You can now run gh api REST calls in sessions that do not have the GitHub CLI installed. Claude Code supplies its own limited gh and sends those calls through the session's GitHub proxy. A new command, claude daemon remote-control stop-shell, stops a background shell command running for a cloud session. Programs driving Claude Code can now read a background task's output with a get_task_output request. Compression of requests sent to the model is available through CLAUDE_CODE_GZIP_REQUEST_BODY_BLOCKS and is off by default. Admins can now set CLAUDE_CODE_TRANSCRIPT_LOCAL_GC through managed policy to turn on local transcript clean-up.
This build holds 7 things that are not switched on yet. Claude Code can offer to move you to a newer model, as a one-line notice or a dialog. A reworked way of saving your shell's functions, aliases and options for Bash commands is also in the code. Both stay off unless Anthropic turns them on remotely. Groundwork has landed for running the SessionStart hooks of plugins that load mid-session. None of this changes what you see today.
Remote sessions no longer forget their background tasks when they reconnect. A stalled live-update connection now times out and reconnects instead of hanging forever. Relaunching after /update now resumes your conversation even when Claude's last reply could not be saved. Messages from Amazon Bedrock guardrails now show up in the reply instead of going missing. A shell command that only names a web address is no longer escalated as if it touched a protected file path. /plugin-types is no longer listed among the command categories.
Written by our agent from the shipped bundle, not by Anthropic.
The /plugin help screen can show /plugin directory [query] for searching the Anthropic Directory in Discover
Claude Code now reads four new startup environment variables and passes them through to cloud sessions; they do nothing unless set
Sessions & agentsA new environment variable, CLAUDE_CODE_MCP_SERVE_TOOL_OUTPUT, appears alongside the other MCP server-mode variables
Claude Code can compress the requests it sends to the model in gzip blocks, controlled by CLAUDE_CODE_GZIP_REQUEST_BODY_BLOCKS and off by default
plugin install --replace explains that it needs a full plugin idRunning plugin install --replace without a name@marketplace id now says that id is needed and that nothing was installed
gh api stand-in for sessions without the GitHub CLITools & outputWant the reasoning? Read walks the 64 entries that probably matter to you, each one opening to what changed and why.
Read this release → Every row →7 more of these are in What probably matters to you, on page 1.
Before judging remote tool calls, Claude Code now also waits a limited time for its first download of feature settings
Unclear How long the time limit for the flag download is was not stated.
When a remote host's reply cannot be verified, Claude is now told to do the rest of the task without that host and say what it could not check
Creating a cloud session now sends a no-monorepo option, takes effort from your choice alone, and rechecks the device link for served tools
Unclear It is not clear how the no-monorepo option is set.
When a turn is handed off to a cloud session, Claude Code can place files listed by name, behind a switch that is off by default
Unclear Which path prefix the file check applies to, and what it means, was not stated.
A control request from a cloud worker vouched for by the service can now be accepted even when its type is not on the allowed list
Unclear It is not clear what the extra check on the request's contents looks for.
Claude Code now records timings for more cloud session setup steps, such as added folders, folder rules, folder skills and restoring folder memory
Unclear Whether the features behind these steps are already in use is not known.
MCP tool calls and resumed tasks now hold a guard on connection error tracking for as long as each call runs
Unclear It is not clear what the guard does for the user or how the tool call path is switched.
Claude Code now runs a tracking step around each MCP tool call and cleans up the task watcher when it finishes
Unclear It is not clear what the tracking step around each call does.
Claude Code now tracks which MCP servers already have a reconnect retry running and logs failed retries
Unclear The visible effect of this change on reconnecting MCP servers is not clear.
Claude Code no longer cleans up dropped MCP connections at the end of device bridge discovery, and handles them when connections finish instead
Unclear The practical effect for users is not known.
Several models in Claude Code's built-in model table have a higher advisor rank, which changes which model is preferred as advisor
Unclear It is not clear which models these entries are, or exactly how the rank is used.
Under one condition, Claude Code now skips its usual model name lookup and returns a fixed name
Unclear The condition and the name it returns could not be identified.
The model picker no longer tells the model-switch check where the turn came from or which MCP sources are involved
Unclear It is not clear whether a confirmation step was removed or only changed.
One model's advisor ranking rose from 5 to 8, and a check that leaves out claude-opus-4-1 gained an extra condition
Unclear Which model the ranking belongs to, and what the new condition tests, are not known.
Claude Code's background service has a new entry for stopping a shell through remote control
Unclear It is not clear how either addition is used or whether users can reach them.
Remote sessions now keep a count of answered replayed refusals and a record of conversation sizes
Unclear It is not clear how either piece of information is used.
Rebuilding the remote control connection now runs an extra reset step, and reloading plugins runs an extra step first
Unclear What the added steps actually do is not known.
Above a set number of attachments, Claude Code now cleans up each downloaded attachment on its own instead of all of them at the end
Unclear What the attachment count limit is, and what it is meant to protect against, is not known.
When an object schema says nothing about extra fields, Claude Code now marks it as allowing them
Unclear Which tools' schemas go through this step is not known.
Tool names are now matched by rank (exact, then partial, then intent), and an account attribute list has a new entry
Unclear It is not clear what the account attribute list is used for.
When cleaning tool input schemas, Claude Code can now drop some anyOf or oneOf alternatives that carry a description
Unclear It is not clear which descriptions trigger the removal, or when this cleanup runs.
A tool's heartbeat progress update now wakes every part of Claude Code waiting for progress, not just one
If a tool's input was rewritten, Claude Code no longer uses the server's classifier verdict for it
Unclear It is not clear what rewrites a tool's input.
The auto mode checker's result now carries a category for why it blocked an action
Unclear It is not clear what uses the block category or whether it is ever shown to users.
When auto mode does not approve an action by itself, it now records a reason, without changing what gets approved
Unclear Where these reasons are shown or logged is not traced.
When choosing a browser fails because the Chrome bridge is unreachable, telemetry now records that cause instead of a general failure
Claude Code's Chrome status now reports whether Claude in Chrome is enabled by default at startup
Unclear What the removed plugin types entry did is not known.
When no browser is connected to Claude in Chrome, Claude Code raises a dedicated error telling you to open Chrome and sign in
Unclear It is not clear whether this replaces an older message in some situations.
A compaction size supplied with client data is now used only when a certain check passes, and otherwise the default applies
Unclear It is not known what the check tests.
Claude Code now waits until a whole batch is saved before compacting the session file, instead of right after the compaction marker
The two messages shown when compaction fails are now assembled from stored text instead of written in place
Unclear The new wording of the two messages is not known.
Claude Code can now apply or discard a stored proposal to change settings, first checking it is still current, writable and untampered
Unclear What creates these proposals and whether anything controls when this is used is not shown.
Claude Code can now tell when a project, local or legacy settings file is a link to another place rather than an ordinary file
Unclear It is not clear what uses the list of linked settings files or what happens when one is found.
When checking whether settings changed, Claude Code now compares normalized settings and which levels differ, instead of raw file contents
Unclear The exact rules for which project-level settings are skipped are not clear.
Reading an artifact now goes through the relay-only network route under one more condition than before
Unclear It is not clear what the new condition is or when it applies.
When an artifact thread is handled in the background, the model is now told to leave pending tasks, agents and next steps to the main session
The description of the event delivery tool can now use a custom warning about untrusted content and extra text, with the old sentence as default
Unclear It is not clear where, if anywhere, the custom wording is used.
Event messages delivered to a session can now end with a second, JSON-mode closing tag that Claude Code accepts and checks
Unclear What switches event delivery into the JSON mode is not shown.
When a local size estimate is far over a limit, Claude Code now answers its token-count check without asking the API
Unclear It is not clear what Claude Code uses this token-count check for.
Claude Code now processes its snapshot of the environment before checking whether the working folder changed
Unclear What the new processing step does to the snapshot is not clear.
When Claude Code cannot tell whether your git commits have been pushed, it now records a reason, and the check can take a specific commit
Unclear It is not clear whether these reasons are ever shown to the user.
Claude Code's check for git changes now reads git's HEAD history file when HEAD does not point at a regular branch
Unclear It is not clear what Claude Code uses this git state summary for.
When a monitor ends, its notification can now carry a different status and summary instead of always saying stopped
Unclear It is not clear what makes a monitor report a status other than stopped.
The messages shown when a monitor ends, fails or is stopped now come from one shared place, and their wording may differ
Unclear Whether the notification wording itself is different from the previous build is not known.
Example notes in the learn, tag and explain side-agent prompt now say "The main agent" instead of "I"
Unclear It is not clear which feature uses this prompt or whether it is switched on by default.
An instruction to use “I” when describing the main agent's decisions has been removed from a summary prompt
Unclear It is not clear which feature uses this summary prompt.
/plugin-types is no longer listed in the table that sorts Claude Code's commands into categories
Unclear It is not clear whether the /plugin-types command itself still exists.
Claude Code's list of git in-progress markers no longer includes REBASE_HEAD, so this check may not notice a rebase
Unclear What Claude Code uses this list for is not known.
Published verbatim by Anthropic for v2.1.287. Text is unmodified from the upstream changelog. Everything else on this page came out of the bundle instead, which is why the two lists don't match.
Of these 106 bullets, 14 name something an entry on this page also names, 40 name something no entry here does, and 52 name nothing specific enough to line up either way. The pairings are made on names both sides wrote down, a flag or a setting or a slash command, so read one as probably the same thing rather than as a fact, and read the middle number as candidates rather than as a miss count.
/plugin enable cc-plugin-you-should-know@builtin (for first-party sessions with telemetry on)
Probably plugin-types-removed-from-the-command-category-map, cloud-session-sync-and-memory-update-tweaks, plugin-background-update-failure-now-reported-for-anthropic, disabled-marketplace-notice, plugin-cloud-session-hooks-consent-display-and-project-sett, plugin-type-contract-now-laid-for-dependent-mods, plugin-types-generated-header-comments-reworded-mcp-types, plugin-screen-telemetry-and-marketplace-network-location-mes, plugin-not-found-errors-now-explain-removal-from-the-marke, plugin-marketplace-dedup-across-directories, plugin-types-removed-from-the-open-state-map-skill-folder, plugin-built-in-marketplace-can-be-turned-on-or-off, plugin-help-lists-a-new-plugin-directory-command, public-plugin-directory-search-and-marketplace-catalogue, plugin-install-replace-now-has-an-error-message-requiring, plugin-tip-advertise-enabling-the-you-should-know-side-ag, switch-installed-plugin-copy-confirmation-in-plugin, plugin-marketplace-community-managed-and-pinned-commit-deta n:<text> filter to the agents view that matches session names and tasks; a filter now shows matches in collapsed sections and Enter opens the first match
No entry names this prompt_text to the OpenTelemetry user_prompt event, a copy of prompt for backends that nest dotted keys; drop or mask it wherever you drop or mask prompt (anthropics/claude-code#70763)
Probably user-prompt-telemetry-now-built-by-a-helper-that-spreads-fie gh api (REST only) for sessions that use Anthropic-managed git on macOS and Linux machines where the GitHub CLI is not installed
Probably github-gh-stand-in-described-in-session-proxy-prompt, built-in-gh-api-stand-in-for-sessions-with-no-github-cli asyncRewake waking Claude over and over with "found issues" notifications when the hook's script file is missing; the broken hook is now reported once
Probably asyncrewake-hook-with-an-unopenable-script-reports-once #96087[BUG] Windows desktop (MSIX): plugin hook scripts invisible to externally-spawned interpreter — every hook ENOENTs and asyncRewake Stop hooks loop forever Open
#95727Desktop app: dangling plugin hook (security_reminder_hook.py missing) fires endless 'Background security review found issues' notifications Open
#96148[BUG] Failing asyncRewake Stop hook causes unbounded model re-wake loop (high token burn) Open
#94634[BUG] Security-guidance plugin hook file missing from session snapshot, causes repeated "No such file or directory" notifications Open
#92234[BUG] Broken security_reminder_hook.py plugin persists after reinstall Open
availableModels list, which could collapse /model to one Opus row
Probably managed-settings-denial-message-for-model-checks /model on a claude.ai login saving the current version's id, so your saved default now follows the newest Fable like Opus and Sonnet do
No entry names this /model, opusplan) rewriting earlier MCP tool announcements, which could drop earlier extended thinking
No entry names this rm (such as one on / or the home directory) losing its always-ask safeguard when the same command also redirected output to a ~ or wildcard path
Nothing to match on claude -p and SDK sessions repeating a model fallback on every later message after the model was switched while a reply was running
No entry names this claude agents after the agent exited and removed the worktree the session was started in
No entry names this /advisor pairing checks: Sonnet 5.5 can now advise Opus 4.7 and 4.8, and advisors the API would refuse are flagged up front instead of being silently dropped
No entry names this __proto__
Probably org-mcp-permission-maps-now-validated-and-null-prototype claude --teleport progress screen, and an MCP form field while it is being checked, to the screen reader again on every spinner frame
No entry names this CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS not removing the structured-output format from session-title and prompt-hook requests, which Bedrock-backed gateways reject
No entry names this --include-partial-messages sending a cut-short reply's message_stop late or never, so apps could show the reply as still in progress
No entry names this claude agents sometimes not showing the permission prompt a background session is waiting on
No entry names this /ultrareview giving advice about .git/info/attributes when the upload stops on a committed .gitattributes it cannot read, such as one saved as UTF-16
No entry names this claude remote-control failing to register behind an HTTP proxy with a misleading "Check your organization permissions" error (anthropics/claude-code#97352)
No entry names this #97352Remote Control cannot register behind an HTTP proxy: some requests are sent as absolute-URI instead of through CONNECT Closed
claude agents changing every second in screen reader mode; they now change at most every 10 seconds
No entry names this API Error: 401 instead of "OAuth token revoked"; in -p mode the error now starts with "Failed to authenticate"
No entry names this /ultrareview upload refusals advising you to copy a variable named by a repository's settings file into your own user settings
No entry names this --output-format stream-json and the SDK not streaming the turns of a context: fork skill run by typing /<skill> as the prompt, as they do for the Skill tool's fork
Probably strict-relay-post-added-and-skill-mention-attachment-type claude plugin marketplace add --sparse and git-subdir plugin installs failing with "transport 'http' not allowed" when the repository is served over plain http
Probably built-in-plugin-directory-marketplace-add-remove-list-and #96134[BUG] Regression since 2.1.174: plugin installation from HTTP marketplace fails with "fatal: transport 'http' not allowed" Closed
--plugin-url download corrupting the session's cached copy of the plugin archive
No entry names this /desktop quoting partial output when opening Claude Desktop timed out or printed too much output; the error now names the cause
No entry names this claude remote-control stopping mid-turn when the Mac went to idle sleep
No entry names this claude hanging or crashing with "Raw mode is not supported" when its input is piped or redirected; it now says why and exits (use -p for piped input)
No entry names this CLAUDE_CODE_SKIP_*_AUTH sending a different Authorization header than real requests when ANTHROPIC_CUSTOM_HEADERS repeats it
Probably custom-headers-parsing-moved-to-shared-helper, self-hosted-remote-environment-and-custom-header-auth-parsin /config: settings that cycle show ‹ › and step both ways with ←/→, narrow terminals stack each value under its label, and PgUp/PgDn page the list
No entry names this /memory: the left and right arrow keys now flip its on/off settings, such as Auto-memory
No entry names this /skill names typed mid-message: Claude is now told they are skills, including disable-model-invocation ones
No entry names this ~ target too
Nothing to match on [1m] suffix (CLAUDE_CODE_DISABLE_1M_CONTEXT=1 keeps 200K)
No entry names this claude agents to arrive as queued messages; slash commands other than /stop sent while a turn is running now run when it ends
Probably remote-tool-calls-background-command-ownership-and-fleet-s Bash allow rules and allowing hooks to prompt for, not run, shell writes to files Claude Code's file tools refuse outright (the Anthropic profile store, the host credentials file)
No entry names this alwaysLoad: false to defer all of that server's tools behind tool search
Probably tengu-mcp-legacy-url-elicitation-default-flipped-to-true-mc CLAUDE_AX_PREPARK_MS=50 to restore the pause
No entry names this /usage or /context command opening the extension's dialog instead of running when picked from the command menu
Probably context-in-hosted-workers-omits-message-breakdown A model matched these bullets to the GitHub issues they fix, so a link can be wrong.
No change to the system prompt since v2.1.286.
Claude Code, interactive mode
1187 documentation changes were recorded within 24 hours either side of this release, nearest first. The closest 12 are below. They're here because they happened near this release in time. That's not a claim that this release caused the edit, or that the page documents anything in it.
The 57 literal strings found in the bundle, with the number of entries that name each one. Picking one searches for it. A name is here because this build's code mentions it, which is not the same as it working or being finished.
What's wrong with this entry?