Claude 업데이트 기록

Radar Ultra

릴리스 42개

2.1.327

CLIGitHub 원문

AI 요약

  • Claude Code is less likely to lose conversation history or answers when resuming after a crash, restart, or backgrounding. Very long transcripts that end early are also retried once.
  • Several safety checks are clearer or stricter: repository changes through /install-github-app now need confirmation, and Claude asks before removing a worktree that may contain unsaved work. Some path and command permissions have also changed, so certain actions may now require approval or be refused.
  • Some fixes help prevent accidental loss or damage: /export asks before replacing an existing file, project MCP changes preserve unrelated .mcp.json keys, and a Remote Control worktree fix prevents stopping the server from deleting ignored files such as .env.
  • You can now pass MCP server JSON through standard input with claude mcp add-json <name> -, keeping tokens out of the command line. This option is not available inside a Claude Code session.
  • /status now identifies several configuration and setup problems, and /config notes when another setting overrides a row. Resumed sessions also show a message if earlier conversation messages could not be loaded.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

6.6395.0

App공식 원문

AI 요약

  • Claude Desktop on macOS now ignores user-level configuration when a machine-level profile exists. Before updating, move any settings you still need—such as retention periods—to the machine profile; a diagnostic report lists settings that were lost.
  • Cowork can no longer convert or render .eps and .ps files or run Ghostscript tools such as ps2pdf, because Ghostscript was removed.
  • Worktree sessions no longer update your local base branch or checkout files in the background. Use git pull to update the branch.
  • Some SSH and WSL file downloads are now allowed by default when remoteFileDownloads is unset, except for blocked file types. Device policy, a configuration server, Admin Console, or managed settings keep downloads limited; a script-installed setting does not, so set it to limited there if needed.
  • The token limit and Usage figures now count Auto mode’s built-in safety checks and stopped requests, so people—especially those using Auto mode—may reach the limit sooner.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.326

CLIGitHub 원문

AI 요약

  • Claude Code strengthens safety checks: more shell and PowerShell commands may ask for approval or be refused, and sandboxed commands have tighter limits on what they can write. Some previously allowed commands may now need approval.
  • Claude Code is more reliable when handling conversations and requests: fixes address lost conversation history after crashes or low memory, failed requests, and cases where Claude’s response or files were not preserved correctly.
  • Many terminal and prompt issues are fixed, including lost typed text, confusing permission prompts, and display problems in long conversations. The update also improves typing, streaming, and resizing responsiveness, especially in long sessions.
  • Remote Control replies now stream to your phone or browser as they are written. The update also fixes several Remote Control issues, including disconnects during replies and sessions showing the wrong activity state.
  • Claude Code gives clearer explanations in several places, including when features are disabled, settings or MCP servers fail, or a sign-in code is rejected. New /status information identifies when a provider or environment variable disables features.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.325

CLIGitHub 원문

AI 요약

  • Several security and permission protections were strengthened: some shell commands now require permission instead of running automatically, sandboxed writes to Claude settings and related files in added folders are blocked, and additional gaps in permission checks were fixed. You can keep sandboxed commands running without prompts by enabling sandbox.autoAllowBashIfSandboxed: true.
  • Sign-in and billing behavior is more reliable: background sessions on a claude.ai login no longer use or bill an API key or token from the shell that started them, and some cases that moved sessions to another sign-in are now refused with an explanation.
  • Fixed sessions losing recent turns after an unclean shutdown, and fixed context-length errors on every turn for many third-party gateways and Foundry, which had prevented compaction.
  • Claude Code now asks before running sandboxed Bash commands by default. To keep running them without prompts, set sandbox.autoAllowBashIfSandboxed: true; managed sandbox settings must set this in the highest-ranking managed source.
  • Plugin prompts now warn that plugins can run code on your computer and access files and the shell. Plugin loading also skips some unsafe paths and links.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

5.11492.5

App공식 원문

AI 요약

  • Sandboxed shell commands now ask for approval more often on macOS and Linux. In Manual and Accept edits, most ask first; Auto mode checks many. This can pause a routine at a prompt. You can restore the previous behavior by setting autoAllowBashIfSandboxed to true in the relevant settings file.
  • Some commands that use variable-based shell allow rules, such as Bash(NODE_ENV*), no longer match those rules and will ask for approval.
  • Workflows in local sessions can pause when you reach your plan’s usage limit, then continue after the limit resets—if it resets within a day and Auto-continue when limits reset is enabled.
  • Fixed a problem that could leave a session busy if another session sent it the same message twice or sent many messages at once.
  • Fixed turns failing when a gateway returned 403 for an expired credential-helper token or a sign-in past its stated expiry; Claude now gets a fresh token and retries.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

5.11492.2

App공식 원문

AI 요약

  • This release includes Claude Code version 2.1.321.
  • Administrators of SSH hosts: the SSH helper is version 87375cc7c944eb9d45c8d9267edc70a4b4827961.
  • The notes report no user-facing changes in General, Code, Cowork, or 3P.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.323

CLIGitHub 원문

AI 요약

  • Claude Code tightened safety checks: some Bash commands now ask for permission more reliably, and the sandbox blocks additional risky file and worktree changes. Automatic mode also falls back to Manual mode if it is unavailable.
  • Sign-in and background sessions are more dependable. Claude Code fixes several cases where sessions could switch accounts or lose authentication, and now stops with an explanation in some cases where it cannot keep the correct sign-in.
  • Added experimental, opt-in Code Mode, where Claude writes JavaScript to use its tools. You can turn it on or off for an organization.
  • Claude Code fixes issues that could lose or misplace conversation content, including after resuming, compacting, or restarting a cloud session.
  • Usage-limit handling is clearer: a menu offers choices when usage credits are unavailable or exhausted, and limit messages in several session types say when the limit resets and how to continue.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

5.11492.0

App공식 원문

AI 요약

  • Fixed several Windows problems: brief freezes when opening sessions or responding to approvals, blank windows after closing while minimized, and Setup offering some non-admin users an install that would remove their working app.
  • Fixed sign-in and saved-work issues, including Claude.ai sign-in being lost after a proxy error, artifact edits disappearing when leaving too soon, and messages being lost after a slow sign-in refresh.
  • Added a Download option for files from Cowork tasks that run on your computer, so you can save a copy wherever you choose.
  • Some browser and preview sites may stop loading if they require a client certificate: the app now sends one only to sites listed in device management. Behind a proxy that requests a certificate for every site, use *.
  • In local Code Auto sessions, Claude can start additional sessions for a task without approval if Auto's check allows it—up to 10 per message you type and 5 at once.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.322

CLIGitHub 원문

AI 요약

  • Fast mode is now available on Vertex in the global, US, and EU regions. If Vertex refuses fast mode, Claude Code falls back to standard speed.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

5.1231.2

App공식 원문

AI 요약

  • The bundled Claude Code version is 2.1.321.
  • The notes report no user-facing changes in General, Code, Cowork, or 3P.
  • The notes report no user-facing changes in General, Code, Cowork, or 3P.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

5.1231.0

App공식 원문

AI 요약

  • Scheduled task cards now explain whether a run failed, needs your input, didn’t run, or is waiting. Tasks may also retry provider or gateway server errors for up to six hours.
  • Claude Code now follows administrators’ MCP server allow and deny rules for servers added in the desktop app. Administrators can also turn off Auto mode without restarting; running sessions and tasks (except WSL sessions) leave Auto mode at the next settings check, usually within 10 minutes.
  • On Windows, Claude now saves its data in a new location for eligible installs and moves existing data there on Windows 10 21H2 or later. The change does not apply to the Microsoft Store version or third-party deployments.
  • Claude now asks before starting a project’s dev server in Auto or Bypass mode on Windows, matching its behavior on macOS and Linux. Cowork’s Auto mode also checks each shell command before running it; blocked commands, or ones the check cannot assess, do not run.
  • Fixed several problems that could interrupt work: edits on network paths could change or remove the wrong content, the Code tab could freeze with many sessions open, and Cowork could lose a comment you were typing when a file updated.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.321

CLIGitHub 원문

AI 요약

  • Claude Fable 5.2 is now the default Fable model. It has a 1M-token context window, and its listed prices are $8/$40 per million tokens, with cache reads at $0.20 per million tokens.
  • Several security and permission checks were fixed: MCP sign-in secrets are no longer sent to another host after a server redirect, and fixes address Bash commands or file operations that could bypass or misapply permission prompts and rules.
  • You can now pass --settings <file> to claude remote-control to apply that settings file to every session it starts. If the file changes, Remote Control will not start another session until you restart it.
  • Fast mode now tries standard speed on the same model first when fast serving is rate-limited, overloaded, or failing, instead of losing replies, switching models, or retrying for minutes.
  • Saved Fable 5.1 model selections now move to Fable on Pro and Max plans immediately, and on Team and Enterprise plans once the organization allows it. You can switch back with /model claude-fable-5-1.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

4.23118.1

App공식 원문

AI 요약

  • Replies can now continue after a connection breaks partway through, keeping the response received so far instead of stopping with a TLS handshake error.
  • On Windows, Claude Code no longer finds programs such as Git through network or relative PATH entries, or through links to network shares. If Git is found only that way, sessions use PowerShell instead of Git Bash.
  • When the sandbox is on, skills in a subfolder’s .claude/skills folder no longer load during the session. To use them, start the session in that subfolder.
  • On macOS, Claude Code now uses a newer signing certificate from the same Anthropic team. Team ID-based rules are unaffected; rules that pin the certificate itself need the new certificate added alongside the old one.
  • Plugin code that calls tools directly no longer shows an approval card. These calls run where only user approval or Auto mode would have stopped them, including connector tools without a toolPolicy entry; calls to tools set to ask are refused.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.319

CLIGitHub 원문

AI 요약

  • Auto mode’s safety checks now behave more explicitly when server checks aren’t available: actions may be denied or sent to you for approval instead of using the built-in classifier. You can keep the built-in classifier with CLAUDE_CODE_AUTO_MODE_SERVER=0; Bedrock, Vertex and Foundry are unchanged.
  • MCP tool calls that run as tasks can now move to the background immediately, with Claude told when they finish. This works even when feature-flag fetching is unavailable; set CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND to keep calls waiting, or CLAUDE_CODE_DISABLE_MCP_TASKS=1 to turn MCP tasks off.
  • You can configure permission prompts to be denied automatically instead of waiting for your response, so the turn can continue. This applies in terminal sessions, subagents and teammates.
  • /tasks now has an “Earlier in this session” list of finished subagents, which you can reopen and message, including after restarting Claude Code.
  • Several reliability fixes address common problems: older MCP servers can connect correctly, /diff works in more cases, and slash commands still run when typed after leading whitespace.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

4.23118.0

App공식 원문

AI 요약

  • SSH sessions are more reliable: they can reconnect after longer network interruptions, keep background work when your sign-in expires, and handle slow connection setup better. A message waiting more than a day is held for you to send manually; on Mac and Linux, reconnection waits up to 90 seconds unless SSH config sets ConnectTimeout.
  • You can minimize the Terminal, Browser, and iOS Simulator panes into a small live view and reopen them later. Moving and resizing the view is available when it sits in a chat corner, but not in third-party deployments.
  • Organizations gained more controls for where Code sessions can run, including settings to disable local or WSL sessions and offer approved SSH connections. SSH connections offered through the new beta setting are still checked against the organization's host allowlist.
  • Claude Desktop imports and exports can now include more of your data, such as projects, memory, scheduled tasks, saved skills, and uploaded plugins. Some imported items arrive turned off; exports include passwords and API keys contained in the included data, so check before sharing.
  • Claude now installs downloaded updates sooner on macOS and Windows, restarting after ten minutes with no Claude window in use and no task running. disableAutoUpdates still turns updates off.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.318

CLIGitHub 원문

AI 요약

  • Claude Code is more reliable when saving settings, sign-in details, and session files on shared network drives. It is also less likely to lose configuration or start without your MCP servers and trusted folders.
  • Several session and conversation problems are fixed: continuing a conversation is less likely to lose earlier messages, interrupted turns are less likely to disappear or be resent, and idle compaction no longer replaces a conversation with a non-summary reply.
  • Sign-in status is clearer: claude auth status and claude -p JSON can show when a stored claude.ai sign-in expires. A one-line warning appears in claude -p text output starting 10 days before expiry; expired sign-ins now give guidance on signing in again.
  • Permission prompts and approved commands behave more safely: keys typed before a slow prompt appears no longer answer it, and commands approved after a session restart run in the requested folder—or do not run if that folder cannot be entered.
  • Remote and cloud sessions have several fixes, including faster reconnection after a pause or sleep and fewer cases where sessions stop during a turn because uploads are refused for several minutes.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.317

CLIGitHub 원문

AI 요약

  • Claude now asks before a web or shell tool call sends your account email address to the network. Scheduled cloud tasks also warn if they include connectors you didn’t request and explain how to remove them.
  • Fixed several session and reliability problems: desktop MCP servers no longer disconnect after the first turn in affected sessions; saved login refreshes are less likely to sign you out on the next run; and sessions are less likely to lose conversation history or recent turns in the listed restart and compaction cases.
  • In cloud sessions, MCP server link requests now appear as cards you can open or decline instead of going unseen. Auto mode also no longer stops for approval when the Artifact tool uploads, writes to a database, or saves a file through the outputs folder.
  • The Claude apps gateway admin API can now report past spend by date range, with totals by developer, day, or identity-provider group. Daily group records are kept for the configured identity-retention period.
  • Added a warning when auto mode has no answer for a subagent’s tool call, and notices in Claude Desktop when your organization’s remote managed settings have not loaded for the session and when they do load.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

4.16336.0

App공식 원문

AI 요약

  • Fixed several message and session problems: dictated words and SSH messages are less likely to be lost, and project or local MCP servers no longer disconnect after the first turn in affected sessions.
  • Scheduled tasks now give notice when they cannot start or need a connector reconnected. Recurring tasks also retry provider outages for up to about three and a half hours, with notices about the first failure and final outcome.
  • Cowork previews are more restricted: HTML and SVG previews no longer show alert, confirm, or print dialogs, and can start a download only shortly after you click in the page.
  • Windows SSH now uses Microsoft OpenSSH 9.4 or newer for hosts new to the device, and OpenSSH is the default connection program. This may require allowing ssh.exe in firewall or endpoint rules; managed settings can keep the built-in transport.
  • Fast mode for Claude Opus 4.8 is deprecated and is scheduled for removal on November 30, 2026. After that, fast-mode requests will return an error rather than switch to standard Opus 4.8.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.316

CLIGitHub 원문

AI 요약

  • Fixed a security issue where --restricted sessions could upload files outside the working directories if a Read allow rule covered them.
  • Fixed model selection on Vertex AI and Amazon Bedrock: if the newest model isn’t available to your account or region, Claude now switches to an older model instead of failing mid-session.
  • Opus 4.8 fast mode is deprecated and will be removed on November 30, 2026. After removal, requests will error rather than fall back to standard Opus 4.8; migrate to Opus 5.5 to keep the speedup.
  • Cloud sessions now show a form card when an MCP server asks you to fill in a form, so you can respond or decline instead of having the request canceled unseen.
  • Fixed prompts with invisible characters so plain prompts send on the first Enter; commands still wait for review.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

2.1.314

CLIGitHub 원문

AI 요약

  • You can now block Write, Edit, and NotebookEdit from changing files outside your working directories, in every permission mode.
  • Very long conversations can be summarized after they sit idle. The summary is used for your next message if the prompt cache has expired; this also works with telemetry off and on Bedrock, Vertex, Foundry, and LLM gateways once the one-hour prompt cache is on.
  • Claude Code handles several common failures more clearly: startup failures with claude -p now show an error and exit with code 1, and sessions retry with lower effort when an organization's limit is below the model's requested effort.
  • Cloud sessions started from the web, desktop, or mobile apps now start even if a setup script exits with an error.
  • Auto mode gives clearer explanations when it blocks a command, including that Claude may retry and that you can review the call in /permissions.

공식 노트를 바탕으로 AI가 정리했습니다. 원문이 우선합니다.

Radar Ultra
OpenAI와 관련 없는 비공식 도구입니다. 공개 게시물을 정리한 것이며 원본 게시물이 기준입니다.