What changed
Claude Code 2.1.304 shipped on October 19, 2026. It adds a guided setup for auto mode, a security check for Managed Agents, and a usage warning when resuming an inactive conversation, and it extends what mods can do in tool.check and settings.read hooks. The fixes span worktrees and deletion safety prompts, compaction and context, subagents, hooks and skills, the sandbox and Bash, LLM gateways, Remote Control, cloud sessions, the VS Code extension and Code Review. The release is large, so the items below are a selection of the changes with the most user impact.
New features
-
/auto-mode-setupA guided setup that reads your project and recent sessions to teach auto mode about your environment. It is now also available on Bedrock, Vertex, Foundry, Claude apps gateway and telemetry-disabled sessions.
-
/claude-api managed-agents-secureCompares a Managed Agents agent’s settings with Anthropic’s published security guide, without changing anything.
-
Usage warning when resuming an inactive conversation (Pro and Max)
On Pro and Max plans, Claude Code now warns when resuming an inactive conversation would use a noticeable share of your 5-hour usage limit. You can resume anyway, or start a new conversation; Claude can look up the previous one if you refer to it.
-
Tool names in the compaction summary
The compaction summary in the transcript view (Ctrl+O) now lists the names of the tools called earlier.
-
More control for mods in
tool.checkA mod’s
tool.checkanswer can now setdecider: "person": only the permission dialog or a host’s callback can approve the ask, and it is refused with the mod’s reason where nobody can be asked. Thetool.checkevent also getse.isReadOnly, set when the tool call is read-only. When a plugin that runs first skips one of your mod’s hooks, the transcript now shows a line naming the event and the plugin. -
[Code Review] Review levels and organization-wide review
Owners choose Quick, Standard (the default) or Deep for the organization, and can give individual repositories their own level. A new admin setting reviews every repository in a GitHub organization, including ones added later, without selecting them one by one.
Key improvements
-
Many subagents at once
Responsiveness and throughput are better when many subagents are running at the same time.
-
Memory and resume in long sessions
Memory use is lower when sending requests in long or resumed conversations, and in SDK and headless mode the first request after resuming a large compacted session is sent sooner. In auto mode, the safety classifier’s context is no longer stored with every request, so session files are smaller.
-
First turn in cloud sessions
A new conversation’s first turn no longer waits for the local (stdio) MCP servers of plugins synced from claude.ai, unless a server sets
alwaysLoad: true. -
Clearer messages
/usagenow says when a rate limit on usage data ends instead of “try again in a moment”. Theclaude -pinvalid-settings warning lists only what stopped a file from loading, and long lists are capped per file. On self-hosted runners, a “not found” from the git host now names the repository and whom to ask. -
[Windows] Faster git start-up
With a standard Git for Windows install, Claude Code now runs its git.exe directly instead of through its cmd launcher.
Changed behavior
-
Tool search behind a custom
ANTHROPIC_BASE_URLTool search is now on by default behind a custom
ANTHROPIC_BASE_URL; a 400 or 422 from the gateway is retried without it. -
Skill list after compaction
Compaction now sends Claude the skill list again afterwards, and
/reload-skillssends only new or changed skills. -
Limits on skill hooks
The skills of one source (a repo, your own folder, or a plugin) can register at most 1,024 hooks and 16,384 matcher characters per session. Under a policy that allows hooks from plugins only, a skill used in the main conversation whose hooks are kept out no longer gets its
allowed-tools. -
Mods’
settings.readhookssettings.readhooks now receive each settings layer and rewrite the merged settings, so a managed mod can make an organization’s setting a ceiling, including thecleanupPeriodDaysthe retention sweep uses. -
Confirmation prompts
The
/artifactsdelete confirmation now shows numbered choices: 1 or Enter keeps the artifact, 2 deletes it, and y, n and Esc still work. The question that turns on the mods Claude writes in a session is now plain text with no link or session id; it stays up when a new turn starts, and Claude tells you what loaded when you turn them on. -
[VSCode] Background sub-agents
Background sub-agents now show as one status line under their Agent card instead of streaming every step into the chat; the full transcript stays in /tasks.
-
Other changes
The /claude-api skill now uses the
antCLI on Claude Platform on AWS instead of falling back to the SDK or curl. The “(research preview)” label is gone from the/fastdialog title on Opus 5.5.
Bug fixes
Data loss & safety prompts
claude --worktreeremoving the worktree and its branch without asking at exit when HEAD was detached or on another branch, which could lose commits.rm -rof a mounted drive or a folder at its top level (/Volumes,/media, WSL’s/mnt/c) not getting the safety prompt thatrm -rf /Applicationsgets.CLAUDE_CODE_TRANSCRIPT_LOCAL_GCdropping messages from the saved transcript when the disk filled up while the file was being trimmed.- Bash and PowerShell deleting a large command’s output file when it could not be copied to the session’s tool-results folder, unless the disk is full.
Compaction & context
- A “Context limit reached” or “Prompt is too long” error after stopping a turn while Claude Code was compacting a conversation the API had rejected as too long.
CLAUDE_CODE_AUTO_COMPACT_WINDOW=200ksilently setting a 100K window; the variable now acceptskandMsuffixes like--autocompact, and warns when a value is out of range or ignored.- Self-paced
/loopwakeups scheduled at the 1-hour maximum firing just after the 1-hour prompt cache expired, so those wakeups no longer re-read the conversation uncached. - Resumed and cloud sessions sending Claude the whole skill list again each time plugins reloaded at startup.
Subagents & agent teams
- Subagents sometimes forgetting their most recent work when they were messaged or resumed right after finishing.
- A plugin keeping a headless run going past
--max-turns, Stop, or ended input by spawning subagents or re-running a finished one. - [Agent teams] Approving a shutdown request from the main session exiting the whole CLI.
Hooks, skills & mods
- A PreToolUse hook’s deny reason being replaced by a timeout error when an Agent SDK callback hook on the same tool call timed out.
- A skill’s
oncehook also removing another skill’s identical hook after it ran. - A failed mod reload removing the tools and commands of the version that stays loaded, and its messages being wrong or silent about whether the previous version is still loaded.
- A click on a mod’s button being lost when buttons with the same label and no
keywere redrawn unchanged. - Mod sharing in the built-in
plugin-authoringskill stopping at the secret scan on machines with managed settings and on Team and Enterprise plans.
Sandbox, Bash & tools
- Sandboxed commands failing to start on Linux when a
sandbox.filesystem.denyReadglob contains brackets that are not a character class, such as a[user-id]route folder. - Unsandboxed Bash commands in Claude Desktop Code tab sessions on macOS and Linux not seeing environment variables exported in shell startup files such as ~/.zshrc.
- An interrupt not ending a Read that waits on an idle pipe or device named like an image.
- Write results and rejected edits showing a file outside the working directory as
../../…; they now show~/…or the full path.
Auth, policy & gateways
context_management: Extra inputs are not permittederrors behind LLM gateways; Claude Code now resends the request without that field.- The organization policy not refreshing until the hourly check after approving an
ANTHROPIC_API_KEYat startup or switching “Use custom API key” in/config. - OpenTelemetry settings delivered through server-managed settings being ignored for the whole session after signing in during first-run setup or the first start after /logout.
- The Consumer Terms notice being skipped for a day or longer after a server error, and
/privacy-settingsoffering choices that may not apply to you after a network or server error.
Terminal display & commands
/effort status,/effort ultracodeand the prompt’s effort indicator showing a guessed effort level whenCLAUDE_CODE_EFFORT_LEVELisautoorunsetand no level is sent.- Notes typed into a multiple-choice question with previews being cut off with ”…” in terminals under about 100 columns, hiding what was typed (#82768).
- Transcript mode’s
{and}landing the prompt on an inconsistent row, or off screen, and Ctrl+C in the agents view clearing a text selection without copying it when pressed before the mouse button was released. claude --helptext for--tools,--resume,--debug,--json-schema,--disable-slash-commands,--system-prompt-snapshotandclaude updatenot matching actual behavior.
Remote Control, SDK & cloud sessions
- Remote Control permanently withholding a desktop conversation’s history when a different account was signed in to the terminal.
- Remote Control failing to start from Claude Desktop when DISABLE_TELEMETRY or DO_NOT_TRACK is set on a paid plan.
- Model switches, messages and interrupts waiting behind a context usage check in SDK and desktop sessions.
- Scheduled Cowork cloud sessions waiting five minutes again for each subagent whose WebFetch needed a URL approval that nobody was there to give.
- A claude.ai Directory plugin’s skills not running as commands when picked from the ”/” menu in Cowork cloud sessions, and folders in the composer’s @ list showing no name in cloud sessions, where picking one added only a space instead of a mention.
VSCode
- One cause of sessions failing to start with “Subprocess initialization did not complete within 60000ms”: time a SessionStart command hook spends running no longer counts toward the limit, up to 10 minutes in total.
- A session closed while it was starting keeping its Claude Code process running until the start ended.
- A mod’s drawings disappearing from a long-open panel until the next turn on machines that also run an older Claude Code.
- Past conversations hiding sessions when
claudeCode.claudeProcessWrapperis set.
The chat panel can now also draw mods for sessions on Amazon Bedrock, Google Vertex AI and Microsoft Foundry, behind an LLM gateway, or with telemetry turned off.
Claude Tag (Slack) & Code Review
- Claude going silent in a thread when its channel becomes shared across Enterprise Grid workspaces mid-task; it now posts a one-time note to @-mention it in a new thread.
@Claude !forktyped outside a thread getting no visible reply; Claude now shows you a private note saying to run it from inside a thread.- [Code Review] Pull requests opened within a minute or two of an admin turning on Code Review being silently skipped; they are now retried and reviewed.
Notes
- If you use
claude --worktree, earlier versions could remove the worktree and its branch without asking at exit when HEAD was detached or on another branch. Updating is recommended. - If you set
CLAUDE_CODE_AUTO_COMPACT_WINDOWwith a suffix such as200k, earlier versions silently set a 100K window.kandMsuffixes now work as intended. - If you use a custom
ANTHROPIC_BASE_URL(LLM gateway), tool search is now on by default; a 400 or 422 from the gateway is retried without it. - If your skills register many hooks, one source is now limited to 1,024 hooks and 16,384 matcher characters per session.
- If you use auto mode,
/auto-mode-setupcan teach it about your project and environment.