Claude Code v2.1.312

Error results now include the reason startup failed#

When Claude Code returns an error result, the error now carries the last recorded reason startup failed

Auto mode explains refused connections to hosts outside the sandbox list#

A new refusal says a host is not in the sandbox's allowed domains and auto mode does not approve connections to other hosts

Unclear It is not clear when this message is shown, or whether it is in use at all.

MCP tool results try smaller forms before giving up#

When building an MCP tool result fails or is too big, Claude Code now tries smaller forms and fails only if none of them fit

Resetting effort now names the actual level instead of saying "auto"#

The message after resetting effort now names the effort level in use rather than always saying "auto"

The near-limit usage warning can now show how much you have used#

The usage warning can now read 'Over N% of your usage limit used' instead of only saying you are close to the limit

Feedback drafts count Claude's replies more accurately#

Feedback drafts Claude queues no longer include request ids from failed replies, and count replies without an id

Send-now hint is hidden if you have bound the send-now key#

A keyboard hint is no longer shown when you have set your own key for chat:sendNow

Interrupted tool results and safety-check refusals are marked#

One display path now marks tool results as interrupted, and refusals caused by safety checks are now labelled as such

Fast mode messages about usage credits are reworded#

The messages shown when fast mode turns off for lack of usage credits now say you are out of credits or they are not available

Windows paths are checked more carefully before writing to Claude Code's own folders#

On Windows, Claude Code now resolves alternate path spellings before writing to its protected folders, and may refuse a write it cannot trace

Unclear It is not clear which of the new checks depend on the server setting, or which run outside Windows.

Advice shown after a rejected publish now depends on whether forcing was refused#

When a publish is rejected because of a conflict, the closing advice now differs depending on whether a forced publish was refused

Unclear Which Claude Code feature this publish step belongs to is not known.

Changed repair of tool input passed to hooks#

Claude Code reworked how it fixes up a tool's input before passing it to hooks, now keeping extra fields and checking more tools

Unclear It is not clear what a hook now receives that it did not before.

Claude Code now watches only one plan file per plan#

Claude Code now watches only the single main markdown file for a plan, not a list of files for that plan

Unclear It is not clear whether plans ever had more than one file in practice, so the effect for you is unknown.

macOS sandbox now blocks commands from listening on the proxy ports#

On macOS, sandboxed commands can no longer bind to or accept connections on the proxy ports, but can still connect out to them

Moving a session to the cloud retries a rejected sign-in in more cases#

When moving a session to the cloud fails with a sign-in error, Claude Code can now retry even if the sign-in token did not change

Credit notice now says "Now using credits"#

The short notice shown when you switch to paying with credits now reads "Now using credits" instead of "Now using usage credits"

Logging out is only confirmed when Anthropic's server really answered#

A token cancellation at logout is no longer counted as confirmed when the success reply lacks an Anthropic request id

One warning now covers every refused tool call from a hook on another machine#

A hook on another machine that changes a tool call now gets the same "rewrote the input, the call was refused" warning for every hook type

Git credential helper for the local proxy now uses 127.0.0.1#

The git credential setting Claude Code writes for its local proxy now names 127.0.0.1 instead of localhost

Tool status marker gains a dim, uncolored neutral look#

The status marker beside a tool call can now show a neutral state: dim, uncolored, and not announced as an error

Sandbox proxy settings use 127.0.0.1 instead of localhost in some cases#

In the sandbox, Claude Code now points proxy variables at 127.0.0.1 instead of localhost when an authentication value is set

New wording while a session moves to the cloud#

The message while moving a session now reads "Moving this session to the cloud…" and says it continues from your branch's last push

One kind of outgoing request no longer follows redirects#

A JSON request Claude Code sends now fails if the server tries to redirect it, instead of following the redirect

Cloud folder syncing watches for changes more reliably#

The file watcher for synced cloud sessions now also watches the parent folder and handles setgid and symlinked folders

Asking for usage in a remote chat now points you to claude.ai#

A remote chat that cannot show usage now says so and links to Settings → Usage on claude.ai

Unclear Which command or chat surface shows this message is not established.

Remote control recovers more precisely from a rejected sign-in#

When starting a remote control session is refused as unauthorised, Claude Code reuses the token it sent and records why

Tool calls to a disconnected remote host now say why it is offline#

When a remote host's tools are withdrawn, tool calls to it return a host_offline error with the reason instead of a generic message

Requests to Anthropic no longer follow redirects to another site#

Claude Code's own requests to Anthropic now stop with an error if they are redirected to a different site

Remote sessions say why a machine is unreachable when it signed off#

When a remote machine has signed off, the unreachable message for its tools now says what it left for, and remote request activity is tracked

Interrupted tool calls are now shown apart from rejected ones#

The display of a refused tool call now knows whether it was interrupted or rejected

Shorter wording in the model-switch warning#

The warning shown when you switch models now ends at "means the full history gets re-read." without "on your next message"

Organization settings errors no longer always say they are retrying#

When organization-managed settings fail to load, the error no longer always says "Retrying" and instead reflects what Claude Code will do

Agent calls cut off by a container restart now resume or report clearly#

If a container restarts mid-turn, Claude Code resumes interrupted agent calls or tells Claude plainly why it cannot

/clear keeps agents' settings and survives a deleted working folder#

/clear now carries over each agent's settings file and falls back to a safe folder if the original one was deleted

Background session commands wait for a sign-in refresh before exiting#

The logs, attach, respawn, rm and stop commands, and --bg, now wait for any sign-in refresh to finish before exiting

Compaction now keeps messages from the matching result#

When Claude Code compacts a conversation, the messages it keeps now come from the same compaction result as the rest of the outcome

Unclear The two values may always have been the same, in which case nothing changes for users.

Fewer needless re-login prompts when a credential was held back on purpose#

A rejected request whose credential was deliberately withheld no longer triggers a token refresh or the sign-in failure path

Policy opt-in no longer overwrites a data-sharing choice you already made#

Accepting the opt-in in the data policy dialog now leaves an existing data-sharing preference unchanged

Paths with Unicode spaces in the middle are now accepted#

A path safety check now allows unusual Unicode spaces inside a path but still rejects them at the end of a path segment

Removing a worktree in an untrusted folder no longer reports success#

If no WorktreeRemove hook runs because the folder is untrusted and the worktree still exists, removal now reports a failure

SDK sign-in recovery only swaps the token when it is actually new#

After a sign-in failure, Claude Code now applies the token from the SDK callback only if it differs from the one already in use

Resumed sessions read back their saved autocompact state#

When a stored session starts again, Claude Code now passes along its saved autocompact state

Effort level is no longer re-sent for unanswered turns#

Claude Code now tracks which turns were answered, so it does not re-send the effort level for turns that got no reply

WebFetch's partial-page note now uses different start, end and length values#

The WebFetch note saying an answer covers only part of a page now builds its character range and page length from different values

Remote sessions ask for credentials again and resend once after being refused#

When a remote session request sent without credentials is refused as unauthorised, Claude Code now asks for credentials again and resends it once

Remote sessions recover from a sign-in error on a connection that had no token#

If a remote session's live connection is refused because it carried no sign-in token, Claude Code now waits for one and reconnects

Skills passed through from MCP now keep their name and description#

When the Skill tool runs a skill passed along from an MCP server, it now carries that skill's name and description

Unclear Where these passed-along MCP skills come from, and whether this is switched on, is not established.

Fixed sign-in refresh handling for pull request calls in bridge sessions#

Pull request calls that fail with an authorization error now check the right response, and some requests echo ticket headers back

Subtitle wrapping and vim dot-repeat get small fixes#

Claude Code now checks whether a subtitle fits on one row, and vim dot-repeat now keeps the column the cursor aims for

Vim mode gj and gk now handle running out of screen rows#

In vim mode, gj and gk, alone or after an operator, now check whether there are rows left to move to and place the cursor accordingly

A second note can tell Claude that staged files are still being copied#

A new note tells the model that staged files are still being copied, behind a remote switch that is off by default

A new remote on/off switch that defaults to off#

Claude Code now checks a new remotely controlled switch that stays off when the server sends nothing; what it controls is not yet visible

Unclear What feature or behaviour this switch controls is not known.

Three new server-controlled switches appear#

Claude Code now checks three new switches that Anthropic controls from its servers; what they do is not known

Unclear What the three switches control is not known.

A hidden feature check now listens to a different remote switch#

A hidden feature check now depends on a different remote switch, and stays off unless that switch is turned on from the server

Unclear It is not known what behaviour this check controls.

Message limits between agents can be restored and reported when cleared#

The counter that limits chains of messages between agents can now be preset, list who was messaged, and report when it is reset

Version numbers containing "-aot." are now recognised as development builds#

Claude Code now treats version strings containing "-aot." like dev and engine builds, and an internal list gains "content-link-long-sessions"

A model check now passes only for Claude Opus 4.5, 4.6 and 4.7#

One model check now passes only for claude-opus-4-5, claude-opus-4-6 and claude-opus-4-7, with no fallback for other models

Unclear Which feature this model check controls is not stated.

Cloud tool helpers report clearer status and a more accurate live count#

Helpers serving tools to a cloud session now number their events and report a live count that ignores helpers being stopped

A restricted sandbox permission check no longer receives the permission mode#

One restricted permission check used with the sandbox no longer passes along your current permission mode

Unclear It is not stated how, or whether, permission decisions differ as a result.

Plan tracking reads tool results in more formats#

Claude Code now finds a plan's link in more kinds of tool result, and some workshop snapshot tracking was removed

Unclear It is not clear what these changes mean for what you see when using plans.

Served channel requests now signal when they start and end#

Requests on a served channel now trigger optional start and end signals around each request

Small changes to web fetch reporting, a skill menu line and memory reading#

Web fetch reports now carry the tool call's id, a skill shows a menu line, and blank memory-limit values are trimmed