Claude docs changes for September 25th, 2026 [diff]

Executive Summary

  • Claude Code plugin docs were overhauled: the old single-page guides were rewritten into short overview pages and split into a new docs-md/claude-code/plugins/ section of 20 focused pages (installing, creating, publishing, org controls, dependencies, measuring cost/usage, troubleshooting, and more).
  • Claude Code 2.1.282 ships a large set of reliability fixes for resumed/continued sessions losing extended thinking or re-sending turns in a changed form, several vim-mode and rendering bugs, a new maxProseWidth setting, telemetry-ignored-variable startup warnings, and a big Slack/Claude Tag and cloud-sessions bug-fix batch.
  • Claude API is resuming billing for pre-output refusals in the bio, frontier_llm, and reasoning_extraction categories (previously every pre-output refusal was free); mid-stream refusals are unaffected.
  • Cache diagnostics graduated from beta to GA on the Claude API — no beta header needed, just send the diagnostics object.
  • Claude Code Projects threads can now run on your own machine through Remote Control instead of only in the cloud, and auto mode's server-side classifier review now covers read-only/sandboxed shell commands and denies actions when the server gives no verdict.

New Claude Code versions

2.1.282

New features

  • Added a maxProseWidth setting that caps the width of Claude's prose in wide terminals while tables and code blocks keep the full width
  • Added a startup notice, and /status and claude doctor entries, listing telemetry variables in a project's settings files that were ignored or that turned telemetry off
  • Added the allowClaudeInChromeWithManagedMcp managed setting to let claude --chrome run alongside an exclusive managed-mcp.json; the error shown when Chrome is blocked now names it
  • Added store.readiness_grace_seconds to the Claude apps gateway so /readyz can stay ready through a short Postgres outage such as a database failover
  • Added a scrollbar to the /feedback drafts list in fullscreen mode
  • [Cloud sessions] Added Claude GitHub App status to Settings › Connectors › GitHub, plus "Open repository"/"Open compare page" links for non-GitHub repositories and support for attaching a repository from a different GitHub owner (such as a fork's upstream) to a running cloud session

Existing feature improvements

  • Changed auto mode to use the server-side classifier by default on a direct Anthropic API connection when telemetry is off (CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out)
  • Changed sandbox.excludedCommands to ignore project and local settings entries when managed settings, --settings, or allowManagedDomainsOnly restrict unsandboxed commands or domains
  • Changed project and local settings to ignore OpenTelemetry variables that turn on export, set its endpoint, or capture content (e.g. CLAUDE_CODE_ENABLE_TELEMETRY, OTEL_LOG_*)
  • Changed Windows/WSL managed settings so an invalid or unreadable admin policy no longer blocks user-writable HKCU and WSL /etc/claude-code settings from applying
  • Changed Skill(anthropic-skills:*)/Skill(claude-ai:*) allow rules, skill/command loading, and MCP servers named anthropic-skills or claude-ai to only apply to content actually synced from claude.ai, not plugins or MCP servers that merely reuse the name
  • Improved the time to resume very large sessions, including ones never compacted
  • Improved /artifacts: aligned titles, whole-item detail truncation, and PgUp/PgDn/Home/End/mouse-wheel support
  • Improved the Windows EBADF resume error and the Claude Desktop unknown-model error messages
  • Updated the claude-api skill: refusal billing links to the new docs, mid-stream refusals bill at normal rates, pre-output refusals count against rate limits, and it recommends ant apply for Managed Agents resources
  • [Cloud sessions] Improved load time for the Routines page and sidebar Scheduled list for accounts with many past scheduled check-ins
  • [Claude Tag] Changed bordered Slack cards (plans, tables, details) to render wide by default, and newly connected Slack workspaces now follow the current default model

Major bug fixes

  • Fixed every request failing with a 400 when conversation history holds web search results the API can't decrypt (e.g. from a turn answered through a third-party gateway)
  • Fixed --continue/--resume sessions re-sending earlier messages in a changed form, which could make the API drop Claude's earlier reasoning, including when thinking was dropped by an immediate slash command or a --tools list missing a previously offered built-in tool
  • Fixed sessions failing every turn with an "Invalid data in redacted_thinking block" error by dropping thinking blocks and retrying once
  • Fixed compaction failing on a refused summarization request by retrying on a fallback model
  • Fixed a failed turn after a safety-related model switch in sessions with thinking off and effort above high
  • Fixed model-switch and Fable usage-credit prompt handling in SDK-hosted and Remote Control sessions
  • Fixed a login error blocking requests for up to a minute after another Claude Code process refreshing auth was closed or killed mid-refresh
  • Fixed CLAUDE.md/rules being read through a repository symlink that reaches macOS's /Network or a /.vol-style kernel path
  • Fixed Bash permission rules with a mid-pattern :* being skipped in settings files while --allowedTools honored them
  • Fixed managed settings ignoring a mistyped boolean lock key (e.g. disableClaudeAiConnectors, allowManagedPermissionRulesOnly) instead of enforcing the locked value
  • Fixed managed permissions, autoMode, worktree, and attribution settings being ignored entirely when one nested value was invalid
  • Fixed skills/commands/skills-directory plugin manifests pre-approving their own tools via allowed-tools under allowManagedPermissionRulesOnly
  • Fixed tool input validation errors naming only one bad parameter when several were invalid in the same call
  • Fixed several rendering/input bugs: pasted multi-line text submitted line-by-line after bracketed-paste reset, blank-screen flash before fullscreen's first frame, garbled rows after the terminal shrank, a stale character after a wrapped CJK/emoji redraw, and cursor placement issues in /skills search and vim mode (>>, r, 2J, dd/dj/dG/p/P, . repeat counts)
  • Fixed /install-github-app continuing remaining steps after "cancelled", and /feedback//bug//share on third-party providers still saving reports after a cancelled save
  • Fixed plugin uninstall deleting a plugin's saved options/secrets when its settings file was unreadable or the installed-plugins list couldn't be read
  • [VSCode] Fixed long replies falling behind the stream, the dictation mic covering the input scrollbar, and Remote Control sessions started locally not opening from the Web session list
  • [Claude Tag] Fixed several Slack/Enterprise Grid bugs: auto-join patterns ignored org-wide, cross-workspace channel routing, retired-model fallback loops, table captions on uploaded files, thread naming in Agents & tools view, and inflated cost/token totals after a worker restart

Claude Code changes

New Documents

The plugin documentation was split out of a handful of long top-level pages into a dedicated docs-md/claude-code/plugins/ section:

overview [Source]

What a plugin is, when to use one instead of a standalone skill/subagent/hook/MCP server, and what an enabled plugin costs in context and permissions.

install [Source]

Install and manage plugins on any surface (terminal, desktop, IDE, cloud), choose an install scope, add marketplaces, and keep plugins updated.

create [Source]

Build a plugin with skills, agents, hooks, and MCP servers, and develop it locally with --plugin-dir before publishing.

create-marketplace [Source]

Build and test a marketplace.json locally before hosting it.

host-marketplace [Source]

Host a marketplace, including authenticating archive downloads with a headersHelper command.

publish [Source]

Share a plugin without a marketplace, or submit it to Anthropic's community marketplace.

manifest-reference [Source]

Full plugin.json field reference: types, defaults, path rules, userConfig, channels, and environment variables.

marketplace-reference [Source]

Reference for every field a marketplace entry accepts, including the command plugin source.

components [Source]

Every component type a plugin can hold (skills, agents, hooks, MCP servers, LSP servers, monitors) with an interactive plugin-directory explorer.

loading [Source]

How plugins are found on disk, cached, auto-updated, and synced from claude.ai, plus Node.js package dependencies.

dependencies [Source]

Declare version-constrained plugin dependencies and bundle a curated plugin set behind one install (mirrors the rewritten plugin-dependencies.md).

relevance [Source]

Add a relevance block to marketplace entries so Claude Code suggests plugins based on session signals (mirrors the rewritten plugin-relevance.md).

cli-hints [Source]

Emit a <claude-code-hint /> tag from your own CLI/SDK to prompt users to install your official-marketplace plugin (mirrors the rewritten plugin-hints.md).

measure [Source]

New page: measure a plugin's token cost and usage, and pick telemetry events for org-wide plugin questions.

org [Source]

Managed-settings controls for administrators: require/allowlist marketplaces, pre-install plugins, and restrict what users can install.

security [Source]

Plugin security and trust model, and Anthropic's official marketplace names.

anthropic-marketplaces [Source]

Distinguishes Anthropic's official, community, and demo marketplaces from Claude Marketplace (claude.com/marketplace).

cli-reference [Source]

Reference for every claude plugin shell subcommand.

code-intelligence [Source]

Ship LSP servers and other code-intelligence components in a plugin.

troubleshooting [Source]

Every claude plugin validate message and load/dependency error with its fix.

claude-apps-gateway-config [Source]

New comprehensive reference for every gateway.yaml option: listener/TLS, OIDC, session, Postgres store, per-provider upstreams (Anthropic, Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry), static upstream headers, model routing, managed policies, and telemetry.

Changed documents

claude-projects [Source]

  • Project threads can now run on your own machine through Remote Control instead of only as cloud sessions, when a task needs something only your computer has; requires Claude Code v2.1.280+ on that machine and doesn't work with Require trusted devices on.

errors [Source]

  • Added The server returned no safety verdict covering auto mode's new denial-on-no-verdict behavior and the 10-in-a-row turn stop
  • Added recovery documentation for role 'system' must precede an 'assistant' message (Claude Code now retries once with the text as a user message) and Invalid encrypted_content in search_result block / Failed to decrypt web search result content

keybindings [Source]

  • Added a Text fields section clarifying that a bound bare letter/digit/Space still types normally in a focused text field (e.g. the "Other" answer to a Claude question)

managed-settings [Source]

  • Documents that claude.ai now itself enforces the admin console's strictKnownMarketplaces/blockedMarketplaces lists when a marketplace is added from claude.ai or Cowork's Customize tab
  • Added a full table of fallback behavior when a managed boolean/allowlist key (e.g. allowManagedHooksOnly, disableCommandPluginSources, crossSessionInbound) is present but invalid, matching 2.1.282's fix

permission-modes [Source]

  • Rewrote Server-side classifier review to cover which session types ask the server by default (now including telemetry-disabled direct-API sessions as of 2.1.282) and the two non-review outcomes (fallback to local classifier vs. denial on no verdict)
  • Documents that read-only and sandboxed shell commands now also wait for server-side review and are blocked if flagged

env-vars [Source]

  • Documents that project/local settings can no longer set CLAUDE_CODE_ENABLE_TELEMETRY and other OpenTelemetry variables that turn on export or capture content (requires v2.1.282); the off-values that still apply are listed separately

settings-reference [Source]

plugins.md, plugins-reference.md, discover-plugins.md, plugin-marketplaces.md

plugin-dependencies.md, plugin-relevance.md, plugin-hints.md, plugin-evals.md

Many other pages (claude-directory.md, admin-setup.md, sub-agents.md, settings-example.md, etc.) changed only in the links they use to point at the new plugins/ paths, with no content change.


API changes

New Documents

managed-agents/quickstart [Source]

New end-to-end tutorial for Claude Managed Agents: install the ant CLI and SDKs, create an agent and environment, start a session, and stream its events, plus links to four framework-specific "build a complete app" quickstarts (Chat SDK, assistant-ui, CopilotKit).

Changed documents

about-claude/models/optimizing-for-cost-and-intelligence [Source]

  • Added a new Show the model elapsed time technique: a system-prompt instruction plus a per-turn elapsed-time message, benchmarked by Anthropic on DRACO, HLE, and an internal physics set, showing meaningful time/cost cuts for a small, benchmark-dependent score change

build-with-claude/cache-diagnostics [Source]

  • Cache diagnostics is now GA on the Claude API: the cache-diagnosis-2026-04-07 beta header is no longer required, sending the diagnostics object is itself the opt-in, and the response field is now null (not omitted) whenever diagnostics wasn't requested

build-with-claude/refusals-and-fallback [Source]

  • How refusals are billed now documents that pre-output refusals in the bio, frontier_llm, and reasoning_extraction categories are billed at normal rates (to deter attempts to circumvent safeguards at scale); other categories, and mid-stream refusals' existing billing, are unchanged. Applies on every platform.

build-with-claude/mid-conversation-system-messages [Source]

  • Mid-conversation tool changes on the Claude API now use a single inline-tools-2026-09-15 beta header (covering both tool-reference changes and inline tool definitions); the older mid-conversation-tool-changes-2026-07-01 header still works for by-reference changes on the Claude API, Amazon Bedrock, and Google Cloud, but not for inline tool definitions

build-with-claude/structured-outputs [Source]

  • The deprecated output_format parameter now requires the structured-outputs-2025-11-13 beta header again; without it the API returns a 400 (previously accepted without a header during a transition period)

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

  • Clarifies that "drop_block" doesn't fix the underlying prefix edit and can increase a session's token usage, since Claude may re-derive the reasoning it lost; recommends alerting on prefix_binding_mismatch entries and fixing the edit instead

manage-claude/compliance-activity-feed [Source]

  • Documents that, as of September 24, 2026, filename/title fields on file, project-document, and artifact activities are always empty/omitted (including on past activities); names must now be looked up by ID through the metadata endpoints with read:compliance_user_data

manage-claude/compliance-sessions [Source]

  • Compliance API local session endpoints are now stable (out of beta) for Claude for Microsoft 365 sessions (Excel, PowerPoint, Word, Outlook), alongside Cowork and Claude Code; Claude Science and Claude in Chrome remain beta

manage-claude/inference-hooks [Source]

manage-claude/cmek-aws-kms [Source]

  • Expanded key-policy and troubleshooting guidance: locating a workspace's compartment ID, the optional EncryptionContext/aws:SourceArn conditions that scope a key to specific workspaces, and diagnosing a denied attach via CloudTrail

managed-agents/self-hosted-sandboxes [Source]

  • Added Go SDK support for self-hosted workers (EnvironmentWorker, work.poller()/HandleItem(), agenttoolset.BetaAgentToolset20260401), alongside Python and TypeScript
  • Documents memory_sync_interval/memory_sync_deletions for tuning how attached memory stores reconcile and handle local deletions, and how to run custom tools or MCP-backed tools through the SDK worker

managed-agents/environments [Source]

  • Added allow_package_managers, which opens outbound access to a maintained list of public package registries/code hosts; required whenever an environment specifies packages, with a note on the prompt-injection/exfiltration risk of host-level (not per-operation) network access

API reference (all languages)

  • The full per-language API reference (api/api/beta/** under cli, csharp, go, java, php, python, ruby, typescript) was regenerated with expanded parameter documentation and constraints for existing and new beta.messages.create fields: request-level compaction, diagnostics (cache diagnostics, now unconditional), speed: "fast", thinking.block_binding drop-on-mismatch controls, response container info, and sparse per-member overrides for agent-team requests. Representative page: api/beta/messages/create.