Claude docs changes for October 6th, 2026 [diff]

Executive Summary

  • Claude Code 2.1.290 is a large release: claude attach/claude logs accept session names, new /claude-api managed-agents-onboard commands, 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_call event ("Validate tool calls"): your AI security server can now return one verdict on Claude's tool calls before any of them run, with tool_info describing who provides each tool.
  • Agent SDK custom tools docs now cover optional parameters, Annotated descriptions and TypedDict schemas in Python, and a new session_state_changed message enabled with CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1.
  • The RBAC Roles API now returns display_name (with name deprecated), and the Agents API raises the tool limit from 128 to 256.
  • Desktop app docs describe a reworked Browser pane and a /code-review card 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> and claude 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 as ant apply files, and /claude-api managed-agents-onboard <quickstart-name> builds a Console quickstart template (such as deep-researcher) with the ant CLI
  • Plugin hooks and mods: serverToolUses on turn.step, agentId on tool.check, ceiling on the tool.check question and verdict, ThemeKey/Color typings, and claude plugin validate now lists gating hooks and whether each has a .catch (gatingHooks under --json)
  • Warnings when a managed settings file is a link to a file outside the managed settings folder, and in /status and doctor when managed settings ignore user-configured sandbox allowRead paths 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 offset to read on
  • /code-review at 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, /effort and /rename sent from claude agents to 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_TRAFFIC also skips the startup connection warm-up; Claude in Chrome browser_batch timeout raised to 90 seconds
  • Security-related changes: project settings can no longer turn on Claude in Chrome or set CLAUDE_CODE_DISABLE_ATTACHMENTS; pyright and more ps forms 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/export prefixes 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 /ultrareview dropping or unfiltered-uploading uncommitted changes in several git configurations
  • Fixed claude --teleport and /teleport deleting files when choosing to stash
  • Fixed many agent view and claude agents issues (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-schema runs 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_usd figures 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 from required, read with args.get()) for Python. [line 12] [Source]
  • Input schema section is now per-language, with .describe() for Zod and Annotated[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_chance example now uses a JSON Schema with hours left out of required. [line 224] [Source]

agent-sdk/python [Source]

  • Documented a third tool() schema form, a TypedDict class with NotRequired keys (with Python 3.10 typing_extensions note), and Annotated descriptions. [lines 124-144] [Source]
  • SystemMessage subtypes without their own dataclass: set CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1 and read message.data["state"]. [line 1589] [Source]

agent-sdk/typescript [Source]

  • New SDKSessionStateChangedMessage (session_state_changed) with running, idle and requires_action states, enabled by CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1. [line 5344] [Source]

agent-view [Source]

  • claude attach and claude logs accept 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> and claude logs <id|name> document name support. [lines 25-32] [Source]
  • --max-budget-usd is 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-review in 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-review card, CI options, SSH dialog). [lines 107-118] [Source]

env-vars [Source]

  • New CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS adds session_state_changed messages to the stream. [line 270] [Source]
  • New CLAUDE_CODE_GZIP_REQUEST_BODIES (set 0 to 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]

  • Claude Code in Slack is listed as Pro and Max only. [line 44] [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; suggested requiredMinimumVersion as a workaround. [line 149] [Source]
  • Session data removal: claude purge on v2.1.288+, claude project purge on v2.1.126-2.1.287. [lines 233-247] [Source]

keybindings [Source]

  • confirm:nextField now documented as toggling fast mode in the /fast dialog; confirm:previousField and permission:toggleDebug are no-op actions that stay valid in keybindings.json. [lines 142-181] [Source]

memory [Source]

  • Subdirectory CLAUDE.md files load on demand when Claude uses Read, Write and other file tools; a Loaded line 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 (was agents-md@builtin before 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@builtin plugin 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: .claude exceptions 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 defaultMode values; @builtin plugin 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; name is deprecated and always equals it. Anthropic-created role names may differ from claude.ai labels and can change, so store the role id. List and retrieve responses show display_name. (Mirrored in beta/organization.md, rbac_roles/list.md, rbac_roles/retrieve.md and the SDK reference pages.)

manage-claude/inference-hooks [Source]

  • Two hook events now exist: prompt and tool_call (with Validate tool calls on). [line 15] [Source]
  • tool_call events 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: type is "tool_call", messages holds only the latest assistant message, one verdict covers all calls, and each tool_use block carries tool_info. [line 241] [Source]
  • tool_info kinds documented with fields and examples: platform, application, third_party (with origin/verified_origin) and client. [lines 255-330] [Source]
  • Request field table: type is "prompt" or "tool_call", and request_id is 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 overview and migration-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 auto alone instead of any/tool tool choice. [lines 32-281] [Source]

models/sonnet-5-5/migration-guide [Source]

  • On Amazon Bedrock, structured outputs aren't available for Claude Sonnet 5.5; describe the format in the prompt or use a non-strict tool and validate in code. [line 1200] [Source]