Claude Code 2.1.285 adds an allowedProviders managed setting, claude --desktop, claude plugin configure, and a CLAUDE_CODE_DISABLE_WEB_FETCH variable. It also puts a time limit on background shell commands and hardens /ultrareview uploads and the Artifact tool.
Permission modes docs now spell out the critical-path rm checks. They cover which targets count, a rewrite guide, and per-mode outcomes, plus a new Amazon Bedrock guardrail option for the Claude Apps gateway.
A new "Open agent view by default" setting (defaultToAgentsView) lets claude start in agent view, and managed-MCP admins can now allow Claude in Chrome alongside managed-mcp.json.
The self-hosted runner docs gain guidance on runner exits and failed starts and on private CAs with Anthropic-managed git. Routine hourly limits and public artifact sharing are also documented.
API docs add a /claude-api preserved-thinking-migration skill workflow and MCP helper functions for SDKs. The tool-runner and streaming docs were also expanded.
Added CLAUDE_CODE_DISABLE_WEB_FETCH environment variable to turn off the WebFetch tool
Added claude --desktop to open the Claude desktop app on the current directory or on a session (with --continue / --resume <id>)
Added claude plugin configure <plugin> to show a plugin's options and save values from stdin with --values-stdin
Added <server>.<key>=<value> to claude plugin install --config so a bundled .mcpb MCP server can be configured at install time
Added allowedProviders managed setting to limit which API providers a machine may use
Added CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES to cap re-sends of a timed-out non-streaming fallback request
VS Code: added a plugin options form in Manage plugins, an on-demand diagnostics tool that reads the Problems panel, and a note on tabs interrupted by a window reload
Claude Tag: added direct messages with Claude for Enterprise Standard or Usage-Based Chat seats that also have Cowork
Existing feature improvements
/resume and claude --resume now open a session running in the background, and claude --resume <id> "prompt" sends the prompt to it as its next turn
Bedrock and Vertex AI sessions fall back to an older model of the same tier when an admin removes access to the default model
Sessions behind a custom ANTHROPIC_BASE_URL now use the 1M context window of models that have one (run /autocompact 200k if your gateway stops at 200K)
Background Bash and PowerShell commands now stop after a time limit (default 30 min, max 2 h) and Claude is notified
claude -p and Python Agent SDK sessions on third-party providers or with telemetry off now start in auto mode when no permission mode is configured
Subagents in auto mode end their run as soon as they hand back a report
Sandbox settings: project settings can no longer widen or turn off an admin-required sandbox or reopen managed read-denies
/tasks folds Claude Code's own background work under one "System tasks" row
claude mcp list now lists WebSocket servers, and claude mcp get hides command, args and env values of plugin-provided stdio servers
/ultrareview on macOS and Linux now requires git 2.31+ for local uploads and refuses symbolic-ref checkouts
The MCP server name widgets is reserved in cloud sessions and on self-hosted runners
Plugin marketplace errors name why a git address was refused, and git URL validation is stricter
Major bug fixes
Fixed Claude Code refusing to start when the OS denies reading the managed settings file (it now warns and continues)
Fixed model switches through set_model (e.g. the Agent SDK's setModel) keeping the old output-token limit and auto-compact window
Fixed fork subagents not keeping the parent's plan mode or dontAsk mode
Fixed several Artifact tool issues: overwriting newer content after a rewind, Artifact allow rules covering files outside the working directories, and auto mode skipping its classifier
Fixed synchronous hooks hanging while a background process kept the hook's output open
Fixed the PowerShell tool skipping deny and ask rules when its command parser failed to start
Fixed a failing API request being retried up to 21 times when streaming kept failing
Fixed redacted logs showing part of a URL password containing @
Fixed plugin and marketplace installs over SSH ignoring GIT_SSH / core.sshCommand
The page was condensed around a single lead example. It now shows capturing the checkpoint UUID and session ID, then resuming with an empty prompt and calling rewind_files() / rewindFiles(). [lines 5-27] [Source]
The rewind must be called on the resumed session, which needs checkpointing enabled in its options too. [line 148] [Source]
New "Open agent view by default" section: turn on defaultToAgentsView in /config so claude with no arguments opens agent view. Passing a prompt starts a regular session. [lines 59-82] [Source]
Artifacts are private until shared. After a public share, Claude asks for approval once per conversation before changing it. [line 42] [Source]
Public sharing (a link viewable without sign-in) is documented and is off by default on Team and Enterprise until an Owner enables it. Public viewers can't see or add comments. [lines 83-106] [Source]
Connector-backed pages can be shared, but connector calls don't run for public viewers. [line 166] [Source]
Admin controls were reorganized under Organization settings (Artifacts, Capabilities, Data and privacy retention). [lines 343-359] [Source]
New "Apply an Amazon Bedrock guardrail" section: a guardrail block (id, version) on Bedrock upstreams. It must be set on all Bedrock upstreams or none, and requires bedrock:ApplyGuardrail. Input tags are not supported. [lines 283-307] [Source]
Non-interactive runs apply pushed settings for that run only without recording approval, and a declined dialog exits the session. [line 757] [Source]
Several desktop policy keys require Claude Code v2.1.281+ on the gateway server. [line 817] [Source]
Telemetry exports from /login sessions carry user.id, user.email and user.groups for per-developer attribution. [line 842] [Source]
New entries: "Failed to start OAuth callback server" (with a claude setup-token workaround), PDF errors including a missing pdftoppm, "Couldn't confirm model with the API", "Can't switch to the default model" (managed deniedModels / availableModels), and the /ultrareview "upload cannot follow that setting" error. [line 1229] [Source]
The "not a recognized model id" section is rewritten. Model switches from the Agent SDK or an app are now confirmed with the provider on the first switch. [lines 2205-2241] [Source]
The minimum-version API error now has a table of how to update each kind of binary (CLI, desktop app, VS Code extension, SDK). [lines 2309-2318] [Source]
The Fable usage-credits consent prompt can time out via dialogExpiry in non-terminal sessions. [lines 696-711] [Source]
New "What a blocked prompt leaves behind" section: a blocked UserPromptSubmit prompt never reaches Claude, but its text is still written to the transcript unless suppressOriginalPrompt is set. It can still appear in local files and prompt history. [line 1348] [Source]
Blocked prompts are no longer described as "erased". [line 834] [Source]
The SessionStartfork matcher now covers conversations moved to the background. [line 1084] [Source]
The agent_needs_input notification also fires for auto mode's classifier-billing notice. [line 2211] [Source]
New allowClaudeInChromeWithManagedMcp device-level setting lets Claude in Chrome run alongside managed-mcp.json. It is otherwise blocked, and claude --chrome exits at startup. [lines 144-152] [Source]
The allowlist bypass now covers three groups: the organization's own servers, built-in servers, and a Claude Tag session's Slack tools. [line 293] [Source]
Fable consent prompt behavior in Agent SDK hosts is clarified, including expiry via dialogExpiry. [line 102] [Source]
Model-switch validation is rewritten: the SDK or an app confirms the ID with the provider. Remote Control checks locally. --model, ANTHROPIC_MODEL and model are not checked up front. [lines 149-156] [Source]
Cowork remote sessions: the server rejects models outside the server-managed availableModels. [line 272] [Source]
Thinking can't be turned off on Opus 5.5, Sonnet 5.5 and Fable. The toggle now shows "Thinking can't be turned off". [line 676] [Source]
A pinned ANTHROPIC_DEFAULT_*_MODEL replaces the picker's 1M rows. Use /model opus[1m] or /model sonnet[1m] to reach the 1M window. [line 860] [Source]
The critical-path rm/rmdir section is restructured into subsections: which paths are critical, other targets that count, nested commands and inline scripts, how to rewrite a flagged command, per-mode outcomes, and time limits. [lines 637-708] [Source]
New targets are documented: positional parameters like $1, sh -c inline scripts, and double-quoted find -exec sh -c scripts. [lines 654-688] [Source]
Auto mode fallback is reorganized. New "No verdict from the classifier" case, "Repeated-block thresholds" subsection, and "How auto mode evaluates actions" heading. [lines 454-489] [Source]
New --ablation <mode> option for claude plugin eval. eval init runs from the plugin root and takes a case name. [lines 409-450] [Source]
Removing a marketplace from its last scope also deletes its cache and uninstalls its plugins, along with their options and data. Use update to refresh instead. [line 663] [Source]
/plugin install <source> in a session reports "marketplace not found" and installs nothing. [line 726] [Source]
Session-only plugins (--plugin-dir) show in plugin list only when the flag precedes the subcommand. [line 791] [Source]
New hourly limits table: scheduled runs 100/hour per account, Run now and API fires 30/hour per routine and 100/hour per account, with no overage. GitHub events have their own per-routine and per-account caps. [lines 362-372] [Source]
Routines belong to an individual account and aren't shared with teammates. [line 59] [Source]
New "Trust a private certificate authority with Anthropic-managed git" section: how GIT_SSL_CAINFO and GIT_SSL_NO_VERIFY are handled for the runner's clone, in-session git, and hooks. Requires v2.1.283+. [lines 163-187] [Source]
New "When the runner exits" section: how to recognize a failed start ([runner:fatal] or error: lines) and how to restart with a growing wait. It covers Kubernetes CrashLoopBackOff and Docker restarts. [lines 520-581] [Source]
The Anthropic git proxy requires --capacity 1, but the sample recipes use --capacity 4. Sessions sharing a container lack per-session isolation. [lines 159-201] [Source]
Keep programs named in GIT_SSH_COMMAND / GIT_ASKPASS out of sessions' write access. A pinned version may be too old for a model. [lines 145-429] [Source]
New /claude-api preserved-thinking-migration subcommand and a five-step workflow (scope, find, measure, fix, report) for finding history edits that invalidate thinking blocks. It runs real, billed API requests. [lines 136-155] [Source]
Beta header cleanup is now listed among the skill's migration tasks. [line 126] [Source]
SDK helpers that convert between MCP types and Claude API types are now listed in a multi-language table: tools, messages, resource-to-content and resource-to-file. [lines 1187-1317] [Source]
Documents inspecting and modifying tool results via generate_tool_call_response() / generateToolResponse(). Ruby has no such hook, so read runner.params[:messages]. [lines 1148-1547] [Source]
The TypeScript and Ruby runners support automatic compaction. [line 1096] [Source]
Memory tool helpers exist in four SDKs. The first view of /memories on an empty store is not an error when using the local-filesystem helper. [lines 284-748] [Source]
Added SDK helpers to stream internally and return the full message: C# Aggregate() and PHP MessageAccumulator. Also documents accumulating partial tool-input JSON deltas. [lines 141-335] [Source]
Ruby note: plain hashes take display: while the typed class spells it display_. The SDK requires streaming when max_tokens is above 21,333, a client-side check. [lines 284-1198] [Source]
ant apply now supports vaults as YAML files in vaults/ (CLI 1.34.0+). Any resource except a skill can be YAML, JSON or Markdown. [lines 9-104] [Source]
Large restructuring of the page (about 1,500 lines changed). No new feature is evident from the diff.
Other changes
The API reference pages under docs-md/api/api/ were regenerated with formatting churn, mostly reindented lists (for example, beta header enums). No substantive changes were found there.
Smaller edits went into managed-agents docs (memory, vaults, self-hosted sandboxes, migration), compliance docs, rate-limits-api, Foundry/Platform-on-AWS pages and the model migration guides.