Source: raw/anthropic-watch-claude-code-tag-v2-1-321.md and raw/anthropic-watch-claude-code-tag-v2-1-322.md (GitHub releases of anthropics/claude-code, fetched 2026-11-11; neither stub was truncated), plus raw/anthropic-watch-claude-code-tag-v2-1-323.md (fetched 2026-11-13; the stub was cut at 20,000 characters, so the full notes are in ai-research/claude-code-release-v2-1-323.md, fetched with gh release view), and raw/anthropic-watch-claude-code-tag-v2-1-325.md (fetched 2026-11-14 UTC; the release body is 19,870 characters, under the stub cap, so the stub is complete). Continues Week 45, which ended with v2.1.319 (November 6 UTC). There is no v2.1.320 release on GitHub (gh release view v2.1.320 returned “release not found” on 2026-11-11), and no v2.1.324 (same check, 2026-11-13). The docs changelog snapshot of 2026-11-14 (ai-research/claude-code-docs-changelog-2026-11-14.md) summarises v2.1.323 as 21 improvements and 195 bug fixes and v2.1.325 as 14 improvements and 88 bug fixes; it groups items differently from the GitHub notes counted below.

Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases · Date: 2026-11-09 to 2026-11-13 (UTC)

Four releases so far this week. v2.1.321 (November 9) adds Claude Fable 5.2 and makes it the default Fable model, moves saved Fable 5.1 picks to it, fixes four permission or sandbox gaps and an MCP OAuth secret leak, and lets mods’ own tool calls skip permission prompts. v2.1.322 (November 10) is one line: fast mode on Vertex. v2.1.323 (November 12) is a large release of 294 items (195 fixes, 48 behaviour changes, 27 additions): it adds Code Mode, an experimental opt-in, tightens Bash permission checks and the sandbox in more than twenty places, and fixes a group of bugs where a background session could move to a different sign-in. v2.1.325 (November 13) fixes a security issue in zstd decoding and changes sandboxed Bash commands to go through the normal permission flow by default, and makes scheduled tasks and /loop ask before running a skill that declares allowed-tools or hooks. This digest covers releases through November 13; later Week 46 releases will be added here.

Key Takeaways

  • Fable 5.2 is the default Fable model (v2.1.321): “Added Claude Fable 5.2 (claude-fable-5-2), now the default Fable model — 1M context, 40 per Mtok with $0.20/Mtok cache reads”.
  • Saved Fable 5.1 picks move to fable (v2.1.321): “on Pro and Max plans at once, on Team and Enterprise plans once your organization allows it”. Run /model claude-fable-5-1 to switch back.
  • Security fixes in v2.1.321:
    • MCP OAuth requests no longer send “secrets such as refresh tokens to another host when a server redirected them”.
    • A read-only Bash command could run “outside the sandbox without a permission prompt”.
    • Bash permission checks could judge a file write “from the wrong folder” and not prompt.
    • Deny rules and protected-file prompts did not apply to the file that a cp or mv into a directory writes. cp and mv spelled by a path (/bin/cp) now get the same path checks.
  • Mods get more room (v2.1.321): mods Claude writes turn on without asking in auto and bypass modes, “even with managed allowUnsandboxedCommands: false or allowManagedDomainsOnly”, and do not load in dontAsk mode. A mod’s own tool calls ($.tool.call, $.mcp.call) “raise no permission prompt, even under your own ask rules or in dontAsk and -p runs; deny rules and organization policy still apply”.
  • Fast mode falls back to standard speed first (v2.1.321): when fast serving is rate-limited, overloaded or failing, Claude Code “now tries standard speed on the same model first” instead of losing replies, switching models or retrying for minutes. The /fast, /config and /model labels now name your current model when it supports fast mode, and “(Opus only)” is gone from the VS Code row.
  • Fast mode on Vertex (v2.1.322): available “in the global, us and eu regions, falling back to standard speed if Vertex refuses it”.
  • claude remote-control --settings (v2.1.321) applies a settings file to every session it starts. If the file changes, no new session starts until you restart claude remote-control. Remote Control sessions also now keep the folder they were started in: “a connected client can no longer move a session to another folder”.
  • Long command echoes are trimmed (v2.1.321): background command notifications and TaskStop results show Claude only the start and end of a command over 1,000 characters, which saves context.
  • Sandboxed Bash and unattended skills now ask (v2.1.325): sandboxed Bash commands “follow the normal permission flow by default” (sandbox.autoAllowBashIfSandboxed: true restores the old behaviour), and scheduled tasks and /loop “ask before running a skill that declares allowed-tools or hooks” (add Skill(name) to permissions.allow). Check unattended jobs after updating.
  • Code Mode (v2.1.323): “an experimental opt-in in which Claude writes JavaScript to call its tools”. The note points to code.claude.com/docs/en/code-mode to turn it on, “or to turn it off for an organization”. That page returned 404 when fetched on 2026-11-13.
  • Permission and sandbox fixes in v2.1.323: more than twenty. Among them: Write and Edit “could write to a path covered by a Read deny rule”; “the Bash sandbox could switch off during a session”; Bash deny and ask rules missed “some commands run through sudo, env and similar wrappers”; an rm could be judged “from the wrong folder” and skip the dangerous-removal prompt; --permission-mode auto could start “in a mode that could allow more when auto mode was unavailable” (it now starts in Manual mode).
  • Background sessions no longer change accounts (v2.1.323): several fixes for a session on a named Anthropic profile, an API key or an auth token “moving to another sign-in” when continued in the background. Such a session “now stops and says so”. The note adds: “reopen any session that was open during the update”.
  • Session-to-session message limit raised to 100,000 (v2.1.323): how many messages each of your sessions on the same machine can send the others without a prompt from you goes “from about 20 to 100,000”; the same change applies to Claude Desktop Code tab sessions.
  • Unknown model IDs are assumed to have a 1M context window (v2.1.323), up from 200K. Sessions on other vendors’ models with a smaller window “can fail every turn until /clear” once full, so set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 for those.
  • “Don’t ask again” no longer saves risky rules (v2.1.323): an allow rule “that looks like it holds a secret, or would be saved to a settings.local.json that git tracks, lasts for the session only and isn’t saved”.
  • The account-email check is narrower (v2.1.321): it now refuses, without asking, “only a shell command that holds your email and names a User-Agent”, so other commands and scheduled runs no longer stop at its prompt.

Release table

VersionDate (UTC)Headline changes
v2.1.3212026-11-09Fable 5.2 added and made the default Fable model; saved Fable 5.1 picks move to fable; MCP OAuth redirect fix; three Bash sandbox and permission fixes; mods’ tool calls skip prompts; fast mode tries standard speed first; claude remote-control --settings; about 80 other fixes
v2.1.3222026-11-10Fast mode on Vertex (global, us, eu) with fallback to standard speed
v2.1.3232026-11-12Code Mode (experimental opt-in); engines["claude-code"] in plugin.json; TypeScript checks in claude plugin validate; more than twenty Bash permission and sandbox fixes; background-session sign-in fixes; session-to-session message limit 20 to 100,000; unknown model IDs assumed 1M context; subagentStatusLine removed; a plugin’s agent setting deprecated; 195 fixes in total
v2.1.3252026-11-13zstd decoding security fix; sandboxed Bash follows the normal permission flow by default; scheduled tasks and /loop ask before running a skill with allowed-tools or hooks; read-only git in an untrusted repository asks first; more sign-in and billing fixes for background sessions; more Bash permission fixes

Models and fast mode

  • Fable 5.2 as above; see Claude Fable 5.2 + Mythos 5.2 for what changed in the model. The fix list also removes a wrong “Draws from usage credits” note that showed after a switch to Fable on Enterprise seats whose plan includes Fable.
  • Plan names: the session-feedback message and the allowedProviders description now name Ultra alongside Pro and Max, forceLoginMethod covers every paid Claude plan, and the “Upgrade to Max” options read “Upgrade to Max or Ultra”.
  • /model picker: “More models” is now an unnumbered row below every other, and older models open under it without renumbering the rest.
  • Fallback model: fixed the fallback model sometimes declining a request or mentioning a stopped response when the first model was blocked before replying.

Permissions, sandbox and auto mode

  • The four fixes in Key Takeaways, plus: Bash reads such as grep -r foo "$PWD/src" no longer ask for permission when the same command with the path typed out runs without asking.
  • /config default mode: it showed Manual as the default permission mode when new sessions start in auto mode; pressing Space on “Auto (default)” now saves Don’t Ask, not Plan.
  • Auto mode and background subagents: fixed auto mode losing a background subagent’s report that had started a turn when the conversation was compacted before Claude answered it. In headless sessions, the desktop app and VS Code, reports from background subagents that wait together are now answered in one model call, as in the terminal.
  • Claude in Chrome policy: Chrome tools that an app running Claude Code provides, such as the Claude desktop app’s Code tab, now follow your organization’s Claude in Chrome policy.
  • Cyber safeguard blocks: the Cyber Verification Program sentence on a block now uses the same words as claude.ai.
  • Windows: Claude Code now skips relative and UNC PATH entries and links onto a network share when finding programs it starts itself, such as git, and picks up PATH changes from settings.

Mods, plugins and skills

  • Mods as above. Headless sessions now reload mods in .claude/skills when their files change; set CLAUDE_CODE_PLUGIN_DIR_WATCH=0 to turn this off. See Claude Mods.
  • Nested skill loading: .claude/skills folders inside ~/.claude/plugins no longer load when Claude works on files there.
  • Managed extraKnownMarketplaces: a declared name can’t be added from another source, an entry from elsewhere is hidden, and plugins the declared marketplace lists are reinstalled from it. On a take-over from an undeclared source, saved data is set aside for cleanupPeriodDays, and options and secrets until claude plugin restore-set-aside.
  • Plugin updates: fixed auto-updates and /plugin marketplace updates downloading a plugin again for each worktree or folder of a repository.
  • Skill and command names holding a Bamum letter that some fonts draw as blank are now refused.
  • Fixes for disappearing project skills, commands and agents: after their folder was removed and came back mid-session (for example with git stash), and in SDK and desktop sessions after Claude entered a worktree. See Plugins and marketplaces.

Hooks

  • A large hook output no longer points Claude to a file that held another hook’s output from the same tool call.
  • A PreToolUse hook’s ask is no longer lost when another hook’s defer finished first and could not be honored; apps that embed Claude Code get the right hook’s ask reason when two hooks asked.
  • PreToolUse mcp_tool hooks on stdio and HTTP MCP servers were skipped when their answer was over 16 MiB; the tool call is now refused.
  • Cowork cloud sessions: a PostToolUse hook can no longer replace the answer to AskUserQuestion or the plan in ExitPlanMode’s result; Claude reads the originals.
  • Remote sessions no longer run UserPromptSubmit hooks on a replayed permission answer after a restart.
  • Windows: FileChanged hooks no longer stop for watchPaths paths after the first shell command when the working directory has a lowercase drive letter.
  • OpenTelemetry hook events in non-interactive sessions now leave two built-in hooks out of num_hooks and num_success after Bash and PowerShell calls. See Hooks.

MCP

  • The OAuth redirect fix above.
  • claude mcp login no longer prints a SEP-2352 issuer warning each time you sign in again to a server that Claude Code registered with.
  • Fixed an MCP server sometimes showing as connected with no tools after a reconnect or directory change, and a project’s MCP server being disconnected after the first message in a Claude Desktop Code tab session when the desktop app has a server or extension of the same name.
  • Fixed a blank screen at startup and empty claude mcp list output when a stored MCP sign-in record was null. See MCP.

Remote Control, cloud sessions and headless

  • Remote Control: --settings and the fixed-folder change above; a message sent with “Queue for later” no longer reaches Claude in the middle of the running turn; a crashed session picks up a new message sooner; /compact no longer shows twice on web and mobile. See Remote Control.
  • Cloud sessions: pkill, pgrep and pidwait no longer match the session’s own parent processes; a restarted session no longer drops a message with a large image or leaves a subagent’s message unanswered; file uploads that get an HTTP 507 from a proxy get a second try; Claude is no longer told that an allowed host is probably not allowed when the proxy could not connect.
  • Background agents are no longer reported as stopped by the user, and refused a resume, when a service interrupted a cloud session and nobody pressed Stop.
  • claude -p: on Linux it now prints an error and exits 1 (in text and json output) when a full NFS export or used-up NFS quota lost transcript lines; it no longer sometimes tells Claude that background commands get 30 minutes when they are stopped after 10; AskUserQuestion denies a host’s permission answer that sets the undocumented afkTimeoutMs field.
  • claude agents no longer ignores --permission-mode, --dangerously-skip-permissions, --model, --effort or --agent typed before agents when a flag such as --verbose was also on the line.
  • Workflows: Claude is now told when you stop or pause a running workflow, so it no longer says the workflow is still running.
  • Claude Desktop Code tab: the number of messages a session can send to your other sessions before it waits for you to type again goes from 10 to 20.

Claude apps gateway

  • policy_mode: layered: managed MCP servers now add up over every matching policy, match: {} included, “so a gateway already on it serves more after the upgrade”.
  • Upstreams: Bedrock, Vertex and Foundry upstreams ignore the ANTHROPIC_*_BASE_URL variables in the gateway’s environment; set base_url on the upstream to use another endpoint.
  • The gateway no longer refuses to start when a desktop setting holds a field newer than the gateway (such as a managed MCP server’s oauth.omitResource); it warns and serves it.
  • Spend limits on a busy replica skip fewer checks and record spend sooner; the spend-limit error shows the reset in local time, such as “resets 4pm (America/Los_Angeles)”; error messages show the gateway’s address without any user name or password. See Claude apps gateway.

Other changes worth knowing

  • Session start git status now reads as unavailable when git status failed (it used to read as clean), and no longer shows the first listed file as staged when it was not.
  • Bash working directory: fixed commands sometimes running in the wrong folder when Claude sent several at once and an earlier one changed folders.
  • Edit and NotebookEdit on network paths (UNC shares, WSL folders, /net mounts): fixed Edit changing the first of several matches or replacing a whole file on an empty old_string, and NotebookEdit deleting or overwriting a notebook’s first cell when the call named no cell.
  • WebFetch no longer passes “No response from model” to Claude as page content; the call fails with a clear error.
  • TLS: a connection that broke partway through a reply no longer ends the turn with “API Error: EPROTO”; the partial reply is kept or the request is retried, “as before 2.1.302”.
  • Memory and speed: lower memory use in long sessions that repeatedly read single lines from large files; claude install finishes 2 to 4 seconds sooner; faster startup on Windows and from faster TLS certificate loading.
  • Context estimates no longer overcount tool definitions when exact token counts are unavailable.
  • Images: in sessions with more images than fit in one request, Claude is told whether it has already seen a dropped image, and repeat reads of an unchanged image are limited.
  • Artifacts: design guidance now treats Inter and Space Grotesk as good default typefaces; the bundled dataviz skill suggests a list of meters for several ratios that each have their own limit; Claude refuses to send a file whose name hides text-direction or control characters.
  • Self-hosted runner: an on-demand runner the server refuses because it is not needed now exits 0 instead of 1.
  • VS Code: a warning above the prompt box when both an auth token and an API key are set.

v2.1.323 (November 12)

The sections above describe v2.1.321 and v2.1.322. Everything in this section is from v2.1.323.

New

  • Code Mode as in Key Takeaways.
  • Plugin version ranges: engines["claude-code"] in plugin.json. Plugins outside their stated Claude Code version range “are not installed, updated or loaded, and install and update pick the newest version that fits”.
  • claude plugin validate now type-checks a plugin’s TypeScript hooks module (each finding is a warning; --strict fails on it) and lists the $ nouns a mod’s engine.create hook adds, withholds or replaces.
  • Usage-credit options menu (stop, switch to usage credits, upgrade) when a model needs usage credits that are turned off or used up.
  • Hook heartbeat: a progress heartbeat every 30 seconds while hooks run long in claude -p and SDK sessions, so a host such as Claude Desktop does not end the turn as hung.
  • .docx is now accepted in Artifact tool asset uploads.
  • New environment variables:
    • CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES=1 turns off subagent progress summaries and the requests that write them (the note says it has been available since 2.1.318).
    • CLAUDE_CODE_RESUME_HINT_COMMAND lets a program that starts Claude Code name its own resume command in the exit hint.
    • CLAUDE_CODE_AUTO_MODE_MODEL is supported in managed settings for Claude Desktop sessions on third-party providers.
    • CLAUDE_CODE_AUTOMOUNT_ROOTS for Linux machines where /home is an autofs map root deeper than one folder below /.
  • Start-up warnings: for Bash allow rules with a * in a variable’s value, such as Bash(PORT=* npm run dev); when a plugin’s language server can’t start; and a one-time notice naming CLAUDE_CODE_PERFORCE_MODE when Perforce settings are found and the mode is off.
  • Mods: isPromptCovered is added to the props a mod’s pane is drawn with.

Permissions and sandbox

  • The fixes in Key Takeaways, plus fixes for:
    • Bash allow rules with an attached wildcard, such as Bash(ls*), skipping the prompt in some cases. A rule that names a variable, such as Bash(NODE_ENV*), now asks.
    • Bash allow rules written with combined sudo flags skipping the prompt.
    • sandbox.excludedCommands letting a command it does not name run outside the sandbox.
    • Sandboxed commands being able to read Claude Code’s file checkpoint copies under the config directory.
    • Project settings being read through a symbolic link that leads to a different folder than the one Claude Code checked.
    • Plan files being written outside the project when plansDirectory is set.
    • A Bash command running without a prompt “for zsh users with certain non-default shell options set”, and some lsof commands running without a prompt.
    • “Don’t ask again” being offered for commands that write to .git or .claude, where the saved rule did not stop the next prompt.
    • claude -p and the Agent SDK running Bash unsandboxed when --settings or SDK settings that turn the sandbox on fail validation. Claude Code now stops at start-up and names the invalid setting.
    • Auto mode behind a gateway on a model whose gateway name contains “haiku”: such a model, other than claude-haiku-5-5, is now refused.
    • Windows: file-reading Bash commands running without a prompt when given a wildcard or a list of files; a .cmd or .bat wrapper beside a program’s .exe being run in place of the .exe.
  • Sandbox changes (macOS, Linux, WSL): write access granted through a symlink in a writable folder that leads elsewhere is ignored (grant the real path); git worktree add, move and remove fail read-only, as do commits in a worktree the session wasn’t started in or given; glab config set without --global or --host fails read-only.
  • Hooks and permission checks: when a tool’s permission check fails with an error after a PreToolUse hook returned “allow”, Claude Code now decides the call “as if no hook had answered”. When it fails on input that a PermissionRequest hook rewrote, Claude Code asks again, or refuses where nobody can answer.
  • Credentials kept from subprocesses: Claude Code’s own ripgrep and most internal git commands now run “without most of its sign-in tokens and API keys”, so a credential helper or hook under such a git command no longer finds them.
  • /security-review now asks before Claude runs it, and skills’ inline git commands stop in bare-repository worktrees and similar folders until you allow them (outside auto mode and bypass).
  • MCP sign-in: the tool Claude uses to finish an MCP server sign-in now asks you to confirm, even in bypass and auto modes. In dontAsk mode, use /mcp.
  • Artifact updates ask first, even under an allow rule, “when people outside your organization can see the artifact or Claude can’t check who can”.
  • Untrusted folders: on self-hosted runners under --trust-workspace false and in VS Code sessions in an untrusted folder, the repository’s settings can no longer disable your hooks or turn plugins on or off.

Sign-in and accounts

  • The background-session fixes in Key Takeaways.
  • /logout in a background session, which had no effect, now signs out every session on the computer, deletes stored MCP logins and plugin secrets, and ends the session.
  • macOS: fixed “Not logged in · Please run /login” coming back right after a successful /login when the keychain was slow to answer.
  • MCP tool calls no longer end in the sign-in server’s error, such as TOKEN_EXPIRED; you are asked to sign in again.
  • /login and claude setup-token errors now say what happened and what to check.

Plugins, mods and MCP

  • Mods: a *, namespace or negated hook on engine.create “now only observes”; to add or withhold a $ noun, hook engine.create by name. claude plugin validate fails a mod whose engine.create hook isn’t written in the one accepted form (installed mods keep loading). Mods can no longer reword Claude Code’s own confirmation questions. A permission prompt from a mod’s tool.check hook now names the plugin that asked. See Claude Mods.
  • Plugins: installs fail when a git submodule isn’t on the commit its repository records; claude plugin install and /plugin show a warning when a plugin’s install command writes to stderr; tool.list plugin hooks now apply the tools they add or remove “to the model and every agent”; marketplaces named npm, pip, uv, cargo, github or gh are no longer refused as reserved names. A plugin’s agent setting is deprecated (it still applies, and now warns).
  • claude.ai chats and Cowork tasks check a plugin file for secret keys before sending it, and ask first if it may hold one.
  • MCP: MCP_SDK_GENERATION=v1 now runs the v2 MCP client in a compatibility mode, “because the v1 MCP client is removed”. Project (.mcp.json) server approvals match the server’s exact name. /mcp shows how many of a server’s tools were left out because the API would reject their input schema, and why. See MCP.

Usage limits and Fable

  • Fable’s weekly limit showed “no reset time and no options menu in the terminal”; fixed. The Fable limit message in -p, SDK, IDE and cloud sessions now says when the limit resets and how to continue.
  • /usage keeps the session and weekly bars for setup tokens and CLAUDE_CODE_OAUTH_TOKEN and says why the rest are missing.
  • Workflows in the desktop app’s Code tab pause at a usage limit and continue after it resets.
  • CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS maximum goes from 64000 to 300000.

Sessions, compaction and context

  • /compact run again, or after a command such as /context, no longer summarizes away a message Claude had not answered yet.
  • A resumed session no longer joins your next message to an old interrupted message.
  • A setting saved in a running session no longer undoes a hand edit made to settings.json moments earlier, and a settings file you delete during a session stays deleted.
  • Finished subagents’ transcripts are freed from memory sooner.
  • ToolSearch now says when a tool Claude searched for by its short name is already available.
  • claude rm and the agents view say a session is “removed from the list, not deleted”; claude --resume still lists it.
  • Split-pane agent teammates check that requests come from their team lead: one that outlives its lead has its asks refused, and plan approvals must be re-sent after /clear or /resume. See Agent teams.

Remote Control, cloud sessions, Claude Tag and VS Code

  • Remote Control: claude remote-control keeps its identity for up to 30 days when restarted in the same folder (it used to reconnect as a new computer after 4 hours); sessions no longer show as disconnected for a minute or more after a network change; the usage-credits question now also shows in the Claude app. See Remote Control.
  • Self-hosted runner: the exit code of an on-demand runner whose work order was superseded is back to 1 (“2.1.321 and 2.1.322 had changed it to 0”).
  • Claude Code on the web (19 items): clearer upload and repository-access errors, a No repository option for new sessions on a self-hosted environment, a warning in the Delete dialog when deleting a session will pause scheduled tasks, and a failure notification when a scheduled task can’t start because its settings require a sandbox that isn’t available.
  • Claude Tag (9 items): sessions started in Slack are added to the Compliance API on Enterprise plans, in beta; a Scope filter on the Network events tab of the Activity page.
  • Code Review: a Download CSV button in admin settings exports review costs per repository by month, week or day.
  • VS Code (14 items): a trust prompt before the first session in a folder you haven’t trusted; live output of a running Bash or PowerShell command under the command’s card.
  • OpenTelemetry: an otlp exporter with an endpoint but no OTEL_EXPORTER_OTLP_PROTOCOL now defaults to http/protobuf; one malformed OTEL_RESOURCE_ATTRIBUTES entry no longer drops the others. See Agent observability.

v2.1.325 (November 13)

Everything in this section is from v2.1.325. There is no v2.1.324.

Security and permissions

  • “Fixed a security issue in how zstd-compressed responses are decoded.” No advisory is linked.
  • Sandboxed Bash now asks like any other command: “Changed sandboxed Bash commands to follow the normal permission flow by default; set sandbox.autoAllowBashIfSandboxed: true to keep running them without prompts.” For managed settings that turn the sandbox on, the key goes in “your highest-ranking managed source”.
  • Unattended runs can now stop for a skill: “Changed scheduled tasks and /loop to ask before running a skill that declares allowed-tools or hooks; add Skill(name) to permissions.allow to run it unattended.” A related change: “Always allow” in Desktop, IDE and SDK hosts offers a rule alone only when saving it would allow that command again, “so a scheduled task that relied on one can now wait for you”.
  • Read-only git outside the project asks first in a repository that isn’t the project’s or one you trusted; name it with --add-dir (not the additionalDirectories setting) to run them without prompts.
  • Bash fixes: some compound commands could run without asking; “a Bash allow rule could run a different program than the one it names”; some deny and ask rules did not apply to certain Bash commands; relative rules such as Read(./.env) stopped covering the project after a cd (they now also apply in added directories).
  • The sandbox now refuses writes to .claude settings, skills, agents, commands and hooks in added directories, as it already did in the launch folder, and fully protects a worktree’s settings, hooks and shared git folder.
  • Artifact tool: some subagents could change or upload to an artifact that people outside your organization can see without asking; some publishes skipped the confirmation for such pages after “don’t ask again”; auto mode’s approval now covers one change only.
  • The account-email check now also refuses your email as a client identity in scripts a command runs, file writes, web fetches and MCP calls.

Sign-in and billing

  • “Fixed background sessions started on your claude.ai login using (and billing) an API key or token from the shell that started the background service.”
  • More cases of a session moving to another sign-in in the background are fixed; such moves “are now refused, with a message saying what works”.
  • Token counting on Amazon Bedrock no longer sends the conversation to Bedrock when allowedProviders does not allow it.

Models, subagents and context

  • Subagents with their own model are no longer moved to the plan-mode model under opusplan or haiku, and no longer size file reads and MCP results with the main session’s model.
  • Subagents cut off at the output token limit are asked for the whole answer again, more briefly, instead of returning only its end.
  • Sessions on many third-party gateways and Foundry no longer stop with a context length error every turn instead of compacting. CLAUDE_CODE_MAX_CONTEXT_TOKENS now also sets the window for unrecognized claude- model IDs such as a gateway’s claude-sonnet-latest.
  • A saved sonnet[1m] setting is no longer rewritten on a new install to a model ID the API rejects.
  • /cost, $.session.usage(), --max-budget-usd and the cost metrics now count a mod’s $.model.complete and $.model.classify calls.

Other changes

  • Code Mode fixes: scripts now respect a worktree entered with EnterWorktree, words typed with a Yes at a permission dialog reach Claude, and resuming a session that ended mid-script waits for your prompt.
  • claude remote-control exits with status 4 when it stops by itself and 69 when it can’t register, so a service manager can restart it.
  • Recent turns no longer go missing from a resumed session after an unclean exit; an approved plan edit that plansDirectory can’t take is saved to the default plans folder with a warning.
  • Plugins: MCP bundles (.mcpb, .dxt) unpack under the plugins directory; plugin loading skips some unsafe paths and links; .claude/skills in --plugin-dir plugins no longer load from a session started above them; pluginDigests is deprecated and ignored.
  • VS Code: prompt suggestions (Tab accepts; promptSuggestionEnabled: false turns them off), a Console sign-in that creates no API key, and an effort selector that respects the organization cap.
  • Claude Code on the web: MCP tool descriptions loaded through tool search are cut at 16,384 characters instead of 4,096.

Breaking and behaviour changes

ChangeVersionWhat to do
Saved Fable 5.1 picks move to fable (Fable 5.2)v2.1.321/model claude-fable-5-1 to stay on 5.1; on Team and Enterprise the move waits for your organization
Mods’ own tool calls raise no permission prompt, even under your ask rulesv2.1.321Use deny rules or organization policy for anything a mod must not do
Mods Claude writes turn on without asking in auto and bypass modesv2.1.321Use dontAsk mode if you do not want them loaded
.claude/skills inside ~/.claude/plugins no longer loadv2.1.321Move any skill you relied on there
policy_mode: layered adds managed MCP servers across all matching policiesv2.1.321Review what a layered gateway serves after the upgrade
Gateway Bedrock, Vertex and Foundry upstreams ignore ANTHROPIC_*_BASE_URLv2.1.321Set base_url on the upstream
Remote Control sessions keep their start folderv2.1.321Start a new session for another folder
Account-email check refuses instead of askingv2.1.321Applies only to a shell command holding your email and naming a User-Agent
Self-hosted on-demand runner exits 0 when refused as not neededv2.1.321Update scripts that treated exit 1 as that case
Self-hosted on-demand runner exits 1 again when its work order was supersededv2.1.323The note says 2.1.321 and 2.1.322 had changed it to 0
Unknown model IDs are assumed to have a 1M context window (was 200K)v2.1.323/autocompact to compact earlier; CLAUDE_CODE_DISABLE_1M_CONTEXT=1 to keep 200K, and set it for other vendors’ smaller-window models
subagentStatusLine setting removedv2.1.323Nothing; settings files that still hold the key load as before
A plugin’s agent setting is deprecatedv2.1.323It still applies; expect warnings at startup, in claude plugin list and claude plugin validate
MCP_SDK_GENERATION=v1 runs the v2 client in compatibility modev2.1.323The v1 MCP client is removed
engine.create wildcard, namespace or negated hooks only observev2.1.323Hook engine.create by name to add or withhold a $ noun
Sandbox: git worktree add, move, remove and commits in other worktrees fail read-onlyv2.1.323Run them outside the sandbox, or start the session in that worktree
Sandbox ignores write access granted through a symlink that leads elsewherev2.1.323Grant the real path
Organization settings and policy requests ignore a proxy set only by ALL_PROXYv2.1.323Set HTTPS_PROXY
A managed-settings policyHelper no longer takes environment variables from user or --settings filesv2.1.323Set what it needs in managed env or the launching shell
Start-up stops when a --settings file isn’t valid JSON and has a sandbox keyv2.1.323Fix the file
Marketplace and plugin installs fail when a git submodule isn’t on its recorded commitv2.1.323Update the submodule pointer in the plugin’s repository
Worktree creation refuses checkouts whose .git folders are links or hold hard linksv2.1.323Use a plain clone
/security-review asks before Claude runs itv2.1.323Approve it when asked (outside auto mode and bypass)
Sandboxed Bash follows the normal permission flowv2.1.325Set sandbox.autoAllowBashIfSandboxed: true (in the highest-ranking managed source if managed settings turn the sandbox on) to keep the old behaviour
Scheduled tasks and /loop ask before a skill with allowed-tools or hooksv2.1.325Add Skill(name) to permissions.allow for skills that must run unattended
Read-only git in an untrusted repository asks firstv2.1.325Add the repository with --add-dir
”Always allow” in Desktop, IDE and SDK hosts may offer no saved rulev2.1.325Check scheduled tasks that relied on one
.claude/skills in --plugin-dir plugins no longer load from a parent folderv2.1.325Start Claude Code in the plugin folder
claude remote-control exit codes 4 and 69v2.1.325Update service-manager scripts

Try It

  1. Update to v2.1.321 or later for the MCP OAuth redirect fix and the three Bash sandbox and permission fixes.
  2. Run /model after updating and check which Fable you are on. If a workflow was tuned for Fable 5.1, pin it with /model claude-fable-5-1 while you compare.
  3. If you rely on ask rules to gate what mods do, add deny rules: a mod’s $.tool.call and $.mcp.call no longer prompt.
  4. Update to v2.1.323 for the further permission and sandbox fixes, then reopen any background session that was open during the update (the release note asks for this).
  5. After updating, review Bash allow rules: start-up now warns about a * in a variable’s value, and rules such as Bash(ls*) or ones that name a variable behave differently.
  6. If you run Claude Code against a non-Anthropic model with a context window under 1M, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1.
  7. After updating to v2.1.325, list the skills your scheduled tasks and /loop jobs call. Any that declare allowed-tools or hooks now stop and ask; add Skill(name) to permissions.allow for each one that must run unattended, and set sandbox.autoAllowBashIfSandboxed: true if you relied on sandboxed Bash running without prompts.
  8. If you run a Claude apps gateway on policy_mode: layered, or with ANTHROPIC_*_BASE_URL set in its environment, check the served MCP servers and upstream endpoints after the upgrade.

Open Questions

  • The security fixes (MCP OAuth redirect, sandbox bypass, wrong-folder permission check) carry no CVE or advisory link in the notes. The same holds for the v2.1.323 permission and sandbox fixes and the v2.1.325 zstd fix.
  • Code Mode: the release note links code.claude.com/docs/en/code-mode, which returned 404 on 2026-11-13. How to turn it on, what it costs in tokens and which tools it covers are not documented in the release note.