Claude docs changes for Sept, 30th 2026 [diff]

Executive Summary

  • 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.

New Claude Code versions

2.1.285

New features

  • 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

Claude Code changes

Changed documents

agent-sdk/file-checkpointing [Source]

  • 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]

agent-view [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 [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]

claude-apps-gateway-config [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]

env-vars [Source]

  • Settings-file env values override shell values only "in most sessions", with a link to the new interaction section. [line 97] [Source]
  • CLAUDE_CODE_MAX_OUTPUT_TOKENS: for a model ID Claude Code can't resolve, the default is 32000 and the cap is 128000. [line 298] [Source]
  • DISABLE_PROMPT_CACHING_OPUS / _SONNET now apply to the default Opus / Sonnet model. [lines 427-428] [Source]

errors [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]

hooks [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 SessionStart fork 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]

keybindings [Source]

  • New Task context. footer:dismiss is now unbound by default. [lines 52-266] [Source]
  • select:* bindings are honored in list panels such as /skills, /mcp and /tasks. [line 370] [Source]
  • Keybinding validation warnings now go to the debug log (--debug). Misspelled modifiers such as ctl+k are dropped. [lines 610-620] [Source]

managed-mcp [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]

model-config [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]

permission-modes [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]
  • Bypass-permissions warning: accepting sets skipDangerousModePermissionPrompt in ~/.claude/settings.json. Declining exits. [line 570] [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]

plugin-evals [Source]

  • Expanded eval documentation, including the --ablation none|with-without baseline comparison (see the CLI reference).

plugins/cli-reference [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]

plugins/troubleshooting [Source]

  • New entry for an installed_plugins.json record this version can't read. [line 630] [Source]
  • The "marketplace not found" entry now distinguishes the <plugin>@<name> form from /plugin install <source>. [line 138] [Source]

routines [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]

self-hosted-environments-deploy [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]

settings-reference [Source]

  • Updated for defaultToAgentsView, allowClaudeInChromeWithManagedMcp and allowedProviders, and a new "How env values interact with your shell" section.

skills [Source]

  • Additions related to the /claude-api skill, including its preserved-thinking migration subcommand.

API changes

Changed documents

agents-and-tools/agent-skills/claude-api-skill [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]

agents-and-tools/mcp-connector [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]

agents-and-tools/tool-use/tool-runner [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]

agents-and-tools/tool-use/memory-tool [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]

build-with-claude/preserved-thinking [Source]

  • Added a pointer to automate the prefix-edit check with /claude-api preserved-thinking-migration. [line 429] [Source]

build-with-claude/streaming [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]

build-with-claude/thinking [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]

cli-sdks-libraries/cli/apply [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]

build-with-claude/structured-outputs [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.