Auto mode is now the built-in starting permission mode for interactive terminal and VS Code sessions on every plan and provider (not just Pro/Max/Team), and critical-path rm/rmdir removals in auto mode now prompt in the terminal with a two-minute countdown instead of always going to the classifier.
Projects gained "Run a thread on your own computer": a project thread can now run locally through Remote Control (from the Desktop app or claude remote-control) when a task needs your filesystem, MCP servers, or local tools.
New managed model-restriction settings deniedModels and availableModelsMatch let admins block a specific model version (e.g. Opus 5.5) even when a broader availableModels entry would otherwise permit it.
Claude Code gained a per-session scratchpad directory for Claude's temporary files (replacing ad hoc use of /tmp), and MCP tool image results are now saved to disk so other tools can reuse the full-resolution file.
Anthropic's plugin community marketplace submission process was replaced by a new "directory" submitted through a developer portal, which lists a plugin on claude.ai, Cowork, and Claude Code together.
Added x-claude-code-prompt-id to the gateway hint headers so LLM gateways can group the requests that serve one user prompt; opt in with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1
Added availableModelsMatch managed setting: with "exact", an availableModels entry allows only the model version it names, so new releases stay blocked until listed
Added deniedModels managed setting to block specific models, even when availableModels allows them
Added MCP tool, WebFetch, and WebSearch outputs to the tool.output OpenTelemetry span event when OTEL_LOG_TOOL_CONTENT=1
Added /doctor prompt-audit (also /checkup prompt-audit) to audit CLAUDE.md files, skills, agents, and commands for prompting patterns written for older models
Added click-to-expand for truncated messages from your other sessions in fullscreen mode
Added path to --plugin-dir load-failure entries in the stream-json system/initplugin_errors, naming the directory that didn't load
Added an opt-in load_test_mode block to the Claude apps gateway config so a deployment can be load tested without calling a model provider
Added a mantle upstream provider to the Claude apps gateway for Amazon Bedrock's Mantle endpoint
Existing feature improvements
Improved the /mcp tool list: shows more tools at once, scrolls with the page keys and mouse, and marks org-blocked tools
Improved MCP tool results: images returned by MCP tools are now also saved to a file so Bash, Read, and other tools can open them
Improved /tasks and other pickers (/help, /hooks, /copy, /chrome, /memory, /ide, /release-notes, /rewind, /diff, /remote-env, /plugin, /skills, /artifacts) with page keys, mouse wheel, and clicks
Improved the compaction spinner: timer starts when compaction begins and counts the summary's tokens as they stream
Improved the Skill tool's reply when a skill's plugin failed to load, so Claude reports the plugin failure instead of calling the skill uninstalled
Improved first-reply and first-request latency, and startup time for claude -p, Remote Control, and cold claude.ai accounts
Changed interactive sessions on third-party providers or with telemetry off to start in auto mode when no permission mode is configured
Changed --system-prompt and --append-system-prompt to accept their text and -file forms together
Changed artifact watching: an auto-armed watch now ends after 3.5 hours with no activity
Major bug fixes
Fixed SDK sessions losing a deferred tool call or finished tool result when a turn ended early
Fixed MCP progress notifications being discarded once a long-running tool call moved to the background
Fixed stdio MCP servers being left running when the session ended while they were still starting
Fixed a brief HTTP 404 from a stateless remote MCP server leaving that server unusable for the rest of the session
Fixed the weekly Fable limit not appearing in /usage and the VS Code usage meters when telemetry is disabled
Fixed dynamic workflows started during a model fallback running every agent on the fallback model instead of retrying the configured model
Fixed DISABLE_PROMPT_CACHING_HAIKU having no effect when Haiku is the session's main model
Fixed several claude plugin bugs: validate accepting names it can't install and skipping outputStyles/lspServers/monitors/themes paths, details showing 0 MCP servers, marketplace remove not naming uninstalled plugins, uninstall removing the wrong of two case-differing plugin ids, and installed_plugins.json losing records or showing no plugins when it holds a bad entry
Fixed /context not counting MCP server instructions
Fixed claude mcp add, add-json, and remove reporting success when the config file couldn't actually be written
Fixed Remote Control being unavailable on paid plans when telemetry is turned off with DISABLE_TELEMETRY or DO_NOT_TRACK
Fixed managed sandbox settings being ignored entirely when one nested value was invalid
Windows: Fixed the PowerShell tool letting cmd /c rd, rmdir, del, or erase delete drive roots and other protected folders
[VSCode] Fixed a session teleported from the web dropping messages sent while Claude was working
[Code Review] Fixed "@claude review" requests going silent when GitHub failed to return the pull request
[Code Review] Fixed billing for a review that stopped at its time limit with nothing verified
Clarified that rm/rmdir removals targeting a critical path are never auto-approved and, in an Agent SDK session in auto mode, are denied by default without reaching canUseTool. [lines 35-45] [Source]
Added the verbatim_prompts option: deliver every prompt as written (client_composed: True) instead of expanding @ mentions or running / commands. Requires Python Agent SDK 0.2.158+ and Claude Code v2.1.248+. [line 810] [Source]
Added a new @anthropic-ai/claude-agent-sdk/core entry point for apps that bundle the Agent SDK, which drops zod/MCP SDK inlining and some root-only exports. [line 44] [Source]
Added prewarm(), SpareProcess, and .claim() (alpha): start a spare Claude Code process before you know which session or folder it will serve. [line 117] [Source]
Added the verbatimPrompts option and client_composed message field, matching the Python SDK's verbatim_prompts. [line 519] [Source]
Added SpareProcess (claim, claimed, exited, close) as the return type of prewarm(). [line 698] [Source]
Documented SDKSystemMessage.plugin_errors, listing which plugin failed to load and why. [line 1626] [Source]
SDKTaskProgressMessage.summary now also carries an MCP server's latest reported progress for a backgrounded MCP tool call, not just subagent progress summaries. [line 5104] [Source]
SDKCommandsChangedMessage is now also emitted when an MCP server's prompts join or leave the command list (v2.1.281+). [line 5255] [Source]
SDKConversationResetMessage gained optional trigger, user_message_uuid, and timestamp fields describing what caused the reset (v2.1.281+). [line 5290] [Source]
Teammates now inherit a --setting-sources restriction from the lead; before v2.1.281, split-pane teammates loaded every settings source. [line 283] [Source]
Added --setting-sources as a configuration flag agent view carries through to dispatched sessions. [line 428] [Source]
claude --bg (and background-session restarts) now check workspace trust for the target directory before starting. [line 428] [Source]
A worktree with only uncommitted changes in a verifiable checked-out submodule no longer blocks "delete again to remove the directory anyway". [line 544] [Source]
Added "Add your own labels": telemetry.resource_attributes lets a gateway operator push fixed OpenTelemetry resource labels (like service.namespace) with every session's telemetry (v2.1.281+ on the gateway server). [line 888] [Source]
Clarified load_test_mode: a replica's CPU-per-request estimate reads lower than production since no request actually leaves for the provider. [line 998] [Source]
In Anthropic-hosted environments, GitHub credentials now stay encrypted on Anthropic's servers and never enter the session VM; traffic goes through a server-side GitHub proxy instead. [line 49] [Source]
Added a "Time limits" constraint pointing to the new cloud-environments time-limits section. [line 351] [Source]
Added a Session scratchpad directory: a per-session temp directory Claude uses instead of /tmp for intermediate files, helper scripts, and drafts, without a permission prompt. [line 1342] [Source]
tool-results/ now also holds full-size copies of images MCP tools return. [line 1314] [Source]
A .claude/rules/ or CLAUDE.md symlink pointed at a network path (UNC share, /net, /Network) is no longer followed, to avoid contacting the host it names. [line 237] [Source]
Added a startup//status warning when an instruction file exceeds the recommended length, or several files together exceed a combined limit. [line 569] [Source]
Added "Run a thread on your own computer": choose Work locally in a project conversation to run that task's thread through Remote Control on a connected machine (Desktop app or claude remote-control, v2.1.280+), instead of in the cloud. Covers connecting the folder, the approval card, auto mode behavior, and the "Lost contact with your folder" and new error states. [line 245] [Source]
Added a Time limits section: command timeouts, SessionStart hook's 600s default, the setup-script ~5-minute caching cutoff, and idle-session expiry, plus BASH_DEFAULT_TIMEOUT_MS/BASH_MAX_TIMEOUT_MS to raise them. [line 351] [Source]
Code Review now automatically retries some interrupted reviews; check the run's summary before manually retriggering. Check-run titles changed to e.g. Code review failed. [line 255] [Source]
New "Subagents and workflows" note: every subagent and workflow-spawned agent sends its own requests, which count toward usage; the attribution breakdown shows the subagent share. [line 344] [Source]
Clarified cloud-session data handling: Anthropic stores the session transcript, and GitHub credentials are encrypted server-side and never enter the VM. [line 88] [Source]
Clarified that committing a project-scope plugin entry turns it on for collaborators but doesn't download it; each collaborator must also run claude plugin install once. [line 125] [Source]
Added CLAUDE_CODE_DISABLE_DANGEROUS_RM_TIMEOUT, CLAUDE_CODE_DISABLE_POWERSHELL_CMD_RM_DENY, and CLAUDE_CODE_DISABLE_SUBSTITUTION_RM_PROMPT to opt out of the new critical-path removal checks. [line 9] [Source]
DISABLE_PROMPT_CACHING_HAIKU now disables caching for the default Haiku model wherever it runs, including as your main model. [line 26] [Source]
With feature-flag fetching off, Remote Control's availability now depends on the specific variable set, rather than blanket unavailability. [line 46] [Source]
Documented dozens of new/changed error messages, including: Couldn't save your login (macOS keychain locked) and Could not refresh your login because another Claude Code process is refreshing it; Part of the response never arrived / The response stream was malformed for dropped or duplicated stream events; new disk-quota/temp-filesystem-full errors (EDQUOT/ENOSPC); a reserved-name notice for skills/commands/workflows named inside the anthropic-skills namespace; Claude Code can't start: your organization's managed settings block the default model; and Desktop-app-specific wording for several auth and usage-limit messages. [line 231] [Source]
Resuming a -p/SDK session interrupted by SIGTERM no longer auto-continues the unfinished turn by default; set CLAUDE_CODE_RESUME_INTERRUPTED_TURN=1 to restore the old behavior. [line 75] [Source]
Added a new warning message when a -p/SDK session's working directory is deleted mid-session; shell commands fail until it exists again. [line 149] [Source]
plugin_errors entries for a failed --plugin-dir directory/archive now include the resolved path (v2.1.283+). [line 165] [Source]
The Resume dialog's Web tab now also lists Remote Control sessions and opens a matching local conversation directly instead of downloading a copy, with clearer download-failure errors. [line 264] [Source]
A message queued with an editor selection attached now keeps that selection even if you select something else afterward. [line 295] [Source]
Pressing Ctrl+C or Ctrl+D twice in most dialogs now closes the dialog instead of exiting Claude Code. [line 148] [Source]
footer:dismiss was removed in v2.1.281 (existing bindings become no-ops). [line 256] [Source]
The rewind menu's message list and the diff file list now use the generic Select actions/bindings instead of their own dedicated MessageSelector/diff:viewDetails actions (v2.1.283+). [lines 271-301] [Source]
New stricter streaming requirements for gateways: deliver the full event sequence without dropping/duplicating/reordering, and relay each response through its final message_delta/message_stop before closing the body. [line 59] [Source]
Added the x-claude-code-prompt-id gateway hint header (v2.1.283+). [line 145] [Source]
Only an SDK host application can register an in-process "type": "sdk" server; such an entry in .mcp.json/settings is now skipped with a clear message. [line 76] [Source]
Duplicate-server detection now normalizes URL case, default ports, and trailing slashes when comparing endpoints. [line 587] [Source]
Added "Images in tool results": MCP-returned images are now also saved full-resolution to the session's tool-results directory (v2.1.283+). [line 1252] [Source]
MCP Apps UI resources (ui:// or text/html;profile=mcp-app) no longer appear in @ suggestions or the resource list tool. [line 1370] [Source]
Prompts from a server named anthropic-skills no longer appear, since that name is reserved for synced skills. [line 1463] [Source]
Added "Block specific models or versions": use deniedModels or availableModelsMatch: "exact" to keep a newer release (e.g. Opus 5.5) blocked even though a broader availableModels entry permits it (v2.1.283+). [line 336] [Source]
Opus 5.5 defaults to effort medium, one level below Opus 5's high default; start at medium rather than carrying over an Opus 5 level. [line 614] [Source]
Auto mode is now the built-in starting permission mode for terminal/VS Code sessions on every plan and provider with Claude Code v2.1.283+, not just Pro/Max/Team. [line 4] [Source]
Critical-path rm/rmdir handling changed (v2.1.281+): in auto mode Claude Code now asks in the terminal with a two-minute countdown (denying elsewhere, such as -p, SDK, VS Code chat) instead of always routing to the classifier; bypassPermissions also gets a timed terminal prompt. [lines 621-656] [Source]
Added new critical-path detection cases: a shell variable followed by a top-level directory name, a variable assigned from a directory-printing substitution in the same command, a backslash-only target, and a target that is entirely a command substitution. [line 656] [Source]
Remove-Item's system-path denial now also covers the cmd built-ins rd, rmdir, del, and erase run through cmd /c (v2.1.283+), matching the Windows PowerShell tool security fix. [line 372] [Source]
claude plugin eval now requires git 2.31+ when git is installed, and stops with a new error before running any case on an older git. [line 21] [Source]
Added guidance to use Anthropic's plugin-dev plugin (/plugin-dev:create-plugin) to have Claude scaffold and validate a larger plugin. [line 120] [Source]
Submitting a plugin to Anthropic's public community marketplace is being replaced by submitting to Anthropic's directory through a new developer portal at claude.ai/directory/manage; a listing there reaches claude.ai, Cowork, and Claude Code together, and requires a paid claude.ai plan. [line 471] [Source]
Added a table detailing exactly which mid-session MCP server changes (connect, drop out, auto-reconnect, deliberate removal) invalidate the cache when tool definitions load into the prefix. [line 100] [Source]
DISABLE_PROMPT_CACHING_HAIKU now covers the main conversation when Haiku is your main model (v2.1.283+). [line 336] [Source]
Remote Control can now be started from the Claude Desktop app's Code tab with /remote-control or /rc, in addition to the CLI and VS Code. [line 248] [Source]
Server mode gained a -d/--debug[=<filter>] flag (v2.1.282+). [line 232] [Source]
DISABLE_TELEMETRY/DO_NOT_TRACK alone no longer block Remote Control unless your organization requires Trusted Devices (v2.1.283+). [line 208] [Source]
/focus on/off is now supported from mobile and web (v2.1.281+). [line 389] [Source]
A schedule set exactly on the hour can now start several minutes late; pick a few minutes past the hour (e.g. 9:07) to start closer to the scheduled time. [line 130] [Source]
Detailed how a startup that reads the cached remote settings now preserves fail-closed (stricter) values, not just dropping invalid entries. [line 188] [Source]
Added availableModelsMatch and deniedModels managed settings for blocking specific model versions (v2.1.283+). [lines 570-833] [Source]
Added maxProseWidth to cap the width of prose in Claude's responses in a wide terminal (v2.1.282+). [line 3174] [Source]
The trailers (git commit attribution) default now names the subagent's model when a subagent makes the commit, not the session's active model. [line 3796] [Source]
The frontmatter name field now sets the actual invocable command for a personal or project skill directory, not just its display label. [line 383] [Source]
Added "Names reserved for synced skills": anthropic-skills and everything inside that namespace is now reserved for claude.ai-synced skills, replacing the earlier claude-ai reservation. [line 231] [Source]
Picking up a newly created top-level skills directory now requires running /reload-skills instead of restarting Claude Code. [line 260] [Source]
Subagent requests now count toward the same usage limits as the main conversation, noted alongside model-selection guidance. [line 4] [Source]
An --agents JSON definition can now have an empty prompt, which leaves the session's system prompt unchanged when selected with --agent (v2.1.281+). [line 208] [Source]
Clarified that ultracode tokens draw on subscription usage limits, and that turning it on already opts out of the "Large workflow" warning, the concurrent-subagent limit, and the auto-mode first-launch approval. [line 143] [Source]
Noted a Microsoft Foundry-specific exception: the Skill version download endpoint (GET /v1/skills/{skill_id}/versions/{version}/content) isn't supported there. [line 137] [Source]
Clarified beta-header requirements for mid-conversation tool changes: inline-tools-2026-09-15 (Claude API) also covers inline tool definitions, and adding an MCP server mid-conversation needs mcp-client-2026-09-15 as a second header on the Claude API. [line 16] [Source]
Updated the feature-availability table to reflect Agent Skills and Files API reaching GA on Claude Platform on AWS and Microsoft Foundry. [line 82] [Source]
Engagement/adoption endpoint data now typically lands by ~13:00–13:30 UTC the following day, earlier than the previously documented ~17:00 UTC. [line 75] [Source]