Source: raw/anthropic-watch-claude-code-tag-v2-1-326.md (GitHub release of anthropics/claude-code, fetched 2026-11-17; the stub was cut at 20,000 of 47,345 characters), with the full notes in ai-research/claude-code-release-v2-1-326.md, fetched with gh release view. v2.1.327–330 from ai-research/claude-code-docs-changelog-2026-11-21.md.
Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.326 · Date: 2026-11-16–20 (UTC)
Five releases this week (v2.1.326–330). v2.1.326 (November 16) lists 367 items: 220 fixes, 60 behaviour changes, 58 improvements, 10 additions, 2 removals and 17 VS Code items. The weight is in permission, sandbox and auto mode changes that make more commands ask or get refused. v2.1.327–330 (November 17–20) add easel charts, Enterprise Admin Console sign-in, the 30-day scheduled-task limit, VS Code /fork and /workflows, and the /code-review scope and /tui classic changes.
Key Takeaways
v2.1.326 (November 16):
- More commands now ask or are refused. Bash permission checks improved so that “some commands now ask first, or are refused where no one can answer”, and the same for Code Mode and PowerShell.
- Auto mode ignores some wildcard allow rules. Rules “that name no program or that start with
$HOME” are ignored so the safety check decides. Wildcard rules whose program is a Windows backslash path are judged too; exact rules keep their allow. - A hook’s
"decision": "block"holds in more cases, “so a hook that always prints one may start blocking calls”. Check any hook that prints a block decision unconditionally. - Sandbox writes are narrower. Sandboxed commands “no longer write to some sensitive home folders by default; some can be allowed with
sandbox.filesystem.allowWrite”. - Sign-in files are protected. File tools refuse writes to
.credentials.jsonand similar sign-in files. - Sessions with no mode set start in auto mode “when your organization’s policy hasn’t loaded”, not Manual.
- Code Mode works with agent teams. “Added support for turning on Code Mode while agent teams are enabled.”
/update-configis much cheaper: “about 13,000 tokens of context instead of 93,000”.- Remote Control CONNECT tunnel: a proxy that refuses CONNECT now stops Remote Control with a clear error.
- Plugins with a malformed display field now load, with their hooks and MCP servers; administrators should check deployed plugins.
v2.1.327–330 (November 17–20):
- Scheduled tasks and
/loopnow run up to 30 days (was 7) while the session keeps running (v2.1.330). /tasksnow shows scheduled tasks,/loopwakeups, and forks — not only background agents (v2.1.329)./code-reviewnarrowed on most models to bugs + CLAUDE.md violations only; cleanup moves to/simplify(v2.1.328). With no target, it now reviews the whole branch (v2.1.329).mcp_toolhooks supportonFailure: "block"— MCP server call failures block the tool call declaratively (v2.1.328).- Auto mode available on Bedrock, GCP, Foundry, and gateway for Opus 4.6 and Sonnet 4.6 (v2.1.328).
- Easel built-in chart mod draws charts under MCP metrics results on Pro, Max, Team and Enterprise (v2.1.330).
deniedMcpServersnow enforced in cloud/runner sessions (v2.1.328); named entries now scope to that connector only (v2.1.330).- VS Code gains
/forkand/workflowscommands;claudeCode.escapeStopsClaudeandclaudeCode.pinPromptsettings added (v2.1.330). attributionProjectbilling field for grouping usage under a billing group (v2.1.327)./tui classicrestores the previous terminal UI layout (v2.1.327).
Additions
- Startup warning for incomplete tool rules: an
--allowedTools,--disallowedToolsor--toolsentry “that doesn’t look like a complete rule” now warns at startup. /statusrow that says when your provider or an environment variable such asDISABLE_TELEMETRYturns features off. Commands that are off for that reason now say why instead of “Unknown command”./help <command>, and/clear --helpand other built-ins now show usage instead of running.- Ctrl+O shows hidden startup notices.
OTEL_RESOURCE_INCLUDE_HOSTaddshost.nameandhost.idto OpenTelemetry resource attributes (off by default).- Mods:
overflowX="scroll"on a mod’sBox, andmaxLinesandwrap="clip"on itsText. managed-mcp.jsonadded mid-session: a message now explains why MCP servers stop loading; it “may apply at the next check; restart open sessions for it to apply in full”./pluginandclaude plugin listnote when a malformedplugin.jsonfield such asauthororhomepageis ignored.- VS Code: a
claudeCode.promptSuggestionssetting turns prompt suggestions off.
Permissions, sandbox and auto mode
- Bash sandbox keeps its state in some cases during a session, until you run
/sandboxor/cd; Bash may ask first. - Environment-variable lines: the sandbox asks before some lines that set environment variables when auto-allow is off; “unattended runs can refuse them”.
- Rewritten commands stay sandboxed: a command that a permission hook or SDK host rewrites while approving stays in the sandbox if the original was checked as sandboxed; the approver can add
dangerouslyDisableSandbox: true. cpandmv: some now ask first. Acprefused only because its source was under anEditdeny rule is fixed.- Very long values: the Bash check asks about some commands with very long values it cannot finish checking, or refuses them “where a deny rule may apply or no one can answer”.
--add-dirand/add-dirin local terminal sessions show the trust prompt for a folder you have not trusted yet;claude --bgrefuses such a folder.- Folder trust prompt now names the project settings files Claude Code will load, and at launch marks files that run commands.
- Managed
autoMode.environmentwithout"$defaults"now replaces every built-in entry in Claude Desktop and Agent SDK apps too. claude mcp serveignoressandbox.excludedCommands./rewindwith the sandbox on (macOS, Linux, WSL) leaves some files untouched and reports them as skipped. A turn whose changes were not tracked now reads “Code changes unknown” instead of “No code changes”.- PowerShell: asks before more commands, some even in bypass mode; some malformed commands are denied.
- Linux and WSL: the sandbox creates some missing
--add-dirfolders before each sandboxed command, so commands that create them can fail, skip or land inside them.
Hooks
updatedInputfrom PreToolUse hooks is kept in more cases when more than one hook answers a tool call.- Exit 127: the error now says the hook’s check may not have run.
- Transcript colours: failures are red, a hook that stops Claude is yellow, and the reason is shown under them.
- Crashed Node or Python hooks show the error, not only the first stack line; Ctrl+O shows up to 10,000 characters of output.
- HTTP hooks ignore
Content-LengthandTransfer-Encodingin theirheaderssetting. HowallowedHttpHookUrlsandhttpHookAllowedEnvVarsapply to managed-settings hooks also changed; the notes do not say how. - Telemetry fix: PreToolUse hooks in managed settings that refused a direct tool call left no OpenTelemetry
tool_decisionevent orpermission_denialsentry; fixed.
Subagents, worktrees and background sessions
- Worktree-isolated subagents refuse, in any repository, some git commands in Bash “that can rewrite branches shared with your checkout; sessions are not affected”.
ExitWorktreerefuses to remove a worktree while a background subagent or shell command may still be running in it, unless Claude passesdiscard_changes.- Subagent plan approval is now two choices: “Yes, and switch to default (ask each time) for this session” and “No”.
- Agent rows say what an agent is waiting on, and
/tasksnames the subagent behind each background command. Stopped and failed agents no longer look like running ones. - More than 20 subagents: token counts of background agents whose transcript is not open update every 5 seconds, to save CPU.
- Background sessions that lose the LLM gateway your shell exported now stop before any model request.
--bareon an unapproved API key: moving a session to the background is refused in most setups.
MCP, Remote Control and cloud
- MCP fixes: parallel tool calls reading another call’s saved result when two long results arrived in the same millisecond; tool search loading another server’s tool for an exact name; prompts disappearing when two servers have nearly identical names; root
anyOf/oneOfshowing only the first branch. claude mcp add-jsonnow says what to change instead of “Invalid configuration: : Invalid input”./mcpshows why a local (stdio) server failed to start.- Cloud sessions: when an MCP call through the session’s proxy is cut off after it was sent, Claude is told the outcome is unknown.
- Remote Control: secrets are masked better in push notifications; Stop on your phone or the web now ends background subagents together with the turn;
--verbosedebug logs go to~/.claude/child-logs/. - Self-hosted runner: Linux and macOS sessions with no
ghget a built-ingh apiby default, via Anthropic;--use-anthropic-gh falseturns it off. Under--trust-workspace false, more repository settings are ignored. - Cowork cloud sessions hold an organization’s OpenTelemetry export until the first turn, “so
claude_code.session.countcan fall”.
Other fixes worth knowing
claude purgehid its prompts when stdout was redirected, “where one unseen Enter could delete a project’s data”; they now show on the terminal.- Settings: a settings file saved with an error mid-session was skipped silently; a warning now names it. Rules added in
/permissionslooked saved when the settings file could not be written. Broken JSON in a settings or.mcp.jsonfile now reports line and column. - Auto mode billing: auto mode could stay on its local classifier, “which counts toward usage”, because of a stale saved value; fixed.
max_tokensover the model limit now retries with the maximum the API reports.- Compaction on Bedrock, Vertex and Foundry, and with telemetry off, sometimes replied to a message instead of summarizing; fixed.
grepin the Bash tool silently skipped some text files; fixed./memoryopened a read-only file with no warning that changes may not save; fixed./model: Default now names your plan’s default model, and choosing it switches you back to that default.- Plugin uninstall also deletes the files extracted from the plugin’s MCP bundles (
.mcpb,.dxt), unless another plugin uses the same bundle./pluginuninstall questions are now lists;y,nanduno longer answer them. - Performance: an idle session wakes the machine less often; the first Bash command on Windows is about a second faster; Claude Desktop’s first turn no longer waits up to 500 ms for local MCP servers whose tools load on demand.
Breaking and behaviour changes to check
| Change | Who should check |
|---|---|
| Bypass-mode tool calls that now ask or are denied | Scheduled and headless runs |
Auto mode ignoring wildcard allow rules with no program or starting with $HOME | Anyone with broad Bash(...) allow rules |
"decision": "block" holding in more cases | Hook authors |
| Sandbox no longer writing to some home folders | Sandboxed builds that write caches under ~ |
| CONNECT tunnel for Remote Control | Networks with restrictive proxies |
Stricter PATH for git in worktree creation, /teleport and IDE checkouts | Repos with git filters or helpers |
| Plan-mode compaction on the answering model | opusplan users watching cost |
--await-initialize refused with --bg or --worktree --tmux | SDK hosts |
Try It
- Run your unattended jobs once by hand on v2.1.326 and read the transcript for new prompts or refusals before trusting the schedule.
- Start with
claude --allowedTools ...as your scripts do and look for the new incomplete-rule warning. - Grep your hooks for
"decision": "block"and confirm each one is conditional. - If a sandboxed command now fails writing under your home directory, add the folder to
sandbox.filesystem.allowWritein user settings. - Run
/statusand check the new row for features turned off by your provider or environment.
Open Questions
- Which commands and folders? The notes say “some commands”, “some sensitive home folders” and “some tool calls” throughout and do not list them.
- What changed for
allowedHttpHookUrlsandhttpHookAllowedEnvVars? The notes say only that how they apply to managed-settings hooks changed. - Which
.\wildcard rules will be restored? The notes say “we plan to restore rules that start with.\” with no date.
v2.1.327–330 (November 17–20, 2026)
Source: ai-research/claude-code-docs-changelog-2026-11-21.md.
v2.1.327 — November 17
26 improvements, 100 bug fixes.
attributionProject: a new billing field for grouping Claude Code usage under a billing group or project identifier. Set in managed settings orsettings.json./tui classic: restores the previous terminal UI layout for users who prefer the older interface before switching permanently to the new one.claude mcp add-json <name> -: reads JSON from stdin — useful in scripts that pipe configuration directly./install-github-appconfirmation: the command now asks for confirmation before changing a repository. Previously it ran immediately on enter.ExitWorktreeasks: before removing a worktree that may hold unsaved work, the command now prompts.- Broken-file list in
/status:/statusnow lists files Claude Code found broken or malformed at startup. - Permission prompt arrow keys: arrow keys in the permission prompt now stop at “No” instead of wrapping back to “Yes”. Prevents accidental approvals from keyboard overshoot.
claude mcp add --scope project: no longer overwrites.mcp.jsonin some multi-layout projects.
v2.1.328 — November 19
19 improvements, 91 bug fixes.
- Auto mode on Bedrock, GCP Agent Platform, Microsoft Foundry, and Claude apps gateway: Opus 4.6 and Sonnet 4.6 now support auto mode on all four providers. Previously auto mode was only available on Anthropic direct.
"onFailure": "block"formcp_toolhooks: the same declarative failure policy already available for command and HTTP hooks now works formcp_toolhooks. When the MCP server call fails, the tool call is blocked without requiring the hook to handle the error explicitly.CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME: environment variable to disable the deferred tool resume behavior./code-reviewscope narrowed: on most models,/code-reviewnow reports only bugs and CLAUDE.md violations. Cleanup findings move to/simplify.deniedMcpServersentries now apply in cloud and runner sessions:"claude.ai <name>"form entries, previously local-terminal only, are now enforced in cloud sessions and self-hosted runners too./bugasks before sending: the/bugcommand now prompts “Yes or No?” before submitting the report to Anthropic.
v2.1.329 — November 19
10 improvements, 91 bug fixes.
/tasksnow lists scheduled tasks,/loopwakeups, and forks: not only background agents. One unified view across all in-flight async work.- Trust-dialog warning about hooks location: when a folder’s hooks are declared in a source other than the user’s own settings (e.g., a project
.claude/settings.jsonyou haven’t reviewed), the trust dialog now says where they come from. - “Server message” row in
/mcp:/mcpnow shows any message the MCP server sent at connection time. Useful for servers that report their version, auth status, or quota at startup. - Code Mode + OpenTelemetry: Code Mode now works with OpenTelemetry export, so code-mode turns appear in OTel traces.
/code-reviewwith no target reviews the whole branch: previously it reviewed only uncommitted changes. Without a target, it now reviews all commits on the branch against the default branch.- Auto mode runs
git statusless often: reduces tool calls and token use in auto-mode sessions.
v2.1.330 — November 20
14 improvements, 98 bug fixes.
- Easel built-in mod: draws charts when asked and automatically under MCP metrics results. Available on Pro, Max and (where enabled) Team and Enterprise on macOS and Linux. Turn automatic charts off in
/config. See Claude Mods for context on the mods system. - Sign-in to Enterprise Admin Console organizations: after
claude auth login, Claude Code now applies the organization’s settings and routes to its cloud provider or gateway when the organization is managed in Anthropic’s Enterprise Admin Console. CLAUDE_CODE_FULL_SIGN_IN_LINK=1: prints the full sign-in URL if the short link shows an error or if a script needs to read the value.- Scheduled tasks and
/loopup to 30 days: recurring scheduled tasks and/loopwakeups now run for up to 30 days instead of 7, as long as the session keeps running. Claude Code says when they will end. Tasks created before this release still end at 7 days. - More
rmcommands require approval: commands aimed at the root, home directory, a top-level folder, or the working folder now ask for approval. Where no one can answer (headless sessions), they are refused. deniedMcpServersname entries scope narrowed: a namedclaude.ai <name>entry now hides only that specific connector, not every connector at the same URL. To block all connectors at a URL, deny the exact URL instead.--start-dirfor self-hosted runner: sets the folder Claude starts its work in without changing where settings and hooks load from.- [VS Code]
/fork: forks the conversation after its last reply. - [VS Code]
/workflows: opens the agent map showing running workflows. - [VS Code]
claudeCode.escapeStopsClaude: turn off so Escape does not stop Claude. - [VS Code]
claudeCode.pinPrompt: lets the message input scroll away instead of staying pinned. - [VS Code]
claudeCode.contextUsageThreshold: controls when the context usage indicator appears.
Notable fixes in v2.1.330:
- Fixed Claude Code freezing for up to several minutes right after compacting a very long conversation.
- Fixed WebFetch freezing the session for minutes on some pages.
- Fixed
ExitWorktreeand the worktree exit dialog deleting a worktree that another running Claude Code session was working in. - Fixed changes to user settings being lost when several Claude Code processes saved them at the same time.
- Fixed the first prompt after a 2+ minute sleep waiting 3 minutes on a dead connection (Linux and WSL).
Related
- What’s New — Week 46 — the previous digest (v2.1.321–325), including Code Mode’s introduction and the sandboxed-Bash permission change.
- Agent guardrails — permissions and sandboxing.
- Auto Mode — the classifier these rule changes feed.
- Hooks — block decisions,
updatedInput,mcp_toolonFailure, and HTTP hooks. - Scheduled tasks — 30-day limit and
/tasksunified view. - Remote Control — streaming replies and the CONNECT tunnel.
- Agent teams — now usable with Code Mode.
- MCP —
deniedMcpServerscloud/runner fix and “Server message” in/mcp. - Code Review GA —
/code-reviewscope changes (bugs + CLAUDE.md only; whole branch). - Claude Mods — easel chart mod introduced in v2.1.330.
- Anthropic SDK releases, November 2026 — Python v2.9.0 and TypeScript sdk-v1.9.0 from the same day.