Short model names such as opus and sonnet now work beyond Anthropic's own API. Fable is now recognised as a short model name too. A new Claude Test panel in the terminal shows each test's pass, fail or blocked mark, with a progress bar and time left. It builds on the groundwork for that panel in v2.1.309. On Vertex and Bedrock, default models can now come from a published model document. Setting CLAUDE_CODE_RETRY_WATCHDOG_SCOPE=capacity limits the automatic retry waiting to capacity errors only.
Several features are in this build but not switched on yet. A /claude-api check that scans your project for retiring Claude model names is built and stays off. Sending a project's documents to Claude as plain text is also off unless switched on remotely. The tool that writes project documents gains a safety check that is off by default. When switched on, it refuses to replace a document whose text differs from what Claude last saw. Groundwork for picking up background agents left behind after a resume is present but hidden.
Starting a cloud session now recovers when this computer is not yet a trusted device. Claude Code re-enrolls the computer and retries, or tells you to run /login. Saving incoming attachments no longer overwrites files that have the same name. Inline images inside tmux now stay in place when the pane moves, resizes or regains focus. A NODE_EXTRA_CA_CERTS file with one blocked certificate no longer causes the whole file to be ignored. The bundled claude-code.d.ts type definitions file is removed.
Written by our agent from the shipped bundle, not by Anthropic.
Short model names such as opus and sonnet are no longer limited to Anthropic's own API, and fable is now recognised as one
Sessions & agentsMCP server settings accept a new connectorId field, stdio servers keep their original configuredCommand, and mcp_call refuses some servers' tools
ExtensionsIn proxied sessions, the built-in gh now refuses to run outside the session's proxy, and a new standalone-proxy mode sets up git and gh
Tools & outputOn Windows, strict sandbox mode now explains why a command was refused, and the CLAUDE_CODE_NANKEEN_KESTREL switch remembers it was on
Interface
CLAUDE_CODE_RETRY_WATCHDOG_SCOPE limits the retry watchdog to capacity waitsSet CLAUDE_CODE_RETRY_WATCHDOG_SCOPE=capacity to keep the retry watchdog waiting through capacity errors without its other retry changes
Want the reasoning? Read walks the 22 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.
Subagent transcripts can be fetched incrementally against the local copy, and new switch-gated code finds launched background agents to restore on resume
Project documents can be sent to Claude as plain text, and project_write refuses to replace a document whose current text Claude has not seen
The built-in Claude Test plugin gets a terminal panel showing test progress, a pane note tool, and new settings for its browser tooling
Unclear It is not clear how the claude-test plugin is switched on or whether you can use it yet.
/claude-api can search the project for deprecated or retired Claude model IDs, behind the tengu_cosmic_sketch flag, with a prompt section only for one test group
Short model names such as opus and sonnet are no longer limited to Anthropic's own API, and fable is now recognised as one
Unclear Whether you can choose fable as a model is not stated.
MCP server settings accept a new connectorId field, stdio servers keep their original configuredCommand, and mcp_call refuses some servers' tools
Unclear It is not clear what Claude Code does with connectorId once it is set.
A session can now be messaged when someone comments on its artifact, and reading those comments can first need your approval
With CLAUDE_CODE_ARTIFACT_ASK_FIRST set, Claude asks a few questions in a form before creating a slides or design artifact
Claude Code now has a second route for claude.ai artifacts, through an MCP server, beside the built-in Artifact tools
Unclear What decides whether a session uses the MCP route or the built-in Artifact tools is not stated.
In proxied sessions, the built-in gh now refuses to run outside the session's proxy, and a new standalone-proxy mode sets up git and gh
Unclear Nothing shown passes "standalone-proxy" to the stand-in, so whether this path is reachable in normal use is unclear.
On Windows, strict sandbox mode now explains why a command was refused, and the CLAUDE_CODE_NANKEEN_KESTREL switch remembers it was on
In Projects sessions, Claude can now receive an urgent reminder after it has stayed silent for a set number of seconds
Unclear It is not clear what switches this reminder on or who supplies its configuration.
On Vertex and Bedrock, default models can now come from a published document per tier, falling back to models mapped in modelOverrides
Claude Code now recognises a goals type of attached context, but turns it into nothing, so nothing changes for you yet
Claude Code can now require that a helper agent's transcript not written on this machine is downloaded and verified before it is loaded
Unclear Nothing shows what switches this verification on, so it is unclear when it applies.
Behind a remote switch, Claude Code can rewrite MCP tool names to a standard form in what it sends and receives, and remembers old names
Unclear It is not clear what the standard names look like or whether existing permission rules for MCP tools are updated to match.
Claude Code can hand Claude a project's docs before its first reply, presented as file reads with a note saying Claude did not make them
Unclear It is not clear what switches this on or whether it runs for anyone in this build.
CLAUDE_CODE_RETRY_WATCHDOG_SCOPE limits the retry watchdog to capacity waits#Set CLAUDE_CODE_RETRY_WATCHDOG_SCOPE=capacity to keep the retry watchdog waiting through capacity errors without its other retry changes
A new message says a session has to agree to limit itself to your trusted devices before it can use a folder on your computer
Claude Code can now recognise tool errors saying a repository must be trusted or that a session has not been granted extra permissions
Unclear What Claude Code does when it recognises one of these errors, and when they can occur, is not shown.
Before sending the conversation to the model, Claude Code can now insert a notice plus a stand-in tool call and its result for a round host
Unclear What a round host is and when this insertion happens is not stated.
Project settings files may set CLAUDE_CODE_MODEL_CATALOG only to "off"; any other value is ignored with a warning
What would fix it?
Smaller changes and internals, grouped as the pipeline found them. Nothing is dropped, it is only further down.
9 more of these are in What probably matters to you, above.
The default model behind each family alias can now come from a document, with modelOverrides and Bedrock and Vertex checks adjusted to match
Unclear Which models have a published-document target is not stated.
Bash now steers Claude from sleep to the Monitor tool, and both Bash and Monitor refuse commands the sandbox policy forbids
Unclear The exact conditions under which either refusal happens are not known.
The artifact-comment auto-react flow gains an MCP hand-off path, drops an early exit and sends a notification when a rule is withheld
Unclear What condition decides whether this new hand-off route is used.
tengu_cobalt_plinth_larch now falls back to on for the artifact type catalog and artifact types, so they are on unless the server or an env var says otherwise
PowerShell hooks whose command is only quoted text now exit with a no-verdict code, and the Git Bash error explains exit code 127
Unclear What Claude Code does when a hook returns the no-verdict code, and which hook events check for it, is not stated.
The sandbox can now block writes when a git worktrees folder cannot be listed, controlled by a remote switch that can change without a new release
Unclear Whether this write block is on or off when the server sends no value is not stated.
Managed git and gh access can now go through a standalone proxy URL, and a new message tells Claude not to bypass TLS checks when setup fails
Cloud sessions get a shorter resume message, a rule for holding back results of certain tools, and a count of project document reads
Starting a cloud session now re-enrolls an untrusted device and retries, and tells you to run /login when that or the trusted-device check fails
The 'highlights' thinking display now only errors on Bedrock, Vertex or a gateway for models released before Claude Opus 4.7, and its reasons are reworded
Calls to the 'remote-browser' cloud browser MCP server that would skip auto mode's safety check or a person's approval are now refused
Self-hosted runners now send and log a list of capabilities when they register, and can prepare repositories at the same time
Anthropic-managed git sessions can now run through a self-hosted agent proxy, except on Windows and in a few other cases
Unclear It is not clear whether the server ever asks for the new kind of governed git that this mode now accepts.
A plugin's terminal pane now draws a full status layout with a link and messages explaining why a dev server failed to start, replacing two plain lines
Unclear Which plugin draws this pane, and whether it is switched on for anyone, is not stated.
The built-in Claude API skill now includes a guide on using Anthropic's models through Google Cloud
When a session goes to the background, it can now keep artifact subscriptions that wake it to react to comments
Unclear It is not stated what causes the built-in artifact tool to be hidden from MCP servers, so it is unclear which sessions this applies to.
Permission rules that deny or ask before the artifact REPL tool runs code now also match a second tool name
Unclear The second tool name the rules now cover is not stated.
Artifact publishing now refuses capabilities that cannot be confirmed as approved in the current Claude Code process
Unclear Whether this check is always on or controlled by a setting is not stated.
Claude now gets guidance for reading claude.ai artifact links through an MCP server, and clearer messages when those tools change
Under a new condition, --watch-artifact uses new notice text and the Artifacts tool asks separately for database reads and edits
Unclear What condition switches this new watching mode on is not stated.
Claude's automatic replies to comments on Artifacts follow new rules, including showing a reply as queued under a new condition
Unclear It is not clear what the new condition is, or when you would see a reply shown as queued.
MCP tools that carry a standard key can now be left out of the tools Claude is offered when a new condition is not met
Unclear It is not clear which MCP tools carry a standard key or what condition decides whether they are kept.
The remote control status line now lists missing sandbox dependencies when the sandbox fails to start, and has a VM-only case
For some MCP servers, Claude Code now reworks tool permissions, skips the overall tool-result size budget and tracks a standard key for grants
Unclear It is not clear which MCP servers this applies to.
In print mode, Claude Code can refuse further MCP server questions after a set number go unanswered in a row; off unless set remotely
The rules for uploading project files were reworked: reads follow symlinks, uploads name resolved paths, and one tool can suggest an allow rule
Unclear Which reads and uploads now behave differently for a user, and how, is not stated.
Claude Code recognises a second, runner-hosted Artifacts MCP setup, extends its guards to it, checks comment reads per artifact, and marks refused servers not connected
Unclear Which additional configurations are now recognised is not known.
Claude Code now warns that sandbox deny paths belonging to another computer are not enforced, with a count and a pointer to /status
Connector search, list and lookup tools can now use alternate prompts that tell Claude what to do when a connector needs you to sign in again
Unclear It is not clear when the new version of the instructions is used instead of the old one.
Model ids are now checked against two settings maps, and the /model picker can swap a disabled entry for an enabled one from the server
Unclear It is not clear what decides when the merging happens, so it is unclear which entries you will now see.
The file permission dialog has a second, fitted layout in a frame, and warns when it could not check where a path leads
Unclear It is not clear what switches the compact layout on, or whether it is on by default.
Moving a session to the cloud with uncommitted or unpushed work no longer fails: /teleport lists what stays behind and asks you to confirm
Claude Code now records a status for the five-hour and seven-day usage limits: allowed, allowed with a warning, or rejected
Unclear Whether this per-window status changes anything shown to you is not stated.
When sending OpenTelemetry data, Claude Code now builds its connection from your own certificate and proxy settings, not the shared setup
Unclear It is not clear which of your certificate and proxy settings now apply to the telemetry connection, or when the strict connection is chosen.
A cloud session not yet limited to your trusted devices now gets its own explanation before it can use a folder on your computer
If /claude-api cannot save its reference files, it now says onboarding from a URL is unavailable, and it ships more reference documents
All tool progress updates are now dropped in the remote and SDK message handler, not only heartbeats and retries
Unclear Whether tool progress is now shown some other way is not stated.
The check that picks out a custom host from ANTHROPIC_BASE_URL now reads the variable directly and returns nothing in some modes
Unclear Which modes cause Claude Code to ignore ANTHROPIC_BASE_URL here is not stated.
Cloud session setup progress now shows "Restored from snapshot" instead of "Restored saved setup"
Scheduled runs with nobody present now refuse artifact writes and uploads from linked, oddly named or uploads-folder files instead of asking
Unclear One further check runs before the refusal, and it is not clear whether it is a feature switch that could be off for some accounts.
When Artifacts are turned off, an MCP server that connects to them is listed as ignored, with a message naming the settings that can cause it
Self-hosted runners now prepare repositories through a scheduler with separate hook and git slots, replacing a simple one-after-another loop
Unclear Whether more than one repository is actually prepared at the same time in practice.
A deny or ask permission rule matched against an MCP tool now also checks your other deny and ask rules and a related parent rule, so it can cover more tools
Unclear Exactly which additional MCP tools a given deny or ask rule now covers is not spelled out.
Settings env is now tracked by whether it came from the repository, and telemetry variables are put back if no settings source still sets them
Under remote control, Claude Code now applies all your local permission rules only when the request is marked as here; otherwise only your deny rules apply
Unclear What location counts as "here" is not spelled out.
The check for reading comments can now call for your approval without a deny or ask rule, through a setting that is on unless turned off remotely
Unclear Which condition decides whether the new setting is consulted is not stated.
Claude Code can now skip task progress events that carry neither a summary nor workflow progress when it writes its output stream
Unclear It is not clear whether this skipping is on by default or only when a program asks for it.
Running steps in the workflow list can now show all their output instead of one line, when an internal check is on
Unclear It is not stated what turns on the full-output view, so it is unclear who sees it.
A connected or cached MCP server can now report its sign-in status as switched_off before any other result
Unclear It is not stated what makes a connected server count as switched off.
Claude Code now filters which plugins get their monitors started, and in some modes only plugins from a source that passes a check are started
Unclear It is not clear which plugin entries are skipped or which modes apply the source check.
Turning off the built-in artifact for MCP now also resets auto-react and wakes, and the failed-watch message drops its remedy when the artifact is hidden
Unless the mode is full, Claude Code now passes on only the soft_deny and hard_deny auto mode settings instead of all of them
If Claude Code cannot compute an edit or write preview in time, the approval dialog now says so and advises denying unless you expected it
On Linux systems where AppArmor blocks the sandbox, Claude Code now runs a test and gives a specific fix instead of a vague failure
Images shown inline in iTerm can now pass through tmux and be redrawn, resent or held back from drawing
Claude Code now checks tmux itself to decide whether kitty or iTerm2 images can show, keeps them in place, and explains how to fix it when they can't
Attaching or reading a file whose path resolves to a different real path, such as through a link, is now refused with the reason resolves_elsewhere
A NODE_EXTRA_CA_CERTS file holding a blocked or damaged certificate now keeps its readable certificates, with new warnings explaining what was dropped
The provider check was removed from the reasons Highlights can give for being unavailable
Unclear Whether a provider restriction still applies somewhere else is not stated.
MCP servers that reach the Artifacts server are now left out of self-hosted runner child sessions where artifacts are disabled or the tool is denied
The check for an API error saying effort is unsupported now matches one more kind of error besides the effort_unsupported message
Unclear What the newly matched kind of error is, and so which models or providers it covers, is not stated.
A remote tool call approved automatically is now rechecked, and it does not run if the target no longer allows automatic approval
Claude Code now refuses to read paths that name a device or process file, or that it cannot resolve well enough to rule that out
Organization survey questions now ignore input for a short moment after appearing, and Settings footer hints can be shorter
When a shell command fits, its permission prompt is now titled "Run command" and drops the "Do you want to proceed?" line
Unclear It is not clear what decides whether the compact layout is used, for example terminal height or a remote switch.
If another MCP server already uses a server's name, its allow rules are ignored and a warning now tells you how to fix it
When the sandbox cannot start, Claude Code now says so plainly and suggests installing tools only when a tool is actually missing
Sandbox deny paths on another computer are dropped, and Claude Code now warns at startup and lists them in claude doctor
On Linux and WSL, rewind now refuses to delete a file if its folder moved or was replaced by a link since the checkpoint
Unclear The default of the server-controlled setting behind this check is not known, so it is unclear whether the check is active.
A connector that fails because you are not signed in now reports that it needs authentication, not that its tools could not be listed
Bash permission checks treat git init, {} and ~ in arguments, and several more git config keys as needing approval
Unclear The new checks depend on an internal switch whose default is not stated, so it is unclear whether they apply for every reader.
The sandbox dependency panel now shows the actual error for bubblewrap and socat instead of just saying they are not installed
After a "prompt too long" error, Claude Code now uses the server's reported sizes to judge how much to trim first, instead of a fixed starting amount
A path that climbs out of its folder through a link's target is now refused unless the check explicitly allows it
When the sandbox is not enforcing, the auto mode warning now says auto mode asks before each shell command it can't check
Model allowlist matching now also tries the part of a model id after its last slash, and matches through catalog names and families
Gateway model discovery and gateway host checks now pass extra connection and certificate options, likely for custom certificate setups
Unclear It is not clear whether custom certificate settings now fully apply to gateway model discovery and managed settings requests.
Claude's Bash instructions now say long leading sleeps are blocked and to use Monitor to stream or poll, or background Bash for one wait
Anthropic has not published official notes for v2.1.310 yet. This section updates automatically when the entry appears in the upstream changelog. Everything else on this page came out of the bundle instead, which is why the two lists don't match.
3 of 30 tool descriptions changed. 1 of 28 tool schemas changed. The appended system-reminder blocks moved: 1 line added.
Claude Code, interactive mode
15 prompt changes in this release could not be quoted from the build, so no entry on this page describes them.
640 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 29 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?