Released: 2026-09-18
When working on projects following common standards, Claude Code now automatically reads AGENTS.md as guidance whenever there is no CLAUDE.md in the root directory.
You can customize the instruction file to read under "Project instructions" in /config. This feature is not yet available on Bedrock, Vertex, or Foundry.
For enterprise Claude apps gateways whose only egress is a forward proxy, setting CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 now hands the proxy the hostname for every outbound request instead of resolving it locally. An optional headers: map has been added to upstream configurations to send static headers to a custom proxy running in front of a provider. When a proxy is configured, the telemetry relay no longer ignores collector hosts or domains listed in NO_PROXY. Loopback connection error messages now clearly suggest setting CLAUDE_GATEWAY_ALLOW_LOOPBACK.
When checking task progress, panels such as /tasks now display a line notifying you that an update is waiting if a background task finishes while the panel is open. Background sessions started with claude --bg no longer unexpectedly exit when a plugin's LSP server exits or closes its stdin. The deprecated TaskOutput tool has been officially removed; Claude now reads background task output files using the Read tool, and the taskOutputMaxChars setting along with the TASK_MAX_OUTPUT_LENGTH environment variable no longer have any effect.
When invoked via automated scripts or the SDK, sessions in headless mode no longer hang indefinitely after encountering an internal error; they now clearly report the error and exit with status code 1. Resuming earlier sessions no longer resets cost and token usage totals to zero, as headless sessions properly save their cumulative totals at exit. The first turn in SDK and headless modes no longer waits on per-directory lookups for CLAUDE.md, speeding up startup. Standalone claude -p runs outside an SDK or IDE have removed the background Haiku auto-title request.
When collaborating across multiple agents, subagent results are now returned to the main agent under a dedicated header with indented formatting, preventing output from being mistaken for the session's own instructions. Resumed subagents and teammates no longer re-render loaded MCP definitions, preserving the prompt caching mechanism. Agent messages arriving mid-turn now appear precisely where they arrived rather than below the command execution row. Fast typing or pasting in claude agents dispatch input no longer produces scrambled text. Pressing Enter on a selected agent panel row now functions properly. Script-computed prompts in cloud workflows are now clearly framed with their source.
When managing extensions, running claude plugin install to reinstall a plugin version currently in use by a session or another program leaves unchanged copies alone instead of corrupting the installed copy; if a newer version exists in the marketplace, the command now proactively notifies you and suggests claude plugin update. Plugins installed from the official marketplace now correctly record their Git commit hash in installed_plugins.json, and updating a pinned-commit plugin updates this record accordingly. A malformed entry in enterprise marketplace rules no longer silently disables the entire marketplace policy. Plugin reload previews no longer keep unpacked archive files until exit, nor do they overwrite the fallback download cache.
When browsing installed plugins on the Installed tab, error messages such as failed plugin updates now strip terminal control characters, ensuring clean, readable text. The /plugin and /skills interfaces no longer crash when a skill or legacy command is named after built-in Object properties like constructor or toString. When every install in a multi-select operation fails, the panel now shows an explicit failure notice instead of closing silently. Removing uninstalled plugins no longer leaves orphaned "failed to load" rows in the interface. MCP servers listed separately from their plugins now clearly indicate which plugin they belong to.
When searches cannot start because the system is out of processes, memory, or file handles, the Grep and Glob tools now return explicit errors instead of incorrectly reporting no matches. Writing to an existing directory path with the Write tool now reports a clear error instead of treating it as declined permission. When matching fails on very large edits containing non-ASCII text or backslashes, the Edit tool now properly reports that the string was not found instead of throwing a regular expression too large error. The Edit tool also fixed an issue where an escaped backslash followed by uXXXX text was mistakenly treated as a \uXXXX escape sequence, which could cause edits of non-ASCII characters to accidentally overwrite escaped backslashes. Escaped control characters in paths are now handled as literal text. Reading PDF pages on Windows when working folder paths are long (about 120 characters or more) has also been resolved.
Previously, when Claude hit a rate limit while publishing, it was told to stop retrying even though nothing had been published. If publishing an artifact hits rate limits, Claude is now clearly informed that nothing was published along with guidance on when to retry, rather than being told to give up completely. When conversations include claude.ai artifact links and the Artifact tool is available, Claude prioritizes reading content with the dedicated Artifact tool instead of using the generic WebFetch tool. Permission prompts for the Artifact tool are now more concise, listing pages and artifacts by title or file name while grouping relevant links after the explanatory text.
With the sandbox on (the guard that runs commands in an isolated area), some commands were slipping out of it entirely when they should not have. When sandboxing is enabled, $TMPDIR no longer expands to an empty string in Bash commands executed outside the sandbox. Under the sandbox.excludedCommands exemption setting, every part of a compound command must match the rule to be exempt from the sandbox, preventing a partial match from bypassing sandboxing entirely. When a command involves high-risk rm deletions, the permission prompt explicitly names the flagged command and suggests adding a ${VAR:?} guard so headless runs can recover. Sandbox prompt instructions on cloud platforms now adopt first-party wording that clearly defines task permission boundaries.
Picking an old conversation back up with --resume could previously crash it, or make every request after it fail. Resuming an earlier session whose history contains plain-string assistant messages or empty text blocks alongside other content no longer causes request failures or crashes. When continuing sessions after /clear, --continue, or --resume, startup hook output no longer causes the first message fragment to be missed, preventing full prompt cache misses. Conversation attachments are no longer re-rendered after a resume or relaunch, fully preserving extended thinking and prompt cache status. Opening worktree sessions with --worktree now reliably loads main project skills even when .claude/skills is untracked by version control. A crash when resuming transcripts containing incomplete hook summaries has also been resolved.
A single malformed value in ~/.claude.json used to be enough to make Claude Code hang or crash on launch. Interactive startup no longer hangs or throws errors for users when customApiKeyResponses or theme values are corrupted. Opening /mcp or the /plugin manage panel with a malformed claudeAiMcpEverConnected field no longer triggers type error crashes. Remote Control session bookkeeping now functions properly even when encountering malformed placeholder records.
Unexpected logouts, stalled or misleading updates: this group fixes problems you can hit while signing in and updating. Running older and newer versions locally no longer causes unexpected account logouts. If a proxy returns an invalid version number, update checks no longer error every 30 minutes; when a minimum or maximum version is set, claude update no longer hangs; a malformed minimumVersion is now ignored. Installations managed via winget or apk no longer falsely report being up to date when version lookups fail. Failed auto-updates now proactively clean up ~/.cache/claude/staging, leaving no large staged downloads behind. Errors following revoked logins now lead with /login, and Console sign-in displays the server's error message when API key creation is refused.
If you like to type your next message while Claude is still busy, that habit no longer costs you. Messages typed while Claude is actively working are now properly processed rather than being ignored by the model. When prompts are entered, invisible Unicode formatting and tag characters are automatically stripped, displaying the cleaned prompt for your review before sending. Loading prompts containing terminal color codes from history or an external editor no longer triggers unrecoverable interface crashes.
Crashing right after launch on slow machines and, in rare cases, a screen that stops updating: neither happens anymore. Launching sessions under heavy system load no longer crashes with unrecoverable interface errors when the initial spinner appears. A rare issue where the screen could stop updating for the rest of the session after an internal rendering error has been eliminated, ensuring continuous live updates. In the fullscreen /resume session picker and other panels covering the prompt area, drag-selecting text with the mouse now reliably displays the copied notice. Startup overflow lines under the logo now concisely show how many notices are hidden.
When running the code review command /ultrareview with nothing to review, the interface now explains the context and offers a command to review your latest commit, while the first commit in a new repository is reviewed in full; in non-interactive sessions, it explicitly refuses when the repository lacks a base branch or shared history. On the Anthropic API, Fable now always appears in the /model list, appearing greyed out only when disabled by organization settings. In Cowork cloud sessions, WebFetch and WebSearch now clearly inform Claude of the exact reason when requests are refused, such as a used-up fetch budget or admin policy restrictions. An issue on Windows where out-of-memory errors immediately following Claude's reply prevented tool calls from executing has also been resolved.
If you use Claude Code in VS Code, this release adds new actions such as signing out, stopping background tasks, and copying responses. A "Sign out" row has been added to the panel menu, along with support for typing /logout in the command menu. The agent map now includes background shells and running tasks, each equipped with a Stop button, and can be opened directly with /tasks. A Copy response button is now available beside AI responses, alongside a typed /copy command. A one-time notice appears when inactive sessions are archived automatically, and an "Unarchive all" action has been added to the Archived sessions group. In environments where plan limits do not apply, the Account & usage dialog and session manager now display costs and token usage.
Preferences you set in VS Code could previously vanish after a restart. Clicking the "General config" menu item now opens settings directly, and typing commands such as /mcp, /hooks, /memory, and /rewind directly opens their respective dialogs. Effort levels adjusted via the slider now persist across sessions on models that already had a level saved with /effort. The mode picker no longer misses the Auto option when saved model settings use differently-cased aliases like "Sonnet". Toggling fast mode with /fast is now correctly saved as the default, remaining active after restarts.
When a Team or Enterprise plan shares cloud environments, the hard part is telling which ones are yours and which ones everyone uses. On Team and Enterprise plans, the cloud environment picker is now divided into Personal and Organization sections, allowing admins to share personal environments with the entire organization. Organization environments open as read-only summaries by default on the Code tab to prevent unauthorized modifications, while editing is consolidated under Admin settings → Cloud environments. Cloud environments saved with Custom network access and no domains no longer silently revert to Trusted mode, and the dialog now requires at least one domain. The admin setting label has been renamed to "Cloud sessions", removing the redundant read-only Mobile row beneath it.
For teams using Claude Tag in Slack, this group of fixes centers on routines and access credentials. Routines created in a Slack channel on an Enterprise Grid org-wide install can now properly read other public channels in their workspace when running. Links on credential presets in Claude Tag access bundles now open each vendor's dedicated credential-setup page instead of generic API documentation. Admins can now point the Pylon credential preset at Pylon's EU host. Google Cloud credential forms in access bundles now explain why a key file is refused, the website and scopes stay locked, and pasted keys are preserved when a replacement is rejected. The network events log in admin settings fixed missing response statuses for connections using AWS signing, client certificates, or a custom CA.
Added AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead; change it under "Project instructions" in `/config` (not yet on Bedrock, Vertex or Foundry)Added `CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1` for Claude apps gateways whose only egress is a forward proxy: every outbound request hands the proxy the hostname instead of resolving it locally
Added an optional `headers:` map on Claude apps gateway upstreams, to send static headers to a proxy you run in front of a provider
Fixed the Claude apps gateway's telemetry relay ignoring a collector hostname or domain listed in `NO_PROXY` when a proxy is set
Improved the Claude apps gateway's loopback error messages to name `CLAUDE_GATEWAY_ALLOW_LOOPBACK`Added a line saying a background task's update is waiting when it finishes while a panel such as `/tasks` is open
Fixed background sessions (`claude --bg`) exiting when a plugin's LSP server exited or closed its stdin
Removed the deprecated TaskOutput tool; Claude reads a background task's output file with Read instead, and the `taskOutputMaxChars` setting and `TASK_MAX_OUTPUT_LENGTH` no longer have any effectFixed `claude -p` and Agent SDK sessions that could hang with no result after an internal error; they now report the error and exit with code 1
Fixed a headless resume (`claude -p --resume`, the SDK, a VS Code extension window reload) starting the session's cost and usage totals at zero; headless sessions now save their totals at exit
Improved session start-up for SDK and headless (`-p`) use: the first turn no longer waits on the per-directory CLAUDE.md lookup
Removed the background Haiku auto-title request from `claude -p` runs launched outside an SDK or IDEFixed messages from other agents (such as a subagent's SendMessage) that arrived mid-turn showing up below the "Ran N shell commands" row instead of where they arrived
Fixed typed or pasted text occasionally coming out scrambled in the `claude agents` dispatch input during key repeat or very fast input
Fixed Enter on a selected agent panel row doing nothing when `keybindings.json` rebinds Enter in the Chat context, for example to `chat:queueSubmit`
Fixed resumed subagents and teammates re-rendering the MCP tool definitions they had loaded, which broke prompt caching for that agent
Changed subagent results to reach the main agent under a header marking them as subagent output, with the result indented, so text in a subagent's result cannot pass as the session's own instructions
Changed workflow scripts' computed `agent()` prompts on Bedrock, Vertex and Foundry to reach the subagent framed as script-authored text, so the safety classifier does not read them as the userFixed `claude plugin install` sometimes failing and breaking the installed copy when reinstalling a plugin version that a session or another program was using; an unchanged copy is now left alone
Fixed one malformed `strictKnownMarketplaces` or `blockedMarketplaces` entry silently disabling the whole enterprise marketplace policy
Fixed plugins from the official marketplace being recorded without their commit in `installed_plugins.json`, and `installed_plugins.json` keeping the old commit after updating a pinned-commit plugin
Fixed plugin reload previews keeping every previewed copy of a plugin archive unpacked until exit, and overwriting the cached `--plugin-url` archive a reload falls back to when its download fails
Improved `claude plugin install` on an already-installed plugin: it now says when the marketplace offers a newer version and names the `claude plugin update` commandFixed `/plugin` not stripping terminal control characters from messages on the Installed tab, such as the error of a failed plugin update
Fixed `/plugin` → Installed and `/skills` crashing when a skill or legacy command is named like a built-in Object property such as `constructor` or `toString`
Fixed `/plugin` closing with no message when every install in a multi-select failed
Fixed uninstalled plugins reappearing as "failed to load" rows in `/plugin` Installed, and Remove not clearing such a row
Improved `/plugin` Installed: an MCP server listed apart from its plugin now shows which plugin it belongs toFixed Grep and Glob reporting no matches when the search could not start because the system was out of processes, memory or file handles; they now return an error saying so
Fixed the Write tool silently ending the turn as a declined permission when the target path is an existing directory; it now reports a clear error
Fixed the Edit tool treating an escaped backslash followed by `uXXXX` text as a `\uXXXX` escape, which could make an edit of a non-ASCII character rewrite an escaped backslash sequence instead
Fixed the Edit tool reporting "Invalid regular expression: regular expression too large" instead of "String not found in file" when a very large edit containing non-ASCII text did not match the file
Fixed a turn ending early with "Path contains null bytes" when a tool call's file path contained `\u0000` written as an escape sequence; escaped control characters now stay as literal text
Fixed PDF page reads on Windows failing when the working folder's path is long (about 120 characters or more)Fixed rate-limited artifact publishes telling Claude to stop retrying; Claude is now told nothing was published and when to send the same publish again
Improved artifact link handling so Claude reads claude.ai artifact links with the Artifact tool instead of WebFetch when that tool is available
Improved the Artifact tool's permission prompts: shorter sentences, pages and artifacts named by title or file name, and links listed after the textFixed `$TMPDIR` expanding empty in Bash commands that run outside the sandbox while sandboxing is enabled
Fixed a `sandbox.excludedCommands` glob exempting an entire compound Bash command from the sandbox when only one part matched; every part must now match
Improved the dangerous-rm permission prompt to name the flagged rm command and suggest a `${VAR:?}` guard, so headless runs can recover
Changed the Bash sandbox instructions on Bedrock, Vertex and Foundry to the first-party wording, which frames the sandbox as the boundary of what the task was givenFixed conversations failing every request with "text content blocks must be non-empty" when an earlier assistant turn held an empty text block beside other content, including after `--resume`
Fixed a crash when resuming a session whose saved history holds an assistant message stored as a plain string
Fixed sessions continued after `/clear` (restart, `--continue`, `--resume`) missing part of their first message when a SessionStart hook printed output, causing a full prompt-cache miss
Fixed a crash ("unrecoverable interface error") when resuming a session whose saved transcript contains a stop hook summary without a well-formed hook list
Fixed project skills from the main repository not loading in `--worktree` sessions when `.claude/skills` is untracked
Fixed attachments recorded earlier in a conversation being re-rendered after a resume or relaunch, which dropped extended thinking and missed the prompt cacheFixed interactive start-up hanging or showing an error for `ANTHROPIC_API_KEY` users when `~/.claude.json` holds a malformed `customApiKeyResponses` value
Fixed a crash ("Type error") when opening `/mcp` or `/plugin manage` with a malformed `claudeAiMcpEverConnected` value in `~/.claude.json`
Fixed a crash at launch when `~/.claude.json` holds a malformed `theme` value
Fixed Remote Control session bookkeeping failing when `~/.claude.json` holds a malformed placeholder recordFixed being unexpectedly logged out when an older Claude Code build (for example an IDE extension's bundled CLI) runs on the same machine as the current one
Fixed update checks erroring every 30 minutes, and `claude update` hanging when a minimum or maximum version is set, if a proxy returns an invalid version; a malformed `minimumVersion` is now ignored
Fixed `claude update` on winget- or apk-managed installs reporting "up to date" when the version lookup failed
Fixed failed auto-updates leaving large staged downloads behind in `~/.cache/claude/staging`
Fixed the error after a revoked claude.ai login blaming an expired Anthropic profile; it now leads with `/login`
Fixed Console sign-in showing only "Request failed with status code 400" when the server refuses to create an API key; it now shows the server's messageFixed a crash ("unrecoverable interface error") when the prompt held text containing terminal color codes, for example a prompt recalled from history or text loaded from the external editor
Fixed messages typed while Claude is still working sometimes being ignored by the model
Improved prompt handling: invisible Unicode formatting and tag characters in a prompt are removed and the cleaned prompt is shown for review before it is sentFixed sessions on slow or heavily loaded machines sometimes exiting with "Claude Code exited after an unrecoverable interface error" when the first spinner appeared
Fixed a rare case where the screen could stop updating for the rest of the session after an internal rendering error
Fixed the "copied" notice not appearing after drag-selecting text in the fullscreen `/resume` picker and other panels that cover the prompt area
Improved the startup notice overflow line under the logo: it now reads "N more notices hidden" instead of "+N more · /status"Fixed a rare case on Windows where a turn could stop with an error such as "Out of memory" right after Claude replied, so that reply's tool calls never ran
Fixed WebFetch and WebSearch in Cowork cloud sessions not telling Claude why a request was refused, such as a used-up fetch budget or an admin policy
Improved `/ultrareview` when there's nothing to review: messages say which case you're in, offer a command that reviews your latest commit, and a new repository's first commit is reviewed in full
Changed Fable to always appear in `/model` on the Anthropic API; it is greyed out only when your organization's settings disable it
Changed `/ultrareview` in non-interactive sessions to refuse when the repository has no base branch or shared history[VSCode] Added a Sign out row to the panel menu, with `/logout` in the typed command menu
[VSCode] Added background shells and other running tasks to the agent map, each with a Stop, and a typed `/tasks` that opens it
[VSCode] Added a Copy response button on responses and a typed `/copy`
[VSCode] Added a one-time notice when inactive sessions are archived automatically, and an "Unarchive all" action on the Archived sessions group
[VSCode] Added the session's cost and token usage to the Account & usage dialog and the session manager where plan limits do not apply (Vertex, Bedrock, Foundry, API key)[VSCode] Fixed the "General config" menu row showing `/config` usage text instead of opening settings, and made typed `/mcp`, `/hooks`, `/memory`, `/rewind` and similar commands open their dialogs
[VSCode] Fixed the effort slider's level not persisting into later sessions on a model that already had a level saved with `/effort`
[VSCode] Fixed Auto missing from the mode picker for conversations opened in an already-used panel when the saved model setting is a differently-cased alias such as "Sonnet"
[VSCode] Fixed `/fast` not saving fast mode as the default, so it was lost when the extension relaunched Claude Code[Claude Code on the web] Added Personal and Organization sections to the environment picker on Team and Enterprise plans, and admins can now share a personal environment with the organization
[Claude Code on the web] Changed organization environments to open as a read-only summary from the Code tab on Team and Enterprise plans, with editing under Admin settings → Cloud environments
[Claude Code on the web] Fixed a cloud environment saved with Custom network access and no domains silently reverting to Trusted; the dialog now asks for at least one domain
[Claude Code on the web] Changed the admin Claude Code setting labeled "Web" to "Cloud sessions" and removed the redundant read-only Mobile row beneath it[Claude Tag] Fixed routines created in a Slack channel on an Enterprise Grid org-wide install failing to read other public channels in their workspace when they ran
[Claude Tag] Fixed the "Learn more" links on credential presets in Claude Tag access bundles to open each vendor's credential-setup page instead of a generic API reference
[Claude Tag] Changed the Pylon credential preset in Claude Tag access bundles so admins can point it at Pylon's EU host
[Claude Tag] Fixed Google Cloud credential forms in Claude Tag access bundles: a refused key file now says why, the website and scopes stay locked, and a rejected rotation keeps the pasted key
[Claude Tag] Fixed the network events log in Claude Tag admin settings showing no response status for requests through connections that use AWS signing, client certificates or a custom CA