now handles an mcp.connect request that takes the name of an MCP server
plugin test command runs a mod's testsExtensionsWant the reasoning? Read walks the 34 entries that probably matter to you, each one opening to what changed and why.
Read this release → Every row →A shell command that is already cancelled when it is created is now cleaned up immediately
If a command was cancelled before it started, Claude Code now returns a cancelled result instead of starting it anyway
Unclear It is likely but not certain that this is the path that runs hook commands.
When the working folder changes, Claude Code now finds the permission prompt tool's MCP server from the listed tools, not only from its parsed name
Unclear The exact effect a user would notice is not clear.
On Windows, the check for whether a host is this machine now ignores letter case and matches shortened forms of the name
Unclear What Claude Code uses this check for is not clear.
When your settings name a hook event Claude Code does not recognize, the validation error again includes its helpful tip
5 more of these are in What probably matters to you, on page 1.
With no API key, Claude Code can now fall back to a bearer token from the environment when fetching policy limits and remote settings
Unclear Which environment variable supplies the token and how the remote switch decides to skip the fallback are not settled.
The message built after compaction can now hold a second, model-only text, though nothing fills it in yet
Unclear It is not clear where, if anywhere, the model-only text is set.
The plan and auto mode check now also looks at the mode that was active before a rewrite, and a line-reading helper sits behind a server switch
Unclear It is not clear what the newly switched helper controls or which part of Claude Code was reworked.
2 more of these are in What probably matters to you, on page 1.
API usage records now carry an optional fallback_credit field, which reports the outcome of a fallback-credit token a request presented
Unclear It is not clear how the fallback_credit value is used.
The current_time tool now joins a set of tools that are handled separately, and connector routing splits into claude.ai and widget cases
Unclear It is not clear what a reader will see differently when current_time is called.
After a restart, Claude Code fills in the remote host for adopted tool calls that lack one, under certain conditions
Unclear The conditions under which the host is filled in are not stated.
When Claude Code loads settings, it can now reuse a frozen copy of a repository's settings instead of reading them again
Unclear It is not clear when settings get frozen.
Compaction now records which artifact versions had been read, error results pass on the API error, and exit status can report ran_nothing
Unclear It is not clear whether any of these changes are visible to someone using Claude Code.
Claude Code gained a handler that checks a repository and one to four worktree paths, with no known way to reach it yet
Unclear How this handler is reached is not established.
Claude Code's cleanup step now clears a stored record about model access as well as the one it already handled
Unclear It is not stated what the model-access file contains or when the cleanup runs.
Usage records now include an empty fallback credit field, and reporting of an offline sync dialog was removed
Unclear It is not clear whether the offline sync dialog itself was removed or what freezing remote settings does.
One kind of request now says whether it goes to ccr-gateway or api, depending on the kind of session
Unclear The conditions that pick each host, and which users this reaches, are not known.
Requests sent through a Google Cloud client now carry the user agent AnthropicGoogleCloud/JS followed by a version
Unclear Which requests go through this Google Cloud client is not stated.
Usage data for auto mode's permission decisions now includes the total wait time, from server wait to the check itself
If the service does not confirm the --max-turns limit, Claude Code now marks the limit as lost and assumes none is enforced
Unclear It is unclear in which runs this happens and where you would see it.
When Claude Code finds your organisation's policy limits unchanged, it now also keeps two credential-related details
Unclear What these details are used for is not established.
Claude Code now reads the current branch by its full git name and strips the refs/heads/ prefix itself
Unclear It is not confirmed why this was changed or whether any reader will see a different branch label.
Notices such as the slow-worktree hint and the auto mode warning are added to the conversation in a new internal way, with no visible effect
Unclear It is not clear whether the new conversation store is in use yet.
The remote connection can now report when an incoming message was read and track messages that arrive more than once
Unclear It is not clear what decides whether read receipts are offered for a given connection.
The code that hands a turn to another host, and that reports on MCP servers added during a session, appears in reworked form
Unclear This may be existing code that moved, and it is not clear whether any behaviour changed.
The connection to remote sessions can now report when incoming messages are read and delivered, and flush streamed output sooner
Unclear What these new reporting options control for someone using a remote session is not clear.
Claude Code can now read the computer name, using macOS's own tool, the Windows COMPUTERNAME value, or the hostname elsewhere
Unclear Where the computer name is used is not clear.
A list of remote-session request types now includes plugin_connect, and a broken check for tool_register was corrected
Unclear It is not clear what this list of request types controls.
Settings served to a remote session are frozen while a device is attached, and calls are now logged with their outcome
Unclear It is not clear whether this is active yet or what, if anything, a user will see.
Remote sessions now measure how long a startup step takes, and queued-message reports count artifact comments
Claude Code gains internal pieces for remote sessions that have no visible effect yet
Unclear How either new piece is used is not known.
Cancelling now also clears an attach that was being held, and a turn now waits for the session to settle before it starts
Unclear It is not clear what the held attach is or which kinds of session this affects.
Remote tunnel connections gain a header carrying a credential version, and a lock check now notices when the lock file has changed
Unclear It is not clear which feature uses the new header and lock check.
Remote sessions gain bookkeeping for the time of the last full sign-in check before model requests, off unless a server value is set
Unclear What uses the recorded time, beyond keeping track of it, is not known.
Claude Code checks for sessions whose CLAUDE_CODE_SESSION_ORIGIN is claude_ai_chat together with a second condition
Unclear It is unclear what this check controls and what exactly changed in it.
Remote sessions can now send their connection status to the server, not only a worker status
Unclear It is not clear whether the server does anything with the connection status yet.
Remote sessions now note when a connection is a resume and report running background tasks
Unclear Where each of these is used, and whether anything gates them, is not clear.
The agent proxy now logs where its certificate authority comes from and runs a reconcile step when it comes from the environment
Unclear When the certificate comes from the environment rather than being fetched is not clear.
Replies Claude Code passes along through its bridge now say when the session was resumed
Unclear Nothing yet shows what uses the two new names or whether the lists control what is allowed.
When setting up a remote connection fails, the wait before retrying is now worked out by shared code with 25% random variation
Unclear Whether the actual waits differ from before is not clear.
Remote session sync status no longer includes a direction such as "through this computer"
Unclear What displays or uses this direction is not clear.
The two settings that decide whether a resumed remote session wakes for restored background agents and shells are still read, both falling back to on
Unclear It is not clear whether this build changes how these settings behave at all.
Claude Code's usage data now includes how long it took to prepare each request before sending it
Usage data about each request to the model now records extra timing figures about Claude Code's own overhead
Requests that write short summaries of tool use are now marked as tool-free helpers, and usage reports on denials note unanswered prompts
Unclear It is not clear whether marking the summary request as tool-free changes how that request runs, or only how it is labelled.
When a memory file is pinned, the usage report no longer reads the file or includes its pinned state
Usage data sent when the home-directory setup recovers now also records whether recovery came after an announced run
Speed data from non-interactive runs now records each round of model requests separately, not just one total per turn
Unclear It is not clear whether this data is sent for every run or only for a sample.
When login is refused because the sign-in provider is not allowed, Claude Code records that separately from refusals by the organization
Unclear It is not clear what causes a login to be refused with provider_not_allowed.
Claude Code now measures how long it takes to resize an image and to shrink it to fit a size limit
Unclear It is not clear where these timings are reported.
A check in how Claude Code reports where a model restriction came from now always gives the same answer
Unclear Whether this check is newly disabled in this release is not established.
Claude Code now measures and reports internally why the end of a conversation is left out of the prompt cache
Unclear It is not clear where this reporting is sent or whether a switch controls it.
Claude Code recognises SSH-forwarded sign-in when a socket is set and exactly one credential variable holds the value ssh-placeholder
Unclear It is not clear which part of Claude Code acts on this check.
Requests Claude Code sends through its Foundry connection now identify themselves as AnthropicFoundry/JS
When Claude runs tools remotely, an approval that arrives after a stop is now recorded, and the end of the work waits for pending questions
Unclear It is not clear whether this changes anything a user would notice.
On Linux, a new directory listing refuses links, unusual paths and folders blocked by read rules, and caps how many entries it returns
Unclear It is not known which feature uses this listing or whether it is switched on.
The Write tool's instructions no longer carry a special warning about files outside the working directory and simply say overwriting an unread file fails
Unclear It is not clear when the removed wording was shown, since the stock Write description captured under this release is unchanged from the previous one, and whether the tool itself behaves differently.
The code that reads --tools and warns about unknown tool names was moved into one place; the warning text is unchanged
Unclear The typo warning may now be skipped for some names that appear in a particular list, which would be a small behaviour change.
Claude Code now handles streamed replies from Anthropic, gateways and Anthropic on AWS through one shared check, and the Bedrock guard is unchanged
Unclear Some requests now also record a timestamp, and what it is used for is not known.
One of Claude Code's requests now includes a list of built-in tools when one is provided
Unclear It is not clear which request carries this list or what supplies it.
A cloud session running without a window now checks this computer's session records only when a condition allows it
Unclear What decides whether the check runs, and what it means for you, is not known.
When Claude Code passes on session options, it now includes the turn limit and the built-in tool list when they are set
Unclear Which part of Claude Code receives and uses these options is not stated.
The schema text for directory_sync now points to a separate document instead of explaining its states inline
A logged warning about searching other project folders for a session's record now says plainly what could not be determined
When a maxTurns turn limit is set, Claude Code now sends it along with a request instead of leaving it out
Unclear Which request now carries the turn limit is not settled.
Downloading files in cloud sessions now records failure reasons and timings for image preparation, writing and hashing
Unclear Where these totals are sent was not established.
Claude Code can build a browser MCP server that runs a bundled Playwright script with node and sets CLAUDE_TEST_PROJECT_DIR
Unclear It is not clear what switches the browser server on or where it is used.
Three new names appear for connecting to MCP, appending to a session and a plain MCP server call, next to existing ones
Unclear What uses these names and whether anything restricts them is not clear.
The link that lets task code call MCP tools can now also connect to a named MCP server, and one callback's errors are logged and ignored
Unclear How this new connect ability is exposed to users is not clear.
If an MCP server is removed before it finishes connecting, the connection attempt now reports a failure
Unclear It is not clear which parts of Claude Code use these new connection steps.
With MCP_DISCOVERY_CACHE on, only dynamic toolbox HTTP servers in cloud sessions that send two identifying headers can use the cache
The bundled zod library was updated, changing how some tool and MCP schemas are converted to and from JSON Schema
A multipleOf divisor of zero or a non-finite number now produces an error instead of being written into a JSON Schema
Claude Code ships a newer version of Zod, the library that checks data shapes, with small changes to validation and JSON Schema output
The library Claude Code uses to check data formats now recognises MAC addresses and credit card numbers
The library Claude Code uses to check that data has the right shape was updated, with little visible effect
The sandbox's file storage on Linux, macOS and Windows can now list directories and optionally flush files to disk as they are created
Unclear What turns the write-through option on is not clear.
disableNoSandbox gains an extra condition#A sandbox check can now pass on a new condition before it looks at the disableNoSandbox setting
Unclear What the new condition tests, and what passing this check allows, are not clear.
The sandbox now keeps two more lists about symbolic links, though what uses them is not yet visible
Unclear What reads these lists, and so what they change in practice, is not established.
The sandbox's default allowed package registries, such as registry.npmjs.org and pypi.org, now come from a shared list
Unclear It is not clear which hosts the shared list contains.
Startup telemetry for waits on remote feature settings now records whether the fetch started early on a warm wake
Unclear It is not clear what the tengu_cobalt_thicket gate itself controls.
Claude Code now logs and measures how long it waits at startup for the model list refresh and MCP policy settings
Claude Code now keeps a set of startup flags, including --print, --background and --version, that it treats as a group
Unclear How Claude Code uses this set of flags at startup is not shown.
Claude Code now measures how long each part of starting a session takes, for internal diagnostics
Unclear Where these timings are sent is not known.
When a queued command is removed, its log entry now includes an ID, and discarded commands pass through an extra step
Unclear It is not clear what the extra step for discarded commands does.
Background hints marked key-only are now de-duplicated separately, so both kinds can appear in one run
Unclear It is not clear what a keyOnly hint shows in the interface.
Records of queued commands in the session log can now include an ID for the command
Unclear It is not clear which parts of Claude Code fill in the command ID.
When sign-in is lost mid-session, fetching feature settings now treats it as a temporary failure rather than a lasting one
Unclear It is unclear what a user would notice differently when sign-in is lost during a session.
How Claude Code decides whether to wait for feature settings at startup was reorganised, and the wait is now recorded
Unclear It is unclear whether how long startup waits has itself changed.
A feature that followed one server-side switch now follows a differently named one, still on by default when the server sends nothing
Unclear It is not clear what feature this switch controls.
Claude Code now takes its list of core skills from a list sent by Anthropic's server, falling back to the last list it received
Unclear What the checks that empty the list look at is not settled.
Claude Code gained another way to drop bundled skills, keeping only those on a kill switch's survivor list
Unclear It is not clear what condition triggers this filter or what the name check tests.
Claude Code's skills sync diagnostics now record how long listing skills takes and, at startup, how long before the first message can go
On Windows, Claude Code adds a new way to list folders and now includes git's longpaths setting among the git settings it checks
Unclear It is not clear what Claude Code does with the longpaths setting or the new folder-listing function.
On WSL, Claude Code reads the system's mount list to tell whether a path sits on a Windows-side drive, and file checks compare link counts and file IDs
Unclear It is not clear what Claude Code does differently once it decides a path is on a Windows-side drive.
ALLUSERSPROFILE joins a list of allowed Windows variables, and a table fixing Opus 4 and 4.1 output at 8192 tokens was replaced
Unclear What the variable list governs, and where the Opus 4 limits now come from, is not stated.
Claude Code now uses each attachment's recorded role, not a fixed list, to decide which attachments become system reminders
Unclear It is not clear which attachment types are handled differently now, or what the accountMemoryWithheld option does.
A failed download that was not a prefetch now always counts; a prefetch counts only if data arrived or it was too large
Unclear What this check feeds into, including whether a counted failure means the download is not retried, is not clear.
Claude Code's saved result for one access check now stores a status, and a hit is logged as remembered instead of cached
Unclear It is not clear what statuses other than accessible can be stored.
Claude Code's Mantle access check now records an "outdated" result as accessible in its usage data, not just an "accessible" one
Unclear It is not clear what the Mantle probe checks or whether anything besides reporting changes.
Published verbatim by Anthropic for v2.1.285. 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 136 bullets, 40 name something an entry on this page also names, 29 name something no entry here does, and 67 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_DISABLE_WEB_FETCH environment variable to turn off the WebFetch tool
Probably new-host-env-var-claude-code-host-prompt-supersedes-record-a, new-claude-code-disable-web-fetch-env-var-turns-off-the-webf claude --desktop to open the Claude desktop app on the current directory, or on a session with --continue / --resume <id>
Probably wide-ranging-desktop-option-rejects-subcommands-claude, update-refusal-message-names-the-two-directories, resume-of-a-live-session-now-hands-off-with-a-prompt, new-claude-desktop-launcher-for-claude-desktop, new-desktop-cli-option, resume-picker-now-offers-to-open-a-live-background-session, desktop-flag-rejected-when-combined-with-background-or-wor, resuming-a-live-background-session-can-now-attach-to-it-and claude plugin configure <plugin> to show a plugin's options and which are unset, or save new values read from stdin with --values-stdin
Probably new-claude-plugin-configure-cli-command-with-values-stdi <server>.<key>=<value> to claude plugin install --config, so a bundled .mcpb MCP server's own settings can be set at install time and it starts without visiting /plugin → Configure
Probably hint-for-bundled-mcp-servers-that-need-configuration, config-keyvalue-parser, mcp-server-policy-checks-extended-to-plugin-provided-and-url, plugin-install-config-reaches-bundled-mcp-server-config, new-claude-plugin-configure-cli-command-with-values-stdi, config-accepts-bundled-mcpb-server-fields allowedProviders managed setting to limit which API providers a machine may use (Anthropic API, a custom endpoint, Bedrock, Mantle, Vertex AI, Foundry, Claude Platform on AWS, or a Cloud gateway)
Probably managed-policy-read-failure-message, login-screen-honours-allowedproviders-managed-setting, wide-ranging-desktop-option-rejects-subcommands-claude, claude-authenticate-joins-an-in-flight-oauth-flow-and-enfo, managed-settings-helper-merge-doc-allowedproviders-is-a-w, settings-validation-for-allowedproviders, provider-not-allowed-startup-refusal, allowedproviders-managed-setting-restricts-which-api-provide, allowedproviders-policy-enforced-on-login-and-setup-token CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES environment variable to cap re-sends of a non-streaming fallback request that timed out
No entry names this claude -p with CLAUDE_CODE_FORK_SUBAGENT=1: a subagent's own Agent call now runs in the foreground, so the subagent gets the child's result
No entry names this GIT_SSH or in your git config's core.sshCommand
Probably plugin-git-clones-now-read-coresshcommand-and-sshvariant, git-ssh-handling-reworked-for-vetted-clones claude plugin disable and enable with a full name@marketplace id changing a settings entry in another letter case instead of the installed plugin's own
No entry names this set_model request (such as the Agent SDK's setModel) leaving the new model on the built-in output-token limit and auto-compact window until restart
No entry names this @, or all of it when the URL writes its @ as %40
No entry names this /teleport fetches taking over the terminal; these fetches now fail fast instead of asking
No entry names this #96858Worktree fetch leaves ssh reading keystrokes from the terminal Open
-p sessions leaving its tools available
Nothing to match on claude -p --permission-prompt-tool: a background subagent's permission request now goes to the prompt tool instead of being auto-denied
No entry names this claude mcp list and claude mcp get, and the not-found error of claude mcp remove, login and logout, printing line breaks and terminal escape sequences from MCP server names and values
Probably mcp-tool-call-stdio-check-and-secret-masking-in-mcp-get python3 -c, node -e) just because they contain =
No entry names this dontAsk mode: a fork now runs under its parent's permission mode and cannot exit plan mode
No entry names this #95155[BUG] Fork subagent performs write actions despite parent session being in plan mode Open
#95154[BUG] Fork subagent performs write actions despite parent session being in plan mode Closed
claude remote-control --help saying --[no-]chrome defaults to the machine's /chrome setting; spawned sessions keep Claude in Chrome off unless --chrome is passed
Probably sandbox-and-gateway-smaller-changes, headless-cli-argument-classification-list /remote-env reading only the newest 20 of an account's environments
No entry names this claude plugin install or /plugin putting it into an installed plugin's cache or data folder when their ids differ only in ., -, @ or (macOS, Windows) capitals; the install is now refused
Probably hint-for-bundled-mcp-servers-that-need-configuration, mcp-server-policy-checks-extended-to-plugin-provided-and-url, new-claude-plugin-configure-cli-command-with-values-stdi ANTHROPIC_AUTH_TOKEN against the Anthropic API never loading the organization's policy
Probably ssh-placeholder-credential-detection agent(), parallel() or pipeline() call that a workflow script awaits later, or not at all, being treated as an unhandled promise rejection, which could end a background session
Probably workflow-swallows-unhandled-promise-rejections some-daemon &) kept its output open; the hook now finishes shortly after its own process exits
No entry names this #92934Windows: synchronous command hooks wait for persistent descendants Closed
modelTimeoutException and serviceUnavailableException errors showing a raw JSON body instead of the error message
No entry names this /autofix-pr and /schedule saying the Claude GitHub App is not installed on a repository whose install status had not been checked yet
No entry names this /artifacts unlinking its file from the artifact, so publishing the same file again created a new artifact instead of updating it
Probably artifact-publish-tool-prompt-no-longer-lists-artifacts-hint Artifact allow rule ("don't ask again") letting the Artifact tool publish a file outside the working directories without asking; add the file's folder with --add-dir for the rule to cover it
Probably network-permission-dont-ask-again-can-persist-to-user-set /cost and SDK modelUsage reporting a turn under the wrong model when the server answered a refusal with a different fallback model than the client expected
No entry names this /ultrareview when the project folder briefly could not be read
Probably workflows-disable-setting-description-clarifies-scope /ultrareview on macOS and Linux failing to upload the working tree from a git worktree whose per-worktree config sets core.longpaths
Probably workflows-disable-setting-description-clarifies-scope /ultrareview uploads including uncommitted changes to credential files whose name has a colon before the extension, such as server:8443.key
Probably workflows-disable-setting-description-clarifies-scope /ultrareview uploads on macOS and Linux running slowly on some unusual file names, and their credential-file check missing file or folder names with many backup or editor marks
Probably workflows-disable-setting-description-clarifies-scope x or r in NORMAL mode no longer breaks a pasted-text placeholder at the end of the prompt
Nothing to match on .mcpb MCP server that still needs configuration: /plugin, the install message and claude plugin install now say so and point to Configure
Probably hint-for-bundled-mcp-servers-that-need-configuration, mcp-server-policy-checks-extended-to-plugin-provided-and-url, new-claude-plugin-configure-cli-command-with-values-stdi #95683[BUG] For marketplace installs the MCP bundle (.mcpb) user_config is never prompted for Open
/ultrareview refusing to upload a checkout on a Linux volume when a changed file's name has a colon or ends in a dot or space
Probably workflows-disable-setting-description-clarifies-scope CLAUDE_CODE_RESUME_INTERRUPTED_TURN re-running a turn that had ended at --max-turns
Probably remote-session-max-turns-is-passed-to-cloud-sessions-and-i, maxturns-validation-for-cloud-sessions-and-attach, headless-resume-of-interrupted-turn-orphaned-task-flush-is, max-turns-reported-as-lost, cloud-session-flag-table-maxturns-no-longer-unsupported-de /ultrareview uploading a linked worktree of a repository rooted at your home folder in some cases
Probably workflows-disable-setting-description-clarifies-scope claude agents to a background session waiting on a permission prompt sometimes approving the pending command
Probably resume-picker-now-offers-to-open-a-live-background-session claude attach, logs, stop, respawn and rm starting a new session with the command name as its prompt when options came before it, such as from a shell alias
Probably resume-picker-now-offers-to-open-a-live-background-session, resuming-a-session-that-is-running-in-the-background-can-ope, resuming-a-live-background-session-can-now-attach-to-it-and claude mcp list leaving out WebSocket (ws) MCP servers; each is now listed with its URL and health status
No entry names this claude mcp get showing no Type, Command, Args, or Environment for stdio servers whose config entry omits the type field
Probably mcp-tool-call-stdio-check-and-secret-masking-in-mcp-get .claude/settings.local.json allow rules being held back outside a git repository when git's trace2 output is configured
No entry names this /claude-api eval runner scaffold and report builder writing through a symlink or hard link planted at an output file
No entry names this /claude-api eval runner scaffold counting responses cut off at max_tokens in the score averages; they are now marked truncated and counted separately
No entry names this showing as literal text in the terminal when a reply uses it to indent text, such as row labels in a markdown table
Nothing to match on /clear or /compact
Probably idle-clear-hint-now-counts-from-other-sessions-turns /btw side question asked of a session hosted by an app such as Claude Desktop now sees the turn in progress, not only the last finished one
No entry names this ping stream event is now sent every 30 seconds on the Anthropic API, Claude Platform on AWS and gateways
Nothing to match on /resume and claude --resume on a session that is running in the background: they now open that session instead of refusing, and a prompt given with claude --resume <id> "prompt" is sent to it as its next turn
Probably update-refusal-message-names-the-two-directories, resume-of-a-live-session-now-hands-off-with-a-prompt, new-claude-desktop-launcher-for-claude-desktop, new-desktop-cli-option, resume-picker-now-offers-to-open-a-live-background-session, resuming-a-live-background-session-can-now-attach-to-it-and _meta['anthropic/alwaysLoad'] to false stays deferred when its --mcp-config, Agent SDK or plugin server is set to alwaysLoad
Probably claude-in-chrome-can-be-served-by-the-remote-session-host-ov, alwaysload-mcp-option-server-tools-can-opt-out-per-tool timeout with run_in_background, default 30 min, max 2 h); Claude is notified when one is stopped
Probably the-bash-tool-description-was-rewritten, the-bash-tool-description-in-the-command-security-review-pro, the-bash-tool-description-in-the-command-team-onboarding-pro, background-shell-commands-get-a-stop-deadline /ultrareview to run when disableWorkflows is on, unless the machine running the review has it set by its own administrator (MDM or the managed-settings file)
Probably workflows-disable-setting-description-clarifies-scope, workflows-kill-switch-claude-code-disable-workflows-and-dis ANTHROPIC_BASE_URL to use the 1M context window of models that have one (Opus 4.7+, Sonnet 5+, Fable); run /autocompact 200k if your gateway stops at 200K
No entry names this claude -p and Python Agent SDK sessions on third-party providers or with telemetry off to start in auto mode when no permission mode is configured, like interactive sessions; --permission-mode still overrides it
No entry names this widgets to be reserved in cloud sessions and on self-hosted runners: your own server under it, or a close spelling such as widgets_, no longer loads, so rename it
No entry names this /ultrareview on macOS and Linux to leave symbolic refs out when uploading a local checkout; a checkout whose current branch is a symbolic ref is now refused with an explanation
Probably workflows-disable-setting-description-clarifies-scope env to no longer set ALLUSERSPROFILE, SystemDrive, or the CommonProgramFiles variables; set them in user or managed settings instead
Probably windows-env-passthrough-list-extended, allusersprofile-added-to-an-env-allowlist-opus-4-max-token /tasks to fold background work Claude Code runs for itself under one "System tasks" row; press Enter on it to show those tasks
Probably background-tasks-dialog-collapses-self-managed-tasks-into-on /ultrareview on macOS and Linux to require git 2.31 or newer to upload a local repository; checkouts made with --separate-git-dir are now refused instead of being uploaded with an older method
Probably workflows-disable-setting-description-clarifies-scope /ultrareview uploads on macOS and Linux to send a partial clone as a working-tree snapshot on git 2.31 or newer, instead of falling back or refusing when git's version looked too old
Probably workflows-disable-setting-description-clarifies-scope /ultrareview uploads on macOS and Linux to refuse, instead of fetching, a partial clone missing some of its working tree's files on older git versions; a clone made without --filter uploads
Probably workflows-disable-setting-description-clarifies-scope claude mcp get to hide the command, arguments, and environment values of stdio MCP servers provided by plugins; variable names are still shown
Probably mcp-tool-call-stdio-check-and-secret-masking-in-mcp-get /claude-api so it can no longer be run from Remote Control clients
No entry names this /config chrome=true to direct you to the /config panel instead of enabling Claude in Chrome by default; /config chrome=false still turns it off when it was on
No entry names this CLAUDE_CODE_RESUME_INTERRUPTED_TURN set, even with Continue After Reload off
Probably headless-resume-of-interrupted-turn-orphaned-task-flush-is CLAUDE_CODE_SUBAGENT_MODEL_FORCE or an agent's own model)
No entry names this .claude/settings.json turns on
Probably project-settings-can-no-longer-set-some-keys MCP_DISCOVERY_CACHE=1, when set in your cloud environment's variables rather than a settings file, to reuse your connectors' tool lists after a session restart; other MCP servers are no longer cached and connect at startup
Probably mcp-discovery-cache-for-ccr-dynamic-http-servers A model matched these bullets to the GitHub issues they fix, so a link can be wrong.
1 of 27 tool descriptions changed. 1 of 25 tool schemas changed. The appended system-reminder blocks moved: 1 line added.
Claude Code, interactive mode
442 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 87 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?