You can now have a new bundled skill write live mods that reload inside your session once you agree. A mod is a plugin that adds panes, status lines, toasts or hooks to Claude Code. Cloud sessions started with claude --cloud can now use the MCP servers on your machine after you allow it. Setting CLAUDE_CODE_REMOTE_TOOLS_FORWARD to true lets Bash, Edit and Write calls run on an attached machine. The new --client-data-url flag loads a signed per-model configuration from Anthropic. Claude Code will not start if that configuration fails to load.
Sixteen additions are in the build but not switched on yet. A GetTask tool that reports on background shell commands stays off until it is switched on remotely. Compacting a long idle conversation before its one-hour prompt cache expires is also off by default. Auto mode gains an option to check Chrome clicks as well as page navigation, but it is not switched on. Remote Control can offer this machine's folder and tools to cloud sessions only behind a server switch that is off.
Among the fixes, Remote Control now works for Pro, Max, Team and Enterprise users who turned telemetry off. If you quit while a local MCP server is still connecting, Claude Code now stops that server instead of leaving it running. MCP servers built on Java MCP SDK 0.17.0 or older now connect after one automatic retry. Git diffs now run with --no-color, so git settings that force colour no longer garble them. Skill names starting claude-ai: are no longer mapped to anthropic-skills:. CLAUDE_CODE_DIR_SYNC_CHAIN is no longer read, because cloud directory sync now always sends changes as linked bundles.
Written by our agent from the shipped bundle, not by Anthropic.
Claude can write plugin mods into a per-session folder that hot-reload after you agree, with a bundled skill and restore on resume
Elsewhere
_host when CLAUDE_CODE_REMOTE_TOOLS_FORWARD is setWith CLAUDE_CODE_REMOTE_TOOLS_FORWARD set, tools gain a _host field for running on an attached machine, with new Bash guidance and classifier rules
--client-data-url flag loads a signed configuration for your model--client-data-url or CLAUDE_CODE_CLIENT_DATA_URL loads a signed per-model configuration from Anthropic, and Claude Code will not start if it fails
The self-hosted gateway gains mantle and anthropicGoogleCloud upstreams, and each upstream's models list now controls which models it serves
Platform & internalsWith CLAUDE_CODE_REMOTE_TOOLS_PIN_STORED_LOGIN set, the serve helper exits with code 78 if its API or sign-in addresses are not Anthropic's own
Want the reasoning? Read walks the 80 entries that probably matter to you, each one opening to what changed and why.
Read this release → Every row →Anything you can use today, anything that visibly changes, and anything worth poking at. One line each, open for detail.
Claude can write plugin mods into a per-session folder that hot-reload after you agree, with a bundled skill and restore on resume
Unclear No skill shipped with this build seems to contain the placeholder, so it is not clear whether anyone can reach this feature yet.
_host when CLAUDE_CODE_REMOTE_TOOLS_FORWARD is set#With CLAUDE_CODE_REMOTE_TOOLS_FORWARD set, tools gain a _host field for running on an attached machine, with new Bash guidance and classifier rules
Unclear It is not clear what normally sets this variable or which machines count as attached.
Sessions started with claude --cloud can call MCP servers on your machine once you allow it, with a new prompt whenever a server's tools change
Unclear Whether a server-side setting also controls this feature was not established.
A Remote Control daemon can now advertise its local folder and tools to cloud sessions, gated by tengu_radiant_heron and a chain of checks
Unclear It is not clear whether this is switched on, as a remote switch can turn it off.
A new idle path can compact a long conversation shortly before the one-hour prompt cache expires, controlled by the tengu_sunny_locket setting
Unclear The default minimum conversation size for idle compaction is not confirmed, and neither is exactly where the idle timer is scheduled.
GetTask, a tool for checking a background shell task, can now be switched on remotely through tengu_violin_rosin and has compact result lines
Unclear Whether Claude Code can insert a GetTask call when the tool itself is switched off is not established.
--client-data-url flag loads a signed configuration for your model#--client-data-url or CLAUDE_CODE_CLIENT_DATA_URL loads a signed per-model configuration from Anthropic, and Claude Code will not start if it fails
Unclear Which behaviours the configuration can change is not stated.
claude daemon remote-control now has its own availability check, help for list/add/remove, and an extra precondition check before trusting a directory
Unclear It is not certain which remote setting decides whether the Remote Control help section is shown.
The self-hosted gateway gains mantle and anthropicGoogleCloud upstreams, and each upstream's models list now controls which models it serves
Claude Code now carries the rules a local machine uses to accept, limit or refuse tool calls forwarded from a cloud session
Unclear What lets a machine start serving these calls was not traced, so whether this is active is unknown, as is whether the code is new or was moved in from elsewhere.
Claude Code can now tell cloud sessions that this machine's folder and tools are available, under a server setting that is off by default
Unclear The other conditions in the decision, such as platform, a local device setting and policy, are not fully known.
Account memory servers can now set whether subagents get full, read-only or no access, and by default only the main agent can change memory
Unclear It is not clear what decides whether an account memory server is attached to a session at all.
Served Bash/PowerShell commands that write Claude Code settings files are denied or sent for approval, and auto mode cannot approve those asks
Served calls from cloud sessions are denied access to credential files, keychains and, under tengu_violin_lining, SSH keys and cloud logins
Claude Code now writes a tsconfig, a .gitignore and type declarations into function-hook plugin folders so editors can type-check hooks
Unclear It is not clear whether the function-hook plugin feature is switched on for everyone or behind a setting.
The self-hosted runner can now split git hosts between Anthropic-managed git and direct access, and its --use-anthropic-git-proxy warning no longer says github.com only
Unclear Which server setting turns on the host split is not established.
Sessions without a project directory now never count as trusted, skip project config lookups, and get a clearer untrusted-workspace warning
Unclear What puts a session into the no-project state is not known.
A new classifyChromeClicks setting can route Claude-in-Chrome clicks through the auto-mode classifier, as navigation already is
Unclear What uses the new setting was not checked.
With CLAUDE_CODE_REMOTE_TOOLS_PIN_STORED_LOGIN set, the serve helper exits with code 78 if its API or sign-in addresses are not Anthropic's own
Unclear It is not clear what normally sets this variable or what triggers the check.
A new remote setting, tengu_tidy_lemon, can make asking your admin for more usage a one-step menu action or give its confirm dialog custom wording
Unclear A second "confirm" style exists alongside the one-step one, and what it shows is not established.
Published artifacts can declare "dynamic": true in capabilities.mcp to call unlisted connectors, with viewers approving each on first use
Unclear The capabilities field only appears when an internal setting is on, and it is not clear who has that setting.
The sandbox accepts a new denyAllNetwork option that cuts off all network access without a proxy; it is not supported on Windows
Unclear Nothing appears to use this option yet, so it may not affect any command in this release.
The footer now shows the connected IDE's selection as a line count or file name, and a selectable item lets you drop it with Backspace or Delete
Unclear This may replace an existing footer rather than add a new one, and whether it is switched on for everyone is not known.
deniedModels and availableModelsMatch restrict which models people can use#Admins can now deny specific models and match availableModels exactly; the default model, /model picker and SDK switches all follow the policy
claude --remote can now offer this machine's tools to the cloud session#When claude --remote starts a cloud session and stays attached, it can now make tools on your own machine available to that session
Unclear It is not clear whether further conditions, such as a server-side switch, can still stop the tools from being served.
--remote can ask before sharing local MCP servers or serving tools unattended#An attached --remote session can now ask whether to share your local MCP servers, and ask again before serving tools in auto or bypass mode
Unclear It is not clear what switches remote tool serving and the unattended-serving feature on, or whether either is on by default.
An artifact's connector declaration can now be marked dynamic, so viewers approve unlisted connectors at first use instead of those calls failing
Unclear It is not clear whether the Artifacts feature itself is available to every account.
A cloud session attaching to your computer can ask whether to offer it your local MCP servers: allow, deny or not now
Unclear The remote switch controlling this was not identified, so it may not be reachable yet.
Headless Remote Control can now offer to run tool calls from its folder and reports how many it is serving
Unclear The setting that controls this and whether it is on by default could not be determined.
The bundled skill for writing function-hook plugins is now described as making mods such as live panes, bands, status lines and toasts, with example files
Unclear What decides whether the "plugin-authoring" bundle is available is not known.
A machine bound to a cloud session now handles its file sync calls, pulling the session's changes in and, when asked, sending the checkout up
Unclear What turns this handler on is not stated.
Claude Code now includes a device tools re-consent dialog and an unattended serving consent dialog
Unclear It is not clear what makes either dialog appear, or whether they are about serving this machine's tools to other sessions.
One schedule variable replaces the two single-interval check-in variables, and setting the old ones no longer changes anything
Unclear Exactly what Claude Code does at each check-in is not shown.
Artifact preview now needs a browser found at startup, unless CLAUDE_CODE_ARTIFACT_PREVIEW is explicitly true
Unclear It is not clear whether artifact preview is included in public builds of Claude Code at all.
The post-turn memory config now takes a role_line_scan option choosing which attachments, instruction files and first window are scanned
Unclear What the role line scan does is not known.
Claude Code carries a built-in copy of its published model list, used by third-party provider sessions until they fetch the current one
Unclear How this built-in copy replaces the older samples and where it is read is not clear.
New instructions tell Claude what to do when an attached machine is unreachable, what exists only on your computer, and what to do with none attached
Unclear What turns the attached-machine feature on is not stated.
Subagents, teammates, forks and background queries are now refused when they try to write to your account memory
Unclear Which account-memory tools are on the exempt list, and whether the memory server itself only connects under a separate switch, are not settled.
When claude-opus-5-5 refuses a request in the frontier_llm category, Claude Code can now fall back to claude-opus-5
Unclear Whether the refusal fallback feature itself is switched on for everyone is not settled.
When a page can ask viewers for any MCP connector, its permission changes say so and Claude is told to warn you those integrations are unverified
Unclear It is not clear where this instruction is given to Claude or whether page publishing is available to everyone.
The built-in fallback for starting sessions in auto mode switched from off to on, so it applies unless the server turns it off
Unclear One of the conditions that must also hold was not identified, so exactly which users can be affected is not settled.
Claude Code's background daemon can now print which folders are serving tools to cloud sessions, and how many sessions each serves
Unclear Which daemon command prints these lines is not confirmed.
Forked agents can now pass on each piece of streamed output as it arrives, and are labelled as a worker by default for memory
Unclear What uses the streamed output or the memory label is not known.
set_max_thinking_tokens can now change thinking_display without resetting the budget, and SDK stream events carry thinking_display
Unclear Only the field's description was seen, so whether Claude Code actually fills it in on stream events in this build is unverified.
CLAUDE_CODE_MEMORY_SUBAGENT_APPEND turns off#Subagents that share context now get a memory note added to their instructions; setting CLAUDE_CODE_MEMORY_SUBAGENT_APPEND stops it
Unclear What sets the memory mode, and whether some hosts set the new variable themselves, is not established.
When a session is attached through remote control, Claude Code can now ask for your consent about MCP servers on the host device
Unclear What the remote switch falls back to when the server sends nothing is not known.
The background tasks dialog groups agent teammates under team headings and can open a View, Foreground and Stop menu
Unclear It is unclear what decides whether the actions menu opens, so it may not be active yet.
API requests can carry a new x-claude-code-prompt-id header naming the prompt behind them, under the same rule as other gateway hint headers
Unclear It is not clear that any request Claude Code makes actually supplies a prompt id yet.
/doctor prompt-audit reviews your project's instruction files#/doctor prompt-audit hands off to the bundled claude-api skill to audit CLAUDE.md, rules, skills and other instruction files
Unclear What the optional path after prompt-audit changes is not stated.
Hooks may still allow or block calls to the account memory server, but a call they try to rewrite is denied and a rewritten result is dropped
Unclear It is not stated when an account memory server is connected in the first place.
The connection used by remote sessions now recognises an "owner present" message type, though no part of Claude Code responds to it yet
Unclear Whether some indirect route installs the handler after all is not ruled out.
Tool requests from a bridged device can now be refused before the tool runs, but no tool uses this yet
Unclear It is not clear whether any tool opts in in a less direct way, or whether the refusal depends on the bridge's enforcement setting.
Claude Code has unused code to accept organization policy limits already fetched by the app that started it
Unclear It was not ruled out that this code is reached from outside under another name.
Each interactive launch now begins a background preview-browser discovery step, which print and SDK modes skip
Unclear What this step looks for, what it is used for, and whether it is switched off internally are not settled.
/status shows a Client data entry when client data is loaded#The part of /status that lists where your settings come from now includes a Client data entry whenever a client-data document is loaded
Unclear It is not clear what a client-data document is or how one gets loaded.
Buttons in pane layouts now accept a role of "dismiss", and any other role value is rejected with an error
Unclear What Claude Code does differently with a dismiss button is not clear.
With CLAUDE_CODE_PLUGIN_DIR_WATCH unset, Claude Code can now watch every local plugin directory not excluded; setting it to false turns watching off
Unclear The rule for which plugin folders are watched is not settled.
CLAUDE_CODE_WEBSEARCH_USE_CCR_PROXY no longer forces fast web search off, and CLAUDE_CODE_WEBSEARCH_CCR_PROXY_FAST is gone
Unclear It is not confirmed whether the other conditions checked before fast mode applies were also changed.
Multi-file artifacts can be published in parts, supporting HTML pages without a doctype are flagged, and same-folder references no longer warn
Unclear The actual file-count and size limits are not stated.
CLAUDE_CODE_GB_BASE_URL gets the same settings checks as the proxy variables#CLAUDE_CODE_GB_BASE_URL joins the proxy, certificate and bridge variables that Claude Code checks when it cleans a settings file
Unclear It is unclear whether the filter keeps or removes these variables, which settings files it runs on, and whether the variable really sets where feature flags are fetched from.
With CLAUDE_CODE_REMOTE_TOOLS_PIN_STORED_LOGIN set, credential-helper settings such as otelHeadersHelper are dropped from user-writable settings files
The debug logger can now turn off its default log file unless a log path or CLAUDE_CODE_DEBUG_LOGS_DIR is set, but nothing calls the switch yet
Unclear Nothing in this build was found that switches the default file off, so this may have no effect yet.
Several keypresses arriving at once can now be drawn one by one without setting CLAUDE_CODE_COMMIT_BETWEEN_KEYS
--system-prompt and --system-prompt-file can now be used together#Passing both --system-prompt and --system-prompt-file no longer stops Claude Code from starting; the two are combined
Unclear The order in which the two values are combined, and what separates them, are not settled.
/memory settings are now a list of toggles#In /memory, Auto-memory, Auto-dream and synced project memory are now toggles in a list above the memory file picker
Unclear Whether the conditions for showing each setting also changed is not established.
The gateway's load test mode now does a small amount of system reading and writing per connection so CPU use matches production
/context lists MCP server instructions as their own category#The /context breakdown now shows how much of the context MCP server instructions take up, next to MCP tools
claude plugin eval now needs git 2.31 or newer#claude plugin eval now checks your git version and refuses to run with git older than 2.31
The footer's "view tasks", "view memories" and "view artifacts" hints now show the key you bound, not always "enter"
Certificate and TLS settings passed through GIT_CONFIG_KEY_n and GIT_CONFIG_VALUE_n now reach git commands Claude Code runs itself
Unclear Which of Claude Code's internal git commands besides the reset receive these settings is not fully known.
deniedModels in the catch-all managed.policies entry is now combined with each specific policy's list instead of being replaced by it
Unclear Which program or command runs this policy merge is not confirmed.
messageSelector: keybinding actions are replaced by select: and rewritten in user config as /rewind, the diff dialog and other pickers move to a shared list
Unclear It is not clear where the rewrite runs, or whether diff:viewDetails is also rewritten in the diff dialog, where it is still a default binding.
DISABLE_PROMPT_CACHING_HAIKU now matches your configured Haiku model#DISABLE_PROMPT_CACHING_HAIKU now also turns off caching for any request whose model resolves to the configured Haiku model
claude plugin list tells you when your plugin records were rebuilt or cleaned up#claude plugin list now prints a note when installed_plugins.json was rebuilt or had unusable entries removed, and says where the old data is
claude plugin validate now checks paths for more plugin parts and flags a .json path that is really a folder
Unclear The full list of checked keys comes from reading the code, not from running the validator.
claude plugin marketplace remove now lists the plugins uninstalled along with the marketplace and says their data is deleted
Unclear Whether plugins were already uninstalled this way before, with only the report being new, is not settled.
In screen-reader mode, w in the /copy picker writes the full response to a file, and the command list shows as a plain column
Unclear How screen-reader mode is turned on is not known.
/hooks matcher list shows sources, matcher and hook count in columns#Each matcher in the /hooks menu now shows a dimmed source badge, the matcher or "(all)", and its hook count as separate parts of a list row
Unclear It is not confirmed that this list is the one in the /hooks menu.
The "Select IDE" dialog in /ide now starts on your current IDE and marks it with a check mark
Marketplace validation now gives an error for plugin or marketplace names Claude Code cannot install, instead of saying they are accepted
What would fix it?
Smaller changes and internals, grouped as the pipeline found them. Nothing is dropped, it is only further down.
28 more of these are in What probably matters to you, above.
Choosing a new conversation from the cold-resume quota warning now tells Claude where the old transcript is, except under HIPAA or when turned off
40 more of these are in What probably matters to you, above.
Four local features that scan your transcripts, including Insights, now run in HIPAA-flagged orgs unless the server restricts them or tengu_hashed_lark restores the deny
Unclear It is not fully clear how the hinted-denied list is filled when no policy has been loaded.
Default Opus, Opus subagents and the picker label add 1M context only where a model's 1M window is opt-in, not native
Unclear It is not known whether any current model counts as having the 1M window built in.
Before forwarding a command, directory sync now checks the machine is still reachable and fails fast, blaming the machine, if it is away
Unclear It is not clear which setups this check runs in, since it depends on the remote machine's version and on sync support that was not traced.
Permission asks marked personOnly now skip the classifier, hooks, the coordinator check and prompt-tool hook races, and guard against accidental keypresses
Unclear Whether the skipped check includes your own permission hooks or only the coordinator's own automation is not clear.
MCP servers whose name clashes with Claude Code's own "memory" server are dropped and recorded as reserved-name refusals, and /mcp no longer looks empty
Unclear Which names count as reserved, what policy turns the refusal on, and what /mcp shows in place of the empty message are not settled.
Claude Code remembers when an MCP server offers no way for a terminal to register, and can leave such servers out of plugins synced from claude.ai
Unclear Exactly which plugin sources count as synced for this check is not settled.
The artifact preview check stubs window.claude.use(), records which capabilities a page asked for, and tells Claude to test those parts after publishing
Unclear What decides whether this preview feature is available is not known.
Organizations under zero or limited data retention no longer get local-to-cloud file sync in cloud sessions and now see a notice explaining why
In Claude Desktop, auto-compact on 1M-context models can now be capped at a 600k or 500k window by a server-side setting
Resuming a session last used under another organisation can now warn that Claude's earlier thinking will be removed and the next reply may be slower
Unclear It is not clear whether only the warning is new, or whether recording which organisation used a session is new as well.
Paid subscribers with telemetry off can pass Remote Control's eligibility check, but remote sessions now refuse with a clear reason when the version floor is unreachable
Unclear What the earlier empty answer led to in practice is not settled.
The spot where plugins can supply their own spinner is now listed for the desktop app as well as the terminal
Unclear Whether the desktop app actually shows a plugin's spinner in this build is not known.
Gateway config validation now rejects forward_user_identity when an Anthropic upstream points at claude.googleapis.com
Syncing a local checkout with a cloud session can now report a failed git index merge, describes retries plainly, and refuses some snapshots
Unclear It is not clear whether folder sync with cloud sessions is available to every user.
When a subagent's definition lists MCP servers, some of those entries can now be dropped and recorded as refused
Unclear Which server names or settings lead to a refusal is not stated.
Refused publishes over a workshop page now tell Claude to read the page's live decisions instead, or to tell you the session can't build on it
Unclear Whether workshop pages are behind a remote switch in this release is not clear.
With live paths on, publishing a live page that saves over itself can now be rejected with the error live_self_saving_page
Unclear It is not clear where the live paths feature is switched on or what its default is.
When no machine is attached any more, Claude is now told to say so rather than doing that machine's work in the current environment
Unclear It is not clear whether this message reaches Claude as a tool result or as system text.
When directory sync supplies your files, setting up a cloud session now creates an empty repository instead of uploading a git bundle
Unclear The exact conditions under which a session counts as served by directory sync are not settled.
When publishing a page, Claude is now told to start dashboards, metrics pages, KPI trackers and scorecards from the Dashboard type if available
The daemon now logs worker kinds it will not start for your account, and starts them if they become available when flags refresh
Unclear It is not clear which flags decide whether each worker kind is available.
With CLAUDE_CODE_AUTO_MODE_CLASSIFY_EDITS set, or for models the server lists, edits allowed by acceptEdits mode now go through auto mode's classifier
Unclear It is not clear what the classifier then does with these edits.
If Claude Code's own "memory" server is configured, any other MCP server whose name works out to "memory" is removed from the server list
Unclear It is not clear what adds Claude Code's own memory server, whether that is switched on for everyone, or whether you are told when yours is dropped.
Self-hosted runners using governed git now build a certificate bundle, move GIT_SSL_* into session git config and pass per-source git environments
Unclear What the per-command git settings contain is not established.
Cloud file sync now recognises compliance and zero-data-retention restrictions, stops for the session and explains why instead of failing generically
When installed_plugins.json holds records or a format this build cannot read, Claude Code keeps them and refuses plugin changes and installs
Plugin hooks modules for command.run now receive only the commands in their runCommands list, and tool.call modules can be skipped on the host
Unclear It is not clear whether ordinary plugins can reach this or only built-in ones.
Sessions run through the SDK or print mode can now reload agents, commands and plugin MCP servers when a watched plugin folder changes
Unclear It is not clear what starts the folder watching or whether a setting or remote switch controls it.
After 20 prompt suggestions go unused in a row, Claude Code shows fewer of them until you use one again
Unclear How often suggestions appear while reduced is not confirmed; it looks like once every 10 turns.
Images returned by an MCP tool are now also saved to disk, with a note giving the image size and file path
Unclear It is not confirmed whether earlier versions already saved these images some other way.
The standalone build can now send API requests through a reusable connection session tied to your current TLS settings
Unclear Whether the runtime bundled with this build actually offers the reusable session is unknown; if it does not, the old path is always used.
When your files arrive in a cloud session, files already at the same paths are moved to .git/claude-trash instead of being overwritten
Unclear Exactly when this set-aside step runs, and whether directory sync is available to everyone, was not established.
When background memory saving writes files, the main conversation is told which ones changed so it rereads them instead of using an old copy
Unclear It is not clear whether background memory saving is switched on for every user.
Claude Code now refuses to write a settings file if it cannot locate the managed settings folder to rule it out
When a sorted query on artifact data fills its limit, Claude is now told that more results may exist and how to fetch them
File sync now records a conflict instead of applying incoming changes to Makefiles, package files, lockfiles, CLAUDE.md and AGENTS.md
Unclear How a recorded conflict is shown to you is not clear.
Any model name containing both "sonnet" and "[1m]" now counts as a Sonnet 1M-context name, apart from some exclusions
Unclear It is not clear what this check controls or which names are excluded.
If Claude created files in a cloud session before your files arrived, yours now win, Claude's copies go to the trash, and you get a notice
Unclear It is not clear which part of Claude Code shows this notice.
The information Claude Code reports about a remote session now includes total cost in US dollars and total token counts, updated as cost changes
Unclear It is not clear which kinds of session send these totals.
When file sync cannot start for a cloud session, Claude Code now names the reason, such as a network drive or an unsupported platform
A watch that Claude Code set up on its own for a published artifact is now retired after 3.5 hours without activity, instead of never
Unclear It is not confirmed that this 3.5-hour timing is the one Claude Code actually uses for artifact watches, rather than one of several timing presets.
Skills synced from your account now show as anthropic-skills:<name> unless they already have that prefix or have their own name override
Unclear The exact name synced skills showed under before is not known.
When a cloud session's file store fills up, directory sync now tells you and Claude that further changes stay in the cloud
Unclear Whether cloud directory sync is available to everyone is not known.
If the main working folder is deleted or reappears mid-session, Claude Code now tells the model the earlier folder no longer applies
Unclear One further condition decides whether the check runs, and what it is was not resolved.
When your machine sends its files to a cloud session, the first upload is now tried once and taken back if it fails
Unclear How this changes the waiting you see in practice is not clear.
A file staged into a session over Remote Control now counts as already read, so Claude can edit it straight away
Directory sync now keeps the committed .gitattributes files it read, and refuses with a fix-it message when a git setting changes attribute rules
Unclear The link between this behaviour and the remote switch is indirect, so it is not certain that switch controls it.
Setting the plugin sync environment variables now turns off one claude.ai plugin sync check instead of turning it on
Unclear It is not clear what the part that also needs the second server-controlled setting does.
Elicitation and ElicitationResult hooks now use the real session and raise an error when that session has no project
Unclear Whether the error is caught, and whether you would see it, is not clear.
When tracing is on, results from MCP tools and some built-in tools are stored on tracing spans as tool.output, with account memory redacted
Unclear It is not clear which built-in tools are included or exactly what differs from the previous build.
When an MCP tool call has moved to the background, progress updates from its server now update the task's status
Unclear It is not settled exactly which MCP calls get this progress display or how often the status can update.
The unknown-skill error and the Skill permission prompt no longer point from a skill's old name or namespace to its current one
Unclear The exact text added about plugin loading errors is not known.
When an organisation's managed policy denies your default model, the replacement is now tagged managed_deny and still labelled Set by your organization
Unclear It is not clear whether the on-screen label reads any differently than it did before.
store_full#When directory sync's store is full, remote commands now run anyway instead of failing, and a full link lane is reported as store_full
Unclear What this means for the command itself or for what a user sees is not clear.
On Windows, cloud-session and teleport git uploads now need a real git.exe, skip .cmd/.bat wrappers, and explain why when git is not found
Unclear This stricter lookup may apply to more than cloud-session uploads.
Restoring a snapshot now gives a specific reason when it was uploaded before the earlier environment had any of your files
Unclear Which kinds of session can reach this restore path is not established.
Attaching a cloud session with folder sync now stops at once with an explanation when the git anchor for syncing is unavailable
Unclear The condition under which this check runs is not stated.
Directory sync now gives up instead of placing files by name when folder anchoring is unavailable, unless CLAUDE_CODE_DIR_SYNC_DISABLE_ANCHORING is set
Unclear It is not clear what, if anything, is shown to you when sync gives up because anchoring is unavailable.
A model from a non-Anthropic provider is now listed only if an extra check for that provider also passes
Unclear What the extra check tests could not be determined.
When a machine attached to the session is offline, Claude is now told not to call it again that turn and to tell you
In auto mode, Chrome actions on a site marked to stay pinned no longer skip all permission checks and are limited to allowed domains
Unclear What causes a site to be marked as one whose host must stay pinned is not known.
When your computer can't be reached, Claude is now told not to call it again that turn, to do what it can in its container and tell you
Unclear Which remote-machine feature these messages belong to, and what switches it on, is not clear.
When a cloud session holds none of this folder's files and this machine cannot sync them, Claude Code suggests starting a new one
Unclear The exact situation that triggers this message is not settled.
If no working folder can be worked out for a subagent, Claude Code now refuses to start it and tells Claude to do the task itself
Unclear It is not clear which kinds of session end up without a folder to give a subagent.
An mcp_tool hook aimed at the account memory server is now refused when that tool is not available in the current session's mode
Unclear Which session modes make which memory tools unavailable is not stated.
If this machine's MCP passthrough approval state cannot be read, no MCP servers are offered to cloud sessions and a one-time notice says so
Unclear It is unclear whether only the notice is new or the handling of an unreadable record too, and what switches passthrough on.
claude mcp add and claude mcp remove now check that the change was saved#After adding or removing a user or local MCP server, Claude Code rereads its config file and fails with an error if the change did not stick
Unclear It is not confirmed that the check runs for both user-scope and local-scope servers.
/mcp now say why when you select them#In the /mcp tool list, organization-blocked tools can be highlighted and selected, and then show that your organization disabled them
cmd as its own key, separate from meta#The keybinding reference now says meta is the alt key and cmd is a separate modifier most terminals never send, with a 3-second chord timeout
Unclear It is not confirmed whether the chord timeout itself changed or was already 3 seconds.
A keybinding like "foo+x", where the part before + is not a real modifier, now gets an error with a suggested fix
claude plugin install and claude plugin uninstall now fail with install_records_unreadable instead of going ahead when their records are unreadable
Unclear It is not clear what the new directory-related failure reasons check.
Claude Code now treats Warp as supporting clickable links, so file and web links in its output can be clicked there
Unclear It is not confirmed whether an explicit hyperlink setting or FORCE_HYPERLINK still overrides this.
When an admin's deniedModels blocks the model a background helper request asks for, Claude Code now sends it to the main conversation model
Running background tasks now report where their interim output is written and when their final result will actually be delivered
Unclear It is not clear where this status text is shown or what asks for it.
Plugin errors in the session's startup message now carry a path field when Claude Code can work one out
HTTP_PROXY#A proxy lookup now checks HTTPS_PROXY, then ALL_PROXY, then git's system proxy, and ignores HTTP_PROXY for every address
Unclear Which Claude Code feature uses this lookup is not settled.
In screen-reader mode, quoted blocks are now labelled as quotations instead of being drawn with a line down the left side
Unclear It is not fully confirmed that screen-reader mode is what switches this on.
Switching to any Sonnet model name ending in [1m] now checks for 1M-context access, except four older Sonnet models
Unclear Where the new "denied_by_managed_settings" and "not_allowed" labels are shown is not known.
If Claude Code quietly refreshed its copy of a file after it changed on disk, Edit now treats Claude's read as out of date
Unclear Not every situation that quietly refreshes a file's record is known.
Settings validation now hides warnings about settings keys that Claude Code has removed
Unclear It is not clear whether removed keys are reported to you some other way.
Claude Code now allows writing Markdown files to a second kind of memory location without asking, and checks more Windows paths for risky patterns
Unclear It is not clear which memory locations the new rule covers or which Windows path patterns the extra list contains.
Before an ultrareview uploads your repository, the notice now says staged and unstaged changes may go too, but untracked files do not
Unclear It is not clear exactly where this text is shown or what controls whether ultrareview can be launched.
In the resume-or-new dialog, "Start a new conversation" now says Claude can look the old conversation up if you mention it later
Unclear It is not clear whether the lookup this option promises is available to every account.
The browser page shown after signing in now uses Anthropic fonts and an animated checkmark or broken-link icon instead of a text label
Plugin load errors that Claude Code reports to SDK programs are now documented and can give the path of the directory that failed
Editable text fields in the settings list can now refuse an invalid value, keep the editor open and show the reason in red
Unclear It is not clear which settings actually check their values this way.
In vim mode, J and visual J now handle spacing like vim and no longer do nothing on the last line
Installing a plugin from GitHub, a URL or a git subfolder now records the exact commit it came from when the source did not name one
Unclear It is not clear how the recorded commit is used after installation.
The /code-review ultra confirmation dialog now ignores selections that arrive during a refusal window, and branch reviews show an extra scope line
Unclear It is not settled that the refusal window is a period of time right after the dialog opens.
If one entry in deniedDomains, denyWrite or denyRead is invalid, Claude Code now withholds the matching allow list instead of applying it
Unclear Which settings sources this applies to, such as managed or personal settings, is not clear.
People watching a session over Remote Control now see API retry messages instead of what looks like a stall
A model sent by an SDK host through apply_flag_settings can now be refused, with a warning naming the default model and the one in effect
Unclear What makes a model choice count as refused under the new check is not clear.
If the session has switched to a backup model, background agents now get the main model back when it is available
Unclear How often a session is actually running on a backup model is unclear.
On Windows, cmd.exe delete commands run from PowerShell are now blocked when they target a drive root, a folder directly under one, or the home folder
Artifact database queries that use order_by now return ordered_by {field, limit} and warn Claude when a full page means more results may exist
Sync now blames staged changes only when a staged file really overlaps incoming commits, and reports other merge failures as index_merge_failed
File sync and teleport now stop with a named attribute_source reason when a local git setting changes which attribute rules git applies
Unclear Whether the FILE_SYNC_STOPPED.md marker file is gone or has only moved is not known.
When Claude Code cannot read its list of installed plugins, plugin errors say so instead of telling you to reinstall
Unclear The exact wording of the added note is not known.
The artifact preview's check for content spilling out of its box now ignores form controls, tiny hidden elements and deliberately clipped text
Claude Code's list of sensitive home-folder paths now covers all of ~/.claude/state instead of a few named files inside it
Unclear What this list controls, whether tool access, sandboxing or sync, is not clear.
The warning for a failed cloud download now says it retries on your next message or Claude's next change
Unclear It is not clear which setting controls whether cloud directory sync is available at all.
When no base folder is available, a relative path now raises an error instead of being resolved against the process's current folder
Unclear It is not clear in which situations Claude Code has no base folder available.
For skills synced from claude.ai, skillOverrides now checks only the skill's name and one other key, so overrides under older names may stop applying
Unclear It is not clear which override keys are no longer checked.
The git fetch Claude Code uses to fetch missing commits now blocks bundle URLs and HTTP redirects
Unclear It is not clear which feature runs this fetch or how often it runs.
When an artifact publish is refused because the latest version was not viewed, the message can now say which version the host is serving
Unclear It is unclear whether the artifacts feature that shows this message is switched on for anyone.
A permission rule that blocks a skill now also matches the extra names a packaged skill can go by, so it may block skills it used to miss
Unclear It is not clear where the packaging alias names come from or when the extra matching is skipped.
If a permission check needs a working folder and the session has none, the tool call is now denied with an explicit message, and retries are too
Going over the file limit in one artifact publish now says nothing was published and tells Claude to send the rest in further publishes to the same URL
Unclear The actual file limits, and what controls whether the artifact publishing feature is available, are not stated.
When directory sync starts on an empty folder, files it set aside now appear in the notice and report, and a later-upload notice can follow
Unclear The exact wording of both notices, and when the later-upload notice appears, is not known.
Chrome file uploads now give a specific reason for a refusal, and hard-linked files get an explanation and a suggested fix
Unclear How the earlier error messages for these cases read is not known.
File uploads now fail at once on HTTP 400, 404 or 422, and on timeouts set not to retry, instead of retrying
After a cloud environment is recreated, restore messages now say files are as of the last upload, and the end-of-last-turn case is gone
claude plugin init warns about names it cannot record instead of failing#The check at the end of claude plugin init now warns, rather than failing, when a name cannot be recorded
Unclear What the check did with such names before this change was not confirmed.
When an upload to a cloud session cannot be used, Claude is no longer told your machine was asked to send the files again
Folder sync now says when a folder could not be opened, suggesting it was moved, renamed, replaced by a link or had permissions changed
When a plugin fails to load, its error record now includes the plugin's path and manifest name where Claude Code knows them
Unclear Whether any screen, such as /plugin or claude doctor, displays the new fields is not settled.
After a background memory pass writes memory files, Claude Code refreshes what it knows about them and tells Claude which ones changed
Unclear Which background passes use this, and whether the memory features involved are switched off by default, are not settled.
The notice for a synced claude.ai skill whose slash name is taken now always says "your claude.ai skill" and folds several clashes into one sentence
Installing a plugin now fails with an error if Claude Code's record of installed plugins holds data this version cannot read
Unclear What exactly makes the plugin records count as unreadable is not known.
The check that spots local or private git remote hosts now covers Docker, minikube, home-router and dev DNS names, and this machine's own hostname
Unclear What uses the result of this check is not confirmed; it may be choosing the source for a cloud session.
A permission check can now be refused with a message saying it needs a working directory and the session has none
Unclear Which sessions have no working directory is not known.
Claude Code now treats localdev.me, local.gd, localhost.direct, ddev.site, lndo.site and lcl.host addresses as local
Unclear It is not clear which feature uses this local-address check.
Claude Code now locates managed settings drop-in folders when deciding whether a file write is protected, and assumes the worst when it cannot
Folder sync now stops with a clear refusal when this machine cannot place its files safely, instead of carrying on in a degraded way
Unclear It is not clear which platforms or conditions make the anchor unavailable.
Cloud file sync now gives specific messages for unresolved git conflicts at session start, restricted connections and empty-start checkouts
The error for required remote managed settings that fail to apply now says up front that the settings could not be loaded
If this machine cannot safely place synced files, folder sync for cloud sessions stops and Claude works directly on this machine instead
Unclear What makes files unsafe to place, and what controls whether directory sync runs at all, is not stated.
allowed-tools entries using CLAUDE_SKILL_DIR or CLAUDE_PROJECT_DIR are dropped, with a warning, when there is no folder to fill them in
Unclear How removing the separate naming for synced skills changes what you see is not clear.
Claude Code's sanitized git operations now allow only https, plus file where permitted, and refuse plain http:// remotes
Unclear Which feature runs these git operations is not stated.
When a remote host fails a liveness check, Claude is now told only that it did not answer, not that the next call checks again
Unclear It is not clear whether other text is added to the end of this message elsewhere.
--remote now tells you when the directory sync offer was put off#If a --remote session's directory sync offer is deferred because its anchor is unavailable, you now see a warning instead of nothing
Unclear It is not clear where the monorepo setting comes from or what sets it.
Claude Code now skips connecting a subagent's MCP server if the session is holding that server name for its own use
Unclear It is not clear which server names count as held by the session.
When Claude reads an artifact, its note can now say that version already counts as viewed, and it adds a short summary of the file
Unclear It is not clear what switches these artifact features on.
When required settings from a Cloud gateway cannot be used, the exit message now starts by saying plainly that they could not be loaded
When a skill cannot be found because its plugin failed to load, Claude is told the plugin could not load, not that the skill is not installed
Unclear A related change to how skill permissions match names for packaged plugins was seen but not worked out.
Plugin evaluation runs now start with a check step, and if it fails the run stops with exit code 1 and the report records it
Unclear That the check concerns git is inferred only from how the report labels it.
Claude is now told that database result limits run from 1 to 1000, default 100, and that sorted queries return one page only
Unclear Which setting makes this database tool available is not known.
When a cloud session deletes or replaces a synced file, the notice now says your copy is kept for at least the cleanupPeriodDays period
Two pickers, for a cloud session's memory files and for hook matchers, now use a different list, with file type and token count shown separately
Unclear How pressing Escape to cancel behaves in these pickers now is not clear.
When a remote machine refuses a tool call, Claude now gets a short error saying it is not accepting calls from cloud sessions right now
Unclear The old message still exists elsewhere, so whether it was replaced or a separate new message was added is not settled.
When a paired remote host stops answering, Claude is now told not to call it again this turn and to report what is blocked
Unclear It is not clear which remote-host feature produces this message or whether that feature is available to everyone.
The startup settings warning now says some values were replaced with their stricter reading, where before it said all were skipped
When the artifact file list can't be fetched, Claude Code no longer skips its file-change check and can still refuse a publish
Unclear It is not clear which switches control the artifact feature or who has it.
In the sandbox, git is now told to use no credential helper for the local sandbox proxy when that proxy requires a token
Unclear Another condition that can stop this setting from being passed has not been identified.
Folder sync for claude --cloud now keeps a turn in the cloud when its store is full and refuses a restore with no base to start from
Unclear The setting that controls this sync was not traced, so who gets this behaviour is unclear.
Claude Code now checks installed_plugins.json before migrating or saving it, and skips old cache cleanup when the file cannot be read
Unclear Exactly what the new check tests in the file is not stated.
In artifact preview summaries, items marked as notes are no longer counted as issues and are listed after them, prefixed "· note: "
Unclear Which preview tool produces these notes is not stated.
On Windows, the "git not found" upload error now says the folder on PATH must contain git.exe itself, since git.cmd and git.bat wrappers are skipped
If installed_plugins.json holds records this version cannot read, the enabledPlugins sync is skipped and the file is left untouched
Unclear Whether you are told when policy-required plugins go unsynced is not stated.
The agent that checks prompt and agent hooks now loses any tool the session's current permission mode does not allow
Unclear The exact check used to decide which tools the permission mode disallows is not described.
A prefixed skill name no longer maps to a synced skill with the plain name; synced skills are also left out of suffix matching
Unclear What the removed prefix is, and whether this matching now happens somewhere else, is unclear.
When Claude Code looks up a repository's git remote URL, it now accepts only https:// (or file:// where allowed) and refuses http://
Unclear It is not known which feature uses this remote lookup.
If your computer is not connected yet, Claude now tries one more time and then tells you, instead of just retrying
The message for later work that cannot be restored now says the uploads never reached storage or were not in a usable form
Unclear Which screen shows this message is not stated.
The Teleport to Repo chooser now has a Cancel item, and the skill override list shows status icons and a locked tag
When directory sync cannot fetch your latest changes, the message is shorter and no longer says a resend was requested
Going over the per-publish file limit now says nothing was published and tells Claude to split the files across several publishes
Claude is now told simply that your latest changes could not be brought into the checkout, and is no longer told a resend is coming
The keybindings troubleshooting table now explains the error for a mistyped modifier, such as ctl in place of ctrl
Unclear It is not clear whether the check that reports this error is itself new in this release.
Plugin dependency error messages now escape and quote names, check version ranges before showing them, and cut long lists short
Claude Code now says when a worktree's git files changed mid-read and suggests trying again, instead of a generic unreadable message
Unclear It is not clear where this message appears, possibly only when setting up remote or cloud sessions.
When an artifact publish is refused and the live file list can't be read, the message now says its file details come from an earlier listing
Unclear It is not clear whether the artifact publishing tool is available to everyone.
Entries in the hooks list now show each hook's type as a separate detail instead of a [type] prefix on the label
Unclear It is inferred, not confirmed, that this list is the one opened by /hooks.
Errors about plugin dependency cycles, cross-marketplace dependencies and missing dependencies now sanitize plugin and marketplace names
When a publish is refused as out of date, Claude is told the saved live version's size and line lengths and to read only what it needs
Unclear The switch that controls this guard and the new check that can hold the message back were not identified.
When an artifact publish is too large, Claude is told to send part now and the rest in another publish, and is given the larger per-version limit
The /plugin marketplace picker shows plugin counts and source per entry, and the cloud-session /skills dialog gets a plain column mode
Unclear It is not known what switches the /skills dialog into its plain column mode.
1 more of these is in What probably matters to you, above.
On exit, still-connecting stdio MCP servers are stopped, new connections are skipped, and late failures report SHUTTING_DOWN instead of real errors
A check that wrongly rejected every git mount push URL in cloud sessions is fixed, so valid URLs can now redirect pushes
Unclear The added check was not identified, so it is unclear which other URLs are now refused.
After a hook or the background memory pass rewrites a file, Claude Code updates its record of that file, so a later edit no longer fails as stale
Unclear It is not clear how often the background memory pass runs or what switches it on.
If a Java MCP SDK 0.17.0-or-older server rejects Claude Code's elicitation support, Claude Code reconnects once without the parts it rejects
Unclear It is not confirmed that the retry removes exactly the form and link parts, or whether the advice text is shown to users in every case.
Claude Code now runs its git diffs with --no-color, so git settings that force colour cannot garble the output
Invalid sandbox settings are now handled one field at a time, falling back to the restrictive value instead of the whole block being ignored
Unclear Which sandbox fields have a restrictive fallback value is not listed.
The rootOnly mode is removed: nested files are no longer skipped and the "files stay in the cloud" warning is gone
Unclear Which Claude Code feature uses root-only sources is not established.
Worktree checks now ignore letter case on Windows and macOS and confirm a match by checking it is the same file on disk
MCP OAuth sign-in now refuses to start when the server URL is missing or invalid, and tells you to fix it
One of Claude Code's file readers now fails with "not a regular file" for pipes, devices and similar paths instead of trying to read them
Unclear It is not clear which parts of Claude Code use this file reader.
Claude Code now ignores a "nobody is watching" report that arrives just after it received something from a viewer, so live output keeps flowing
Unclear How long the window after incoming traffic lasts is not known.
Remote Control diagnostics now mark the feature-flag check as fine on paid plans, since Remote Control is decided by plan there
Unclear It is not clear which command shows these diagnostic lines.
An HTTP MCP server that returns 404 during a tool call without a session is now treated as a normal error, not an expired session
Access refusals caused by compliance or zero-data-retention restrictions are now treated as "unavailable" rather than as unauthorised
Unclear Exactly which remote feature this applies to is not known.
When a finished result marked as an error is rebuilt for SDK programs, it now keeps its message text instead of an empty string
Unclear It is not clear when Claude Code rebuilds a result this way, for example when resuming or replaying a finished session.
When choosing between remote, MDM and file managed settings, a source with no usable values no longer wins over one that has them
Claude Code's worktree location check now tolerates case differences on case-insensitive systems, avoiding false "Cannot create worktree" errors
Unclear Exactly what the on-disk confirmation checks is not settled.
When run headless or through the SDK, the final result's usage total now includes assistant replies that arrived without stream events
Unclear It is not settled that this corrects under-reported usage in practice, or which sessions hide informational system messages.
A Remote Control session now keeps reporting that it needs action while a held permission prompt is still unanswered, instead of running or idle
Unclear One check that can end this step early was not resolved, so it may not apply in every case.
The notification saying Claude is waiting for you is now sent at most once for each time Claude finishes, not again when its timer restarts
MCP tool progress now checks for a progress handler before calling it, and can also be forwarded to a background progress reporter
Unclear It is unclear when the background progress reporter exists, so it is not known whether any progress is shown for background calls yet.
Subagent progress, subagent totals and workflow agent counts now ignore usage from assistant messages marked isUnmetered
In launches set up not to write to disk, Claude Code now skips creating the plans folder instead of always trying to
Unclear It is not fully confirmed that the condition which skips the folder is the no-disk launch mode.
The "Now using usage credits for" notice now matches model names regardless of upper or lower case, so it appears in cases it used to miss
After switching accounts, usage-limit wrap-up hints from the previous account no longer carry over to the new one
The "Current week (model)" usage rows now show nothing instead of failing when the server sends limits in an unexpected shape
In vim mode, the cursor now lands in the right place after an insert that ends with a combining accent mark or a newline
If a permission decision returns an empty set of replacement arguments for a nested tool call, the original arguments are now kept
Unclear Whether the same fix applies to ordinary, non-nested tool calls is not known.
If Claude Code fails to save a notification, including one confirming an MCP server's request for input, the error is now caught
If the notification sent after answering an MCP server's request for input fails, the error is now caught
/loop treats a folder named loop.md as no file#/loop now ignores a folder called loop.md and handles project and user loop.md files in the same way
Unclear It is not clear how the earlier version handled a folder named loop.md or other read errors.
In vim NORMAL mode, a restored draft puts the cursor on its last character, and newlines you insert are recorded with your typing
Unclear That the recorded newline is meant for vim's repeat command is inferred from a name, not confirmed.
In vim mode, new lines typed with Enter or Ctrl+J during insert mode are now recorded so the . repeat command can replay them
Unclear It is not confirmed that . now replays the new line correctly, because the code that stores and replays recorded inserts was not examined.
When an agent hook fails, the logged message and the 'Error executing agent hook:' output now carry the error that actually happened
8 more of these are in What probably matters to you, above.
Handling for unreadable or newer memory items is built into background memory consolidation but cannot run in this build
A new check can now keep Remote Control off at startup even where a persistent remote session used to always switch it on
Unclear What the new check actually tests is not identified.
The progress indicator during conversation compaction now updates while the cache-reusing summary runs, instead of sitting still
The shadow logging that measures idle sessions now records when Claude Code starts and stops waiting for your input, not only when work runs
Unclear It is not clear whether the start-of-wait events are also held back by the same flag as the end events.
Code that tells a wake keeper when an MCP task moves to the background is included, but is turned off at build time
Claude Code gained an unused option to store bundled skill files in another location; they stay on local disk
Unclear It was not ruled out that some other part of Claude Code sets this option.
Cloud session startup now has a switch, fixed off, that would skip choosing an environment and detecting the git branch
Unclear It is not clear which feature this switch belongs to.
Opening a promo claim page can now add your organization's ID to the link, when a remote setting switches that on
Two internal names for loading and rendering plugins were added to a list; nothing in this build uses them yet
Unclear It is not clear what the internal list is for.
3 more of these are in What probably matters to you, above.
Sessions set up by the runner get a governed git host list in CLAUDE_CODE_AGENT_PROXY_GIT_HOSTS and configure proxy git access for each host
Unclear It is not confirmed that CLAUDE_CODE_AGENT_PROXY_GIT_CONFIG is what turns this setup on.
A SendMessage already approved by auto mode is forwarded to Claude Desktop without being checked again, provided nothing about it changed
Unclear What Claude Desktop does with the forwarded ID is not clear.
In cloud runner environments, the per-session trust bundle for governed git is now deleted when the session ends
Unclear What turns governed git on is not stated.
Policy limit checks now skip re-downloading unchanged data and can accept a check already fetched by a host app
Unclear Nothing shows which host app, if any, currently hands over a prefetched result.
In Cowork, memory context entries with an anchor are no longer shown directly; the latest unanchored version's content is inserted before the anchor
Unclear What triggers this code and whether it is switched on anywhere is not stated.
Usage data from auto mode decisions now notes whether a Chrome action was a pointer press and whether a click was routed
Unclear It is not clear whether auto mode's actual decisions changed or only what it records.
When there is no project, Claude Code now uses a new fallback to decide which folder holds subagent transcripts
Unclear It is not clear whether transcript locations change for ordinary sessions that do have a project.
Questions from the AskUserQuestion tool can now be shown as a title with separate body lines instead of one block
Unclear It is not known which parts of Claude Code ask for the split, so it may not be visible yet.
Claude Code now recomputes installLocation from each entry's source when it loads known marketplaces and installed-plugin entries
Unclear It is not clear how the location is recalculated or when it differs from the saved value.
Claude Code's sampled check of SDK output now keeps one example per message type and checks them later, when the session is idle
Confirming a usage-credit request to your admin now records whether the eligibility check and pending-request lookup both succeeded
Unclear How Claude Code shows or uses this recorded result is not known.
A model removed from the model list by an administrator's managed policy is now recorded as "managed_denied" instead of being counted as allowlist-filtered
Unclear It is not clear whether this reason is ever shown to users or only recorded in diagnostic data.
Six more Windows NTSTATUS error codes, for memory, open-file limits, device and timeout failures, now map to specific POSIX errors
Unclear How Claude Code uses these codes, for example whether it retries on them, is not settled.
Claude Code now allows a diagnostic event for file uploads on managed-agent tasks
Unclear It is not clear what in Claude Code sends this event or when.
In remote sessions, usage metrics can now go to the session's own metrics address, signed with CLAUDE_CODE_SESSION_ACCESS_TOKEN
When WebSearch fails, the usage data Claude Code reports now says whether it was in standard or extended mode
When a turn starts, Claude Code now tags it as started by a person or started automatically
Unclear Where this tag is used after it is recorded is not settled.
For HIPAA-flagged organizations, or before policy limits load, usage metrics keep only known values and replace custom names with placeholders
Usage data sent about folder syncing and artifact publishing gains a few extra fields, with no change to how either behaves
Once a session knows its request to resume thinking is refused, later refusals are no longer each reported in usage data
The environment information Claude Code reports now carries a version, a base version and a build time
Unclear It is unclear whether the version and build time describe the runtime Claude Code runs on or the Claude Code build itself.
Claude Code now caches whether metrics are enabled separately for each account and organisation, so switching accounts cannot reuse another's answer
Metrics sent to Anthropic keep only allowed details when a HIPAA policy applies, or when your organisation's policy hasn't loaded yet
Unclear It is not clear which attributes are on the allowed list.
When starting another program is refused, the error report now uses the refusal reason as its message
Unclear Whether this is visible to users at all is unclear.
Uploads to the Files API can now include an expiry time when one is given
Unclear Nothing in Claude Code appears to pass an expiry to this upload yet, so it may not be used.
In internal builds only, a feedback share link can now also be set when the upload is finished on the host machine
In diskless mode, Claude Code no longer reuses a saved temporary folder path and works it out again on each use
Unclear It is not fully settled that this applies only when the diskless launch option is on.
A check based on earlier evidence no longer clears a record that health data covered by HIPAA was seen
Unclear It is not clear which feature keeps this record or when it is used.
In diskless mode, the task list now guards its changes with in-memory locks instead of lock files on disk
Unclear How Claude Code decides it was launched diskless is not firmly established.
The feedback dialog still skips one extra request when a server-side flag is on, the same as in the previous build
Unclear It is not clear what the skipped second request does.
Claude Code no longer loads its standalone security dialog in non-interactive sessions
Unclear It is not fully settled that the check used here is the one that tests for an interactive session.
When a local machine refuses to run tools for a cloud session, Claude Code sends the server a short refusal notice with a reason code
Unclear The refusal report was already present in the previous build, and which part of it differs in this release is unclear.
On Windows, Claude Code can now report "served-profile-dirs-unreadable" as the reason remote tool serving is off
Unclear It is not clear what remote tool serving covers or where this reason is shown.
When a call to a connected machine fails because it is not connected, the error now also records that the machine is not answering
Unclear Where the new marker is read and what message it produces is not clear.
If Claude Code cannot read a switch for its tool-host feature, it treats the switch as off and logs an error
Unclear It is not clear what the tool-host feature does or which switches it reads.
When the bridge is not connected, the failure is now flagged as the machine not answering, the same way as a failed liveness check
Unclear It is not clear what part of Claude Code reads the new flag, or whether it changes how these failures are shown or retried.
Failed host file operations now report whether the storage backend refused them, with an error code, or simply could not perform them
Unclear Which file operations reach these new errors is not traced.
Starting the device bridge now always waits for a readiness signal, creating one when none was supplied
Unclear What the readiness signal waits for is not known.
Headless runs, subagent checks and a path helper now cope with a session that has no working directory instead of always expecting one
Unclear What puts a session into the state of having no working directory is not known.
Claude Code's activity tracking now records idle time spent waiting for your input, separately from its working periods
Unclear It is not clear what reads these activity records or whether a setting controls them.
When a new remote session was started because of the monorepo flag, Claude Code no longer records its settings scope as "project"
Unclear What the "project" scope controls later in the session is not stated.
With no working folder, Claude leaves the folder and git lines out of its setup notes, and MCP servers are not given a folder
Unclear How a session ends up with no working folder is not known.
hostEnvironment joins the list of session metadata fields, next to the version, git repo and remote session ID
Unclear It is unclear whether this list marks fields to keep or fields to strip, and what the host environment value contains.
When a remote session's result goes out before its transcript, the result no longer carries the final answer text
Unclear It is not clear what setting makes Claude Code send the result early.
Before offering file sync, Claude Code can now measure git history size for repositories that started empty, and checks untracked size first
Unclear It is not clear when the "created_empty" starting point is used or what the first-upload limit is.
When the remote store is full, the result is now reported as kept on this machine rather than as a failure
Unclear It is unclear which sync or transfer feature this result belongs to.
When work is kept on this machine instead of sent, the failure now carries its actual reason instead of always saying it was over the cap
Unclear It is unclear which feature this result belongs to.
On Linux, a file identity check now treats only missing-file and permission errors as a mismatch and reports any other error
Unclear Which everyday actions rely on this check is not stated.
Cloud folder sync now refuses with the reason "unreadable" when its private admin folder cannot be created because of the attribute source
Unclear It is not clear what the attribute source is or whether this sync path is used by readers.
Artifact startup can request a lighter boot and switches to a full boot if the lighter one comes back incomplete
Unclear Which features ask for the lighter boot is not known.
An internal artifact check can now request a lighter response from the server and skips some record-keeping when it gets one
Unclear It is not clear what requests the lighter response or when.
A publishing feature now records rounds skipped over a weakened sync point as skipped rather than upload-only
Unclear It is not clear which Claude Code feature these publishing journals belong to.
After writing an artifact, Claude Code skips the extra read-back request when the write response already echoed the content
Unclear It is not known which storage backends send the content back with the write, so it is unclear how often the read-back is skipped.
Startup usage data now records where the precompute compaction setting came from, and reports it as on when it is not set
A context window set by an experiment can now qualify for the exemption that 1M-token windows get from some automatic compaction checks
Unclear Exactly which checks the exemption skips, and which experiment sets the window, is not known.
Automatic compaction now reports how much of its summary has streamed in, using the same progress events as a normal reply
Unclear It is not clear whether or how this progress is shown on screen.
When compaction begins, Claude Code now resets the response length counter and switches the display to a requesting state
Unclear How the display actually renders these changes is not confirmed.
The routine that clears old session data now empties a checkout's trash but keeps the git links a synced checkout still depends on
Unclear It is not clear what the three new cleanup steps remove, or which entry is kept.
Periodic cleanup of ~/.claude/shares now also removes temporary .zip.tmp files, not just finished .zip files
Unclear What the extra check tests before a temporary file is removed is not known; it may check that the process that created the file has ended.
A cleanup routine now removes set-aside and unreadable copies of installed_plugins in the plugins folder instead of zip files in the shares folder
Unclear Which code writes these set-aside and unreadable files, and when the cleanup runs, is not known.
Fetching the published plugin catalog and its signature file can now tell caches not to serve stored copies
Unclear It is not clear when Claude Code actually asks to skip the cache.
Plugin evaluation runs that fail during git setup are now recorded as git preflight failures, not as other failures
Unclear Where the git setup check runs is not known.
Four plugin failure types, such as an invalid plugin ID, are now reported as validation errors in Claude Code's error reports
Unclear Whether the new category also changes any message shown on screen was not checked.
A Remote Control machine now re-registers itself when its tool-hosting details, Claude Code version or platform change, not only after an error
Unclear It is not clear whether this behaviour sits behind a setting or a server-side switch.
Claude Code's background daemon now records how many tools each remote control worker is serving and updates its state when that number changes
Unclear It is not clear where, if anywhere, this count is shown to you.
Remote Control no longer gives project-thread child sessions their usual grants when a component called the floor cannot be reached
Unclear What the floor refers to is not established.
Resuming a forked skill now looks it up only by its exact name, no longer through synced or namespaced skill names
Unclear It is not clear whether this lookup was removed or moved somewhere else.
Account-level skills drop the field that marked them as made by Anthropic or by a user, changing how their origin is decided
Unclear It is not clear how Claude Code now decides whether an account skill comes from Anthropic.
Skills loaded from the synced-skills folder no longer carry a class label; their server details are looked up by skill name instead
Unclear What used the class label, and so what behaviour changes without it, is not clear.
Checks on PowerShell commands now return the list of paths they accepted instead of a simple yes or no
Unclear What, if anything, this changes in the permission prompts a user sees is not confirmed.
The message shown when Remove-Item is blocked on a protected system path can now be replaced with a more specific reason
Unclear It is not clear which cases, if any, now show a custom message instead of the standard one.
On Windows, busy, too-many-open-files, out-of-memory, timeout and similar errors are now reported instead of quietly switching to a weaker fallback
Unclear It is not known what happens next once one of these errors is reported.
Claude Code now checks each background worker type separately to decide whether it may start, instead of only allowing heartbeat
Unclear Which worker kinds the new check allows is not known.
Live supervisors no longer get an activity timestamp set directly, and custom skills and agents no longer record whether Anthropic or a user made them
Unclear The last-activity update may simply have moved into one of the new routines, in which case supervisor behaviour is unchanged.
Client data can now turn off a model feature that is on by default, where before it could only turn features on
Unclear Where the client data comes from is not settled.
Claude Code's save of the current project's config now reports success or failure instead of returning nothing
Unclear Which parts of Claude Code act on the new success or failure result is not known.
Confined evaluation runs now set GIT_PROXY_COMMAND to empty, so git cannot run a proxy program passed down from outside
Unclear It is not settled whether GIT_PROXY_COMMAND was already cleared elsewhere, or what the extra git config values now contain.
When passing certificate settings to git, Claude Code can now set the trusted certificate file per site and per full address, not only globally
Unclear It is not clear where this is used, and it may be a renamed internal rather than a change you would notice.
A timer-based wake-up for MCP task workers now appears in Claude Code's list of known event sources, next to the existing wake-up
Unclear The code that triggers this wake-up was not found, so it is not clear whether it runs.
If an MCP server fails to connect after shutdown cleanup has begun, Claude Code no longer logs that server's error output
Unclear It is unclear whether this log line ever reached you on screen or only went to the debug log.
Claude Code now has the label "Background memory extraction" for a background memory process, next to the existing consolidation label
Unclear What starts the memory extraction process, and where the label is shown, is not settled.
A standing memory note is now marked as not saved when the server says the memory store has reached its memory or size limit
Unclear It is not clear where this is shown to you, or whether the memory feature is available to every account.
Every way of loading a saved session now reads a dev mods folder value and keeps it on the session
Unclear It is not clear what sets this folder or what feature uses it.
While hook modules are still loading, Claude Code now checks each command against those modules instead of applying one blanket check
Unclear What Claude Code does with the result of this per-command check is not stated.
The AskUserQuestion screen now works out whether a question is a confirmation Claude Code itself requires
Unclear What this changes on screen is not known.
Claude Code's check on shell commands joined together now inspects two lists of commands at once and looks at different lists for output redirects
Unclear What the new check tests for, and so which commands are treated differently, is not known.
Skill names starting claude-ai: are no longer mapped to anthropic-skills:, and plugin skills can now be reached by shorter aliases
Unclear It is not known whether anything still produces claude-ai: skill names.
Chained pushes in cloud directory sync are now always on, and CLAUDE_CODE_DIR_SYNC_CHAIN is no longer read
The bundled claude-test plugin drops its read-only code-mapping agent and now lists claude-test as a command it may run
Unclear It is not clear what listing a command under runCommands allows the plugin to do.
The spinner row no longer shows a percentage bar while a conversation is being compacted
Unclear The progress bar may have moved somewhere else rather than being removed.
Permission allow rules written for a synced skill's former name no longer carry over when the skill is renamed
Unclear It is not confirmed whether this matching moved elsewhere rather than being dropped.
An optional setting called anthropicNamespaceReserved was removed from the options Claude Code accepts when it starts an agent
Unclear It is not clear whether any behaviour for reserved names in plugin agents ended along with this option.
The status line beside the spinner no longer shows how far compaction has got as a percentage
Unclear It is not clear whether the compaction percentage moved somewhere else rather than being removed.
Published verbatim by Anthropic for v2.1.283. 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 94 bullets, 36 name something an entry on this page also names, 15 name something no entry here does, and 43 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.
x-claude-code-prompt-id to the gateway hint headers so LLM gateways can group the requests that serve one user prompt; opt in with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1
Probably new-x-claude-code-prompt-id-request-header-behind-the-gatew availableModelsMatch managed setting: with "exact", an availableModels entry allows only the model version it names, so new releases stay blocked until listed
Probably switching-to-the-default-model-now-checks-managed-availablem, managed-deniedmodels-now-also-matches-the-host-managed-provi, default-model-steps-down-to-an-allowed-model-when-managed-se, managed-policy-deniedmodels-now-restricts-the-model-picker, new-managed-settings-deniedmodels-and-availablemodelsmatc deniedModels managed setting to block specific models, even when availableModels allows them
Probably switching-to-the-default-model-now-checks-managed-availablem, managed-deniedmodels-now-also-matches-the-host-managed-provi, side-queries-to-a-denied-model-are-sent-to-the-main-loop-mod, default-model-steps-down-to-an-allowed-model-when-managed-se, managed-policy-deniedmodels-now-restricts-the-model-picker, new-managed-settings-deniedmodels-and-availablemodelsmatc, managed-policies-deniedmodels-from-the-catch-all-policy-is tool.output OpenTelemetry span event when OTEL_LOG_TOOL_CONTENT=1
Probably mcp-and-selected-tool-outputs-recorded-on-tracing-spans /doctor prompt-audit (also /checkup prompt-audit) to audit your CLAUDE.md files, skills, agents and commands for prompting patterns written for older models
Probably doctor-prompt-audit-runs-the-claude-api-skills-prompt-au path to --plugin-dir load-failure entries in the stream-json system/init plugin_errors, naming the directory that did not load
Probably sdk-init-plugin-errors-is-now-public-gains-path-and-do, validator-errors-on-plugin-and-marketplace-names-claude-code load_test_mode block to the Claude apps gateway config: requests are built and signed but not sent upstream, and clients get a canned reply, so a deployment can be load tested
Probably gateway-load-test-mode-simulates-kernel-io mantle upstream provider to the Claude apps gateway for Amazon Bedrock's Mantle endpoint
Nothing to match on result.usage
No entry names this /mcp no longer offers Authenticate for such servers
Probably mcp-tool-list-org-blocked-tools-can-be-focused-and-expla, mcp-mcp-servers-refused-for-a-reserved-name-no-longer-fall, new-tengu-starry-locket-gate-drops-unregistrable-mcp-servers, user-configured-mcp-servers-named-memory-are-refused-when /usage and the VS Code usage meters when telemetry is disabled
No entry names this /model accepting Sonnet 4.6 or Sonnet 5 with [1m] when the id carried a date or -v1:0 suffix, in the cases where the plain id was refused
Probably switching-to-the-default-model-now-checks-managed-availablem, sonnet-1m-alias-detection-broadened, sonnet-1m-context-availability-check-now-covers-any-sonnet, default-opus-gets-the-1m-suffix-only-where-1m-context-is, opus-subagents-can-get-the-1m-context-variant, managed-policy-deniedmodels-now-restricts-the-model-picker /model picker showing a hardcoded Haiku version and price when ANTHROPIC_DEFAULT_HAIKU_MODEL pins a different model
Probably switching-to-the-default-model-now-checks-managed-availablem, managed-policy-deniedmodels-now-restricts-the-model-picker, disable-prompt-caching-haiku-now-matches-the-configured-haik DISABLE_PROMPT_CACHING_HAIKU having no effect when Haiku is the session's main model
Probably disable-prompt-caching-haiku-now-matches-the-configured-haik #93295[BUG] `DISABLE_PROMPT_CACHING_HAIKU` is silently ignored when Haiku is the main model — undocumented main-model guard, absent on `_SONNET`/`_OPUS` Closed
claude plugin validate saying Claude Code accepts a plugin or marketplace name it cannot install; such names in marketplace.json now fail validation
Probably claude-plugin-validate-checks-more-manifest-paths-flags-j claude plugin validate passing plugins whose outputStyles, themes, monitors, or lspServers paths are missing or point outside the plugin directory
Probably claude-plugin-validate-checks-more-manifest-paths-flags-j claude plugin details showing 0 MCP servers for plugins that declare their servers in plugin.json
No entry names this claude plugin marketplace remove not saying which installed plugins it uninstalled with the marketplace; it now lists them
Probably removing-a-marketplace-now-uninstalls-its-plugins-and-lists claude plugin uninstall removing the other of two installed plugins whose ids differ only in case, with its options and secrets, when the one named had no enabledPlugins entry at that scope
Probably plugin-installuninstall-refuse-when-install-records-are-unr, plugin-enabled-state-sync-leaves-an-unreadable-installed-plu installed_plugins.json showing no plugins when it holds a record under an invalid plugin id; such a file loads again
Probably installed-pluginsjson-unreadable-or-foreign-records-are-ke, plugin-install-refuses-when-installed-pluginsjson-is-in-an, installed-pluginsjson-the-file-is-checked-before-migration, plugin-enabled-state-sync-leaves-an-unreadable-installed-plu, claude-plugin-list-reports-a-rebuilt-or-cleaned-up-install installed_plugins.json being rewritten, losing records, when it holds a record this version cannot read; claude plugin commands now name the record and say how to recover
Probably installed-pluginsjson-unreadable-or-foreign-records-are-ke, plugin-install-refuses-when-installed-pluginsjson-is-in-an, plugin-installuninstall-refuse-when-install-records-are-unr, claude-plugin-init-self-validation-warns-instead-of-failin, installed-pluginsjson-the-file-is-checked-before-migration, plugin-enabled-state-sync-leaves-an-unreadable-installed-plu, claude-plugin-validate-checks-more-manifest-paths-flags-j, claude-plugin-list-reports-a-rebuilt-or-cleaned-up-install, claude-plugin-eval-requires-git-231-or-newer, removing-a-marketplace-now-uninstalls-its-plugins-and-lists /context not counting MCP server instructions: they now appear as their own row and count toward the total
Probably context-shows-mcp-server-instructions-as-a-separate-categor #94832[BUG] /context reports 0 tokens for MCP server instructions that cost 653 per turn Open
claude mcp add, add-json, and remove reporting success when the user or local config file could not be written, for example inside a sandbox
Probably claude-mcp-addremove-now-re-reads-the-global-config-to-co #95016claude mcp add reports success when the config write is denied (sandboxed Bash) Open
cmd an alias of meta, which could produce cmd+ shortcuts most terminals never send
Nothing to match on keybindings.json silently accepting a misspelled modifier such as ctl+k; it now warns in the debug log and suggests the fix
Probably message-selector-keybindings-migrated-to-generic-select-ac footer:openSelected was rebound or unbound in keybindings.json
Probably message-selector-keybindings-migrated-to-generic-select-ac, footer-view-tasksmemoriesartifacts-hints-follow-the-foot GIT_CONFIG_COUNT environment pairs
Probably claude-plugin-eval-requires-git-231-or-newer, git-tlsssl-settings-passed-through-git-config-key-n-env-var git asking credential helpers to store the sandbox proxy's login, which printed "failed to store"
No entry names this #96307Sandbox proxy credentials are sent to Git credential helpers on macOS (`fatal: failed to store: 100001`) Open
sandbox settings being ignored entirely when one nested value was invalid; the invalid value now fails closed and the rest of the block still applies
Nothing to match on DISABLE_TELEMETRY or DO_NOT_TRACK
Probably remote-control-no-longer-needs-the-feature-flag-for-subscrib, remote-sessions-refused-with-an-explicit-reason-when-telemet /remote-control menu cutting its QR-code hint mid-word in narrow terminals
No entry names this . dropping a Shift+Enter newline, leaving the cursor inside an accented letter, and repeating an older change after 3J or Visual-mode J on the last line
Nothing to match on V then p now lands on the first non-blank
Nothing to match on J joining lines with different spacing than Vim (such as a space before ) or after a tab), and 3J or Visual-mode J on the last line not moving the cursor as Vim does
Nothing to match on cmd /c rd, rmdir, del or erase delete drive roots, the home folder and other folders that Remove-Item refuses
Probably cmdexe-rdrmdirdelerase-on-protected-paths-blocked-in-pow, powershell-protected-path-removal-denial-accepts-a-custom-me, powershell-guard-blocks-cmd-c-rd-rmdir-del-and-erase-aime /mcp tool list: it shows more tools at once, scrolls with the page keys and mouse, and marks tools your organization blocked with a warning icon
Probably mcp-tool-list-org-blocked-tools-can-be-focused-and-expla, mcp-mcp-servers-refused-for-a-reserved-name-no-longer-fall, new-tengu-starry-locket-gate-drops-unregistrable-mcp-servers, user-configured-mcp-servers-named-memory-are-refused-when /tasks: rows show a status icon, the name and whole facts, the title and key hints stay on screen with many tasks, and the list gains paging keys, the mouse wheel and clicks
No entry names this /help, /hooks, /copy, /chrome, /memory, /ide, /release-notes, /rewind, /diff, /remote-env, /plugin and other pickers with page keys, mouse wheel and clicks
Probably plugin-not-cached-not-installed-errors-now-say-when-th, a-bundled-hooks-modules-commandrun-hook-now-fires-only-for, plugin-marketplace-picker-and-cloud-session-skills-dialog, memory-settings-rebuilt-as-a-toggle-list, screen-reader-mode-copy-picker-and-slash-command-list-rend, ide-picker-uses-a-new-list-with-a-current-ide-check-mark, hooks-matcher-list-redrawn-as-list-items-with-source-badge, rewind-picker-rebuilt-on-the-shared-scrolling-list-component /skills and /artifacts, to draw their pointer dim while the search box has the keys, so only one pointer is highlighted
Probably plugin-marketplace-picker-and-cloud-session-skills-dialog prompt-audit on Claude Code configuration: stale paths, stale commands and contradicting instruction files now lead the report, and thinking keywords that Claude Code documents are kept
Probably doctor-prompt-audit-runs-the-claude-api-skills-prompt-au installed_plugins.json that cannot be read at all: its contents are kept in a file beside it before it is rebuilt, and claude plugin list names that file
Probably installed-pluginsjson-unreadable-or-foreign-records-are-ke, plugin-install-refuses-when-installed-pluginsjson-is-in-an, installed-pluginsjson-the-file-is-checked-before-migration, plugin-enabled-state-sync-leaves-an-unreadable-installed-plu, claude-plugin-list-reports-a-rebuilt-or-cleaned-up-install claude -p and Claude Code Remote no longer load the interactive UI, and the auto-mode classifier's rules and the Artifact tool load on first use instead of at launch
No entry names this permissions.defaultMode still overrides it
Probably tengu-harbor-willow-now-defaults-on-sessions-can-start-in-a /ultrareview launch dialog to say that reviewing a local branch may upload uncommitted changes to tracked files
No entry names this /model picker's Opus row and the Default model's name to drop "(1M context)" where Opus already has a 1M context window; the window is unchanged
Probably switching-to-the-default-model-now-checks-managed-availablem, managed-policy-deniedmodels-now-restricts-the-model-picker --system-prompt and --append-system-prompt to accept their text and -file forms together; the file's text comes first
Probably procselffd-identity-check-no-longer-swallows-unexpected-e, cloud-memory-file-picker-and-hook-matcher-list-moved-to-a-ne, win32-dirsync-anchor-transient-io-errors-now-propagate-ins, system-prompt-and-system-prompt-file-can-now-be-given-to Skill(anthropic-skills:<name>) deny rules to also block that skill when Claude Desktop delivers it as a plugin, and Skill(skill:<name>) denies to match the skill's alias and display name
No entry names this /rewind and /diff lists to move on the same keybinding actions as every other list (select:); messageSelector:/diff:* rebinds still work
Probably a-bundled-hooks-modules-commandrun-hook-now-fires-only-for, rewind-picker-rebuilt-on-the-shared-scrolling-list-component /workflows run list to size itself like other lists: half the terminal inline, and it keeps its title on screen when the prompt shows below
No entry names this claude plugin eval to require git 2.31 or later when git is installed; a run on an older git is refused with a message naming the version
Probably claude-plugin-eval-requires-git-231-or-newer pre-push hook, ignore a writable system core.hooksPath, and not sign commits without --configure-git
No entry names this GIT_SSL_CAINFO and GIT_SSL_NO_VERIFY under Anthropic-managed git: the runner's own git always verifies Anthropic's git route, and warning lines say what applies where
Probably runner-git-helpers-take-scoped-configenv-push-on-release-r, self-hosted-runner-builds-a-certificate-file-for-anthropics, runner-moves-git-ssl-cainfo-git-ssl-no-verify-into-session claude-ai name: skills, commands, workflows and MCP servers' skills and prompts so named load again, and Skill(claude-ai:*) rules are ordinary prefix rules
Probably the-claude-ai-skill-prefix-is-dropped-and-plugin-skills-g A model matched these bullets to the GitHub issues they fix, so a link can be wrong.
1 of 27 tool descriptions changed. The appended system-reminder blocks moved: 5 lines added.
Claude Code, interactive mode
813 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 96 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?