Claude docs changes for October 6th, 2026 [diff]
Executive Summary
- Claude Code 2.1.290 is a large release:
claude attach/claude logsaccept session names, new/claude-api managed-agents-onboardcommands, a refilling WebSearch budget, many permission and sandbox hardening fixes, and mod/plugin hook additions. 2.1.291 fixes two regressions (lost permission answers in cloud sessions, lost last messages on quit). - Inference hooks gain a new
tool_callevent ("Validate tool calls"): your AI security server can now return one verdict on Claude's tool calls before any of them run, withtool_infodescribing who provides each tool. - Agent SDK custom tools docs now cover optional parameters,
Annotateddescriptions andTypedDictschemas in Python, and a newsession_state_changedmessage enabled withCLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1. - The RBAC Roles API now returns
display_name(withnamedeprecated), and the Agents API raises the tool limit from 128 to 256. - Desktop app docs describe a reworked Browser pane and a
/code-reviewcard with "Walk through in diff" and "Apply fixes"; the legacy Claude in Slack page is now limited to Pro/Max workspaces not connected to Claude Tag.
New Claude Code versions
2.1.290
New features
claude attach <name>andclaude logs <name>: part of a session name works in place of the id/claude-api managed-agents-onboard <url>sets up the Managed Agents pattern a page describes asant applyfiles, and/claude-api managed-agents-onboard <quickstart-name>builds a Console quickstart template (such asdeep-researcher) with theantCLI- Plugin hooks and mods:
serverToolUsesonturn.step,agentIdontool.check,ceilingon thetool.checkquestion and verdict,ThemeKey/Colortypings, andclaude plugin validatenow lists gating hooks and whether each has a.catch(gatingHooksunder--json) - Warnings when a managed settings file is a link to a file outside the managed settings folder, and in
/statusand doctor when managed settings ignore user-configured sandboxallowReadpaths or allowed domains - Claude apps gateway sign-in approval page has a Deny button
- [Claude Tag] Fast mode in Slack with
!fast/!fast off, and an optional Path prefixes field for custom connections in an access bundle - [VSCode] Manage plugins dialog can review and run a marketplace's install or update command; screen reader announcement when a message is queued
Existing feature improvements
- The interactive session's WebSearch budget now refills over time (100 calls/hour, set with
CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR) instead of ending after 200 calls - WebFetch now reports how much text was unread past 100,000 characters and takes an
offsetto read on /code-reviewat medium effort also reports cleanup and CLAUDE.md convention findings on models without tuned review settings (including Opus 5.5 and Sonnet 5.5)/model,/effortand/renamesent fromclaude agentsto a busy background session apply right away- Background sessions waiting on a scheduled wakeup (
/loop) are left running through updates and low memory - Faster responsiveness while resuming large sessions;
/and@suggestion lists show a ❯ pointer on the selected row CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFICalso skips the startup connection warm-up; Claude in Chromebrowser_batchtimeout raised to 90 seconds- Security-related changes: project settings can no longer turn on Claude in Chrome or set
CLAUDE_CODE_DISABLE_ATTACHMENTS;pyrightand morepsforms now ask for permission;!shell commands with raw control characters are refused - Claude apps gateway: minimum PostgreSQL version lowered from 14 to 11, certificate-expiry warnings, nicer sign-in pages
- Claude Tag channel instructions limit is now 8,192 characters instead of bytes
- [VSCode] Message timestamps are shown by default
Major bug fixes
- Fixed requests failing behind proxies/gateways that reject a Claude Code beta header with a status other than 400
- Fixed long sessions with hundreds of images getting stuck on "Request rejected as unprocessable by the model"
- Fixed scheduled tasks (
/loop, reminders) not returning on resume after compaction, never firing after a hand-off to the background, and firing extra runs on resume/respawn/fork - Fixed several permission bypasses: auto-approved read-only commands whose wildcards the shell would expand, zsh variable-name differences, deny/ask rules missed after
declare/exportprefixes or a PreToolUse hook input rewrite, Read deny rules not applying to pasted image paths, and symlink swaps during image/@-mention reads - Fixed plan mode letting the auto mode classifier approve non-read-only connector tools with a server-pushed ask policy, and plan mode not being restored on
--continue/--resume - Fixed
/ultrareviewdropping or unfiltered-uploading uncommitted changes in several git configurations - Fixed
claude --teleportand/teleportdeleting files when choosing to stash - Fixed many agent view and
claude agentsissues (Esc confirming restarts, Ctrl+X deleting whole sections, undeliverable commands replayed on restart, replies refused after a crash) - Fixed freezes and slowdowns: very long messages, secret scan/masking on long text, rewind menu with large pastes, transcript expansion with non-ASCII output
- Fixed unbounded memory use when an HTTP MCP server sends a very large response
- Fixed headless
--json-schemaruns exiting non-zero after the structured output was delivered
2.1.291
Major bug fixes
- Fixed a regression in 2.1.290 where cloud sessions could drop answers to permission prompts
- Fixed a regression in 2.1.288 where the last messages of a session could be lost when quitting
Claude Code changes
New Documents
None.
Changed documents
agent-sdk/cost-tracking [Source]
- Added an "estimates, not billing" anchor and a note that combined
total_cost_usdfigures are client-side estimates. [line 9] [Source]
agent-sdk/custom-tools [Source]
- New "Make a parameter optional" guidance:
.optional()in Zod for TypeScript; JSON Schema form (omit fromrequired, read withargs.get()) for Python. [line 12] [Source] - Input schema section is now per-language, with
.describe()for Zod andAnnotated[type, "description"]for Python field descriptions. [lines 28-31] [Source] - Added httpx install instructions (uv and pip) for Python examples that make HTTP requests. [line 38] [Source]
- Run instructions now separate TypeScript, Python (uv) and Python (pip) commands. [line 190] [Source]
- The
get_precipitation_chanceexample now uses a JSON Schema withhoursleft out ofrequired. [line 224] [Source]
agent-sdk/python [Source]
- Documented a third
tool()schema form, aTypedDictclass withNotRequiredkeys (with Python 3.10typing_extensionsnote), andAnnotateddescriptions. [lines 124-144] [Source] SystemMessagesubtypes without their own dataclass: setCLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1and readmessage.data["state"]. [line 1589] [Source]
agent-sdk/typescript [Source]
- New
SDKSessionStateChangedMessage(session_state_changed) withrunning,idleandrequires_actionstates, enabled byCLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1. [line 5344] [Source]
agent-view [Source]
claude attachandclaude logsaccept part of a session name (v2.1.290+). [lines 739-751] [Source]- Session row example and icon-shape description updated (
✻means running or needs your input); voice dictation reply works in hold mode. [lines 111-136] [Source] - Added v2.1.290 entry to the version history table. [line 972] [Source]
cli-reference [Source]
claude attach <id|name>andclaude logs <id|name>document name support. [lines 25-32] [Source]--max-budget-usdis checked against a client-side cost estimate that can differ from billing. [line 99] [Source]
desktop [Source]
- Browser pane controls reworked: Dev servers menu, Keep cookies, Clear browsing data, Browser tools toggle, and a first-click prompt for opening links in the built-in browser. [lines 107-118] [Source]
- Code review is now
/code-reviewin the prompt box, with a Code review card offering "Walk through in diff", "Fix this one" and "Apply fixes". [lines 156-163] [Source] - CI options renamed "Auto-fix CI & address comments" and "Auto-merge when ready", opened via CI in the status bar. [lines 169-172] [Source]
- SSH connection dialog fields and menu path updated; the "Disable Bypass permissions mode" admin setting entry was removed. [lines 657-664] [Source]
desktop-changelog [Source]
- Contains the same desktop updates as above (Browser pane,
/code-reviewcard, CI options, SSH dialog). [lines 107-118] [Source]
env-vars [Source]
- New
CLAUDE_CODE_EMIT_SESSION_STATE_EVENTSaddssession_state_changedmessages to the stream. [line 270] [Source] - New
CLAUDE_CODE_GZIP_REQUEST_BODIES(set0to disable gzip of API, telemetry and artifact publish bodies). [line 298] [Source] - Beta tracing endpoint is now specified as an OTLP/HTTP endpoint. [line 178] [Source]
feature-availability [Source]
fullscreen [Source]
- Clicking a file path on Linux/WSL needs a file manager providing
org.freedesktop.FileManager1. [line 98] [Source]
hipaa-setup [Source]
- Versions older than v2.1.285 ignore
allowedProviders; suggestedrequiredMinimumVersionas a workaround. [line 149] [Source] - Session data removal:
claude purgeon v2.1.288+,claude project purgeon v2.1.126-2.1.287. [lines 233-247] [Source]
keybindings [Source]
confirm:nextFieldnow documented as toggling fast mode in the/fastdialog;confirm:previousFieldandpermission:toggleDebugare no-op actions that stay valid inkeybindings.json. [lines 142-181] [Source]
memory [Source]
- Subdirectory CLAUDE.md files load on demand when Claude uses Read, Write and other file tools; a
Loadedline appears and they don't show under Memory files in/context. [lines 138-140] [Source] - The built-in AGENTS.md plugin ID is
cc-plugin-agents-md@builtin(wasagents-md@builtinbefore v2.1.285). [lines 376-390] [Source] - Debugging tips for on-demand CLAUDE.md loading. [line 564] [Source]
claude-md [Source]
- Same on-demand loading and
cc-plugin-agents-md@builtinplugin ID updates as in memory. [lines 138-140] [Source]
permission-modes [Source]
- Auto mode default-block list is merged into one list (no longer split at v2.1.200). [lines 367-376] [Source]
- Protected directories:
.claudeexceptions now list worktrees, the session's plan files, background session scratch directories, and project/subagent memory markdown files. [lines 626-632] [Source]
plugins/mods/overview [Source]
- Mods work in the Desktop app from v2.1.286 (terminal needs v2.1.287); added how to check the version in each. [line 89] [Source]
settings-reference [Source]
- In cloud sessions Claude Code honors only a subset of
defaultModevalues;@builtinplugin config example updated to the new AGENTS.md plugin ID. [lines 1631-1643] [Source]
slack [Source]
- The earlier Claude in Slack now only answers Pro and Max accounts in workspaces not connected to Claude Tag; Team/Enterprise should use Claude Tag. [lines 3-9] [Source]
- New troubleshooting entries for the "This workspace isn't set up for Claude Tag yet" and "legacy Claude in Slack bot is retired" notices. [lines 166-190] [Source]
API changes
New Documents
None.
Changed documents
api/beta/agents [Source]
- Maximum tools across all toolsets raised from 128 to 256 for Create and Update Agent (also mirrored in the per-language SDK reference pages and
beta.md). [line 418] [Source]
api/beta/organization/rbac_roles [Source]
- RBAC roles now expose
display_name;nameis deprecated and always equals it. Anthropic-created role names may differ from claude.ai labels and can change, so store the roleid. List and retrieve responses showdisplay_name. (Mirrored inbeta/organization.md,rbac_roles/list.md,rbac_roles/retrieve.mdand the SDK reference pages.)
manage-claude/inference-hooks [Source]
- Two hook events now exist:
promptandtool_call(with Validate tool calls on). [line 15] [Source] tool_callevents may omit some claude.ai-internal tools; Claude Tag measurement requests are not sent to your endpoint. [lines 82-84] [Source]
manage-claude/inference-hooks-configuration [Source]
- New Validate tool calls setting (off by default, requires Enforce verdicts, takes about a minute to propagate). [line 92] [Source]
manage-claude/inference-hooks-endpoint [Source]
- New "The tool call frame" section:
typeis"tool_call",messagesholds only the latest assistant message, one verdict covers all calls, and eachtool_useblock carriestool_info. [line 241] [Source] tool_infokinds documented with fields and examples:platform,application,third_party(withorigin/verified_origin) andclient. [lines 255-330] [Source]- Request field table:
typeis"prompt"or"tool_call", andrequest_idis a per-frame identifier. [lines 161-170] [Source]
models/fable-5/introducing-claude-fable-5-and-claude-mythos-5 [Source]
- Removed statements that Claude Mythos 5 lacks the safety classifiers Claude Fable 5 has (also in fable-5
overviewandmigration-guide). [line 15] [Source]
models/fable-5-1/overview [Source]
- Recommended starting model changed from Claude Opus 5 to Claude Opus 5.5 (also in
whats-new-fable-5-1); the safety classifiers divergence bullet was dropped from the Fable 5.1 migration guide. [line 21] [Source]
models/opus-5-5/migration-guide [Source]
- On Amazon Bedrock, structured outputs aren't available for Claude Opus 5.5: use system prompt instructions for prefill replacement, and
autoalone instead ofany/tooltool choice. [lines 32-281] [Source]