Plugin changes made in the /plugin menu now apply when you close it — Claude Code runs /reload-plugins for you, so installing, enabling, disabling and uninstalling no longer leave the session stale. This rewording lands across a dozen pages
Task-tracking tool availability is now stated as an allowlist rather than a denylist: TaskCreate/TaskGet/TaskUpdate/TaskList/TodoWrite ship by default only on Claude 3.x, Opus 4–4.7, Sonnet 4–4.6 and Haiku 4.5. Every other model, including a model ID Claude Code doesn't recognize, needs an opt-in
WorktreeRemove hooks gain decision control for the first time: a non-zero exit now fails worktree removal, and a refused session delete can be forced through with claude rm <id> --force-remove-worktree <worktree-id>
The Claude apps gateway can now push its contracted pricing rates to signed-in clients as the modelPricing managed setting, and a managed policy can send session telemetry straight to your own collector instead of through the gateway relay
Auto mode denials now name the classifier rule that matched, in square brackets ([Data Exfiltration], [Production Deploy]), in place of the fixed Blocked by classifier text
Added pricing: delivery to signed-in clients through managed settings, so /cost and telemetry match the Claude apps gateway spend meter
Added the gatewayInternalNetworks managed setting, letting administrators allow /login to a gateway on their organization's own public IPv4 block
Added a gateway startup warning when access_control.allow_cidrs is empty, plus a one-time warning on the first request from a public address
Added claude self-hosted-runner --remove-session-state to delete per-session directories under <base-dir>/_sessions/ when a session ends
Added --json to claude plugin install, uninstall, update, enable and disable, and errorDetails/noteDetails to claude plugin list --json rows
Added configDirectory to claude auth status --json
Added browser-tab icons for published artifacts, chosen by Claude to match each page
Existing feature improvements
Improved /plugin: installing, enabling or disabling a plugin now takes effect when you close the menu, so /reload-plugins is no longer needed afterwards
Improved --continue / --resume: the conversation appears immediately instead of waiting for SessionStart hooks, and the first message no longer re-reads the whole transcript
Improved fullscreen mode so adding or removing a prompt line repaints as fast as typing a character
Improved auto mode denials: the message Claude receives names the blocking rule and asks it to try a safer method and finish unrelated work before stopping to ask
Improved startup time in projects with .claude/workflows/ scripts by no longer parsing each script to list them
Changed the task-tracking tools to be offered only on Claude 3.x, Opus 4.0–4.7, Sonnet 4.0–4.6 and Haiku 4.5; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 elsewhere
Changed Bedrock, Vertex and Foundry sessions to deliver environment and settings details as attachments and to keep the tool list byte-stable, matching first-party sessions
Changed plain WebFetch deny and ask rules to stop applying to Artifact tool reads and updates; use an Artifact rule instead
Major bug fixes
Fixed every turn failing with HTTP 400 on third-party Anthropic-compatible endpoints since 2.1.265, caused by a regex in the Artifact tool's input schema
Fixed WebFetch hanging indefinitely on a server that never finishes the response; fetches now fail after 300 seconds, tunable with CLAUDE_CODE_WEBFETCH_DEADLINE_MS
Fixed sustained high CPU usage from a busy loop in long-running idle sessions and from rapid terminal focus reports during a session recap
Fixed deny and ask permission rules not applying on symlinked directories (/etc, /tmp, /var on macOS; /bin on Linux) when a path was given by its real location
Fixed a Read or Edit deny rule not applying when an env -C, eval or similar unanalyzable command shared the line
Fixed plugin, marketplace, /mcp, /plugin and claude mcp output leaking tokens, passwords and ${VAR}-resolved secrets
Fixed prompt caching and extended thinking breaking mid-session for SDK sessions using excludeDynamicSections
Fixed a running session silently switching to the organization's default model when another Claude Code process refreshed a stale model-access entry
Fixed workload identity federation via a profile failing mid-run with 401 … jti reused
Fixed /compact mangling text containing $ sequences, and fixed SDK prompt suggestions and /rename sending the pre-compaction conversation
Fixed a respawned in-process teammate picking up tools or a system prompt from a same-named agent file in an untrusted folder
Fixed PermissionRequest hooks not firing in --print mode, and policy-helper warnings not printing on headless runs
Dropped the separate Fable 5 minimum version; both Fable models are now described as simply requiring Fable access, with only Fable 5.1's v2.1.257 floor called out. [line 91] [Source]
System-prompt recording now applies on Bedrock, Google Cloud's Agent Platform and Foundry too. The old carve-out for sessions that don't fetch feature flags is now written as history: before v2.1.268 those sessions rebuilt the prompt every request and snapshot had no effect. [line 346] [Source]
Task-tool availability restated as an allowlist — Claude 3.x, Opus 4–4.7, Sonnet 4–4.6, Haiku 4.5 — with every other model, including unrecognized model IDs, requiring an opt-in. Applies from Claude Code v2.1.268 (TypeScript SDK v0.3.268). [line 2820] [Source]
Same allowlist rewrite, with the page's opening paragraph inverted to lead with where the tools are provided rather than where they're withheld. [lines 5-23] [Source]
pending_permission_requests is now always present on a successful initialize response and empty when nothing is pending, rather than optional. Earlier versions could omit it, so hand-rolled wire-protocol parsers should read a missing field as an older CLI, not as proof nothing is pending. Requires v2.1.268. [lines 659-663] [Source]
The Artifact tool's favicon is now described as marking the artifact in the user's gallery rather than as the browser-tab icon. [line 3103] [Source]
Task-tool availability rewritten to the same allowlist form in both places the page states it. [line 2786] [Source]
A named Agent call no longer launches a teammate when the call is a fork or passes isolation itself — the two exceptions to teams forming automatically. [lines 213-215] [Source]
Teammates inherit the lead's permission mode except dontAsk mode, which is now explicitly not inherited. [line 269] [Source]
Teammate subagent scopes narrowed from "project, user, plugin, or CLI-defined" to "project, user, or managed". [line 251] [Source]
New guidance for a session delete refused because git or a WorktreeRemove hook couldn't remove the worktree: the message now names the cause, reports how the hook ended (exited 1) and quotes the start of its stderr. Pressing Ctrl+X twice, or passing --force-remove-worktree, deletes the directory anyway — but only when Claude Code can confirm it's a linked worktree under .claude/worktrees/ with no uncommitted tracked changes, no nested repository and no other session claiming it. [lines 542-546] [Source]
claude rm <id> --force-remove-worktree <worktree-id> added to the shell command table, requiring v2.1.268. [line 691] [Source]
Version history gained a v2.1.268 row describing what the refusal used to show: only worktree could not be removed (WorktreeRemove hook failed) or git's error, with the hook's stderr going to the debug log alone. [line 928] [Source]
Dropped the v2.1.171-and-earlier caveats around AWS_REGION resolution; reading the AWS config files is now simply how it works. [lines 234-245] [Source]
Claude now picks a browser-tab icon matching what the page is, such as a chart or a calendar, separately from the artifact's title and emoji — and you can ask for a specific one. [line 53] [Source]
The "no backend" limits were trimmed: the claim that an artifact can't store form input and that MCP connectors are its only route to outside data has been dropped from both the intro and the constraints table. [lines 26-277] [Source]
Denial reasons now name the matched rule in square brackets — [Data Exfiltration], [Production Deploy] — replacing the description of a fixed Blocked by classifier string produced by an internal severity scale. [lines 366-376] [Source]
Getting a failed channel server's stderr trace now requires restarting with claude --debug --dangerously-load-development-channels server:webhook, not just reading the debug log. [line 161] [Source]
Both install walkthroughs now point at "Apply plugin changes without restarting" instead of telling you to run /reload-plugins. [lines 42-118] [Source]
New section "Messages sent mid-turn not checkpointed": a message you queue while Claude works joins the running turn, so it gets no checkpoint and never appears in the rewind menu. To undo it, rewind to the prompt that started the turn — which discards the work Claude did before your message arrived too. [lines 85-89] [Source]
New "Send the rates to signed-in clients" section: with v2.1.268 on the gateway server, pricing is injected into served managed policies as the modelPricing setting, so matched developers see contracted rates in /usage, the status line and OpenTelemetry. A policy opts out with modelPricing: {}, and one that sets its own modelPricing is left whole. [lines 467-473] [Source]
New "Export directly to your collector" section: setting OTEL_EXPORTER_OTLP_ENDPOINT in a managed policy's env block sends telemetry straight to your collector, bypassing the relay, subject to five checks — the endpoint must come from the gateway itself, use https:// (or loopback http://), resolve to a /v1/<signal> path, not be the gateway's own host, and no otelHeadersHelper may be configured anywhere. Requires v2.1.265 on each client. [lines 772-796] [Source]
New "When a destination fails" section: the gateway never buffers or retries telemetry, pauses a destination in 30-second stretches after five consecutive failures, and treats 400/413/415/422/431 as refused payloads that neither advance nor reset the failure count. [lines 797-803] [Source]
With forward_user_identity: true, a 429 answering a request that carried the developer's email no longer fails over — it's a per-user denial, so your proxy's per-user budget holds. Before v2.1.267 every 429 failed over. [lines 121-207] [Source]
pricing now also works with a managed: block, not just admin:, and override rates gained an upper bound of 10000. [lines 438-455] [Source]
Desktop and Cowork sessions now stamp telemetry with user.email and user.groups; an over-long or comma-bearing group list is omitted rather than truncated. Requires v2.1.265 on the gateway and Claude Desktop 1.24012. [lines 711-717] [Source]
The three pushed OTEL_*_EXPORTER selectors are now set per-signal to otlp or none based on whether a forward_to destination enables that signal; before v2.1.265 all three were pushed as otlp. [lines 751-755] [Source]
The telemetry-destination entry was broken into sub-bullets and now allows a policy to name your own collector as the endpoint. [lines 378-381] [Source]
The network-access bullet gained a default-allowed-domains anchor and now links out to the access levels, default allowed domains, and the traffic that bypasses the allowlist. [line 287] [Source]
A .claude/rules/ symlink pointing outside the working directory is now treated as an external import: the rules don't load until you approve external imports, and then only those without a paths field. Because that approval is only ever requested for an @path import, shared rules are better kept in ~/.claude/rules/. [lines 240-242] [Source]
claude rm rewritten around the two flags a refusal can print: --discard-unpushed <commit>@<worktree-id> (v2.1.260) for unpushed commits, and the new --force-remove-worktree <worktree-id> (v2.1.268) for a worktree git or the hook couldn't remove. [line 40] [Source]
--system-prompt-snapshot now takes effect on Bedrock, Agent Platform and Foundry; the feature-flag carve-out is recast as pre-v2.1.268 history. [line 158] [Source]
Cloud environment management documented as a dropdown in the prompt box, with Add cloud environment to create one and a hover gear icon to edit or archive your own. [lines 630-635] [Source]
The comparison table's /loop persistence cell changed from "Restored on --resume if unexpired" to "Restored on --resume, with exceptions", linking to the new limitations. [line 18] [Source]
Closing the /plugin menu now triggers /reload-plugins automatically for installs, enables, disables and uninstalls; if the reload would invalidate the prompt cache it warns and leaves the changes pending for /reload-plugins --force. If Claude is still responding, the reload waits. [line 386] [Source]
New list of the plugin changes that still need a manual /reload-plugins: a claude plugin command run in another terminal, edits to a --plugin-dir plugin, an auto-update that asks for a reload, and a held --plugin-dir change. [lines 388-393] [Source]
Records that before v2.1.268, menu changes and non-activating installs stayed pending until you ran /reload-plugins yourself. [line 395] [Source]
Added CLAUDE_CODE_WEBFETCH_DEADLINE_MS, bounding how long WebFetch waits for a page including redirects. Defaults to 300000; 0 removes the limit, and any non-integer spelling silently keeps the default. Requires v2.1.268. [line 381] [Source]
CLAUDE_CODE_ENABLE_TODO_TOOLS reworded to "get the task-tracking tools on every model", matching the allowlist framing. [line 268] [Source]
New macOS guidance for EPERM in ~/Desktop, ~/Documents, ~/Downloads or iCloud Drive: it's the terminal app being blocked by privacy settings, confirmed by ls failing with Operation not permitted even under sudo. Fix by quitting the terminal with Cmd+Q and reopening, or granting the folder in System Settings > Privacy & Security > Files and Folders. [lines 2171-2177] [Source]
The GitHub-connect URL for ultrareview moved from claude.ai/code/onboarding?step=alt-auth to claude.ai/connect-github. [lines 2531-2538] [Source]
WorktreeRemove gains decision control: any non-zero exit code now fails worktree removal if the directory still exists afterward. The worktree and the background session both stay, and the refusal reports how the hook ended and quotes its stderr. [lines 2954-2957] [Source]
New "MCP tool hook fields" guidance on when an mcp_tool hook actually runs: SessionStart at launch fires before MCP servers are available and its mcp_tool hooks are skipped (logged as no MCP client context), Setup skips them always, and only a SessionStart re-fired by /clear or compaction runs them. Use a type: "command" hook for anything needed from the first turn. [lines 558-584] [Source]
stopReason now stays in the conversation, so Claude sees it if the conversation continues — previously documented as not shown to Claude. [line 922] [Source]
PermissionDenied's reason field rewritten for the bracketed rule names, and the SubagentStop example reason changed to [Irreversible Local Destruction]. [lines 2151-2157] [Source]
AskUserQuestion and ExitPlanMode are now offered in -p runs only when the run has a permission host, such as an Agent SDK canUseTool callback or a --permission-prompt-tool MCP tool. [lines 1768-1783] [Source]
CLAUDE_ENV_FILE variables set by CwdChanged and FileChanged hooks now persist only until the next CwdChanged event, when Claude Code clears them. [lines 2695-2815] [Source]
Write's structured output field changed from success: true to type: "create" in both the PostToolUse example and the PostToolBatch comparison. [lines 1939-2107] [Source]
permission_suggestions clarified: some dialogs, such as the one for file edits, don't read the array at all and derive options from the request itself. [line 1827] [Source]
Plugin changes made in the VS Code dialog now apply immediately to open sessions in that window instead of prompting for a restart; if the originating session can't reload, the dialog offers a retry or a restart. [line 241] [Source]
The task list section rewritten around the allowlist: the list stays empty on any model outside it, including unrecognized model IDs, unless you opt in. [line 594] [Source]
Two more default ctrl+x chords documented — ctrl+x ctrl+a and ctrl+x tab in Chat, both requiring v2.1.260 — and added to the unbind example. [lines 471-495] [Source]
Where to put a gateway's own token now depends on the provider, and the page warns that without the skip-auth variable Claude Code strips any Authorization header that ANTHROPIC_AUTH_TOKEN, apiKeyHelper or ANTHROPIC_CUSTOM_HEADERS would add. [lines 379-386] [Source]
New note that a set AWS_BEARER_TOKEN_BEDROCK overrides your gateway token even with CLAUDE_CODE_SKIP_BEDROCK_AUTH set. [line 386] [Source]
New Agent Platform guidance on per-model regions, pinned model versions, and declaring capabilities for a model ID your Claude Code version doesn't recognize. [lines 428-432] [Source]
Notes that on the Bedrock and Agent Platform routes Claude Code now limits beta headers and request fields to what the provider accepts. [line 375] [Source]
A failed-connection message now shows the host as written in configuration, so a ${VAR} in the host is no longer expanded in the error — the reverse of the previous behavior. [line 270] [Source]
New constraints on OAuth client secrets: settable only when adding the server, never re-prompted or read from MCP_CLIENT_SECRET at login, and changeable only by removing and re-adding the server. [lines 820-821] [Source]
Plugin MCP servers now connect or disconnect when the plugin change applies, rather than requiring /reload-plugins. [line 468] [Source]
user.email and the other identity attributes can now also come from a cloud session's own credentials, not just your sign-in. [lines 458-1225] [Source]
The retry attempt attribute now notes that some causes carry their own smaller retry budget — AWS or Google Cloud credential loading is retried at most twice. [lines 1200-1202] [Source]
New --json option documented on plugin install, uninstall, update, enable and disable, all requiring v2.1.268. --json can't be combined with --prune on uninstall. [lines 993-1128] [Source]
New "JSON result format" section: parse only the last line of stdout, since a marketplace-declared command may print ahead of it. Always carries command, outcome (ok/failed) and message; a usage error prints no result line and exits 1. [lines 998-1004] [Source]
plugin list --json rows gained parallel errorDetails and noteDetails arrays giving each diagnostic's type and the names it refers to. [line 1147] [Source]
New "Resuming a session" section explaining that a resumed conversation keeps the system prompt it started with by default, so an upgrade or different --append-system-prompt text doesn't invalidate the cache until compaction or a new conversation. [lines 229-233] [Source]
"Changing output style" simplified: the feature-flag split is gone, and switching styles mid-session now always arrives as a message in the conversation, keeping the cache. [lines 209-213] [Source]
The layer table no longer lists a Claude Code upgrade as something that changes the system prompt layer. [lines 21-25] [Source]
Compaction now switches a resumed session to the current system prompt, rebuilding that layer once. [line 160] [Source]
Mid-conversation system context is now sent uncached when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS is set. [line 47] [Source]
Cache scoping reworded: the working directory, platform, shell, OS version, branch and recent commits are now carried by the conversation rather than embedded in the system prompt. [lines 285-287] [Source]
The /rc active indicator now sits at the end of the working-directory line in the startup header under fullscreen rendering, and in the footer otherwise. A failed connection also adds a warning line with the reason to the conversation. [lines 121-125] [Source]
A self-paced /loop is now explicitly not restored on resume — run /loop again to restart it. Only CronCreate-scheduled tasks come back. [line 202] [Source]
Approving sandbox.credentials or sandbox.network.tlsTerminate now also covers the sandbox.network.allowedDomains entries delivered alongside them, and the dialog reappears when an administrator changes one of those entries. [line 251] [Source]
New copyOnSelect global config setting (default true) controlling automatic clipboard copy on mouse selection in fullscreen rendering and agent view. Ignored in settings.json — it belongs in ~/.claude.json. [lines 5808-5824] [Source]
A failed startup helper now reports its stdout when stderr is empty, and a timeout names the timeoutMs limit while including none of the helper's output. [line 5684] [Source]
context: fork given a prominent warning that despite its name it does not fork the current conversation: the subagent starts with no history, so the skill's instructions must stand alone. Fork the conversation instead when the task depends on history. [lines 652-654] [Source]
Command files in a subdirectory of .claude/commands/ documented: the path becomes a :-separated prefix, so .claude/commands/frontend/component.md is invoked as /frontend:component. [line 362] [Source]
disableBundledSkills no longer described as sparing /doctor — it now turns off bundled skills outright. [lines 23-244] [Source]
A missing description now falls back to the first non-empty line of markdown content rather than the first paragraph. [line 312] [Source]
Several synced-skill protections now carry a v2.1.228 minimum: name-collision checks, display-text sanitization, literal handling of ! and @ in synced bodies, and the refusal to run !`<command>` from a synced skill. [lines 211-612] [Source]
Carries the same set of skill changes: the context: fork warning, the /frontend:component subdirectory naming rule, the disableBundledSkills rewording, the description fallback, and the v2.1.228 floors on synced-skill handling. [lines 23-654] [Source]
Important clarification: a disallowedTools entry with a specifier, such as Bash(git push *), removes the whole tool from the subagent rather than only the matching commands. To keep Bash and block specific commands, use a permissions.deny Bash rule instead. [lines 277-442] [Source]
Theme tokens retargeted: promptBorder now covers the input box border in every mode rather than Manual only, autoAccept and planMode no longer own their borders, and planMode extends to plan messages and dialogs. [lines 205-216] [Source]
"Task tool availability" rewritten as the allowlist, explicitly covering custom model names served through an LLM gateway. Applies from v2.1.268. [lines 505-520] [Source]
Auto-backgrounding simplified: only commands starting with sleep are now excluded. The former carve-outs for any command running git and for compound commands Claude Code can't fully parse are gone. [line 175] [Source]
WebFetch's five-minute download deadline is now tunable with CLAUDE_CODE_WEBFETCH_DEADLINE_MS on v2.1.268. [line 533] [Source]
New section on copied text not reaching your clipboard over SSH: outside tmux Claude Code falls back to an OSC 52 escape sequence, /copy reports success whether or not it arrived, and iTerm2 ignores it until you enable clipboard access while Terminal.app doesn't support it at all. Work around it with your terminal's native-selection key (Fn in Terminal.app, Option in iTerm2) or CLAUDE_CODE_DISABLE_MOUSE=1. [lines 91-101] [Source]
Same two changes as the IDE integrations page: sessions open in your preferred location, and plugin dialog changes apply without a restart. [lines 47-241] [Source]
The thinking.type.disabled 400 error now has two documented message forms. Claude Mythos Preview — the only always-thinking model that still accepts extended thinking — returns a different message from Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5, which are told to use thinking.type.adaptive with output_config.effort. [lines 472-484] [Source]
Both error message forms spelled out here too, with the explanation now pointing at the per-model configuration table instead of repeating which models reject "thinking.type.enabled". [lines 59-71] [Source]
The endpoint setup flow was reordered so the signing secret now exists before the first connection test. The dialog (renamed Set up endpoint) asks only for the URL, then reveals the secret once, then reopens with custom headers and Test connection. [lines 36-53] [Source]
New "Signing secret required" failure result, telling you to click Generate secret under Request signing before testing. [line 65] [Source]
The Rotate secret button reads Generate secret when the organization has no secret yet. [line 134] [Source]
Connection tests are now signed, since the setup flow generates the secret first. The unsigned-request exception narrowed to organizations that enabled Inference hooks before secrets were required. [line 291] [Source]