What changed
Claude Code 2.1.299 shipped on October 14, 2026. It adds idle compaction in the terminal, an environment variable that changes how @ file mentions work, and /effort and /fast in agent view. Some changes can alter how existing settings behave, notably MCP allow rules and when UserPromptSubmit hooks run. The fixes span third-party gateway API keys, the sandbox, MCP, cloud sessions and Remote Control, and wrapping of CJK and emoji text, and two regressions from 2.1.296 are fixed. The release is very large, so the items below are a selection of the changes with the most user impact.
New features
-
Idle compaction in the terminal
A 200K+ token conversation left idle for about 54 minutes is compacted before its one-hour prompt cache expires, so your next message doesn’t reprocess it all.
CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENSchanges the 200K-token minimum, and"idleCompaction": falsein settings.json turns it off. Both need v2.1.290 or later. -
CLAUDE_CODE_DISABLE_FILE_MENTIONSWith this environment variable, Claude gets the paths of
@file mentions to read with its file tools, not the files’ contents. -
/effortand/fastin agent viewSet what new sessions start with. A bare
/model,/effortor/fastthere now opens a picker. -
New
/usagewarnings/usagenow warns about max effort, workflows, always-on context and large tool results. -
A one-time note in plan mode
Shown when plan mode won’t ask before most edits or commands because bypass permissions is available.
-
--disallow-toolsforclaude plugin eval--disallow-toolsand adisallowed_toolscase field keep specific built-in tools, such as Agent, out of eval runs. -
--shutdown-drain-secforclaude self-hosted-runner orchestratorOn SIGTERM the orchestrator stops polling and gives the poll in flight and running spawn hooks up to that many seconds to finish.
-
Turning off the attribution block behind a Claude apps gateway
Set
CLAUDE_CODE_ATTRIBUTION_HEADER=0in theenvof a Claude apps gateway’scodepolicy to turn off the attribution block in Claude Desktop’s Code tab. The attribution block in Claude Code sessions behind a gateway gets a new field, and the same setting removes it. -
Mod API additions
Mods installed by managed settings can read the bytes of an image or document attached to the prompt through
$.prompt.mediain aprompt.submithook, and check how the session is signed in with$.session.identity(). A mod’sTextgetssize,weight,monoandtabular, and itsBoxgets named spacing steps (xstoxl);boldis deprecated in favor ofweight. A mod’sskill.prompthook now receives the skill’s frontmatter, source, plugin name and file path, and thesession.appendhook gets aqueuedoor.
Key improvements
-
Memory and disk use for @-mentioned PDFs
Their bytes are no longer kept in the session transcript.
-
Startup warning for an unrecognized model ID
Now one short line under the welcome banner with a docs link, instead of a long raw message.
-
“Unknown skill” errors
The Skill tool now suggests a skill’s listed name when Claude wrongly adds the
anthropic-skills:prefix to it. -
/reload-skillsNow says it re-sends the full skill list to Claude, which is how to get skills back after compaction.
-
Worktree creation errors
When a worktree can’t be created because its base commit doesn’t resolve, the error says how the lookup ended and the likely reasons, such as a repository without a commit. Creating a worktree at a bare-repository layout’s root now explains the layout instead of showing an unrelated git error.
-
Artifact watching
A local session can now watch up to 25 published artifacts at once, up from 10.
-
Background agent notifications
They now say when an agent is only waiting on its own background work, got a message too late for its result, or was stopped by a usage limit.
-
Errors for an expired claude.ai login
When a cloud session is refused because your claude.ai login was rejected or has expired, the error now tells you to sign in again with
claude auth login. -
[VSCode] Startup
In workspaces without a Python environment, Claude starts while the Python extension loads. In a new window on a folder you’ve opened before, the model picker and slash command menu fill in right away.
Changed behavior
-
MCP allow rules reach only the server they name
mcp__ano longer allows a server nameda__bora_. Writemcp__<server>__*with the server’s whole normalized name. Agenttools:lists likewise take only the named MCP server’s tools; list a server whose name contains__asmcp__<whole name>__*. -
UserPromptSubmithooks run before expansionThey now run before a skill or custom slash command expands, so a hook that blocks the prompt also stops its embedded shell commands. A skill’s own frontmatter
UserPromptSubmithook first runs on the prompt after the one that invokes the skill. -
The resume command printed at exit
It names the session by its ID, no longer by its name, so that nothing in a session name can run when the line is pasted. For a worktree session it uses
-wand-rif the long flags don’t fit the terminal. -
HTTP redirects on MCP requests
Requests to claude.ai’s connector proxy and to Anthropic-hosted MCP servers no longer follow HTTP redirects, nor do MCP requests through the session’s MCP proxy in cloud and self-hosted runner sessions. Redirects by a connector’s own server are unaffected.
-
Skill and command names with hidden characters
A skill or command whose file or folder name has hidden, direction-changing or blank-like characters,
<or>is no longer offered until renamed. -
Plugin hooks that can’t be reached
Actions a plugin hook can block are now refused, not let through, when the hook can’t be handed them or can’t hand back its answer; a tool that already ran has its result withheld. Plugin hooks and workflow scripts refuse values they hand back that nest deeper than 28,000 levels (workflow scripts allowed 100,000).
-
claude plugin testA call to a mod’s served
$method that hasn’t answered within ten seconds now fails, as in a session. Inui.rendertrees, each mod’s Box or Text root carriespart: { plugin }, so exact-tree assertions must include it. -
MCP protocol version and tool description length
Cloud session and Remote Control MCP connectors negotiate MCP protocol version 2026-07-28 by default on installs that fetch no flags;
MCP_PROTOCOL_NEGOTIATION=legacyopts out. Tool descriptions and server instructions from Anthropic’s first-party MCP servers atapi.anthropic.comreach Claude in full instead of being cut at 4,096 characters. -
Permissions in Cowork cloud sessions
In auto mode, a
PermissionRequesthook can no longer approve shell commands or subagents. Outside auto mode, an approved-site Claude in Chrome action that the safety classifier flags now asks you instead of being refused, unless managed settings apply. -
/btwin non-interactive sessionsSide questions return a timeout error when still unanswered after 11 minutes.
-
/ultrareviewupload errorsThey no longer suggest
git add ., which could stage or commit untracked files such as.env. -
Other changes
WebFetch page reads pick their effort as web searches do. The
turn.startmod hook also hears subagents’ and teammates’ turns; tell them apart bye.agentId. The cyber safeguard notice no longer asks people to apply to the Cyber Verification Program.
Bug fixes
Gateways & API keys
- Feature flag, startup, fast mode, profile and feedback requests to Anthropic carrying the API key configured for a third-party gateway (
ANTHROPIC_BASE_URL); Anthropic API keys are still sent. - The server-managed settings request to Anthropic carrying a third-party gateway’s API key when a managed settings file sets
ANTHROPIC_BASE_URL. - Streams through a custom
ANTHROPIC_BASE_URLbeing aborted after 180 seconds with no data arriving; they now wait 300 seconds, whether or not feature flags are loaded. - Sessions on a gateway-specific model name behind a custom
ANTHROPIC_BASE_URLending withPrompt is too long · automatic compaction failednear 200K tokens instead of sending the request. - The Claude apps gateway serving models that a policy’s
availableModelsMatch: exactordeniedModelsblocks; it now answers 400 and leaves them out of its model lists. forceLoginOrgUUIDandforceLoginMethodfrom MDM or the managed-settings file being hidden or replaced by server-managed settings.
Sandbox & permissions
- A sandboxed shell command being able to write into a
--plugin-dirfolder, whose code Claude Code reloads and runs outside the sandbox. - Sandboxed commands on macOS being able to read or write a denied path by its real location when the denied path passes through a symlink.
- Sandbox
denyReadandallowReadglobs with braces, parentheses or backslashes in a folder’s name being read differently by Claude Code’s settings checks than by the sandbox; globallowReadandallowWriteentries are also re-checked when their starting directory is replaced by a symlink. - Plan mode running a write (an MCP tool’s, a file’s, a shell command’s) that a
PreToolUsehook or a mod’stool.checkhook allowed. - rm prompts on shell
-cscripts that don’t runrm. An rm built at run time in a script Claude Code can’t read no longer prompts unlessCLAUDE_CODE_STRICT_INLINE_SHELL_RM_PROMPT=1. - An agent teammate started without a description keeping part of a credential from its prompt in the task description shown for it.
- The Write and Edit tools creating a mod in a session’s mods folder without asking in accept-edits mode when the config directory’s path has no
.claudein it. - Worktree-isolated sessions accepting some awk programs the isolation guard should refuse.
MCP
- MCP connections over HTTP hanging for 30 seconds and failing when a server never ends the body of its reply to a notification (#99014).
- MCP OAuth sign-in, token refresh and sign-out not sending configured
headers/headersHelpervalues, which broke servers behind an authenticating proxy. claude mcp list,claude doctorand/statusreporting an empty.mcp.jsonas invalid.- A settings file being ignored when
enabledMcpjsonServersordisabledMcpjsonServersheld a single server name as a plain string. claude mcp list,claude mcp getand subagent MCP servers starting a server that a policy check blocks; servers added to a running headless or SDK session are also no longer contacted before policy checks run.- MCP servers named
constructororprototypenever being connected, and tools of a server namedconstructorortoStringasking for permission in auto mode and bypass permissions mode. - Deny and ask rules for an MCP server (
mcp__<server>) not applying when the server’s normalized name contains a double underscore. - MCP tool calls in headless and SDK sessions sometimes failing or running twice when a background reconnect closed the connection they were running on.
Sessions, cloud & Remote Control
- Cloud and headless sessions running an approved tool call twice if a new message arrived as the session resumed after a restart (regression in 2.1.296).
- Edit and Retry in Cowork and claude.ai cloud sessions failing on a message stopped while its
UserPromptSubmitor modprompt.submithooks were still running (regression in 2.1.296). - The claude.ai view of a Remote Control or cloud session freezing when a large batch of events could not finish uploading over a slow connection.
- A Remote Control prompt being lost when another queued prompt in the same batch failed to deliver; files of 30 to 35 MiB, and files in long-running sessions, now also arrive over Remote Control.
- A cloud session sometimes ignoring the answer to a permission prompt or a question when it was given while the session’s worker was being replaced; replies cut short when the worker was replaced right after a compaction or the output token limit are also fixed.
- High memory use after resuming a cloud or Remote Control session that had run many subagents.
- Replies cut off at the output token limit ending the turn silently, with no retry or error, after a request fell back to non-streaming mode.
SendMessagenot finding a subagent by name after the CLI restarted and resumed the session (names given before the last compaction still need the agent ID).- Background sessions started with
--safe-modesilently leaving safe mode when the session was woken or restarted.
Scheduled tasks
- Scheduled wakeups (ScheduleWakeup) not firing while a background task was running in Agent SDK, IDE and
claude remote-controlsessions. - Scheduled tasks being announced for the wrong minute; the scheduling tools now report each job’s actual next run time.
- Scheduled tasks and loops firing again after every turn, or firing early, during the repeated hour when daylight saving time ends.
Memory & debug logs
- Memory growth under heavy debug logging (
--debug,--debug-file), and a hang on exit when the debug log is on a disk or network mount that stops accepting writes. - Memory growing without bound when debug output goes to stderr and the parent process stops reading it.
Files, images & display
- Read garbling UTF-16LE files with a byte-order mark, and Write dropping byte-order marks or silently converting Windows-1252, Shift-JIS, GBK or UTF-16BE files to UTF-8 (now refused).
- Wrapped text mixing CJK or emoji with Latin words losing its last line or leaving a blank row at some terminal widths, along with blank rows in very narrow terminals and text cut beside wide characters.
- Bold, italic and other nested text styles running over table borders or being lost on wrapped rows.
- Long conversations losing newly attached images on every turn after one oversized image was refused by the API; images attached while Claude is working (VS Code, desktop app, SDK) are also no longer sent at full size.
- Long pastes in terminals without bracketed paste losing all text after an invisible control character, such as a mis-decoded curly quote.
claude -psaying nothing when a settings file fails validation; text-output runs now list the file and the invalid entry on stderr./usage, the approaching-limit warning and the status line sometimes showing more than 100% used (#92238).
Plugins & mods
- Installed mods that kept running in an open session after being turned off remotely; they now unload between turns.
- A mod that submits a prompt at the end of every turn keeping a headless run going past
--max-turns,--max-budget-usd, an interrupt, or the end of the host’s conversation. - Mods’ hooks being handed values with
__proto__keys silently removed, which could skip atool.checkguard on a tool call another mod made. - A user-installed mod’s
$.env.setchanging the environment or shell prefix that command hooks from managed settings start with. /reload-pluginsnot loading a plugin installed from another terminal a moment earlier.claude plugin evalpassing runs that the API refused; a run that ends on an API error now scores 0 and makes the suite exit 1.
Windows
- The process-liveness check leaving orphaned PowerShell processes behind on busy machines until memory ran out.
- PreToolUse and PermissionRequest hooks that set no shell failing open when Git for Windows isn’t installed; a guard that fails now blocks the call.
- Bash commands failing with “command not found” for a whole session when Git Bash did not report its PATH at startup.
- Sandboxed commands failing on WSL2 with
bwrap: Can't mkdir /mnt/c/Program Files/ClaudeCodewhen the sandbox can write to the Windows drive.
VSCode & cloud sessions
- [VSCode] An unsent prompt and its attachments being lost when switching to another session and back.
- [VSCode] Unused blank chats keeping a background Claude process running after starting a new conversation.
- [Cloud sessions] A memory spike that could crash a session when many of its existing output files were re-sent to it at once.
- [Cloud sessions] Saving an edited routine silently moving it to a different environment when its own had been deleted or wasn’t in the list.
- [Remote Control] Sessions whose computer can’t be reached still appearing to send your messages; your text now stays in the box and Send is off until it reconnects.
Claude Tag & Code Review
- [Claude Tag] Claude in Slack staying silent on untagged posts that a channel’s own instructions tell it to answer, in channels with guests.
- [Claude Tag] Claude in Slack reporting a reply to another channel’s thread as posted when it isn’t a member there; it now says the reply wasn’t delivered and how to fix that.
- [Claude Tag] Claude in Slack sometimes ending a turn without posting its answer when a self-hosted runner’s image has no python3.
- [Code Review] The monthly spend-limit comment and check-run card sending admins to Claude Code settings; they now link the Usage page, where that limit is set.
Notes
- If you write MCP allow rules as
mcp__<server>, they now reach only that server. Tools of a server whose name merely starts the same way (such asa__b) may prompt again; addmcp__<server>__*with the server’s whole normalized name where needed. The same applies to agenttools:lists. - If you check prompts with a
UserPromptSubmithook, it now runs before a skill or custom slash command expands. A hook that relied on the expanded text should be re-checked. - If you saw tool calls run twice in cloud sessions, or Edit and Retry fail in Cowork, after 2.1.296, both regressions are fixed here.
- Idle compaction only applies to 200K+ token conversations left idle in the terminal for about 54 minutes. Set
"idleCompaction": falsein settings.json to turn it off. - If you use a third-party gateway (
ANTHROPIC_BASE_URL), earlier versions sent its API key with some requests to Anthropic. This release fixes that. - If you followed the old
/ultrareviewadvice to rungit add ., check that untracked files such as.envweren’t staged.