Claude Code v2.1.314 更新加入工作目錄外寫入限制並修正多項問題

What's changed

Added permissions.blockWritesOutsideWorkingDirectories: Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode
Added an optional groups list to Claude apps gateway upstreams: only developers in the listed IdP groups are sent to that upstream, on failover too
Added guardrail_trace to Claude apps gateway Bedrock upstreams, to log which Amazon Bedrock guardrail filter fired on a request, without the matched text
Added idle compaction to VS Code and Claude Desktop's Code tab: a 200K+ token conversation left idle is summarized, and the summary is used on your next message if the prompt cache has expired
Fixed sessions failing turn after turn with "Invalid effort level" when an organization's admin caps a model's effort: Claude Code now lowers the effort, retries, and says which level it uses
Fixed claude -p exiting with code 0 and no output when startup failed; it now prints the error and exits 1
Fixed Bedrock sessions with many images failing every turn and /compact with "400 Input is too long." instead of removing older images and compacting
Fixed Microsoft Foundry sessions failing when the resource has no deployment for the default Sonnet, Opus or Haiku model: Claude Code now falls back to an older model of that family
Fixed repeated refused requests behind a gateway that doesn't accept tool search's tool_reference blocks: the request is sent again once without tool search, and the session remembers
Fixed TLS connections failing with UNABLE_TO_GET_ISSUER_CERT when the system certificate store holds a renewed root certificate that is cross-signed with its predecessor
Fixed MCP servers that also work without sign-in showing "needs authentication" after their stored sign-in stopped working; they are now used signed out
Fixed claude mcp logout reporting success when the MCP sign-in was not removed, for example while the macOS keychain is locked over SSH
Fixed extra sign-in renewals when several sessions share an Anthropic profile login and the credentials lock is busy
Fixed claude ultrareview and claude remote-control asking the workspace trust question in a CI job that has a terminal; with CI set they now refuse an untrusted folder at once
Fixed an administrator's TLS trust settings in managed settings not being applied in Claude Desktop sessions
Fixed model switching being refused on machines with managed plugin settings when one of your own --plugin-dir or --plugin-url plugins had a broken manifest or failed to download
Fixed the Bedrock and Vertex model upgrade offer coming back at every start when managed settings pin the model
Fixed forceRemoteSettingsRefresh set only in remote managed settings no longer applying once the stored copy had expired, which let sessions start without policy when the server could not be reached
Fixed "Continue automatically at usage limit" continuing more than three times without a prompt from you when a task restarted itself between waits (a /goal check-in, a background or scheduled task)
Fixed background and cloud sessions started with Ultracode turning it back on at their next start after it had been switched off
Fixed the last broken reply and its retry prompts staying in the conversation after Claude gave up on a tool call that could not be parsed
Fixed a PreCompact hook with very large output making compaction fail or repeat; each hook's output is now cut at 10,000 characters
Fixed background diff refreshes briefly taking the repository's index lock, which could make a git command run at the same moment fail with "Unable to create '.git/index.lock': File exists"
Fixed markdown tables jumping between grid and stacked layouts on half-typed rows while streaming, and long tables falling back to raw pipes
Fixed the /resume picker showing the text of CLAUDE.md or CLAUDE.local.md as a session's title when the session only ran slash commands
Fixed /btw answering from the main conversation when asked from a subagent's open transcript
Fixed pasting an image over SSH to a Linux machine doing nothing silently; the prompt now says to copy the file over and paste its path
Fixed claude remote-control sessions that failed to start when the session service was slow for a moment: the worker's registration is now retried with growing waits
Fixed model switches over Remote Control that take over 8 seconds getting a "still pending" error instead of the final result
Fixed the sandbox.excludedCommands setting description saying listed commands always run outside the sandbox; it now says every command in a call must match an entry
Fixed shell and MCP calls that auto mode, a deny rule or a PreToolUse hook blocked, or that otherwise did not run, reading as if they ran in a collapsed tool group
Fixed sandbox protections not applying to a worktree entered by claude -p --resume or --continue
Fixed hooks matching Bash not running on shell commands started by the Monitor tool; such hooks can block or ask, but not approve or rewrite those commands
Fixed a settings file excluded with --setting-sources still supplying some settings, such as auto mode rules and saved permission consents
Fixed auto mode reporting a call as blocked, and listing it in /permissions, when its safety check gave no verdict
Fixed a Claude apps gateway Vertex upstream hanging for minutes when a Google token endpoint accepts the connection and never answers
Fixed the Claude apps gateway's start-up warning for a deprecated Claude Desktop key that it still serves as written saying "stops being accepted" after the key's end date
Fixed /upgrade opening the browser and starting a new login for plans billed through the App Store or Google Play; it now says where to change the plan
Fixed --resume, and rewinding past a switch to another model after a refusal, moving a session on a model name Claude Code doesn't recognize to a different model
Fixed requests failing when a gateway or cloud provider refuses one of Claude Code's beta headers by name: the request is sent again without it
Fixed model IDs spelled with a dot behind an LLM gateway, such as claude-opus-4.8, being treated as an older model with the wrong context window, output limit, effort, thinking mode and cost
Fixed Remote Control clients showing a stray "Compacted (ctrl+o to see full summary)" row after /compact
Fixed the cloud sessions tip and the Remote Control idle notice showing when an organization's policy turns those features off
Fixed a resumed claude.ai cloud session making a model call nobody asked for when the previous worker stopped just after receiving a new chat's files
Fixed WebFetch in Cowork cloud sessions silently dropping a PDF's text past 100,000 characters; it now says how much was unread and where to read on
Fixed Claude being told only that a command was declined after you stopped it with Esc or Stop in a cloud session or the IDE; it now also hears that the command may have run
Fixed blank lines at the top of a code block disappearing when the code block opens a quote
Fixed a blank line appearing above some ⎿ rows in the transcript view (Ctrl+O)
Fixed vim mode dw, dW, yw and yW carrying on past the end of the line, and made w and W stop on empty lines
Fixed x and Enter in the agent panel sometimes acting on the wrong agent when another agent started or finished
Fixed pasted text and dragged-in files being lost in Claude's question dialog unless the "Other" row had focus
Fixed pasted text being lost when a row in the prompt footer was selected
Fixed vim mode p over a character selection ignoring whether the yanked text was whole lines
Fixed vim mode aw and iw taking the line break at the end of a line
Fixed ↓ selecting footer items hidden by the ? help menu, the command list or a paste hint
Fixed Esc on a command's panel, such as /btw, also interrupting the subagent whose transcript is open
Fixed the welcome header in fullscreen mode showing the session's model while viewing a subagent or teammate that runs on a different model
Fixed the fullscreen transcript being left mid-conversation after "Restore conversation" in the rewind menu while scrolled up
Fixed a rare case where a deleted session could reappear in the resume list as an almost empty entry
Fixed prompts containing tags such as <bash-stdout> or <local-command-stdout> not being acknowledged to a local stream-json host
Fixed sessions staying untitled when the first prompt failed (for example on a usage limit) and the conversation continued afterward
Fixed the skill file watcher checking every removed skill kept in ~/.claude/skills/.trash, which added file system load that grew with each removal
Fixed the session name from /rename disappearing while Claude asks for permission or a plan decision, so side-by-side sessions stay identifiable
Fixed background sessions that fail to start being reported as a deleted daemon binary when the binary is still on disk
Fixed the tip and idle reminder about push notifications appearing when the push settings in /config are hidden or your organization has turned off Remote Control
Fixed ~/.claude/skills/.trash and ~/.claude/plugins/.trash growing without limit: each now holds about 1024 entries, whatever cleanupPeriodDays says
Fixed the Bash edit diff listing a whole file as changed when a CRLF file under git's line-ending conversion was re-saved unchanged or had one line edited
Self-hosted runner: Fixed an idle session moving to another runner instead of pausing when the runner's request to pause it lost its connection before Anthropic answered
Improved memory use in long headless and SDK turns that run many subagents: the parent session no longer keeps each subagent's tool results until the turn ends
Improved prompt history: Ctrl+R now matches text inside collapsed [Pasted text] pills
Improved the /design-sync upload prompt to show the folder, the project and the files it will upload or delete
Improved /permissions review tab, now called Recently blocked: the footer says what Esc will do with marked rows
Improved auto mode's blocked calls: messages name the command, say Claude may retry and auto mode checks again, and point to /permissions to review
Improved permission prompts: a long reason takes fewer rows, so it pushes the command being approved off the screen on fewer terminals
Improved the cloud session error shown while a request refused for your claude.ai login is still being retried: it now also suggests claude auth login
Improved the behavior of stored claude.ai, Team, and Enterprise logins when MDM and remote settings differ on base URL or apiKeyHelper
Improved /config in agent view to say how to stop agent view opening by default
Improved session auto-titles: a session that starts with a skill or a prompt-based command such as /security-review now gets a title from it instead of staying untitled
Improved --resume speed and memory use on damaged transcripts
Improved /context speed in sessions with ten thousand or more MCP tools
Self-hosted runner: Improved the error shown when a session's branch isn't on the git remote: it now covers a never-pushed branch, says how to fix each case, and no longer suggests retrying
Updated the /claude-api skill: its Datadog setup for API tracing now sends the dd-llmobs-enabled: true header, as the Console does, and its ant install instructions drop the outdated macOS xattr step
Updated the bundled dataviz skill to advise printing y-axis tick labels in full under 10,000, and horizontal ones in full under a million where they fit
Updated the bundled dataviz skill to answer donut and pie chart requests with ranked bars or a segmented bar that prints its values
Changed Claude's file tools to ask before editing the local copy of your organization's server-managed settings, as they do for your settings files
Changed /ultrareview to check workspace trust before it runs git, as claude ultrareview does
Changed the Bash tool's shell to clear any inherited which alias or function and to stop passing exported which/unsetenv functions to commands, so which is the system command or zsh builtin
Changed /permissions Recently blocked to refuse approving a call you can't read in full, including one the terminal is too short to show
Changed auto mode in Cowork cloud sessions: a hook's allow of an MCP tool call now counts as no answer, so the classifier or a permission prompt decides
Changed hook if conditions such as Monitor(tail *) to match the Monitor tool's shell commands as Bash(tail *) matches Bash's
Changed the claude-api skill to ask to read a Managed Agents workflow run's workflow before changing the agent to improve that run
Changed flagged requests to retry on another model, not end in a refusal, when ANTHROPIC_DEFAULT_OPUS_MODEL, availableModels or deniedModels leaves only an older Opus model, or none
Changed idle compaction to also run for subscribers with telemetry turned off, and on Bedrock, Vertex, Foundry and LLM gateways once the one-hour prompt cache is on
Changed FORCE_HYPERLINK=false (also no and off) to turn hyperlinks off instead of on
macOS: Changed git grep in most commands that Claude runs to use the C locale, to avoid a reported memory leak in macOS's regex library on files that aren't valid UTF-8
[VSCode] Fixed "Clear authentication" for MCP servers reporting success when the saved sign-in could not be removed
[VSCode] Fixed missing spinner and Stop button while a queued message's hooks run
[VSCode] Fixed session group changes that cannot be saved, such as a delete, failing silently; an error notification now appears
[VSCode] Fixed Cmd+F / Ctrl+F not opening the find widget in the plan preview tab
[VSCode] Fixed the voice dictation shortcut (Ctrl+D) doing nothing under non-Latin keyboard layouts such as Russian, Greek or Hebrew
[VSCode] Fixed missed one-time saved scheduled tasks being deleted without asking when the extension opened without a chat
[VSCode] Changed CLAUDE_CODE_DISABLE_CFC_PROMPT to also stop the extension adding the Claude in Chrome instructions to a session's messages; the browser tools stay available
[Cloud sessions] Added an Include completed button to the Scheduled tasks page when finished one-off tasks are the only ones hidden, so you can show them without opening the Sort menu
[Cloud sessions] Fixed recently used self-hosted environments never appearing among the recent environments on the new session card
[Cloud sessions] Fixed cloud sessions reporting uncommitted changes every turn after Claude moved a branch to a commit it had pushed from another worktree
[Cloud sessions] Fixed the self-hosted environment details dialog in admin settings squeezing its content out of view in very short windows or at high zoom; the dialog now scrolls
[Cloud sessions] Fixed the permission mode picker staying on Auto after a running cloud or Remote Control session refused the switch; it now shows the mode in use and tells you
[Cloud sessions] Improved scheduled tasks: runs in Anthropic's cloud now load the plugins enabled on your claude.ai account
[Cloud sessions] Improved Test connection for GitHub Enterprise in admin settings to also connect the way cloud sessions do, and to warn you when that second test fails
[Cloud sessions] Improved cloud session transcripts: a rename row now names the new title, and repository and environment lists show as a table instead of one line of JSON
[Cloud sessions] Changed cloud sessions from the web, desktop and mobile apps to start even when a setup script exits with an error, instead of failing to start
[Remote Control] Fixed the permission mode button in Remote Control sessions showing the wrong mode until you typed; while your computer is online it now shows the real mode on open
[Claude Tag] Added a dismissible link to the Claude Tag API docs on the Bundles and Channels tabs in admin settings, for Enterprise organizations with Slack connected
[Claude Tag] Fixed @Claude !restart giving up after one failed workspace check in channels shared across Enterprise Grid workspaces; it now retries, then asks you to run it again
[Claude Tag] Fixed a second confirmation DM and channel note when you connect your Claude account from Slack twice from the same link within a few minutes
[Claude Tag] Fixed follow-up messages in Slack showing no sign of progress while Claude's session waits for a self-hosted runner; Claude now posts a wait notice in the thread
[Claude Tag] Fixed Claude sometimes not answering a tag in its own thread in Slack Connect channels where both organizations have installed Claude
[Claude Tag] Fixed Retry on plugin and skill pages in Claude Tag admin settings discarding unsaved access changes after a failed load; your edits now stay in place
[Claude Tag] Fixed the repositories control in Claude Tag admin settings showing the same action twice; its menu now lists only what the button doesn't already do