The haiku model name now picks Claude Haiku 5.5 on Anthropic's own API. Bedrock, Vertex and other providers stay on Haiku 4.5. Haiku 5.5 has a 1M-token context window and its own pricing. Desktop admins can now start Claude Code through a corporate launcher with the claudeCodeProcessWrapperEnabled setting. New inferenceIdpOidc and inferenceIdpAuthFlow settings extend company sign-in to more setups, including Bedrock proxies. Admins can also set a Microsoft Foundry base URL for Claude Code sessions from the desktop app.
Four features are in the build but not switched on yet. One is a startup check for team and enterprise users. It would block startup if organization policy or settings fail to load. Another would tell Claude which of its settings files a shell command changed. New keyboard actions for moving between tabs are listed but do nothing yet.
Waiting permission requests no longer time out while you are answering an input prompt. Large repositories stall less when starting a cloud session. MCP servers no longer show an old error after their tool list refreshes successfully. Remote Control now shows the session as running again after it reconnects mid-turn. Sessions that Claude Code starts on its own now keep your Chrome on or off choice. Away and back signals from people viewing a remote session are no longer passed on.
Written by our agent from the shipped bundle, not by Anthropic.
The MCP clientSecretHelper must now print a JSON object, managed config adds OIDC sign-in and session retention, and Workspace becomes Capabilities
Settings & config
isDeferred$.tool.register accepts an isDeferred true or false option, so plugin tools can load only when found through tool search
/claude-testClaude Test gains sign-in-dependent access and messages for when it is still loading or its browser helper failed to load
ElsewhereWith CLAUDE_CODE_PEWTER_OWL_TOOL set, the affected tool path now reports whether the turn ends and what to tell you, not a response ID
New inferenceIdpOidc and inferenceIdpAuthFlow settings extend company sign-in to more setups, including Bedrock proxies
Want the reasoning? Read walks the 42 entries that probably matter to you, each one opening to what changed and why.
Read this release → Every row →Anything you can use today, anything that visibly changes, and anything worth poking at. One line each, open for detail.
Claude Haiku 5.5 joins the model list and the haiku alias now resolves to it on Anthropic's API, while Bedrock, Vertex and other providers stay on Haiku 4.5
Unclear Whether Haiku 5.5 can be picked as a model yet is not known.
The MCP clientSecretHelper must now print a JSON object, managed config adds OIDC sign-in and session retention, and Workspace becomes Capabilities
Unclear It is not clear whether these options can be used from the command-line Claude Code or only from the desktop app.
project_write can now upload a file outside the working directory if this session read all of it, and Project names that differ only in Unicode form match
Unclear Whether these Projects tools are available to readers in this build or held behind a switch.
Managed settings are now prefetched by default for first-party users, fetch failures are classified, and a dark-launched org-config check is measured
Unclear It is not settled whether anything in this release uses the verdict to actually stop a session.
Claude Code's list of keyboard actions gains strip:tabNext, strip:tabPrevious and strip:viewAll, with no change in behaviour yet
isDeferred#$.tool.register accepts an isDeferred true or false option, so plugin tools can load only when found through tool search
Admin policy now applies a bare Bash rule to PowerShell too, offers Auto mode in Code unless autoModeEnabled is false, and treats an empty folder list as none
Backgrounding a session with the left arrow carries queued prompts to the background job or blocks the move, and passes CLAUDE_BG_CARRIED_PROMPTS_SHA256
CLAUDE_CODE_HOST_SKILL_CATALOG refuses "." and ".." and drops its child-session check, CLAUDE_CODE_DESKTOP_SKILL_SWITCHES is read, and some startup details are skipped
Unclear What the skill switches do, and in which runs the environment details are skipped, is not known.
/claude-test#Claude Test gains sign-in-dependent access and messages for when it is still loading or its browser helper failed to load
Unclear Whether /claude-test is available for a given account is decided remotely and is not identified.
With CLAUDE_CODE_PEWTER_OWL_TOOL set, the affected tool path now reports whether the turn ends and what to tell you, not a response ID
Unclear What a reader who has set CLAUDE_CODE_PEWTER_OWL_TOOL sees differently when the turn now ends or carries on.
Claude Code now checks a signed model catalog, the details behind the model picker, every 5 to 15 minutes over the network
Unclear It is not settled whether this catalog check runs in the Claude Code command-line tool or only in the desktop app.
Models can now carry a higher price for prompts over a token threshold, used in cost accounting and shown in the pricing line
Unclear It is not stated where Claude Code shows or uses this pricing, so it is unclear which cost figures you see would change.
Automatic moves to a larger-context model and Fable-style model handling now check how the account is billed and whether extra usage was agreed
Unclear It is not clear what this model-specific behaviour does for a user who gets it.
New inferenceIdpOidc and inferenceIdpAuthFlow settings extend company sign-in to more setups, including Bedrock proxies
Unclear These settings are read by the desktop app's settings readers, and it is not clear whether the Claude Code command line also honours them.
A new managed setting, claudeCodeProcessWrapperEnabled, makes the desktop app start Claude Code through the launcher named in processWrapper
With inferenceGatewayManagedClaudeCode on, Desktop runs Claude Code in gateway mode for gateway sign-ins, turning off usage analytics
Admins of Claude Desktop third-party deployments get settings for the built-in browser and for how long sessions are kept
The managed MCP config schema gains a policy-only transport, an OIDC redirectHost option, and checks for cloud-metadata hosts and unsafe helper arguments
Admin-managed settings gain a Built-in browser group, a desktop target for requireFullVmSandbox, and the $schemaVersion and $schema keys
start_task can now upload attached files#Calls to an MCP tool named start_task now check their input and upload attached files first, on remote HTTP servers
Unclear Which MCP servers offer a start_task tool is not known, so it is unclear who can use this.
manifest.json#Some plugins can now list their skills in a manifest.json that switches individual skills off; a broken file loads none of them
Unclear It is not clear which plugins count as account-managed, so it is unclear which plugins read manifest.json.
A new coworkVmIpv6Enabled setting gives the Cowork VM IPv6, credential helpers can return headers, and pin changes apply without restart
Organization-managed settings for Cowork and third-party deployments gain controls for scheduled tasks, keep-awake, plugin MCP servers and sharing
Unclear It is not clear how many of these settings the Claude Code command line reads, as opposed to the desktop app.
Admins of third-party Claude Desktop deployments get managed settings for default model effort, model catalog and Chrome password managers
Each model in the model list can now set maxEffort, the highest reasoning effort level offered for it
Unclear It is not clear whether the cap is applied anywhere other than Claude Desktop.
Desktop enterprise settings gain a Microsoft Foundry base URL, passed to Claude Code sessions as ANTHROPIC_FOUNDRY_BASE_URL
Hooks that ran somewhere other than the tool can no longer change its input or result: input changes become a denial, result changes are dropped
Claude Code can now hold hooks until a pre-check finishes, and can mark a hook's result as having already run somewhere else
Unclear It is not clear what the hooks now wait for or which sessions this affects.
A new remotely controlled switch, off by default, sits alongside the new outcomes for restoring settings
Unclear It is not clear what this switch controls.
The list of policy settings groups now includes a label called "Workspace VM network", next to "Network proxy"
Unclear Nothing shows which settings, if any, belong under this label.
Claude Code's settings list now has a "Built-in browser" group inside the sandbox settings
Unclear Which settings belong to this group and whether they do anything yet is not stated.
Whether creating a cloud session sends your machine's settings now depends on several consent signals, with a notice step in the terminal
Unclear Exactly which combination of these choices causes settings to be sent.
Admin settings for the Claude Desktop app gain sshTransport, desktopHome and dangerousMaxVersion
In organization policy settings, a group entry can now restate the settings of a named item, and each item may appear only once per group
CLAUDE_CODE_THINKING_DISPLAY_UPDATES can switch off thinking display updates#Thinking display updates are on by default, and CLAUDE_CODE_THINKING_DISPLAY_UPDATES can override that
Unclear It is not clear what thinking display updates show, or what value turns them off.
The gateway external identity provider setting now describes a redirectHost: "localhost" option, where the return address used to be fixed at 127.0.0.1
Unclear Whether this setting applies to the Claude Code command-line tool or only to the desktop app.
A credential helper script can now be given up to 32 arguments and a separate path to use on Windows
Unclear It is not clear whether these fields affect the command-line Claude Code or only the desktop app.
Admin policy for the desktop app gains a live-apply rule for plugin marketplaces and a per-server tool policy override in managedMcpServers
For url marketplace sources, a credentialHelper may print HTTP headers as a JSON object instead of only an Authorization value
Unclear It is not clear whether helpers could already print headers before, or whether this is newly possible.
A new disableLocalConfigCache admin setting keeps only the organization ID and restrictions on disk, alongside other Claude Desktop policy changes
Task events and task lists gain subagent_type, naming the agent type or main-session, and an awaited flag
What would fix it?
Smaller changes and internals, grouped as the pipeline found them. Nothing is dropped, it is only further down.
10 more of these are in What probably matters to you, above.
Claude Code now rejects compressed responses from MCP servers in tool containers, and MCP connect and tool listing were reworked
Unclear It is not clear what the new encoding option changes for a given server.
Subagent results now carry canContinueAgent, and background-agent notices only suggest messaging the agent when the caller can actually do that
When an account is using overage, the 1-hour prompt cache is turned off and a cache-touch step is skipped
Unclear It is not clear which other condition must also hold for this to apply.
Remote settings sync now keeps credential-looking files and rules on this machine, reports them, and logs dropped credential rules as warnings
Unclear Whether this screening is always on is not known, so it is unclear whether every user's sync holds these files back.
Device hooks registration now checks a tool-serving flag, and reports "not switched on for your account" rather than failing when the flag says off
When local settings and CLAUDE.md files are sent to a cloud session, anything that looks like a credential is now held back
Claude Code now loads MCP resources and cached tools step by step, and has a new marker for trimming large MCP results
Unclear Whether large MCP results are actually cut short in this build, and how much of a result a reader would lose.
If the managed MCP server list cannot be read, every MCP server is now blocked; entries can also be policy-only
Unclear What Claude Code does when the managed MCP server list fails to be read is not shown.
Publishing an artifact from an artifact type can be approved automatically in some cases, and a hint now suggests a motion type for videos
Unclear Whether the artifact type catalog is switched on, and the full set of conditions for automatic approval, is not known.
The design artifact quickstart now suggests a motion type for animations and demo videos when available, and a new path can skip the type-create approval prompt
Unclear What the skipped approval prompt now lets through is not known.
Messages waiting in the queue are no longer run while a session hand-over is under way, and the queue is checked again afterwards
Unclear What triggers a session hand-over is not known, so it is unclear when a user would see the queue pause.
Tools from SDK MCP servers set up at the project level no longer pass extra context to the check that approves actions in auto mode
The loop that checks for new sessions to start now limits requests to free capacity, spaces out polls and waits for running hooks before exiting
In a diskless session, command, script, HTTP and mcp_tool hooks are now held back and not run, and an error is logged
Unclear What makes a session diskless, and so whether your hooks are affected.
Synced skills now load only when Claude Code can read the host's skill catalog, and nothing loads when it cannot
Self-hosted runners now accept an optional slot count in the server's hints about starting work, and log it
In cloud sessions, a PermissionRequest hook that is a command, script, HTTP or MCP tool hook now causes the request to be denied
Claude Code does not forward hooks to a device on Windows or when CLAUDE_CODE_DISABLE_HOOK_FORWARDING is set, and records when hooks are held back
On Anthropic's own API, Claude Code may use Claude Haiku 4.5 instead of the default small model when a condition is met
Unclear It is not clear what condition decides when Claude Haiku 4.5 is used.
Hooks from your computer are only offered to a claude --cloud session when the feature is switched on for your account
Claude Code now warns if a settings file on your computer changed while a command from a cloud session ran, and says which settings layer changed
Claude Code now checks queued commands when they are taken from the queue and again before they start, and drops the first one if it steps aside
Unclear Nothing shows what makes a queued command step aside, so which prompts this affects is unknown.
The /run skill now picks its prompt based on whether its code is loaded, slow to load or not loaded, and shows a different notice in those cases
Subscribers who get the 1-hour prompt cache can be moved to a 5-minute cache for certain models when a server-side switch is on
Unclear Which models the shorter 5-minute cache can apply to is not known.
New instructions tell Claude to spell out paths and names literally in commands, since your machine may question lines it cannot fully read
Unclear It is not clear which sessions include these instructions, possibly only remote or cloud ones.
The claude.ai skills sync now waits longer between checks when you are idle, and wakes up again when you return
The consent that decides whether your local settings are sent to a cloud or remote session is now worked out differently
Unclear Exactly which consent choices now lead to settings being sent and which do not.
Check-ins about silent background shell commands now come once per agent, list every quiet command and say when the next one is due
When a call to an MCP server gets no answer, Claude is told it may have finished, still be running or never started, and to check before repeating
Claude Code now refuses a worktree whose .git file leads to another repository than the one its trust and settings are kept under
If the connection drops mid-call, Claude Code no longer resends an MCP tool call whenever the server may already have it, and reports an error
Claude now advises signing the Chrome extension into the same organization as the conversation, and warns that logging out erases its shortcuts
Claude Code now treats overage consent as needed whenever your plan has a limit end date, even if that date has passed
Unclear It is not clear whether this makes the overage consent prompt appear more often in practice.
Claude Code now looks up a pricing entry from the model and the usage details before working out an estimated cost
Unclear Which kinds of usage get a different price is not known, so it is unclear whose cost estimates will change.
Claude's instructions for building artifacts now say to pin each library to an exact version at least two weeks old
A compaction summary now ends with a note that the messages kept after it were not seen when it was written, so tasks it lists may already be done
The 'Gateway auth scheme' setting is now marked as running on the desktop app, alongside its existing third-party scope
modelPrefer1mContext and disableConfigDeprecationWarnings are now marked to take effect as soon as a remote policy changes them
Unclear It is not confirmed that the hot-applied mark makes these changes take effect without a restart.
Claude Code can now say it refused to start because required organization settings could not be loaded, or were refused for this sign-in
Moving a session to the cloud now reports when it has moved, asks to confirm, and cleans up if the request fails before sending
Managed settings now explain what a retired setting did, and a local address on a fail-closed setting falls back to the safe value
Unclear It is not settled whether this applies to the Claude Code command-line tool or only to code bundled with the desktop app.
After a long gap or a forced redial, a remote session's connection can now announce "returned" instead of always "reconnected"
Unclear Whether a setting or server switch controls the new "returned" announcement is not known.
The note on what was not sent to a cloud session now lists credential-like deny or ask rules not in force there, plus text settings
Artifact patch edit requests can carry a side_reply request class set from the server, invalid model lists now switch the feature off, and related small changes
Claude Code now treats bypass permissions mode as accepted if the bypassPermissionsModeAccepted setting is set or another condition is met
Unclear Which other condition counts as acceptance, and whether the agreement prompt is now skipped more often, is not clear.
Claude Code now passes bypass permissions mode on to other sessions only if your user settings actually show you agreed to it
Cowork admin changes to the banner and userPluginUploadsEnabled now apply to sessions that are already running
In remote sessions, a turn cut off by a rate limit is no longer auto-resumed ahead of a message you have already queued, when a remote switch is on
Unclear What the remote switch falls back to when the server sends nothing is not known.
Artifact creation through a gateway relay now waits and retries once when busy, and says when approval or policy refused it
Claude Code no longer hands work off to the background on a timer while you have unsent text or a question is open
Claude in Chrome file uploads now reject credential files like .env, .pem, .key and .ssh, plus device paths and over-long paths
Claude Code now explains when left-arrow backgrounding is cancelled by unsent text or delayed by a question waiting for your answer
Hosted managed config now rejects an MCP server's OAuth client secret unless it is a Google client, and points others to oauth.clientSecretHelper
Unclear It is not clear how much of this checking applies to the command-line tool rather than to Desktop and the hosted control plane.
While teleport is sending, its screen can now show a question for you to answer and a custom progress line
Unclear It is not clear what questions or progress lines teleport actually shows, or when it asks them.
When Claude Code sets up a provider's environment, it now also blanks CLAUDE_CODE_USE_ANTHROPIC_AWS, CLAUDE_CODE_USE_ANTHROPIC_GOOGLE_CLOUD and CLAUDE_CODE_USE_GATEWAY
When Claude Code packs your configuration for a cloud environment, files and rules that look like credentials are dropped and counted
On third-party providers, Haiku model names are no longer skipped by the check that sets up the Sonnet default model
Unclear What the check now does for a Haiku model name is not known.
A change in a task's awaited flag now triggers a background_tasks_changed message, and awaited tasks are no longer tracked as background work
Published verbatim by Anthropic for v2.1.293. 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 56 bullets, 9 name something an entry on this page also names, 15 name something no entry here does, and 32 name nothing specific enough to line up either way. The pairings are made on names both sides wrote down, a flag or a setting or a slash command, so read one as probably the same thing rather than as a fact, and read the middle number as candidates rather than as a miss count.
claude-haiku-5-5), now the default Haiku model on the Anthropic API — 1M context, $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K)
Probably haiku-alias-moves-to-claude-haiku-55, haiku-alias-now-defaults-to-claude-haiku-5-5-with-third-par, claude-haiku-5-5-model-family-recognised, claude-haiku-55-model-added-with-vertex-region-override agentType to the subagentStatusLine payload, so scripts can tell custom subagent types apart
No entry names this isDeferred to $.tool.register for mods: false lists the tool's schema in the prompt from the start instead of behind tool search
Probably mods-can-register-deferred-tools, toolregister-accepts-isdeferred ← moved the session to the background; if a queued message can't move, ← now stays put and says so
Nothing to match on /model effort ←/→ wrapping past the highest or lowest level, which could accidentally save Low as a model's default effort
No entry names this #99212/model picker: pressing Right on the highest effort level wraps to Low and saves it as the default Closed
/tui disconnecting Claude in Chrome in a session started with --chrome, and ignoring --no-chrome
Probably spawned-sessions-now-forward-chrome-no-chrome SendMessage in sessions, including resumed ones, where a host, a permission rule or a --tools list removed that tool
Probably resumed-subagent-run-flag-documented-for-sendmessage-waits, orphaned-background-agent-resume-messages-depend-on-whether, agent-subagent-report-and-messaging-prompt-strings-refactore, restarted-background-agent-notices-adapt-when-the-agent-cann #92183[BUG] Desktop app disallows SendMessage, so subagents cannot be messaged or resumed Closed
--agent sessions being told a built-in tool was disabled for the whole session when only their own tool list left it out
No entry names this /clear
Probably safety-stop-count-field-on-results claude logs, stop, kill, rm and claude daemon status, stop, uninstall sometimes signing you out when your login had expired or was about to
No entry names this worker being shown as "Agent" when it starts, and the agent detail dialog's title losing the agent type once the agent finishes
Probably tool-not-found-errors-tailored-for-the-agent-tool-and-disabl, background-agent-completion-notice-omits-send-it-a-message, subagent-results-now-report-cancontinueagent Artifact("(unprintable path)") while a publish call was still streaming in
No entry names this /ultrareview upload on Linux wrongly refusing some repositories, such as one inside another checkout, over a settings file that "could not be parsed" while a sandboxed command was running
No entry names this /ultrareview upload's refusal over split-index files advising a git command that could leave git unable to read its index
No entry names this claude remote-control
No entry names this #92661PushNotification reports "Remote Control inactive" (no_transport) in sessions served by `claude remote-control` Open
#99781PushNotification reports "Remote Control inactive" in sessions started with claude rc Closed
classic.* events being skipped while the plugin hooks worker restarts, which left settings hooks to answer without them
No entry names this claude plugin test failing for mods that call $.session.append; tests can read the appended rows back with the new mock.session
No entry names this claude plugin eval refusing every Bash-granting run on Macs with Docker Desktop (links under ~/.docker/bin); the refusal now names which part of a credential store held the link
No entry names this #93368`claude plugin eval` refuses every Bash-granting evaluation under Docker Desktop's default `~/.docker` layout; `DOCKER_CONFIG` does not bypass Open
#94308`claude plugin eval` refuses every Bash-granting run on a Mac with Docker Desktop (symlinks in ~/.docker/cli-plugins) Closed
desktop policy sets Claude Desktop's built-in browser keys, such as builtinBrowserEnabled
Probably desktop-3p-config-built-in-browser-policy-keys-added #100016[BUG] Claude gateway rejects builtinBrowserEnabled in policy desktop block Open
claude agents offering bypass permissions that background sessions then ignored when consent was saved only in .claude/settings.local.json or a --settings file; it now asks for consent first, and a session that ignores bypass shows a short notice that stays
No entry names this ← backgrounding a session after 10 seconds while the prompt held unsent text (it now cancels the move) or a question was waiting for your answer
Nothing to match on ← had just been pressed to move the session to the background
Nothing to match on /feedback returning to the drafts list after Ctrl+O or Ctrl+Z while a report was sending, leaving the send impossible to cancel
Probably feedback-transcripts-redact-account-memory-compact-summaries claude purge stopping silently (exit 0, or a hang in a terminal) when a file or folder could not be deleted; it now deletes the rest, lists what it could not delete, and exits 1
No entry names this >> and << on a line of only spaces leaving the cursor past the end of the line, where a following x deleted nothing
Nothing to match on V then d) the cursor lands on the first non-blank, and . after it acts on the cursor's line
Nothing to match on /loop wakeup or scheduled task; Claude is no longer told, and the session stays asleep
Probably org-managed-config-schema-gains-sharing-endpoint-scheduled claude_code.at_mention logging to emit at most 100 agent and 100 MCP-resource events each time a prompt is read
No entry names this A model matched these bullets to the GitHub issues they fix, so a link can be wrong.
1 added and 1 removed, of 218 lines, about 9 words, in the prompt 14 of 27 arms receive. 3 other prompts also changed. 1 new prompt appeared. The appended system-reminder blocks moved: 1 line added.
Claude Code, interactive mode
No prompt capture for v2.1.293, so this release's prompt surface is unknown.
718 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 37 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?