Claude docs changes for August 26th, 2026 [diff]

Executive Summary

• Claude Code 2.1.246 adds a startup warning for Bash allow rules whose wildcard sits before the subcommand (e.g. Bash(git * main)), since such rules can approve options like git's -c that run arbitrary commands, plus a new Auto mode tab in /permissions for editing classifier rules without hand-editing settings files.
• /goal idle check-ins are now capped at three per goal between prompts instead of running indefinitely, and /code-review can now start itself automatically on more platforms (Bedrock, Vertex AI, Foundry, the Claude apps gateway) since it now follows feature-flag availability instead of a hardcoded exception list.
• The permissions docs got a rewrite of Bash wildcard matching with a full table of examples, clarifying exactly where * can go and what it does and doesn't match.
• Anthropic split its single sprawling model migration guide into one page per model (Opus 5, Sonnet 5, Haiku 4.5, Fable 5/Mythos 5); the old combined page is now just an index linking to them.
• The Linux desktop app documents Cowork's virtualization requirements (KVM, QEMU, /dev/kvm access) and new troubleshooting for unmet dependencies and Cowork startup failures, and admins gained a disableDesktopLocalSessions managed setting to force developers onto SSH or cloud sessions.

New Claude Code versions

2.1.246

New features

• Added a startup warning for Bash allow rules with a wildcard before the subcommand (e.g. Bash(git * main)), since they also match options inserted before the subcommand
• Added an Auto mode tab to /permissions for viewing and editing auto mode classifier rules
• Added the turn's completion time to the end-of-turn duration line, e.g. ✻ Sautéed for 23s · done 6:05 PM

Existing feature improvements

• Improved /cd: the new directory's project settings, hooks, .mcp.json servers, skills, and agents now take effect right after the move instead of on --resume
• Improved Bash tool latency on bash shells by replaying snapshot functions without a base64 subshell per function
• Improved subagent results: a subagent that stops at its maxTurns limit now returns its output marked as partial, with a hint to continue it via SendMessage, instead of appearing finished
• Improved non-interactive sessions (-p, SDK, cloud sessions) to automatically continue a response cut off mid-stream by a server error, connection loss, or stall instead of ending with an error
• Improved attribution of usage telemetry to your organization for workload identity federation sessions and other edge cases
• Changed /code-review so Claude can also start it on its own on Bedrock, Vertex AI, and Foundry, through the Claude apps gateway, and when telemetry or non-essential traffic is disabled
• /goal: Changed idle sessions to start at most three check-ins on long-running background work per goal; your next message allows three more
• Changed claude install and claude update to defer a pending managed-settings consent prompt to the next interactive session instead of prompting mid-command
• Changed OpenTelemetry plugin events for plugins synced from claude.ai: plugin_id_hash now reflects the plugin's real marketplace, and enabled_via is admin-install for admin-installed plugins

Major bug fixes

• Fixed a severe transcript slowdown when a diff contained a very long single line (e.g. a base64 string); such lines now render truncated with a marker
• Fixed fullscreen mode showing a blank transcript after resizing the terminal, and erratic fullscreen scrolling including jump-to-bottom getting stuck mid-transcript
• Fixed background sessions failing to open after 45 seconds when the starting directory had been deleted, the machine had slept, or the host is slow to start processes
• Fixed background sessions failing to open with an EACCES error when another Claude Code process was re-installing the npm package at that moment
• Fixed markdown rendering being disabled for a whole message when its first 500 characters contained no markdown
• Fixed MCP tool calls interrupted by an incoming message in headless/remote sessions being reported as "completed with no output" instead of an explicit interrupted error
• Fixed MCP tool arguments being sent as JSON strings when the parameter's schema is empty ({}), instead of their real type
• Fixed the background retention sweep removing git worktrees under .claude/worktrees/ that you created yourself when an old background-session record pointed at them
• Fixed auto mode tool calls being denied as "temporarily unavailable" on very large sessions by scaling the safety-check deadline with prompt size
• Fixed the Write tool reporting "Out of memory" or freezing for a long time after overwriting a very large existing file, even though the file had been written
• Fixed resumed sessions failing every turn with a 400 when the saved history contains tool blocks the Anthropic API does not accept (typically written by a third-party API proxy)
• Fixed telemetry and metrics requests to Anthropic carrying the API key configured for a third-party gateway (ANTHROPIC_BASE_URL); a credential is now only sent to its own host
• Fixed sessions that ended in plan mode resuming outside plan mode in the VS Code extension, and in claude -p --continue/--resume with a permission prompt tool, when no permission mode was set
• Fixed Bash permission checks to always require approval for malformed commands with a dangling && or || operator
• Fixed --strict-mcp-config sessions prompting to approve .mcp.json servers they would never load, which left background sessions waiting at startup
• Fixed memory growing with session length in the fullscreen and Ctrl+O transcript views
• Fixed the UI stopping with a render error on the first tool call when a third-party Anthropic-compatible endpoint (ANTHROPIC_BASE_URL) streams a tool_use block without an id
• Fixed curl -fsSL https://claude.ai/install.sh | bash failing with "Raw mode is not supported" for some Team/Enterprise users with server-managed settings
• Fixed the command sandbox's filesystem configuration not respecting --setting-sources

Claude Code changes

Changed documents

amazon-bedrock [Source]

• Before running the awsAuthRefresh command, Claude Code now first makes an STS GetCallerIdentity call to confirm credentials are actually expired, sent through your proxy configuration; before v2.1.239 it hung at startup on proxy-only networks. [line 113] [Source]
• New CLAUDE_CODE_DISABLE_BEDROCK_CONTENT_TYPE_DEFAULT guidance for a streaming response with a missing/empty Content-Type header: Claude Code now defaults to decoding it as the binary event-stream format instead of erroring. [line 463] [Source]

artifacts [Source]

• Softened wording on comment reading in the first session after install/upgrade: Claude "might not be able to" read comments yet, rather than definitely can't. [line 94] [Source]

auto-mode-config [Source]

• New "Edit rules from /permissions" section: the Auto mode tab (v2.1.246+) lists allow/soft_deny/hard_deny/environment entries per scope, lets you add/edit/delete rules and toggle built-in defaults, and edit environment as a document in your editor. [lines 16, 242-250] [Source]

claude-platform-on-aws [Source]

• New sales/pricing banner added at the top of the page ("Deploying Claude Code across your organization?"). [line 3] [Source]

code-review [Source]

• Removed the hardcoded list of sessions where Claude could never auto-start /code-review (cloud providers, Claude apps gateway, privacy env vars). As of v2.1.246, Claude auto-starts it wherever feature-flag fetching reaches it; before that, it depended on an Anthropic-side flag and sessions without flag fetching only ran it when typed. [line 323] [Source]

commands [Source]

• /cd's directory-suggestion autocomplete now requires Claude Code v2.1.206 or later. [line 46] [Source]
• /permissions description now mentions the new Auto mode tab. [line 93] [Source]
• /skills: removed the "then Enter to save" step for cycling a skill's visibility. [line 120] [Source]
• / menu highlighting rules (matching from the start of a command name/alias or a word within it) now require Claude Code v2.1.236 or later; before that, pressing Enter after a typo ran the closest match instead of submitting as typed. [lines 148-149] [Source]

costs [Source]

• Goal check-ins are now capped at three idle check-ins per goal between prompts (previously uncapped), reducing token spend from indefinite idle check-ins. [line 270] [Source]

desktop-linux [Source]

• New "Cowork requirements" section: documents hardware virtualization (KVM), QEMU/UEFI firmware packages, and /dev/kvm//dev/vhost-vsock access needed to run Cowork's VM on Linux. [line 14] [Source]
• New "Unmet dependencies" troubleshooting section for libc6 version mismatches and downloading the wrong-architecture .deb. [line 124] [Source]
• New "Cowork isn't available" troubleshooting section mapping each Cowork error message to its fix. [line 135] [Source]

desktop [Source]

• New "Local sessions on managed devices" section documenting the disableDesktopLocalSessions managed setting, which grays out the Local (and Windows WSL) environment option and defaults new sessions to SSH. [line 562] [Source]

env-vars [Source]

• New CLAUDE_CODE_DISABLE_BEDROCK_CONTENT_TYPE_DEFAULT variable to opt out of the new default binary-decode behavior for Bedrock streams missing a Content-Type header. [line 213] [Source]
• Clarifies that flag fetching is also off through the Claude apps gateway and on Bedrock/Claude Platform on AWS/Google Cloud/Foundry (unless a host sets CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST); removed the note that flag-off sessions can never auto-start /code-review (see code-review.md change above). [line 462] [Source]

errors [Source]

• New "Streaming response ended before any complete data was received" entry: a warning (shown once per session) now surfaces this proxy/gateway condition instead of Claude Code silently retrying without streaming. [line 1042] [Source]
• New "Has a wildcard before the rest of the command" configuration warning for Bash allow rules like Bash(git * main), explaining the -c/--exec-path-style risk and how to narrow the rule. [line 2367] [Source]

fast-mode [Source]

• The VS Code extension now supports fast mode, following the fastMode setting and offering a "Toggle fast mode" command, reversing the prior "not supported in VS Code" note. [line 10] [Source]

feature-availability [Source]

• Removed the "/code-review runs when you type it; Claude doesn't start it on its own" partial-support bullet from the Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry tabs, matching the code-review.md behavior change above. [Source]

goal [Source]

• Idle check-ins are now capped at three per goal between your prompts (the third one says check-ins are paused until you prompt again); before v2.1.246 they were uncapped. [lines 17, 127] [Source]

index [Source], overview [Source]

• Clarifies that the desktop app bundles Claude Code, so installing it means you don't need the CLI separately. [line 82] [Source]

keybindings [Source]

• scroll:lineUp/scroll:lineDown now default to wheelup/wheeldown instead of being unbound, and those two key names are now documented as available bindings. [lines 325-326, 385] [Source]
• Key parsing is now fully case-insensitive: K and ctrl+K both behave the same as their lowercase forms; write shift+k to require Shift explicitly. This replaces the old rule where a bare uppercase letter implied Shift. [line 367] [Source]
• Validation now also warns about invalid action values (not a string or null); removed the "terminal multiplexer conflicts" warning category. [line 469] [Source]

managed-settings [Source]

• Clarifies that some server-managed changes now wait for the next launch rather than applying to the running session or on next poll (see server-managed-settings.md change below). [line 78] [Source]
• Adds "Desktop local-session" to the list of managed-only keys not covered by the permission/plugin/delivery table, referencing the new disableDesktopLocalSessions key. [line 222] [Source]

monitoring-usage [Source]

• plugin.scope gains a new "community" value, and enabled_via gains a new "admin-install" value for plugins set to required/auto-install in Organization settings; before v2.1.246 these reported as "user-install" or "seed-mount". plugin_id_hash for claude.ai-synced plugins now incorporates the marketplace name claude.ai reports (or synced otherwise). [lines 843-844] [Source]

permission-modes [Source]

• Splits out the "first session after install/upgrade" case: a non-interactive session can now pick up feature flags in time to start in auto normally, whereas an interactive terminal session still defaults to default until flags arrive. [line 64] [Source]

permissions [Source]

• Notes that the /permissions dialog now also shows auto mode classifier rules under a new Auto mode tab when auto mode is available. [line 58] [Source]
• Rewrote the "Wildcard patterns" section: explains that Claude Code now warns at startup about a rule whose * sits before the subcommand, and adds a full table of example rules with what they match/don't match (e.g. Bash(ls *) vs Bash(ls*), trailing-* word-boundary behavior). Removed the duplicate, less-detailed wildcard explanation from the Bash-specific subsection below it. [lines 128, 148-159, 190] [Source]

quickstart [Source]

• Narrows the auto-mode-by-default claim to interactive terminal sessions specifically, dropping the earlier mention that it also applied to VS Code sessions. [line 151] [Source]

server-managed-settings [Source]

• New behavior for Claude apps gateway clients: they fetch settings from the gateway and wait for that fetch before the session starts, and exit at startup if it fails (with a claude auth exemption), rather than fetching asynchronously like other sessions. [line 126] [Source]
• Clarifies that some settings updates (OpenTelemetry exporter config, the model key, and removing an env variable) now only apply at the next launch, not to the running session; a tampered cached-settings file for one of these keys also survives a fresh server fetch until relaunch. [lines 112, 150, 248] [Source]

settings-reference [Source]

• New disableDesktopLocalSessions managed setting: grays out on-device Code sessions (and Windows WSL) in the desktop app, defaulting new sessions to a configured SSH connection; the terminal CLI ignores it. [lines 65, 4403-4417] [Source]

statusline [Source]

• Status line scripts now run immediately when you save a config change to the command itself, skipping the usual 300ms debounce used for other update triggers. [lines 121, 136] [Source]
• Documents that cost.total_cost_usd correctly resets to $0 on /clear, noting it used to carry over before v2.1.211. [line 161] [Source]

terminal-config [Source]

• Notes /terminal-setup's GPU acceleration change has applied since v2.1.157. [line 27] [Source]
• Clarifies that userMessageBackground, bashMessageBackgroundColor, and memoryBackgroundColor theme tokens apply in both the default and fullscreen renderers, not fullscreen only. [line 224] [Source]

voice-dictation [Source]

• The hold space to speak footer hint now only shows for your first three sessions with voice dictation enabled, instead of persisting forever. [line 47] [Source]
• Hold mode's footer now shows listening… (was a live waveform) once recording starts, plus a new mic-level-reactive cursor bar; tap mode's footer now shows ● REC · tap to send instead of a waveform. [lines 53, 71] [Source]
• The repeated-failure pause now triggers on three capture failures within a 10-second window (was an unspecified threshold), and the missing-audio-tool error message changed to Voice mode requires SoX for audio recording. [lines 144, 147] [Source]

API changes

New Documents

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

New standalone page covering migration to Claude Opus 5 from Opus 4.8, 4.7, and 4.6/earlier: thinking-on-by-default, the effort cap on disabling thinking, pricing (unchanged from 4.8), Priority Tier unavailability, the 1M context window, mid-conversation system messages/tool changes, refusal handling, and per-source checklists. This content was previously part of the single combined about-claude/models/migration-guide.md page.

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

New standalone page covering migration to Claude Sonnet 5 from Sonnet 4.6, 4.5/earlier, and Haiku 4.5: the new tokenizer (~30% more tokens), adaptive thinking on by default, removed sampling parameters, and new real-time cybersecurity safeguards. Split out from the old combined migration guide.

models/haiku-4-5/migration-guide [Source]

New standalone page covering migration to Claude Haiku 4.5 from Haiku 3.5/earlier: sampling parameter restrictions, updated tool versions, the refusal stop reason, and new rate limits. Split out from the old combined migration guide.

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

New standalone page covering migration to Claude Fable 5 and Claude Mythos 5 from Mythos Preview and Opus 4.8: always-on adaptive thinking, no prefill, $10/$50 per-MTok pricing, 30-day-minimum data retention, and safety classifier refusals (Fable 5 only). Split out from the old combined migration guide.

Changed documents

about-claude/models/migration-guide [Source]

• This page shrank from ~980 lines to a short index: it's now just a bulleted list of links to the four new per-model migration guides above, plus a "Get help" section. All the detailed migration content moved to those pages. [lines 1-13] [Source]

Note: several other API pages (agent-skills/claude-api-skill, claude-prompting-best-practices, prompting-claude-opus-4-8, prompting-claude-opus-5, refusals-and-fallback, the opus-4-5/opus-4-6/opus-4-7/opus-4-8/sonnet-4-5/sonnet-4-6 model overviews, opus-5/whats-new-opus-5, sonnet-5/whats-new-sonnet-5, models/overview, and release-notes/overview) changed today only to repoint existing links from the old combined migration guide to the new per-model pages above; no wording or content changed, so they're omitted here.