Claude docs changes for October 2nd, 2026 [diff]
Executive Summary
- Claude Mods: Claude Code 2.1.287 introduces mods, which are plugins made of JavaScript/TypeScript event handlers that run inside Claude Code. They can draw panes and bands, redraw built-in UI, intercept tool calls and add commands. A new 10-page
plugins/mods/guide set covers them, along with admin controls (prependPlugins/appendPlugins). - Sandboxing docs reworked: the sandboxing guide was reorganized around "what the sandbox restricts", credential masking,
excludedCommandsrules and a new "admin-required" sandbox that ignores repository-level sandbox settings. - Self-hosted sandboxes for Managed Agents: the single guide was split into five pages covering workers, custom tools, memory, operations and a reference.
- New Admin API endpoints: organization analytics (usage, cost, per-user reports, summaries, chat projects), spend limits (set, list, effective) and RBAC group member add/remove.
- New diff panel and usage-limit auto-continue in interactive mode, plus a big expansion of the error reference.
New Claude Code versions
2.1.287
New features
- Added Claude Mods: plugins can now modify deeper behavior of Claude Code
- Added "You should know", a built-in mod where a side agent flags things you or Claude might miss (
/plugin enable cc-plugin-you-should-know@builtin) - Added an
n:<text>filter to the agents view that matches session names and tasks - Added
prompt_textto the OpenTelemetryuser_promptevent - Added URL prompts from MCP servers on the 2025-11-25 protocol, for example to sign in (servers that stop connecting may need
"bareElicitationCapability": true) - Windows: added a startup warning when denying the Bash tool also turns off PowerShell
- Self-hosted runner: added a built-in REST-only
gh apiwhere the GitHub CLI isn't installed - VS Code: added "Run in background" for running commands and sub-agents, and background shell/Monitor output in agent map cards
Existing feature improvements
/configis easier to navigate: cycling settings step both ways with ←/→, narrow terminals stack values, PgUp/PgDn page the list- Clearer plugin marketplace errors, plugin dependency notes, and retry of unfinished plugin installs
- Opus 4.7+ and Fable now use a 1M context window by default on Bedrock, Vertex, Foundry and the Claude apps gateway (
CLAUDE_CODE_DISABLE_1M_CONTEXT=1keeps 200K) - MCP
alwaysLoad: falsenow defers all of that server's tools behind tool search - Better handling of large MCP tool results (less memory, smaller session files)
- Automatic model switches after a flagged message keep your current effort level
- Replies from
claude agentsarrive as queued messages; waiting permission prompts show oldest first - Whole-tool
Bashallow rules and allowing hooks now prompt for shell writes to files the file tools refuse (profile store, credentials file) - Files sent from cloud/Remote Control sessions are retried once on transient failures and large files stream from disk
- Windows: faster Bash tool
Major bug fixes
- Fixed a dangerous
rmlosing its always-ask safeguard when the command also redirected output to a~or wildcard path - Fixed Remote Control missing messages for minutes after an unanswered reconnect request, and failing to register behind an HTTP proxy
- Fixed switching between Opus 5.5 and Sonnet 5.5 dropping earlier extended thinking
- Fixed Bedrock Guardrails mid-response blocks ending the turn with an API error
- Fixed a folder's CLAUDE.md being attached twice after resume or compaction
- Fixed cloud sessions losing earlier conversation when restarted during compaction
- Fixed
/advisorpairing checks (Sonnet 5.5 can advise Opus 4.7 and 4.8) - Fixed Windows interactive
claudehanging with piped input - Many screen reader mode fixes
Claude Code changes
New Documents
Mods admin [Source]
How organizations manage mods: running org mods before or after user-installed ones with prependPlugins / appendPlugins, and controlling what mods may do.
Mods API [Source]
API a mod can call, including adding a /command or a tool that runs your function with no Claude turn.
Mods create [Source]
How to ask Claude to write a mod, or write one yourself, and load it with --plugin-dir.
Mods events [Source]
Event hooks a mod can handle: guarding or changing tool calls, following a turn, and so on.
Mods gallery [Source]
Example mods with code.
Mods interface [Source]
Drawing panes and bands with tabs, buttons and text fields, and changing what Claude Code already draws (tool rows, spinner, dialogs).
Mods overview [Source]
What mods are, how they differ from settings hooks, skills and MCP servers, how to install them, sample mods, trust considerations, and where mods run (CLI and Desktop Code tab).
Mods reference [Source]
Reference for mod files, hooks and calls.
Mods test [Source]
How to test a mod.
Mods troubleshoot [Source]
Troubleshooting mods that don't load or behave as expected.
Changed documents
desktop [Source]
- Added an "Auto mode availability" section: available to all Anthropic API users on Opus 4.6+, Sonnet 4.6+ or Fable.
env-vars [Source]
- New
CLAUDE_CODE_DISABLE_AUTH_REFRESH_LOCKandCLAUDE_CODE_DISABLE_WEB_FETCHvariables. [line 228] [line 263] [Source] CLAUDE_AX_PREPARK_MSdefault changed from50to0. [line 188] [Source]ANTHROPIC_FOUNDRY_RESOURCEnow refuses a URL or host name. [line 163] [Source]
errors [Source]
- Many new entries: advisor less capable than main model, Foundry resource must be a resource name, piped input in interactive mode, GitHub IP allow list/SSO/Conditional Access blocks,
/recapin routines, WebFetch preflight failures, network-share publish failures, and expanded sign-in and retry guidance.
hooks [Source]
- Notes that plugins can register function hooks (mods). [line 9] [Source]
/hooksbrowser description simplified; added anAll eventsentry. [lines 670-678] [Source]
interactive-mode [Source]
- New diff panel (v2.1.287+) in fullscreen rendering:
askbutton to attach a file's diff, per-turnsourcepicker, andCtrl+X Bto change the comparison base. The classic renderer keeps a diff dialog. - Documented auto-continue after a usage limit resets, including the case where the computer slept.
model-config [Source]
- New "Effort level after a fallback" section: the fallback model keeps the flagged request's effort level. [lines 517-528] [Source]
memory [Source]
- Auto memory toggle can't be turned back on in background sessions or sessions started by another Claude Code session. Clarified lazy loading of subdirectory
CLAUDE.mdfiles and path-scoped rules.
permission-modes [Source]
- Server-side auto mode review now applies to
-p, Agent SDK, VS Code and desktop sessions on any plan. [line 321] [Source] - A mod handling
tool.checkcan approve actions before the classifier. [line 495] [Source] - Reading another organization's public artifact needs approval, which
dontAskmode can't give. Directories loaded with--plugin-dirare now protected. - Removed "as of v2.1.198" version qualifiers from classifier block rules.
permissions [Source]
- Documented how mods extending permissions interact with ask rules, PreToolUse blocks, the classifier and deny rules.
plugins/loading [Source]
- New rules for installing a plugin's npm/Bun dependencies from a lockfile: registry packages only, pinned versions,
httpsdownloads, no overrides or patches.
prompt-library [Source]
- Prompts reworked to use fill-in
{slot}placeholders with example values and "teaches" descriptions.
sandboxing [Source]
- Reorganized with "What the sandbox restricts" and "What runs outside the sandbox" sections (file/web tools, hooks, MCP servers,
!commands, excluded commands, unsandboxed retries). - Explained how each permission mode handles the unsandboxed retry, and added strict sandbox mode.
- Added
excludedCommandsmatching rules and "admin-required" sandbox, where repository settings are ignored. - Credential masking now covers environment variables and credential files, with
injectHosts, TLS termination and macOS/Linux differences.
settings-reference [Source]
- New
prependPluginsandappendPluginssettings for org-run mods. [line 556] [Source] - Managed permission-only setting now also makes v2.1.282+ ignore
allowed-toolsfrontmatter in untrusted skills. [line 1339] [Source] - Sandbox settings now note limits on project/local scopes;
failIfUnavailablebehavior clarified;excludedCommandsskipsgit clone/init/worktreewith outside paths. [line 1817] [Source] - AWS credential re-signing pairing rules documented. [line 2386] [Source]
claudeInChromeDefaultEnablednow also applies to the VS Code extension. [line 583] [Source]
API changes
New Documents
Add RBAC Group Member [Source]
POST /v1/organizations/rbac_groups/{rbac_group_id}/members adds a user to an RBAC group. SCIM-provisioned groups can't be modified this way.
Remove RBAC Group Member [Source]
Removes a user from an RBAC group.
Analytics reports [Source]
New organization analytics endpoints under /v1/organizations/analytics/: summaries (activity summaries), usage_report, cost_report, user_usage_report and user_cost_report (per-user token usage and cost), and apps/chat/projects (per-project chat activity with cursor pagination). Each has a list page.
Spend limits [Source]
New POST /v1/organizations/spend_limits to set (upsert by scope and period) a spend limit, plus list and effective (list effective spend limits) endpoints.
Self-hosted sandboxes: custom tools [Source]
Serve custom tools from a self-hosted sandbox worker, and wrap an MCP server inside your network as custom tools without a tunnel.
Self-hosted sandboxes: memory [Source]
Attach memory stores to self-hosted sandbox sessions: prepare the host, configure sync, handle read-only stores and conflicts.
Self-hosted sandboxes: operations [Source]
Read queue depth, stop sessions and workers without losing work, and fix common failures.
Self-hosted sandboxes: reference [Source]
ant CLI flags, environment variables, host requirements, filesystem paths and SDK helper options for workers.
Self-hosted sandboxes: workers [Source]
How workers claim work (always-on or webhook-triggered) and where sessions run (one process or one sandbox per session).
Changed documents
Self-hosted sandboxes [Source]
- Trimmed to an overview; detailed content moved into the five new sub-pages above.
Beta API reference (messages, memory stores, organization, SDKs) [Source]
- Added the
spend-limit-reads-2026-09-26beta header value across the beta endpoints and the language SDK references (Go, Java, Python, TypeScript, Ruby, C#, PHP, CLI). - The spend limits, increase requests and organization analytics (users, skills, connectors) references were substantially regenerated; much of the remaining churn in the beta pages is field reordering.