You can now set maxProseWidth to wrap Claude's text replies at a fixed column width, while code and tables keep the full terminal width. Claude Code now retries once on the same model when a safety stop cuts off a response. This is on for everyone by default. Set CLAUDE_CODE_DISABLE_REFUSAL_RETRY to turn it off. Cloud sessions can now start from a linked git worktree on macOS and Linux instead of being refused. CLAUDE_CODE_GZIP_REQUEST_BODY_LEVEL sets a gzip level from 1 to 9 for requests Claude Code already compresses. The self-hosted gateway gains store.readiness_grace_seconds, which keeps it reporting ready through short Postgres outages.
This build holds 20 entries that are not switched on yet. Web search has a quick "standard" mode and a thorough "extended" mode, but both stay off unless configured. You will later be able to attach a local git folder to a running cloud session and keep files in sync. That switch is off by default. A dialog that warns how much of your 5-hour usage limit resuming an old conversation would use is built but off. CLAUDE_CODE_REMOTE_TOOLS_SPECULATIVE_CLASSIFIER is now read, but setting it changes nothing yet.
/model can now switch to claude-fable- models that need credits. It shows a confirmation step instead of rejecting them as invalid. Conversations no longer fail when the API rejects old web search results. Claude Code removes those results and retries. Cancelling a remote message that was already delivered now reports the failure instead of claiming success. The startup tip about dynamic workflows and the ultracode keyword is gone, and so is autocomplete for MCP server URL templates after a slash.
Written by our agent from the shipped bundle, not by Anthropic.
Messages relayed from a project's coordinator session now hold back later queued prompts, and Claude can be told one is waiting
Sessions & agentsGateway operators can enable load_test_mode, which sends no requests to the model, returns canned replies and can create test users
Platform & internalsNew store.readiness_grace_seconds keeps the gateway's /readyz ready through short Postgres outages, and a load_test_mode block serves canned replies
Platform & internalsThe post-turn memory config now takes an extract_transport option, "tools" or "text", defaulting to "tools"
Sessions & agentsWhen switching models with /model, a 'credits required' answer for a qualifying model no longer fails as an invalid model
Sessions & agents
ultracode removedElsewhereclaude doctor no longer checks strictPluginOnlyCustomization in managed settingsPlatform & internalsWant the reasoning? Read walks the 45 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.
New code attaches a local git checkout to a cloud session with file sync, and attach can prepare the link while reading the stream position
Unclear It is unclear what the other condition checked alongside the switch requires.
A new WebSearch mode choice, system hint and remote prompt addendum are built in but only switch on through tengu_sleepy_shore or an environment variable
Unclear It is not clear what value CLAUDE_CODE_WEB_SEARCH_FAST_ARG must hold to count as enabled.
Opus 5 leaves the main model list and quick select, Opus 5.5 gets a new description, and picker labels and examples follow the current aliases
Unclear It is not clear which screen this list feeds, such as the model picker or a fallback catalog.
Cloud session creation gains a persistentVolume option that is hard-coded off and sends nothing, and the dir-sync prompt notes it applies to every worktree
Unclear It is not confirmed that every way of creating a cloud session takes this value only from the launch options.
Messages relayed from a project's coordinator session now hold back later queued prompts, and Claude can be told one is waiting
Unclear It is unclear whether ordinary sessions without a coordinator ever reach this path.
Skills, MCP servers and MCP prompts named claude-ai or anthropic-skills are refused, and synced skills gain classes, aliases and namespace rules
Unclear It is not clear what the internal check is or which marketplace qualifies, so it is unknown when this applies.
Resuming a large, long-idle conversation can show how much of your 5-hour usage limit it will use, offering a fresh conversation instead
Unclear Which accounts the server sends this dialog's configuration to is not known.
Artifacts written by others get a new approval-based shared read mode, summary-only reads for other organizations and wider untrusted-content labels
Unclear It is not clear what the further check looks at, or whether reading with approval is new in this release.
SDK hosts now receive session_metadata system messages listing a session's published artifacts, the summary cloud sessions already get
Unclear When Claude Code actually sends this message is not known.
The usage-credits upsell no longer takes its wording from the server; tengu_saffron_ladder now only shows or hides a fixed 'Upgrade to Max' option
Unclear Whether the Fable-limit message itself lost its remotely set subline entirely is not settled.
Memory file writes can stamp a claudeCodeItem metadata block, and code to hide that block on recall exists but never runs in this build
Unclear It is not clear whether anything in the public release installs the hook this code needs.
Gateway operators can enable load_test_mode, which sends no requests to the model, returns canned replies and can create test users
New store.readiness_grace_seconds keeps the gateway's /readyz ready through short Postgres outages, and a load_test_mode block serves canned replies
When a safety classifier stops a response, Claude Code now retries once on the same model unless CLAUDE_CODE_DISABLE_REFUSAL_RETRY is set
Unclear It is unclear exactly which other sessions skip the retry.
The post-turn memory config now takes an extract_transport option, "tools" or "text", defaulting to "tools"
Unclear What the "text" setting actually changes in memory extraction is not known.
The built-in claude-api skill now ships a preserved-thinking migration guide, a causes page and two helper scripts
New, off-by-default recording writes each API request's system prompt, tools and message changes into the session transcript file
Unclear It is not known what reads these records back, or what value of CLAUDE_CODE_ELEGANT_MEADOW turns recording on.
When switching models with /model, a 'credits required' answer for a qualifying model no longer fails as an invalid model
Unclear It is not clear which account condition applies or what the confirmation step shows.
The code that records a desktop-app approval of a settings-file edit is now filled in, but a check that always says no keeps every edit going to /settings-review
Unclear It is unclear whether the permission-card route is reached some other way today.
In auto mode, Claude Code now falls back to Anthropic's server-side classifier for permission decisions when the server sends no value for that setting
When an internal switch is on, background memory extraction also reads the messages you typed and runs as its own agent
Unclear It is not known what controls this switch or whether it is on for anyone.
A confirmation dialog for edits to account-level memory is registered, but only appears once its code is loaded, and that code starts out empty
Unclear Nothing shows what loads the dialog's code, so it is unclear whether any user can reach it.
Starting a remote cloud session now forwards a persistent volume setting from whatever requested the session
Unclear What sets this value, and what it does for a user, is not known.
Claude Code can build a newer device-link proof that lists capabilities, but nothing supplies the list, so the old proof is still sent
Unclear Code loaded separately from the main program could still supply a capability list.
Claude Code's memory extraction can now also be started by a separate feature, but that feature is not included in this build
CLAUDE_CODE_REMOTE_TOOLS_SPECULATIVE_CLASSIFIER is now read, but the code around it always gives the same answer, so setting it changes nothing
Unclear The helper for the remote setting may be called somewhere under another name.
A new environment variable, CLAUDE_CODE_REMOTE_TOOLS_SPECULATIVE_CLASSIFIER, is read, but the check it feeds always answers no, so nothing changes
Unclear What the classifier path would do once enabled is not shown.
Two tool names, list_project_files and read_project_file, now appear in a tool list tied to remote Claude Code sessions
Unclear It is not clear whether this list allows these tools, blocks them or only displays them, or which users reach it.
New allowClaudeInChromeWithManagedMcp policy lets admins permit Claude in Chrome under managed-mcp.json, and blocked setups now explain why
Unclear When the separate Chrome-specific message for the --chrome flag is shown is not clear.
CLAUDE_CODE_GZIP_REQUEST_BODY_LEVEL sets a gzip level from 1 to 9 for requests Claude Code already compresses
Unclear Whether CLAUDE_CODE_GZIP_REQUEST_BODIES can switch compression on without the remote switches is not settled.
maxProseWidth setting caps how wide Claude's text replies run#Set maxProseWidth to wrap paragraphs, headings, lists and quotes at a fixed column width while code and tables keep the full terminal width
A per-model setting and CLAUDE_CODE_FORCE_MID_CONVERSATION_SYSTEM now keep mid-conversation system turns on even when stored settings say off
Unclear Only one extra use of the environment variable is confirmed, so this may be existing code that moved rather than new behaviour.
/artifacts panel works out actions per artifact and adds an actions menu#In /artifacts, each row's actions now follow that artifact's own permissions, and accessibility mode opens an "Actions for …" menu
Unclear How the accessibility mode that shows the actions menu is switched on is not stated.
claude remote-control help now lists the --[no-]chrome option#The help text for claude remote-control now documents --[no-]chrome, which the command already accepted
A new store.readiness_grace_seconds setting keeps the gateway's /readyz check reporting ready for a set time while Postgres is not answering
Unclear The full wording of the new log lines is not known.
Starting a cloud session from a linked git worktree now works on macOS and Linux instead of being refused
Unclear Whether linked worktrees already worked for some users before this release is not known.
Claiming a session through the Agent SDK now also fetches fresh startup information by default, so no separate call is needed
Choosing the upgrade option in the model consent dialog now runs /upgrade straight away instead of just showing a message
Unclear Which users are shown the upgrade option in the dialog is not known.
/config now uses toggles#In /config, the Agents view submenu now shows toggles that change each setting directly instead of a list you pick from
Unclear It is not clear which settings the Agents view submenu contains.
/install-github-app can be stopped part-way and lists what it already did#Pressing Esc during /install-github-app now stops after the current step and reports which steps were already completed in the repository
/status explains when only managed permission rules apply#When an organisation limits permission rules to managed settings, /status now shows a row listing any allowed tools that were ignored
Unclear It is not confirmed that this panel is the one opened by /status.
Claude Code now offers its modifier-Enter send-now shortcut based on what your terminal can actually send, not on a fixed list of terminal names
Unclear It is not clear exactly which terminals gained or lost the shortcut.
The task list footer now shows whatever key you bound to collapse it and hides some hints when the terminal is narrow
Unclear It is not known which hints are left out on narrow terminals.
claude remote-control accepts --debug and -d instead of failing#claude remote-control no longer stops with an unknown-argument error when given --debug, -d or --debug=…
Unclear Whether the flag actually turns on debug output for remote control or is only tolerated is not known.
With CLAUDE_CODE_SIMULATE_PROXY_USAGE set, Claude Code no longer puts caching markers on system messages
What would fix it?
Smaller changes and internals, grouped as the pipeline found them. Nothing is dropped, it is only further down.
11 more of these are in What probably matters to you, above.
Attaching a folder to a cloud session now asks before keeping it in sync, and Claude is told your files replace its own at the same path
Unclear It is not clear which remote switch controls whether this attach-and-sync path is available.
15 more of these are in What probably matters to you, above.
When Claude Code enrolls the computer as a trusted device on its own, it now shows a notice instead of writing only a debug log
Unclear How the interactive screen displays the notice is not settled.
With CLAUDE_CODE_PROJECTS_SESSION set, WebFetch can skip refetching the same page and return a note pointing at the earlier result instead
Unclear Who sets CLAUDE_CODE_PROJECTS_SESSION is not stated; it may be a hosted environment rather than users.
Child sessions Remote Control starts for a project thread now get autoCompactEnabled turned on, behind a new switch that falls back to on
Unclear It is not clear which product creates these project-thread child sessions or how many people use them.
Behind the tengu_mutable_teapot gate, always-ask rules from mcpServerPolicy are removed from what the auto-mode classifier is sent
Unclear Whether ask rules from MCP server policy are still enforced on your machine is not settled.
set_max_thinking_tokens with thinking_display highlights now returns an error and changes nothing when highlights is unavailable, instead of silently falling back
Unclear Only the request's description was checked, not the code that enforces it.
Published artifacts whose link to your organization can't be confirmed now get the same restricted handling as external or public ones
Unclear It is not clear whether the artifacts feature itself is switched on for all users.
Model picker entries for sonnet, opus, fable and their 1M-context variants now come from the bundled catalog, and Fable is no longer hard-coded as "Fable 5.1"
Unclear The display name the catalog currently gives Fable is not stated.
The schema for Claude Code's stream-json output now lists two more message types: a post-turn summary and an active goal message
Unclear It is not clear whether Claude Code was already sending these messages before the schema listed them.
After a folder on your computer is attached to a remote session, Claude Code now waits for that computer to announce itself before reporting which tools are available
Unclear The longest time Claude Code will wait for the announcement is not known.
For attached clones, directory sync now leaves files on the other machine and waits on a switch before going ahead
Unclear What controls the switch, and whether it is a remote setting, is not known.
Cloud sessions no longer get this computer's tools over the device bridge when trusted devices are required or cannot be checked
Unclear What the "off switch" the notice asks you to remove is, and when the bridge connection type applies, is not settled.
Session 'allow' rules for editing files by path pattern are now dropped unless the path is a skill folder in a .claude directory
Unclear It is not clear whether the rule filtering is active in this build.
The bundled Claude Test plugin's read-only explorer agent, used to map an app's code on a first run, differs from the previous build
Unclear It is unclear whether this build adds the explorer agent or removes it, and how the claude-test plugin is switched on.
Claude Code now picks managed-settings sources differently on Windows, macOS and WSL, and a source with only unusable values no longer blocks the next one
Unclear The exact order in which these policy sources now take priority is not settled.
Artifact consent is now tied to your account, rechecked per action, and reads prompted by new-comment notifications need a fresh approval each time
Unclear What extra content is added to a new artifact's result, and which settings make the Artifact tool available, are not settled.
A project's .claude settings can no longer set OpenTelemetry endpoints or turn telemetry on, only off, and a warning points to /status for details
On Windows, Claude Code now resolves AppData, Local AppData, Start Menu, Programs and Startup folders, not just Documents
Unclear What the extra folders are used for once found is not known.
MCP connect, sign-in and disabled-server messages change wording in some sessions, and reconnecting a disabled server now explains it is disabled
Unclear It is not known exactly which kinds of session drop the hint.
Under allowManagedPermissionRulesOnly, allowed-tools in untrusted skill, command and plugin frontmatter no longer grants tool permissions
Unclear Only the setting's description is known to have changed, so it is not settled whether the enforcement itself is new in this release.
The Artifact tool now lets Claude publish straight from a type_url it was already given, skipping quickstart, with design systems in the publish result
Claude Code can now refuse to rewrite a hard-linked file in place, but the general switch is off and only one write path uses it
Unclear Which of Claude Code's file-editing actions uses the path where the refusal is active is not settled.
Cloud session uploads and the private git directory copy now refuse when git attribute or ignore settings come from somewhere they cannot carry along
Unclear Which commands lead to this upload, and whether anything further up limits who reaches it, is not known.
Artifact read_db and write_db now refuse edits prompted by new-comment notifications and fail with an explanation when nobody can approve
Unclear It is not clear whether such a write is refused outright or you are asked to approve it.
In headless and SDK use, Stop, handoff and attachment prefetch requests can now be answered before earlier input finishes processing
Unclear Exactly which kinds of earlier input may be overtaken is not fully settled.
PostToolUse hooks run in managed-hooks-only mode now receive the duration of the tool call; before, none was passed
Unclear It is not known how the duration appears in the input a hook receives.
When a subagent calls a tool on certain MCP servers, the request now includes that subagent's ID, its type, and whether it is an observer
Unclear It is not clear exactly which MCP server qualifies for this extra information.
WebFetch permission checks can now match rules for reading artifacts, and artifact reads can be checked against WebFetch rules
Unclear It is not clear which checks supply the extra context, or exactly which rules cover which tool in each case.
In some repositories, trust now counts only if the top folder itself was accepted, so subfolders may ask for trust again
Unclear It is not clear what the further check tests, so which repositories are affected is unknown.
prompt.fill now says why it did not fill the prompt#Plugin hooks calling prompt.fill now get a refusal reason such as "dialog" or "no_composer" when the prompt was not filled
On Windows, resuming a session that fails with an EBADF read error now explains that security software may be intercepting file reads
Unclear Whether the explanation is shown on screen or only attached to the error behind the scenes is not settled.
An invalid strictPluginOnlyCustomization value in managed settings is now treated as true, and bad payloads are rejected with an error
The error for a repository setting that affects git in a cloud session now says to remove or move it, then start a new cloud session
On Vertex, WebSearch and a related beta header are now enabled by a single comparison against claude-opus-4-0 instead of fixed model lists
Unclear Exactly which Vertex models the new check lets through is not known.
Messages you send through Remote Control now show as "Your message" or "Your message via Slack" in the session inbox
Bad values in managed settings now fall back to the strict choice by one shared rule, and one bad enabledPlugins entry no longer voids the rest
Unclear It is not confirmed that every key's new restrictive fallback matches the one it used before.
When the cloud session reports it has detached from file sync, pulling changes now ends sync and tells you the copies are now separate
Unclear It is unclear under which conditions this check runs.
When Claude Code is launched in diskless mode, it no longer writes debug log files or updates the latest-log link
Unclear It is not clear how a diskless launch is started.
After a sandbox block, Claude may now say /sandbox can change the sandbox settings rather than manage restrictions, depending on the environment
Unclear It is not known which condition chooses the "change the sandbox settings" wording.
Before resending a live artifact, Claude Code now checks the published version and serves the newer copy if it has moved on
Unclear It is not confirmed that this check is called from the artifact resend path.
The rate-limit warning notice is skipped during a grace period, and warnings about the 5-hour limit can be turned off from the server
Unclear It is not clear what starts a rate-limit grace period.
The confirmation after subscribing a session to pull request updates now lists CI green alongside CI failures, reviews and close/reopen
Unclear Only the confirmation text changed in Claude Code; whether CI green events are actually delivered depends on the server.
The ultracode effort label now reads "ultracode · xhigh + workflows · /effort" and points to the /effort command
Tools left off a host's launch list are now described to Claude as switched off for the session rather than refused by a permission rule
Unclear Which apps fill in the list of left-out tools is not known.
Creating an artifact from one of Anthropic's own artifact types now includes a note about which design system to use
Unclear Which action by Claude leads to this new creation step is not stated.
Metadata from MCP servers is still forwarded, but any keys starting with com.anthropic/ are now stripped out first
Restoring an earlier cloud session's directory sync now fails fast on refused links, retries less, and records which laptop branch the work came from
Unclear It is not clear what controls whether this folder sync is active at all.
The remote agent that watches a pull request is now told a passing CI notice means there is nothing to fix
If starting a cloud session from local files is switched off, Claude Code now says the session will start without this folder's files
Unclear It is not clear what decides whether starting from local files is switched off.
The command-line plugin install path can now tell the installer to replace a copy of the plugin that is already installed
Unclear The name of the command-line option that turns this on is not confirmed.
When a host app shows permission prompts, Claude Code only accepts its approval for a request it actually sent to that host, and uses it once
Unclear It is not clear whether the previous build already let each answer count only once.
/focus is now listed on hosted sessions#The /focus command is now marked to be advertised on hosted sessions, where before it was not
Unclear Exactly where being advertised on hosted sessions makes the command appear is not confirmed.
When Claude Code is launched to write nothing to disk, it now skips saving its global and project config and config backups
Unclear What starts Claude Code in diskless mode is not stated.
A Skill(...) allow rule no longer silently approves a skill that uses a name reserved for skills synced from your claude.ai account
Unclear Which name prefixes count as reserved is not settled.
The spend-limit layer records zero cost in load-test mode and now reports failed spend checks and failed spend writes
Unclear Which product ships this spend-limit proxy is not identified.
The promotion slash command now fetches a fresh claim status and explains pending, expired, claimed-elsewhere and already-taken cases
Unclear The name of the slash command and whether it is available to everyone are not known.
A cloud session's sync can now report as detached, meaning its copy and your local folder are no longer kept in step
Unclear What causes a session to become detached is not known.
If Claude calls a synced skill by its former namespace, the Skill tool now suggests the skill's current full name instead of only reporting it unknown
Loading a saved session now skips work on older compacted history, reads in batches, and logs lines it cannot read instead of dropping them silently
Unclear What condition turns off the new deferred reading after a compaction marker is not known.
Directory sync now applies the .gitignore in each subfolder, not just the top-level one, and leaves out a folder whose .gitignore it cannot use
On Google Vertex, WebSearch is now enabled for any model except claude-opus-4-0 instead of only six listed model families
Unclear It is not clear whether the exclusion covers only claude-opus-4-0 exactly or other model names that start with it.
Admin blockedMarketplaces and strictKnownMarketplaces settings are enforced more consistently, including empty lists and child processes
Output from commands run while Claude is responding is now added after the response instead of landing in the middle of it, behind tengu_foamy_spring
Unclear It is unclear whether this behaviour always applies or only when that remote switch is on.
The Bash tool now cleans up after each call, deleting foreground commands' output files unless a background task still owns them
Unclear A second place that does the same cleanup may belong to a different tool.
When no key in a remote managed-settings payload can be applied, Claude Code now keeps the settings already in force and stops retrying
A plugin marketplace refused because its name looks like Anthropic's now fails with a clear error and instructions, not a generic load failure
Under a strict sandbox policy, project settings can no longer exclude commands from the sandbox, and adding an exclusion goes to user settings or is refused
Unclear The debug message lists user settings as trusted too, but it is not confirmed that user settings entries are kept.
Your own artifacts published outside this session are now marked as data, and pages with terminal control characters are never inlined
Unclear It is unclear what controls whether the artifacts feature is available.
A memory file attached to a conversation is now marked as partially shown when part of it was removed, not only when it was too long
Unclear It is unclear what kind of content counts as a removed block in a memory file.
Routine cleanup of old files now includes the access-audit folder and skips or stops on folders that are links or get swapped mid-cleanup
The sandbox now treats the todos, statsig, logs and access-audit folders under the Claude config directory like the folders it already protected
Unclear It is not confirmed that this treatment means blocking writes rather than some other restriction.
In Desktop Code tab sessions, session-only permissions to edit the .claude folder are now dropped unless they cover a single skill
Unclear The exact condition Claude Code checks to decide that a session is a Desktop Code tab session is not known.
When the API refuses a request, the error message now includes a Message ID line after the Request ID
Syncing files to a cloud session now warns when ignore-file rules had to be widened or re-include rules were skipped
Unclear It is not known whether cloud-session file sync is available to everyone or held back for some accounts.
Uninstalling a plugin now fails with a clear message if a settings file still turns it on or cannot be checked
Claude Code now checks a plugin's name before installing it and refuses invalid ones instead of saving them to its install records
A managed deny or ask list whose rules are all invalid is now marked unreadable rather than quietly treated as empty
Unclear What Claude Code does with a list marked unreadable, beyond warning about it, is not settled.
Cloud session file sync now checks a project's git data and refuses, with an explanation, when objects are tampered or trees use an old format
Unclear The size limits for the combined .gitignore files are not known.
When you accept a path suggestion, Claude Code now always inserts @path, or @"path" if it has spaces, instead of a separate / form
Unclear The / form may now be produced somewhere else, so it is unclear whether suggestions that start with / actually behave differently.
claude --cloud refuses repositories that could steer file sync's git#Cloud file sync now refuses repos with a symbolic link inside .git, or git settings forced by a settings file in the repo
When Claude calls a tool with the wrong capitalization, the error now names the correctly spelled tool
Cloud file sync now switches off for a session when a settings file in the repository sets PATH, a loader variable, or the temp or config folder
The dialog for choosing how files are handled in a remote session has been replaced with a new one that knows the repository and its worktrees
Unclear It is not known what the new dialog looks like or what makes it appear.
If file sync is still uploading when you exit, Claude Code now only warns you; the rest goes up with your first message next time
Unclear It is not known what causes a sync to end with the "detached" reason.
Publishing an artifact without an address now reuses one already reserved for it, and large updates get a pre-check before upload
Unclear The size limit and where the reserved slug comes from are not known.
Dialogs asking you to open a link now ignore keystrokes that arrive right after they appear, and such dialogs reserve screen space
Unclear The length of the window during which keystrokes are ignored is not known.
The output of /recap is now limited to a readable line width instead of stretching across the whole terminal
Stopping a running MCP task now tells the worker behind it that it was stopped, instead of only marking the task as killed
Unclear It is not clear what the worker does when it is told the task was stopped.
An invalid value in managed settings now becomes its most restrictive value, and a block with an unreadable deny rule loses its allow grants
Unclear It is not confirmed that the sandbox settings go through this new handling in every case.
When .git/info/attributes stops cloud changes from being applied locally, Claude Code now warns you and says to move the lines into .gitattributes
Unclear It is not settled whether earlier versions showed a different message when this skip happened.
If none of the settings a Cloud gateway serves can be applied, Claude Code now exits and tells you to ask your administrator
The first file upload to a cloud session now handles linked git worktrees, and refuses with a reason when the setup looks unsafe
Unclear It is not clear whether linked worktrees were already turned away elsewhere before, so this may be new support or only added safety checks.
Background requests that get a 503 server error are now retried once in more cases, usually after a short wait
Unclear How long the wait before the retry lasts is not known.
In MCP URL and form dialogs, pressing Enter just after the selection reaches Accept now does nothing for a short moment
Unclear How long the pause after reaching Accept lasts is not stated.
Claude Code no longer walks up your project folders looking for CLAUDE.md files when both project and local setting sources are turned off
Unclear It is unclear whether this has any effect a user would notice.
When Claude sends a tool invalid input, the error now explains every kind of problem, lists a limited number, and counts the rest
Claude Code now refuses a plugin marketplace whose name looks like one of Anthropic's own and tells you how to rename or remove it
Unclear The check that decides a name looks like Anthropic's is not described, only the message shown.
The hover-to-show scrollbar in scrollable areas now works only in fullscreen mode with full mouse tracking, where the terminal reports hovering
Unclear How the scrollbar looks when hover-reveal is off is not stated.
PDF read failures are now sorted more carefully, so a file name containing words like 'password' no longer produces a misleading error
Claude Code now records kitty keyboard support whenever the terminal replies to its check, without needing a second condition too
Unclear It is not clear what the second condition was, so it is unclear which terminals were affected before.
Uploading to a cloud session from a branch with no commits now fails early unless the staged files are empty or match an existing branch
Unclear It is not clear whether an upload that goes ahead because the staged files match a branch starts the cloud session from the right point.
When uncommitted work can't be captured because git stored a folder in an old format, the upload now says so and how to fix it
. repeat now accepts a count#In vim mode, typing a number before ., such as 3., now passes that count to the repeat instead of ignoring it
Unclear It is not settled how the count is applied, for example whether it replaces the count of the original change.
When a remote pull request review is refused for a repository, /code-review now suggests running the review locally instead
Claude Code now runs its own git commands with LC_ALL and LANGUAGE set to C, so their output is in English whatever your system language
Scrolling lists in Claude Code's terminal interface now draw a scrollbar when they run past the screen, unless they turn it off
Unclear It is not clear which lists relied on the old default and so now gain a scrollbar.
Claude Code now loads a plugin's code files in parallel batches, skips duplicates reached through symlinks, and names file-descriptor exhaustion
:#Permission rules that combine with the : prefix form, or put :* mid-rule, now come with a warning explaining how they match
A blanket allow rule for the artifact tool no longer auto-approves its comments action, which now gets its own permission check
Unclear The artifact tool's name as users see it, and whether it is available to everyone, is not known.
Managed settings made only of fail-closed substitutes no longer count as authored policy, and remote managed settings now keep more raw values
Unclear It is unclear what else in Claude Code depends on whether a source is marked as written policy.
A git worktree whose .git/worktrees entry has another device or mount over it is now refused with foreign_mount
Unclear It is unclear which feature runs this check first, cloud file capture or starting work in a worktree.
Artifact tool labels over the length limit are now tidied and cut to size, and the limit comes from a shared constant instead of a fixed 60
Unclear The exact length limit and the wording of the note Claude receives are not known.
The Artifact tool now tries to reserve a slug before uploading, falls back to a server-assigned one, and records the publish path and omitted files
Unclear Which "not found" errors on republishing are now handled differently is not stated.
Git snapshots for cloud sessions now name tampered objects and legacy trees with actionable fixes instead of a generic git error
In cloud sessions a turn now starts on the files already there instead of waiting on a laptop upload, and Claude is told the files are still coming
Unclear It is not confirmed whether your own machine has also stopped reporting that an upload is in progress.
When file sync cannot set up its private folder because of your config home, it now says so and tells you to check it is writable
Publishing the same artifact content again after its live version has changed now fails with a dedicated reason and guidance
Choosing a prompt command that has a URL template from autocomplete now fills in the command followed by a space, like other commands
Unclear It is not known whether the removed behaviour was ever switched on for anyone.
Path checks now treat "My Documents", "Application Data" and "Local Settings" as the same as Documents, AppData\Roaming and AppData\Local
Unclear Which checks, such as permission or deny rules, use this matching is not known.
Cloud file sync now names two failures: nested ignore files too large, and a linked worktree's git files changing while read
When earlier work cannot be restored in a remote session, Claude Code can now say a later upload could not be read from storage
Installing or switching a plugin now stops on a stale listing, an unrecordable plugin id, or fewer installed copies than shown
Unclear When the listing counts as not fresh, and when the fewer-copies case arises, is not settled.
Workflow keywords in the prompt keep their highlight color but no longer have a shimmer animation
Unclear What a related prompt-box border decoration that was also removed used to display is not settled.
Lists in menus and settings now check at click time whether they are disabled, and ignore mouse clicks while they are
Unclear It is unclear when the scrollbar is now hidden, and which dialogs use the new list that fits its height to the screen.
Two places that wrap text before it reaches Claude now return nothing when the text contains control characters
Unclear It is not known which tool or feature uses these two wrappers.
The message for checkouts that cloud file sync skips now lists linked worktrees only on Windows or WSL, alongside submodules and separate git folders
Unclear Only the message changed visibly, so it is not confirmed that linked worktrees outside Windows and WSL are now actually synced.
In GitHub Actions, the sandbox now works out the .claude folder from three starting locations and write-protects each copy it finds
Unclear It is not clear which three starting locations are used.
Text longer than 2000 characters is now trimmed, keeping the start and adding a note saying how many characters were left out
Unclear It is not established what this trimming feeds into, so its link to title generation is a likely reading only.
When Claude Code runs inside Claude Desktop, the built-in health check now says updates come with Desktop instead of looking up the latest version
Unclear Which command or skill runs this diagnostic prompt is not stated.
The settings schema now explains that sandbox.excludedCommands is a convenience, not a security boundary, and that its patterns merge across sources
When no one can answer a permission prompt, the refusal to read an artifact now separates unconfirmed ownership from someone else's artifact
Unclear It is unclear exactly what condition triggers the new refusal and how the hook-based ask for WebFetch is decided.
Folder sync now looks for planted shortcuts in more parts of the git folder and treats a missing object store as a replaced one
Unclear It is not clear whether the earlier version already checked some of these locations in another way.
Which environment variables from a project's settings get dropped now depends on what your launch environment, flag and policy settings set
Unclear It is not clear whether this makes Claude Code drop more project environment variables or fewer.
Two upload failures, a file that changed mid-check and an old-style tree, now show the unreadable-file message instead of a git error
Unclear It is not clear what produces the legacy_tree failure.
Display names on incoming messages are now cleaned of brackets, angle brackets, quotes and unusual symbols before reaching Claude
Unclear It is not clear which messaging feature this applies to or whether it is switched on for everyone.
The scan that keeps a repository's settings inside it now refuses paths that reach a network location through a symlink
Unclear Which commands or environments run this scan is not confirmed.
A hook tied to a project folder that is no longer a usable working copy no longer runs and shows a warning
Unclear How a hook becomes tied to a particular project folder is not known.
When the policy server sends settings but none of them can be applied, the error now says so and asks your administrator to fix them
The warning Claude sees on artifact type pages now says the page was written by the type's publisher, not you or the user
Plugin uninstall now adds notes about saved data that was kept and older local settings that were skipped
Unclear The exact wording of the new notes is not known.
When a folder's starting snapshot for the session is missing from this machine, Claude Code now says so in plain words
Unclear What causes the starting snapshot to go missing is not stated.
If your consent to folder sync is withdrawn, syncing now stops straight away instead of waiting for a second check
Unclear What removing the upload heartbeat changes in practice is not settled.
When a command's output is lost, Claude Code can now tell you your disk quota is full or the disk is out of space, and suggest a fix
Allow and deny rules for skills now also match a second spelling of the skill's name, and so do per-skill overrides
Unclear What the second spelling is, for example with or without a prefix, is not stated.
After fetching managed settings, Claude Code can now reapply them, on success or on a failure that carries rulings
Unclear The condition that triggers the reapply, and what exactly gets reapplied, are not stated.
The message shown when file sync stops now says neither the cloud copy nor your local folder loses any files, and they become separate copies
Unclear Only the message was seen to change; whether the cloud copy is actually no longer emptied is not confirmed.
When text is put into the prompt box from your input history, each tab in it is now replaced with four spaces
Unclear It is not confirmed that this applies only to recalling earlier input rather than to other ways text is placed in the prompt.
When an MCP server asks for your input, pressing Enter on Accept too soon after it gets focus no longer submits the form
Unclear The length of the waiting period is not known.
Directory sync now refuses snapshots with tampered or old-format git data by name and copes better with file names that are not valid UTF-8
Directory sync now says when it skipped a .gitignore because the file was busy or because it holds too many rules
A turn held for over a minute by directory sync is now always reported as waiting on the sync service, never on your machine's upload
Unclear What the new "unsent" outcome means in practice is not settled.
When checking a plugin, Claude Code now refuses paths from a repository that lead to /net, /net/<host>, or /Network or /home on macOS
Policy settings can now switch on certain features only when they come from an administrator-owned or parent-supplied source
Unclear Which features are switched on this way is not known.
When a linked worktree is refused, the advice now says to start from an ordinary clone of the repository
Unclear Which feature shows this message is not identified.
Text that imitates Claude Code's own tool-result, tool-error or sandbox tags is now neutralized before Claude reads it
Unclear It is unclear exactly which kinds of text pass through this cleaning, and whether some of these tags were already caught by an older rule.
Installing a plugin with a malformed id now explains which part is invalid, which characters are allowed, and who can fix it
Summarized thinking blocks now show a faint " · summary" hint, or none on some models, rather than " · summarized"
Unclear Which models declare the capability that hides the hint is decided by the server and is not known.
When working out a file's real location, Claude Code now treats macOS special paths such as /.vol and /.nofollow the same way it treats network paths
Unclear It is not confirmed that this was done to close a way around Claude Code's permission or sandbox checks, or which checks depend on it.
/hooks where that menu is missing#When the /hooks menu is not available, the settings skill now tells you to start a new session or check the settings file instead
Unclear It is not clear exactly which environments count as having no /hooks menu.
Two new messages explain why a Windows computer's tools aren't offered to a cloud session, and the profile-folder message names more folders
Claude Code now refuses to accept a proposed settings change in place when the settings file is hard-linked to another file
Unclear What Claude Code shows you when an accept is refused as hard_linked is not known.
Installing a plugin now fails with an error if any plugin it depends on has an ID that doesn't pass the plugin ID check
Unclear The exact wording of the error message shown is not known.
In a diskless launch, Claude Code now keeps the MCP discovery cache off for the whole session regardless of other settings
Unclear How a user starts Claude Code in diskless mode is not known.
Folder-sync sessions now say plainly when their starting snapshot is gone, instead of reporting it as a read error
The "Unknown command" message now shortens names at 200 characters instead of 512, and the check that follows runs more often
Unclear It is not clear whether dropping the alias step changes which commands Claude Code recognises.
claude plugin install in replace mode can now fail with "replace_unavailable"; other refusal codes are mostly switched off
Unclear It is not clear what condition makes an install fail with "replace_unavailable".
Sessions attached to a local folder can now explain why file sync stopped, such as mismatched branches or sync being off
Unclear It is not clear where these explanations are shown.
When a synced skill loses its name to another command, the message now says "Anthropic's skill" where it applies, not always "your claude.ai skill"
In cloud sessions, Claude Code's proxy now leaves the certificate store alone when it already trusts the right certificates, not deleting and re-adding them
When an API error carries no request id of its own, Claude Code now reads one from the x-amzn-requestid response header
Files Claude changed in a cloud session are not applied locally when .git/info/attributes can't be read normally or could switch off git-crypt or git-lfs
Unclear It is not clear which sync paths use this check, or whether it is switched on for everyone.
One row of the Clawd mascot logo now has three spaces between its block characters instead of one
Unclear Which version or pose of the logo this row belongs to is not known.
2 more of these are in What probably matters to you, above.
If the API rejects old web search results in a conversation, Claude Code now removes them and retries instead of failing
Cancelling a message the remote client already delivered is no longer reported as a success, and the case is logged
In vim mode, commands such as dd, p, o and >> now act on the actual line of text rather than the wrapped screen row
Unclear How the new join for J in visual mode differs from the old one is not confirmed.
When a slow local MCP server rejects the first connection over a version mix-up, Claude Code now asks again on the same process before restarting it
Output from certain slash commands typed while Claude is still working now appears after the turn ends instead of in the middle of the reply
Unclear Whether every kind of command that runs mid-turn has its output held back is not confirmed.
:* in the middle no longer fail to load#Bash permission rules like Bash(foo:bar), with : not at the end, are no longer rejected as invalid
Unclear How a rule with :* in the middle is now matched against commands is not established; it may be read as a literal colon followed by a wildcard.
Output from a local slash command typed while Claude is responding is now added after the current turn, not spliced into it
Unclear The exact point at which the queued output appears has not been confirmed.
Assistant messages assembled mid-stream are marked isUnmetered and left out of context, token and cost figures until the real message arrives
Unclear It is not clear which situations produce these placeholder replies in the first place.
A stale OAuth refresh lock left by a crashed Claude Code process can now be removed once its recorded owner is proven gone
Unclear It is not verified whether the check for a gone process behaves differently on Windows.
Claude Code now restores the terminal's paste mode along with its other modes, so pastes are not read as typed keystrokes
Unclear What causes Claude Code to restore the terminal modes is not settled.
Changing project or local settings in a session that has no working directory now fails with an error instead of quietly doing nothing
Unclear Which kinds of session run without a working directory, such as headless or remote ones, is not confirmed.
When Claude Code saves the conversation record, it now keeps waiting until all queued writes are written, including ones added during the save
When a session's cost figures are saved and restored, token totals now come back along with the cost
Unclear Where the restored token count is shown is not known.
Pulling a queued prompt back into the input now adjusts the cursor position when vim mode is in NORMAL mode
Unclear The exact adjustment made to the cursor position is not confirmed.
In /plugin, removing a plugin that failed to load now uninstalls it even when settings still have it switched on
Unclear It is unclear what happened before when settings still enabled a failed plugin you tried to remove.
Cancelling /feedback while it saves a feedback bundle to disk now stops the save instead of reporting a late result
If installed_plugins.json cannot be read again during an uninstall, the plugin's options, secrets and data folder are kept
Unclear It is not confirmed which command option produces the machine-readable output.
3dd do nothing on the last line#In vim mode, a line command with a count above one, such as 3dd, now does nothing when the cursor is on the last line, as in Vim
Unclear Whether using the cursor position also changes where the cursor ends up after these commands is not established.
In vim mode, leaving insert mode now moves the cursor back one whole character, so emoji and accented characters are not split
The GitHub Actions setup flow now checks for cancellation before each step and keeps a record of the changes it has already made
Unclear It is not clear where, or whether, the progress log is shown to you after a cancelled or failed run.
Skill overrides now also check a synced skill's alternate names, and setting one from the skill view saves it under the main name
Unclear It is not clear what the other spellings of a synced skill's name are in practice.
r with a count no longer overwrites past the end of the line#In Vim mode, a counted r such as 5rx near the end of a line now does nothing instead of overwriting the line break
On WSL, Claude Code now matches paths against every mounted Windows drive and explains when it cannot read the mount list
Approving an artifact watch now records the approval correctly instead of replacing Claude Code's internal state with unrelated data
Unclear It is not clear when this approval step runs or which part of the state it is stored in.
Claude Code's terminal display now tracks lines scrolled off screen and redraws everything when a change touches them
Unclear The exact display problem this fixes has not been confirmed.
Claude Code now treats the API's "invalid data in redacted_thinking block" error the same way as a thinking signature error
Unclear What Claude Code does after classing the error, such as removing the thinking and retrying, is not confirmed.
In vim mode, 3J now joins three lines instead of four, matching Vim, and the cursor lands at the last join
Running quickstart for a second artifact type in one session no longer overwrites the record of the first
Unclear What Claude Code does with this record is not stated.
A queued slash command is now handled with the knowledge that modules which add commands may still be loading
Unclear It is not clear whether such a command is now held until loading finishes or sent on in some other form.
In print mode, a second copy of the answer to a waiting permission prompt is now dropped instead of being applied again
When several requests are refused as not signed in at once, every one of them now receives the same outcome from the single sign-in refresh
Unclear It is not settled whether this actually changes the login prompts a reader sees.
Shifting lines left or right in the input box now leaves empty lines untouched instead of giving them two spaces
Unclear It is not confirmed which keys or editing mode trigger this shifting.
With CLAUDE_CODE_REPO_CHECKOUTS unset, Claude Code no longer assumes a checkout when it cannot work out the current folder
Unclear Which everyday uses of Claude Code reach the case where no working folder can be found is not known.
On WSL, a Windows-side managed policy that is broken or unreadable now wins instead of falling back to the Linux managed-settings.json
Unclear It is not clear exactly which policy source outranks the Linux file when inheritance is off.
When drawing a list item, the terminal renderer now skips blank space at its start and processes quoted passages inside it before drawing them
Unclear It is not clear what processing quoted passages inside list items now get, and so exactly how they look.
Claude Code now creates the folder for a lock file before creating the lock file itself
Unclear It is not clear which feature this lock file belongs to.
Claude Code now splits error output on Windows line endings as well, so that output is classified instead of lumped under "other"
Unclear It is not clear which program's error output this classification applies to.
14 more of these are in What probably matters to you, above.
The claude.ai skills sync interval and retry backoff are now set remotely, and a new shared-round mode lets sessions reuse each other's skill listings
Unclear It is not confirmed whether the 10-minute fallback matches how often skills were synced before.
The directory-sync journal now keeps links-stopped and basis-branch state, and sync can end with a 'detached' reason when peerKeepsFiles is set
Unclear It is not clear when the option is turned on, so it may not apply to anyone yet.
One time-limited API request can now pass its remaining time as a limit on rate-limit waits, when a setting is turned on
Unclear It is unclear which request this applies to and which configuration supplies rateLimitFailFastEnabled.
New file-sync checks for unrecognized working copies and alternate repository roots ship in this build but cannot trigger, because their detector is a stub
Unclear It is not fully certain that the detection always reports nothing in this build.
Remote SDK sessions can prepare attached files while still waiting to take over a turn, behind a switch that is off by default
Unclear What exactly happens to the attachments while the handoff is still waiting is not known.
Loaders for plugin commands, skills and agents now first check for content handed in from outside, but nothing in this build ever hands any in
Unclear The slot might still be filled in some way that was not found.
Claude Code now checks for one more settings source for hooks and allow rules, but in this build that check always adds nothing
When switched on remotely, a first message over 2000 characters is cut to its first 1500 before Claude Code generates a session title from it
A warning for marketplace entries whose names look official exists in the code but cannot trigger in this build
Unclear It cannot be fully ruled out that the new warning is reachable in some case.
The Discover plugins screen now calls a load-time reporter, but in this build that reporter does nothing
3 more of these are in What probably matters to you, above.
Whether tool inputs are repeated back to the API is now decided once per session, and resumed sessions keep that and other prompt-format choices
Unclear When the server sends no value, nothing appears to be saved and echo stays unset, and what that means in practice is unclear.
Claude Code now sends an atisPin value for feature-flag evaluation and re-fetches flags when it changes, behind the new tengu_serene_teacup switch
Unclear It is not clear where the "atis" value comes from or what it stands for.
Notifications coming from a session inbox are now labelled as their own kind of task notification
Unclear It is not clear whether anything you can see depends on this new kind.
Buttons built with Claude Code's JSX UI toolkit now accept a variant of "primary" or "secondary", and any other value is an error
Unclear It is not clear who can use this toolkit, for example plugins or generated artifacts.
When the advisor is enabled, forked agents now get a one-time instruction never to consult the advisor tool, except for compaction forks
Unclear One further condition on whether the advisor counts as enabled was not resolved.
The built-in Claude Test plugin's agents cannot use a browser tool named claude_test_show_run; only the main conversation may call it
Unclear It is not clear what the show-run tool does, or whether the Claude Test plugin is available to everyone.
Skills synced from the server now carry creatorType, which decides whether a skill counts as Anthropic's or the account's
Unclear It is not clear where this origin is shown to users.
Hooks defined in settings can now carry the project folder they belong to through to the moment they run
Unclear It is unclear where the project folder gets attached to a hook, or how running the hook uses it, for example whether it sets CLAUDE_PROJECT_DIR.
The Skill tool now loads a skill's definition first when the skill supports it, and turns down a skill whose definition still is not loaded
Unclear Whether any skill in this build actually provides its definition this way, so that this path is reached, is not settled.
The signal for being in a usage-limit grace window now follows the latest main or subagent response and clears when no grace use is reported
Unclear Only the description of the signal changed visibly, so it is not confirmed that the code computing it changed the same way.
Fullscreen, mouse and surface mode switches are held and written with the next drawn frame, and flushed when leaving fullscreen
Unclear Whether this visibly reduces flicker or tearing in fullscreen views has not been confirmed.
A tool call left waiting on a permission now runs only once Claude Code confirms its saved record was deleted; otherwise it is held back
Unclear The check is skipped in some situations, and it is not clear how often that happens.
Claude Code now records message-format settings before preparing tool details, and reorders attachments in bare or CLAUDE_CODE_SIMPLE mode
Unclear The condition that switches this step on and its visible effect are not known.
A new check writes and deletes a small test file next to a task's output to confirm the location is writable
Unclear Where this check is used is not known.
A documentation table bundled with Claude Code now groups rejected redacted_thinking data with thinking signature errors
When Claude Code fills in a prompt on request, its reply can now also carry a refusal message
Unclear Which request this reply answers, and which program reads the refusal, is not known.
A session setup option has been renamed from replPrimitiveGrant to primitiveToolGrant
Unclear Whether this option is part of the public SDK or only used inside Claude Code is not known.
The SDK result format gains an optional internal field timing each step between a message arriving and being queued
Unclear It is not clear whether Claude Code currently fills in this field.
Sampled checks of SDK output against its expected format now wait until the first result is written and the session is idle
Requests to count tokens now go through the client's own fetch function when one is configured, rather than always the default
Unclear Where the custom fetch function comes from, and so which setups are affected, is not settled.
In headless mode an approved parked tool call now runs only after its saved record is confirmed deleted; an env var skips the wait
Unclear It is not confirmed that earlier builds ran approved parked calls without waiting.
Session snapshots now save the request layout, raw tool-input echo choice and context placement, and restore them when you resume
Unclear What each saved layout field changes in the actual request sent to the model is not settled.
When third-party transcripts are excluded, session history now strips API request lines and replaces unreadable lines with a size-only placeholder
Unclear Which feature collects this session history is not stated.
A remote run's status can now say it was detached because the machine stopped, instead of only saying it ended
Unclear It is not clear where this status is shown to a user.
A per-model reminder setting now starts from the value recorded in the session, and a server value can still replace it
Unclear It is not clear what the reminder is or exactly when the recorded value is used.
Chain reports can now record "earlier" as a stop reason, and one kind of message block now caps how wide its text runs
Unclear It is not clear which Claude Code feature produces these chain reports, or which message block gets the narrower text.
Reading transcript files now reports the original error even if closing the file also fails, and can read a chosen part of a file
Unclear Whether this has any effect a user would notice is not known.
Claude Code's Git snapshots now tell Git to ignore its cached summary of folders and build the snapshot from the files themselves
Unclear It is not clear which Claude Code feature uses these snapshots.
Claude Code's git attribute check can now read only committed .gitattributes files, ignoring your personal and system-wide attribute settings
Unclear Which parts of Claude Code use the new committed-only mode is not stated.
When checking whether a Git folder has moved, Claude Code now also compares its creation time, so a recreated folder counts as moved
Unclear Which feature relies on this moved check is not known.
One of Claude Code's own git diff calls now always sets colour, path prefixes and submodule display, so your Git settings cannot change it
Unclear It is not clear which feature uses this diff, or which of these options it already passed before.
When Claude Code writes a git config file for a copied repository, it now quotes values and writes valueless keys on their own
Unclear Which Claude Code feature writes this config, and where the second group of entries comes from, is not known.
When loading plugins from a folder, Claude Code now reports once per run whether the desktop wrapper marked their reserved skill names
When a request to the API is gzip-compressed with a set level, Claude Code's usage reports now record that level
Unclear It is not known where the compression level is chosen or whether a setting controls it.
Claude Code now logs a step-by-step timing breakdown for queued messages, plus whether a read-ahead option was on and used
Unclear It is not clear which setting switches the attachment read-ahead on.
Claude Code's usage report on held tools now also counts the ones left out at launch
Unclear It is not clear which tools count as held or as left out at launch.
Claude Code's usage reports now label "Could not write image" failures separately instead of grouping them under "other"
Unclear It is not confirmed which helper tool produces this error, though the wording suggests a PDF-to-image converter.
When republishing an artifact, Claude Code can now save an exact copy of the current HTML and tell Claude to build the next version from it
Unclear Which remote switch controls this check, and whether it is on by default, is not known.
Guidance for installing a design system now says to send tokens.json in the first call that publishes the deliverable
When an artifact's source cannot be shown or saved, Claude is now told a re-read may ask for your approval and not to republish from a partial copy
When Claude resends refused artifact content after confirming the live version, it is now told the next resend will be published
Unclear Whether the artifact feature is available to every user is not known.
A message filter that already dropped some memory context now also removes system messages marked persist: false
Unclear Which part of Claude Code uses this filter is not stated.
If building session context fails, Claude Code now keeps the conversation's existing delivery format instead of switching
Unclear It is not clear what the announced and inline formats look like in practice, or whether the one remote switch is the only thing controlling this.
When a request branches off a conversation, Claude Code now marks the cache just before the point where it diverges
Unclear It is not clear which requests supply the point of divergence, so how often this applies is unknown.
In cloud sessions with file sync, saving a capture point now queues an upload and reports it as queued instead of waiting for a snapshot
Unclear It is not known which sessions have file sync turned on.
Claude's note about when command output from an attached machine shows up locally now always uses the "not necessarily by your next step" wording
Unclear It is not clear which sessions used to receive the "never within one script" wording.
When a folder's .gitignore can't be read because it is busy, Claude Code now treats that as temporary, not as unreadable
Unclear It is not clear what Claude Code does differently with a momentary problem, such as whether it tries again.
The remote switch for the model catalog now has separate settings for serving and publishing it, and still falls back to off
Unclear Which visible parts of Claude Code the two "published" readers drive is not settled.
Claude Code's per-model capability details now include whether a model supports extended thinking and whether it accepts a temperature
Unclear It is not clear what uses this record or whether any user sees it.
Claude Code now rejects its built-in model list as stale if it maps a model family to a different model than this build
Unclear It is not clear what Claude Code uses for the model list after rejecting the stale one.
A plugin agent can now fetch its definition at launch, and Claude Code shows an error if that definition does not match
Unclear It is unclear whether any plugin agent actually uses this loading path yet, or whether it is switched on.
When fetching the plugin catalog, Claude Code can now ask for details of each plugin's components and keep them
Unclear It is not clear which part of Claude Code asks for these component details.
The screen for managing installed plugins gains an empty key handler that nothing calls yet
Unclear Whether anything calls the handler was only checked approximately.
Claude Code now records how many permission rules were dropped when the rules given to auto mode's safety check exceed a size limit
Unclear Whether rules were already being dropped before, or only the record of it is new, is not settled.
When permission rules are too large and get trimmed before an automatic permission check, Claude Code now reports sizes and counts
Unclear It is not clear where this report is sent.
The Edit tool now passes the updated file through the same cleaning and shared-item check that memory saves use before writing it
Unclear Which files this applies to, and whether the earlier transformation already did the same job, is not clear.
Saving to a personal memory store now merges the new text with a related shared item, re-checks it, and then updates that shared item
Unclear What the shared item is, and who else can see it, is not settled.
The prompt box now records whether its text was typed key by key and passes a marker along when such a prompt is sent
Unclear It is not known what Claude Code does with the marker after the prompt is sent.
The part of Claude Code that shows queued messages gains connections to a per-session store that is always empty, so nothing changes
Unclear What the store would hold is unknown.
The request Claude Code sends to link a device to a remote tool host no longer includes an issued-at time and is now laid out as separate fields
Unclear What the binding request is used for, and why the timestamp was dropped, is not clear.
The check that flags unapproved file paths now has a branch for remote calls, but remote calls still skip it
Unclear It is not clear whether the new branch for remote calls can run at all.
A settings change made in a Claude Code process with nowhere to save it now logs a warning that it is kept in memory only
Unclear It is not clear what makes a process have no storage.
Claude Code now treats GIT_CONFIG_GLOBAL and GIT_CONFIG_SYSTEM as environment variables that point to a single file
Unclear What Claude Code does differently with variables sorted as file paths is not known.
ultracode removed#The startup tip pointing users to dynamic workflows and the ultracode keyword no longer appears
Unclear It is not settled whether typing ultracode still triggers anything in this build.
Prompt autocomplete no longer suggests MCP server URL templates for input like /server:...://
After a command runs on another machine, directory sync no longer tells Claude mid-turn how and when changed files will arrive
Unclear The note may have moved to another place rather than being removed.
A check that refused bypassPermissions when claiming an already-started session on internal builds, and said to start cold, has been removed
Unclear It is not clear whether the same restriction is now enforced somewhere else.
claude doctor no longer checks strictPluginOnlyCustomization in managed settings#claude doctor has stopped warning about invalid or unrecognized strictPluginOnlyCustomization values in managed-settings.json
Unclear Whether this check moved somewhere else in Claude Code is not known.
Claude Code removed handling for an internal MCP needs-sign-in attachment that already produced nothing
Unclear Whether this attachment is still created anywhere is not known.
The autoAcceptShimmer colour has been removed from every built-in theme
Unclear It is not confirmed what appeared on screen in this colour before it was removed.
Published verbatim by Anthropic for v2.1.282. 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 86 bullets, 27 name something an entry on this page also names, 15 name something no entry here does, and 44 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.
maxProseWidth setting that caps the width of Claude's prose in wide terminals while tables and code blocks keep the full width
Probably new-maxprosewidth-setting-limits-the-width-of-prose-in-res /status and claude doctor entries, listing telemetry variables in a project's settings files that were ignored or that turned telemetry off
Probably new-startup-warning-when-a-projects-settings-set-telemetry, claude-doctor-no-longer-validates-strictpluginonlycustomiz, status-shows-when-permission-rules-are-limited-to-managed-s allowClaudeInChromeWithManagedMcp managed setting to let claude --chrome run alongside an exclusive managed-mcp.json; the error shown when Chrome is blocked now names it
Probably chrome-install-prompt-is-now-also-skipped-when-a-managed-mcp, chrome-policy-refusal-in-remote-control-now-uses-a-shared, claude-in-chrome-can-run-alongside-managed-mcpjson-when-dev store.readiness_grace_seconds to the Claude apps gateway so /readyz can stay ready through a short Postgres outage such as a database failover
Probably gateway-readyz-grace-period-through-postgres-outages-and, gateway-storereadiness-grace-seconds-keeps-readyz-reporti, bundled-server-gets-storereadiness-grace-seconds-so-readyz /feedback drafts list in fullscreen mode; it appears while the mouse is over the list
Probably feedback-bundle-mode-can-now-be-cancelled-mid-save --continue, --resume) re-sending earlier messages in a changed form, which could make the API drop Claude's earlier reasoning
Probably resume-failures-on-windows-explain-ebadf-transcript-read-err /model, /rename, /artifacts or another immediate slash command was used while Claude was working
Probably model-credits-required-429-no-longer-always-blocks-a-mod, model-catalog-entry-drops-quick-select, artifacts-panel-rebuilt-on-a-list-component-accessibility-m, model-accepts-claude-fable-models-when-validation-returns --tools list that leaves out a built-in tool offered earlier in the conversation
No entry names this data in redacted_thinking block" API error; Claude Code now drops the conversation's thinking blocks and retries once
Probably error-code-table-in-a-bundled-doc-now-covers-redacted-thinki, redacted-thinking-errors-are-now-classed-as-thinking-signa /model with a full Fable model id stopping at an API error instead of opening the usage-credits prompt when the plan needs usage credits that aren't turned on yet
Probably model-credits-required-429-no-longer-always-blocks-a-mod, model-catalog-entry-drops-quick-select, model-accepts-claude-fable-models-when-validation-returns /Network via .. or a /.vol-style kernel path, or a rules link to macOS's /home being listed
Probably plugin-eval-refuses-paths-that-land-on-net-and-other-automo, path-canonicalisation-treats-macos-vol-file-nofollow :* being skipped in settings files while --allowedTools honored them; they now work from every source, with a startup warning on how they match
Probably allowmanagedpermissionrulesonly-now-also-ignores-allowed, status-shows-when-permission-rules-are-limited-to-managed-s disableClaudeAiConnectors or allowManagedPermissionRulesOnly; the lock now applies and startup names the key
Probably allowmanagedpermissionrulesonly-now-also-ignores-allowed, allowed-tools-in-skills-and-plugin-commands-ignored-under-al, coordinator-no-longer-advertises-skill-tool-grants-from-untr, status-shows-when-permission-rules-are-limited-to-managed-s permissions, autoMode, worktree and attribution settings being ignored entirely when one nested value was invalid; the rest of the block now still applies
Probably managed-settings-invalid-lock-values-and-nested-permission #88564[FEATURE] Settings schema validation failures should result in more fault-tolerant behavior Open
--add-dir skills, commands and skills-directory plugin manifests pre-approving their own tools via allowed-tools under managed allowManagedPermissionRulesOnly
Probably allowmanagedpermissionrulesonly-now-also-ignores-allowed, allowed-tools-in-skills-and-plugin-commands-ignored-under-al, coordinator-no-longer-advertises-skill-tool-grants-from-untr, status-shows-when-permission-rules-are-limited-to-managed-s SessionStart hook
No entry names this claude remote-control --debug failing with "Unknown argument: --debug", although Remote Control's own eligibility error says to run with --debug
Probably claude-remote-control-accepts-debug-d-and-debug-in /install-github-app saying "cancelled" and then still pushing the branch and saving the API key secret; leaving now stops the remaining steps and reports what was already done
Probably install-github-app-can-now-be-stopped-part-way-and-reports / in /skills moving the skill list instead of reaching the search box
No entry names this /skills search box to the skill list while typing, which could hide the caret and put IME input in the wrong place
No entry names this /skills and /mcp, being two columns narrower outside fullscreen mode, where the scrollbar can never appear
Probably claudeai-mcp-connector-error-drops-the-mcp-hint-in-some-se, missing-mcp-authentication-the-model-is-told-differently-ho, mcp-disabled-server-status-gets-project-scoped-wording-in, mcp-reconnect-now-tells-you-when-a-server-is-disabled-with · separator in the /tasks dialog footer when the stop-all-agents shortcut is unbound in keybindings.json
No entry names this >> indenting empty lines, r with a count longer than the line changing text, 2J joining one line too many, and a count on the last line (2dd, 2>>) shifting or deleting it
No entry names this dd, dj, dG or a whole-line p/P it lands on the first non-blank, yy no longer moves it, and Esc after an emoji no longer leaves it inside the emoji
Nothing to match on . when repeating x, s, p, d or c, and whole-line p/P, o, O, J, >> and << acting on the wrong line when a line above wraps
Nothing to match on x did nothing
Nothing to match on /artifacts: titles line up in one column, details are dropped whole instead of cut mid-word, and the list supports PgUp/PgDn, Home/End, the mouse wheel and clicks
Probably artifacts-panel-rebuilt-on-a-list-component-accessibility-m claude-api skill: pre-output refusal billing now links to the How refusals are billed docs, mid-stream refusals bill at normal rates, and pre-output refusals count against rate limits
Probably claude-api-skill-adds-preserved-thinking-migration-docs claude-api skill to recommend ant apply for keeping Managed Agents resources as version-controlled files
Probably claude-api-skill-adds-preserved-thinking-migration-docs CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out)
No entry names this sandbox.excludedCommands to ignore project and local settings entries when managed settings or --settings set allowUnsandboxedCommands: false, or managed allowManagedDomainsOnly: true
Probably sandbox-project-settings-excludedcommands-are-ignored-whe, sandboxexcludedcommands-setting-now-documented-in-the-schem, sandbox-exclude-command-refused-when-unsandboxed-commands, remotecloud-session-launch-forwards-the-autocompact-setting, remote-control-project-thread-child-sessions-launched-with-a CLAUDE_CODE_ENABLE_TELEMETRY and OTEL_LOG_*
Probably project-settings-can-no-longer-set-opentelemetry-exporter-en managed-settings.json) keeps user-writable HKCU and WSL /etc/claude-code from applying
Probably claude-doctor-no-longer-validates-strictpluginonlycustomiz, wsl-managed-settings-a-windows-side-policy-file-that-fails, claude-in-chrome-can-run-alongside-managed-mcpjson-when-dev Skill(anthropic-skills:) and Skill(claude-ai:) allow rules to cover only skills synced from claude.ai, not plugins or other skills that merely use such a name
No entry names this anthropic-skills or claude-ai namespace to no longer load; a plugin so named still loads but yields name ties to synced skills
Probably skills-and-mcp-servers-whose-names-use-claudeais-reserved, forked-synced-skills-get-an-anthropic-namespace-reservation, mcp-prompts-from-servers-with-reserved-names-are-no-longer-l, synced-skill-overrides-and-namespaces-new-copper-heron-off, skillcommand-lookup-also-matches-claude-aianthropic-skill, skill-tool-suggests-the-new-namespaced-name-when-a-skill-is, plugin-marked-as-an-account-skills-wrapper anthropic-skills or claude-ai to list no skills or prompts (their tools still work); rename the server in your MCP configuration to list them again
Probably skills-and-mcp-servers-whose-names-use-claudeais-reserved, forked-synced-skills-get-an-anthropic-namespace-reservation, mcp-prompts-from-servers-with-reserved-names-are-no-longer-l, synced-skill-overrides-and-namespaces-new-copper-heron-off, skillcommand-lookup-also-matches-claude-aianthropic-skill, skill-tool-suggests-the-new-namespaced-name-when-a-skill-is, plugin-marked-as-an-account-skills-wrapper ultracode visuals in /effort and the prompt input to plain styling (no ripple, border flourish or keyword glimmer) and removed the dynamic-workflows spinner tip
Probably ultracode-effort-label-shortened A model matched these bullets to the GitHub issues they fix, so a link can be wrong.
No change to the system prompt since v2.1.281.
Claude Code, interactive mode
854 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 66 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?