Released: 2026-09-17
When system memory is running critically low, the terminal now displays a visible warning with guided steps to free memory or restart safely. Background commands that were previously stopped after idling for thirty minutes under mild memory pressure now only stop when memory is critically low, with the debug log clearly recording the reason.
Non-interactive mode can now set the maximum wait time for connecting MCP (Model Context Protocol) servers during the first turn via CLAUDE_CODE_MCP_STARTUP_WAIT_MS. When using --strict-mcp-config with an empty --mcp-config, the first turn no longer gets held up waiting to time out on incidental servers. In streaming JSON sessions, deferred tool search tools no longer stall first-turn startup. Environments such as Bedrock, Vertex, and Foundry now adopt the v2 client and the latest protocol negotiation by default. Invalid SDK server entries in configuration files are skipped with a warning.
When calling external extension tools, if a request is refused by the server due to insufficient permissions, the error message now clearly names the missing permissions and points to /mcp re-authentication instead of misreporting an expired sign-in. Connection error logs and the login tool description no longer leak resolved secret strings from environment variables. Servers using legacy HTTP and SSE protocols now connect properly even when returning error codes. Streamable tool calls are no longer bound to a fixed timeout and won't time out early. Prompts and resources update immediately when a server sends list-changed notifications.
System telemetry now adds an effort attribute to LLM request traces. Through the new managed settings resolved event, administrators can clearly track configuration sources and policy status, and setting the environment variable OTEL_LOG_MANAGED_SETTINGS=1 automatically redacts sensitive setting values and digests in logs. When exporting raw request bodies to local files, the newly generated index.jsonl file and event attributes accurately link each model response to its corresponding request file and transcript message.
The Claude apps gateway now supports adding store.connect_timeout_seconds in its configuration to lengthen the database connection timeout. If the database connection experiences a slight delay during boot, the gateway actively retries the connection up to three times, and clearly points to the configuration location if unreachable. When the server receives a shutdown signal, it no longer forcefully cuts open streams, instead waiting up to twenty-five seconds for in-flight requests to complete transmission smoothly.
The Claude apps gateway now issues an active load warning when open connections exceed two hundred and fifty-six, and includes identity information in telemetry data. Under high load, the spend-limit check is simplified from four database round trips to one, effectively reducing timeout risks; if the database drops a connection during the check, the system properly handles the exception. When a user triggers sign-in rate limits at /login, the interface clearly explains the refusal reason, while gateway logs flag which limit was hit and suggest which setting to change.
When a corrupted transcript leads to tool call errors, sessions attempt to self-heal or provide a hint to use /rewind to end retry loops. Long-running tasks driven by hooks automatically compact the conversation when context overflows again after previous compaction. For instance, an active /goal will not lose its ongoing objective when resuming a compacted session. Ordered lists entered in messages are no longer forcefully renumbered. Stop prompt hooks that trigger repeatedly now display with a concise label.
When managing agents using the claude agents command, launch parameters such as --model, --effort, and permission modes are no longer lost after an auto-update relaunch. On Bedrock, Vertex, and Foundry platforms, subagents configured with the Opus model correctly recognize and retain their model settings. Messages sent by a subagent to the main session remain fully preserved in the desktop transcript after a relaunch, and a subagent's concise progress summary is no longer accidentally overwritten by a runaway multi-paragraph reply.
In headless mode and SDK sessions, multiple background tasks finishing concurrently are now handled in a single model call rather than sending individual requests one by one. Resuming an interrupted background agent avoids leaving behind incomplete batched tool calls, and resuming work with CLAUDE_CODE_RESUME_INTERRUPTED_TURN properly reports previously unfinished background tasks. Status notifications when background agents wait on tasks are more accurate. Script output and exit status from the Monitor tool are delivered merged into a single notification.
When plugin policy is centrally managed remotely, the system no longer frequently rewrites the manifest file on startup, preventing the desktop version from repeatedly reloading plugins across all open sessions. After plugins reload, terminal commands only re-source the profile if executable directories changed, effectively eliminating command execution stalls. Standalone plugin directories no longer mistakenly inherit version numbers from enclosing Git repositories; plugins loaded from archive files avoid stale extracted files after multiple reloads; and repository clones retain Git LFS pointers by default to save bandwidth.
The Artifact tool now ensures the session reads the latest content before updating, eliminating the dilemma where an approved update is subsequently refused. In team collaboration and cloud environments, reading a teammate's artifact is no longer misjudged as disabled network access. If not signed in to a cloud account, the terminal clearly notifies on the first attempt and stops futile retries sooner. Publishing based on an older version is proactively intercepted with the newer page provided to merge, and new versions published remotely in local environments no longer interrupt current conversation turns.
When executing terminal commands involving special variable assignments or loops, the system now properly triggers permission prompts. In worktree-isolated sessions, potentially dangerous operations with nested shell expansions are directly refused. In files containing multi-byte characters, the preview position in Edit permission prompts no longer drifts from the actual modification location. Furthermore, the system performs stricter safety checks before removing an agent worktree containing submodule checkouts to prevent data loss.
In fullscreen mode, collapsed teammate and agent messages now support click-to-expand for full reading. Clickable links pointing to local file paths in the terminal have been corrected to adhere to the file:// protocol standard, opening files smoothly when clicked inside terminals like VS Code. In AskUserQuestion preview mode, entered notes attach accurately to the currently highlighted option, and pressing Enter to submit no longer drops the selected item. Launching inside editor terminals on Linux under Wayland no longer unexpectedly opens extra blank editor windows.
Error hints in desktop sessions now suggest slash commands such as /usage-credits instead of invalid command-line flags. When organization policy disables fast mode, entering /fast on in non-interactive environments explicitly reports that it has been disabled. The /schedule command preserves prompt message roles when saving routines. The /status command correctly displays failure details for the API key helper and unifies terminology for related features under cloud sessions. The /code-review command switches to leaner inline review prompts, avoiding spawning excessive review subagents for standard models.
When a language server publishes project-wide diagnostics across thousands of files, the system no longer causes noticeable per-turn slowdowns. Self-hosted runners now keep retrying upon transient failures during token refreshes, and actively fetch new tokens following authentication errors, preventing sessions from frequently breaking. The runner skips read-only repositories refused by the git host during access checks rather than failing session startup. The first turn of a cloud session no longer misses tools from connecting MCP servers.
After reloading the window in VS Code, the system automatically continues interrupted steps with a label in chat, and messages sent while running a tool are no longer lost upon reload. The Customize menu adds Memory and Instructions entries to view auto-memory toggles and edit instruction files. A new claudeCode.lockEditorGroups setting prevents the extension from locking editor groups, resolving unintended layout splitting when reopening tabs or starting new sessions.
When opening a new tab via shortcut in VS Code, the input box automatically gains focus so you can type immediately without clicking. Overlapping settings writes from the extension now include safeguards to prevent ~/.claude/settings.json corruption or dropped settings. Asking side questions with /btw right after starting a conversation no longer mistakenly displays another session's history. The brief freeze when first querying global gitignore files is resolved, with standard XDG config paths supported. Sign-in and sign-out states in terminals reflect immediately after changing configuration directories in environment variables.
Edit diff boxes in chat now fit the exact number of rows displayed, no longer getting cut off at the bottom under wrapped lines or specific panel widths. Under High Contrast themes, inline code and code blocks are once again clearly readable. Screen readers now clearly announce message senders (such as "You" or "Claude") along with the names of executed tools. The plan review card supports scrolling through comments, preventing the feedback button from being obstructed. Plugin management and MCP dialogs now support full keyboard navigation. Session names no longer shift horizontally when typing queries in the session picker.
In the diff view for cloud sessions on the web, a new branch picker lets you freely compare changes against any branch instead of being restricted to the base branch. When GitHub token renewals briefly fail, git operations no longer throw service unavailable errors; and after credentials refresh, commits no longer experience temporary signing errors. Sessions marked as read no longer revert back to an unread status.
Editing a routine no longer triggers unintended double firing or automatically re-enables a paused routine. When saving a routine fails to link with GitHub due to repository trigger limits, the toast message explicitly details the failure cause. When the owner's GitHub connection is lost, the routine skips the current run and continues retrying for up to seventy-two hours instead of disabling the routine on the first failed check. When subscription pauses place a routine on hold, the notice clearly guides users to manually re-enable it after restoring the subscription.
A Guests setting has been added to Claude Tag admin settings, allowing administrators to specify access rules upfront when adding channels. When other Slack apps or bots mention @Claude, Claude wakes up and replies properly even after days of channel inactivity; and mentions during new channel creation are delivered smoothly once Claude joins. Significant follow-up updates sent minutes later are posted as new notification replies instead of silently editing previous messages. Monitoring for cross-workspace shared channels and environment name displays have also been resolved.
When searching for messages within a single Slack channel, search failures are resolved and matching results are returned accurately. When stopped by a safety filter, the system explicitly explains the reason and no longer cancels running background work. Email addresses in replies no longer show a redundant mailto: prefix. In busy threads, live progress checklists are capped at two thousand characters and updated at most every fifteen minutes, while automatically updating older checklist links. Editing a Slack canvas in specific channels no longer repeatedly appends redundant guest notes.
Automated Code Review now posts each finding using short, plain sentences, allowing developers to understand issues at a glance. When GitHub or internal services experience transient startup errors, the system automatically waits and retries to prevent unexpected interruptions. Upon re-review, resolved threads close normally without staying open due to appended lower-severity notes. If a review is skipped due to organization limits, check-run cards and comments directly link to the admin page for prompt resolution.
Added a visible warning when memory usage is critical, with steps to free memory or restart safely
Fixed background commands being stopped after 30 idle minutes on machines under mild memory pressure; they're now stopped only when memory is critically low, and the debug log says whyAdded `CLAUDE_CODE_MCP_STARTUP_WAIT_MS` to bound how long the first non-interactive turn waits for connecting MCP servers (`0` = don't wait)
Fixed `--strict-mcp-config` with an empty `--mcp-config` holding the first non-interactive turn for up to `MCP_TIMEOUT` on incidental MCP servers
Improved startup in `--input-format stream-json` sessions: the first turn no longer waits up to 2s for still-connecting MCP servers whose tools tool search defers; they arrive on a later turn
Changed Bedrock, Vertex, Foundry and telemetry-disabled installs to use the v2 MCP client and MCP 2026-07-28 negotiation with direct HTTP servers by default, as other installs already do (opt out: `MCP_SDK_GENERATION=v1` or `MCP_PROTOCOL_NEGOTIATION=legacy`)
Changed `"type": "sdk"` MCP entries in `.mcp.json`, settings, plugins and agent files to be skipped with a warning: only an SDK host application can register in-process serversFixed MCP servers configured as `http` that only speak legacy HTTP+SSE failing to connect when they answer the first request with 422 or another 4xx error
Fixed Streamable HTTP MCP tool calls timing out after about 5 minutes even when a longer per-server `timeout` was set
Fixed MCP prompts and resources not refreshing when a server sends list-changed notifications without declaring `listChanged`
Fixed MCP tool calls refused with 403 insufficient_scope being reported as an expired sign-in: the error now names the missing permissions and points to `/mcp` re-authentication
Fixed MCP connection errors and the MCP login tool's description showing secrets resolved from `${VAR}` placeholders in MCP configsAdded `effort` attribute to the `claude_code.llm_request` OpenTelemetry trace span, matching the `api_request` event
Added `claude_code.managed_settings_resolved` OTel event: managed-settings sources and policy helper state; redacted settings and digests with `OTEL_LOG_MANAGED_SETTINGS=1`
Improved `OTEL_LOG_RAW_API_BODIES=file:<dir>` output: a new `index.jsonl` and `request_body_id` / `message.id` event attributes link each response to its request file and transcript messageAdded `store.connect_timeout_seconds` to the Claude apps gateway config to lengthen the Postgres connect timeout (default 5 seconds), and improved the boot error when the database is unreachable to point to `store.postgres_url` and the configured timeout
Fixed Claude apps gateway cutting every open stream on SIGTERM: it now lets in-flight requests finish for up to 25 seconds before exiting (`CLAUDE_GATEWAY_DRAIN_TIMEOUT_MS`)
Improved Claude apps gateway boot: it now tries the first Postgres connection up to three times before exiting, so a database that is reachable a few seconds late no longer fails the bootAdded `enduser.sub`, the IdP subject, to the telemetry Claude Desktop and Cowork send through a Claude apps gateway
Added a Claude apps gateway warning when a replica has more requests open than the 256 it sends upstream at once, and a startup log line showing that limit
Fixed an unhandled promise rejection in the Claude apps gateway when Postgres drops a connection during a spend check
Improved the Claude apps gateway's spend-limit check under load: it now takes one database round trip instead of four, so fewer checks time out on a busy gateway
Improved Claude apps gateway sign-in rate limit errors: `/login` now explains the refusal, and the gateway log says which limit was hit and which setting to changeFixed sessions getting stuck endlessly retrying "unexpected tool_use_id" 400 errors: corrupted transcripts now self-heal where possible, and otherwise a clear error (with a `/rewind` hint) ends the loop
Fixed hook-driven sessions (such as an active `/goal`) ending with "Prompt is too long" instead of compacting when the context overflowed again after a reactive compaction
Fixed an active `/goal` being lost when resuming (`--continue` / `--resume`) a session that had compacted
Fixed the transcript renumbering ordered lists in your own messages (typing "3. 2. 1." displayed "3. 4. 5."); numbers and "N)" markers now show as typed
Fixed Stop prompt hooks re-sending their whole prompt on every block in a conversation; repeat blocks now name the condition with a 500-character labelFixed `claude agents` losing `--model`, `--effort`, `--permission-mode`, `--allow-dangerously-skip-permissions` and `--agent` after an auto-update relaunch
Fixed subagents with `model: "opus"` on Bedrock, Vertex or Foundry leaving the session's model when its id has no recognizable model family (unless `ANTHROPIC_DEFAULT_OPUS_MODEL` is set)
Fixed a message a subagent sends to the main session disappearing from the Claude Desktop transcript after a relaunch
Fixed a sub-agent's progress summary being replaced by a runaway multi-paragraph replyFixed a resumed background agent keeping half of an interrupted tool batch when one of its calls was approved with a message
Fixed a local `claude -p --resume` started with `CLAUDE_CODE_RESUME_INTERRUPTED_TURN` not reporting background tasks the previous process left unfinished
Fixed background agent notifications claiming the agent had no live background work when it was still waiting on its own background task and would resume
Fixed headless and SDK sessions making a separate model call for every background task that finished; completions already queued are now answered by one call
Improved Monitor tool notifications: a script's final output and its exit now arrive as one notification instead of two, saving a model turnFixed a plugin or marketplace directory with no git repository of its own taking its version from an enclosing git repository, such as a git-managed `~/.claude`
Fixed `installed_plugins.json` being rewritten on nearly every start-up when plugin policy comes from remote managed settings, which made Claude Desktop reload every open session's plugins
Fixed the Bash tool re-sourcing the shell profile (a multi-second stall on the next command) after every plugin reload; it now does so only when the plugins' `bin/` directories changed
Fixed plugins with a top-level `$schema` in `hooks/hooks.json` showing an "unknown key" notice
Fixed a plugin loaded from a `.zip` being served from a stale extraction after several overlapping reloads
Changed plugin and marketplace clones to leave Git LFS files as pointers instead of downloading them; `git lfs pull` in the checkout fetches themFixed the Artifact tool asking you to approve an update to an artifact that it then refused because the session had not read the latest version
Fixed Cowork and claude.ai cloud sessions with network access on treating reads of a teammate's artifact as if network access were off
Improved Artifact tool errors: when you are not signed in to claude.ai the terminal now says so on the first attempt, and Claude is told to stop retrying a rejected call sooner
Improved artifact publishing: a publish built on an older version is stopped before it is sent, with the newer page to merge
Changed artifact watching in local sessions: a new version published elsewhere no longer starts a turn; Claude learns of it from a later Artifact tool resultFixed Bash permission checks for commands that loop over or assign certain special shell variables; these commands now ask for permission
Fixed worktree-isolated sessions accepting Bash commands with certain nested shell expansions; these are now refused
Fixed the Edit permission prompt preview sometimes showing a different location than the approved edit in files with multi-byte characters
Improved safety checks before removing an agent worktree that contains submodule checkoutsAdded click-to-expand for collapsed teammate and agent messages in fullscreen mode
Fixed clickable links to local file paths doing nothing in VS Code and other terminals that require a `file://` URI
Fixed AskUserQuestion preview notes being attached to a previously chosen option instead of the highlighted one
Fixed AskUserQuestion preview mode dropping the highlighted option when submitting a note with Enter
Fixed extra empty editor windows opening at startup on Linux under Wayland when running inside the Cursor or VS Code terminalFixed error hints in Claude Desktop sessions to suggest slash commands like `/usage-credits` instead of CLI flags that cannot be used there
Fixed `/schedule` saving a routine's prompt without its message role when Claude writes the routine in the shape that listing routines returns
Fixed `/status` not showing the `apiKeyHelper` failure that its own error banner told you to check
Fixed `/fast on` in non-interactive sessions reporting on and then turning off under an organization's managed fast mode policy; it now says the organization has disabled it
Changed `/code-review` to use leaner inline review prompts for every model that has no tuned settings of its own, instead of spawning many review subagents
Changed the `/status` GitHub line to read "Cloud sessions", and `/web-setup`, `/ultrareview`, and teleport messages to say "cloud session" instead of "Claude Code on the web"Fixed a per-turn slowdown when a language server publishes project-wide diagnostics for thousands of files
Fixed self-hosted runner sessions failing every turn with a 401 after a few failed token refreshes, until the next scheduled refresh; the runner now keeps retrying, and fetches a new token after a 401
Fixed the first turn of a cloud session sometimes starting without the tools of an SDK-hosted MCP server that was still connecting
Changed self-hosted runners to skip a read-only repository the git host refuses at the access check instead of failing the session start[VSCode] Added continuation of the step a window reload interrupted, labeled in the chat, with a Claude Code: Continue After Reload setting to turn it off
[VSCode] Added Memory and Instructions entries to the Customize menu: Memory shows the auto-memory toggles, the saved memories and the memory folders, and Instructions edits the CLAUDE.md files
[VSCode] Added a `claudeCode.lockEditorGroups` setting to stop Claude from locking the editor groups it opens in
[VSCode] Fixed a message sent while Claude was running a tool disappearing from the conversation after a window reload
[VSCode] Fixed reopening a closed Claude tab splitting the editor layout when its locked group still holds another Claude tab and a file
[VSCode] Fixed New session opening another locked editor group whenever a file tab shared the group with your Claude tab[VSCode] Fixed a `/btw` side question asked in a new conversation's first seconds occasionally showing another session's side-question history
[VSCode] Fixed a brief freeze when the extension first looks up your global gitignore file
[VSCode] Fixed sign-ins and sign-outs made in a terminal not showing until a reload after `CLAUDE_CONFIG_DIR` changed in the Environment Variables setting
[VSCode] Fixed overlapping settings writes from the extension leaving `~/.claude/settings.json` unparseable or dropping a setting
[VSCode] Fixed Open in New Tab (Ctrl/Cmd+Shift+Esc) sometimes leaving the new tab's message box unfocused, so typing went nowhere until you clicked it
[VSCode] Changed the default global gitignore file to `$XDG_CONFIG_HOME/git/ignore` when `XDG_CONFIG_HOME` is an absolute path[VSCode] Fixed the Manage Plugins enable toggle and MCP servers dialog rows being unreachable from the keyboard
[VSCode] Fixed Edit diffs in the chat being cut off at the bottom at some panel widths and for long wrapped lines; diff boxes now fit the rows shown
[VSCode] Fixed session names shifting sideways in the session picker while typing a search query
[VSCode] Fixed the plan review card cutting off its Send feedback button and reason field when a plan has several comments; the comment list now scrolls
[VSCode] Fixed inline code and code blocks in chat replies being unreadable under the High Contrast themes
[VSCode] Improved screen reader navigation of the conversation: each message is announced as "You" or "Claude", with the tool name for tool steps[Claude Code on the web] Added a "Compare against" branch picker to a cloud session's diff view, so you can diff its changes against any branch instead of only the base branch
[Claude Code on the web] Fixed git operations in cloud sessions failing with "service unavailable" when GitHub's token renewal briefly errors
[Claude Code on the web] Fixed commits in cloud sessions occasionally failing with a signing error for a few minutes after the session's credentials refreshed
[Claude Code on the web] Fixed sessions sometimes flipping back to unread right after you mark them read[Claude Code on the web] Fixed editing a routine occasionally making it fire twice or re-enabling a routine that had just been paused
[Claude Code on the web] Fixed the toast after saving a routine whose GitHub trigger couldn't be linked to show the reason, such as a per-repository trigger limit, instead of only "edit to retry"
[Claude Code on the web] Changed routines to skip a run and retry for up to 72 hours when the owner's GitHub connection is missing, instead of switching the routine off at the first failed check
[Claude Code on the web] Changed a routine's on-hold notice: when your subscription is paused it now tells you to turn the routine back on yourself instead of promising an automatic resume[Claude Tag] Added a Guests setting to the Add channel and Add workspace forms in Claude Tag admin settings, so owners can pick Inherit, Allow, Channel only or Restrict up front
[Claude Tag] Fixed Claude not answering when another Slack app or bot @mentions it; the tag now gets a reply and wakes Claude in a channel it had stopped following after days of inactivity
[Claude Tag] Fixed Claude missing another app's message that tagged @Claude right after a new Slack channel was created; it's now delivered once Claude has joined
[Claude Tag] Fixed Claude folding a follow-up sent minutes after its last Slack message into it as a silent edit; late updates such as blockers now post as a new reply that notifies
[Claude Tag] Fixed Claude refusing to watch an Enterprise Grid channel shared with the whole organization when asked from another workspace in the grid
[Claude Tag] Fixed the Environment picker in Claude Tag admin settings showing a raw environment ID instead of the environment's name for archived or app-created environments[Claude Tag] Fixed Claude's Slack search failing with an error whenever it searched within a single channel; it now returns that channel's matching messages
[Claude Tag] Fixed a safety-filter stop silently resetting a Slack thread's context when nobody was waiting; Claude now always says so and no longer cancels background work still running
[Claude Tag] Fixed email addresses in Claude's Slack replies rendering with a visible `mailto:` prefix; they now show as the plain, clickable address
[Claude Tag] Improved Claude's live progress checklist in Slack: capped at 2,000 characters, reposted at most every 15 minutes in busy threads, with older "Latest task list" links updated
[Claude Tag] Removed the repeated guest-attribution note Claude appended to a Slack canvas each time it edited one in a channel using the "Channel only" guest setting[Code Review] Fixed re-reviews occasionally leaving a fixed finding's thread open when the new review also filed a lower-severity note under it
[Code Review] Fixed rare reviews ending with "Code review encountered an error" when GitHub or an internal service failed transiently at launch; they now wait and retry
[Code Review] Improved how Code Review words each posted finding: short plain sentences that say who is affected, where the code goes wrong, and the fix up front
[Code Review] Improved the check-run card and PR comment when a review is skipped because of an organization limit: each cause now links the admin page that fixes it