Claude docs changes for August 19th, 2026 [diff]

Executive Summary

• The Workbench is now called Playground across the Claude Console and API docs — same tool, new name, referenced consistently in dozens of pages
• Claude Code 2.1.235 adds an opt-in spellcheck setting that underlines misspelled words in the prompt input using a locally installed aspell, hunspell, or ispell
• Workload Identity Federation token exchange failures now return a uniform 401 authentication_error (Authentication failed) instead of 400 invalid_grant, with deny reasons visible on a new authentication history page instead of hidden server-side
• /goal now clears itself on unrecoverable errors (auth failure, exhausted credits, context overflow, unavailable model) and periodically asks Claude to check on long-running background work that's holding up evaluation
• Files API storage limit doubled from 500 GB to 1 TB per organization, and both the Files API and Skills API now log create/delete/download operations to the Compliance API Activity Feed

New Claude Code versions

2.1.235

New features

• Added an optional spellcheck setting that underlines misspelled words in the prompt input as you type, using your installed aspell, hunspell, or ispell

Existing feature improvements

• Improved memory and CPU usage while cloud sessions such as /ultrareview or /autofix-pr run in the background — their event streams are no longer re-scanned and re-rendered on every update
• Improved permission dialogs: display text and "don't ask again" options now always match what a grant would cover, and "don't ask again" is withheld when contents cannot be fully displayed
• Improved the embedded grep in native macOS/Linux builds: pathological patterns now fail fast instead of exhausting memory, and -m N with -A/-C prints correct context
• Improved the context-limit error to say when auto-compact is off and point to /config to re-enable it

Major bug fixes

• Fixed whole-prompt-cache invalidation when a language server disconnected or reconnected mid-session
• Fixed nested markdown list items misaligning at depth 3+ and added a hanging indent to wrapped list items in the terminal UI
• Fixed prompt input highlights (slash commands, keywords, mentions) appearing shifted by one or more characters in some multi-line prompts
• Fixed Shift+Tab inside the permission prompt's comment field approving the edit and granting session-wide edit permission instead of closing the field
• Fixed the Agent tool advertising a general-purpose default in sessions where that agent is unavailable: an omitted subagent_type there now gets a clear error listing the available agents
• Fixed notebook cell delete/replace approval dialogs silently omitting the existing cell content when the notebook or cell could not be read; the dialog now says why
• Fixed slash commands run while Claude is responding showing HTML entities instead of the actual characters
• Fixed the prompt footer not showing the "Update installed" restart notice after a background auto-update
• Fixed the expanded task list (ctrl+t) always starting collapsed when resuming or relaunching into a session that still has open tasks
• Vim mode: NORMAL mode and cursor position are now preserved when toggling the detailed transcript (ctrl+o) or closing a panel
• Dialogs: arrow keys and Enter pressed in quick succession now select the option you navigated to instead of the previously highlighted one
• SendMessage now refuses messages too large for cross-session delivery up front instead of silently dropping them
• Remote Control: claude rc now applies the same enterprise-gateway availability check as interactive startup
• [VSCode] Fixed focus jumping between open Claude tabs on its own when a window with several Claude panels is restored or reloaded

Claude Code changes

Changed documents

agent-loop [Source]

• Minor wording clarifications around CLAUDE.md summarization instructions and error handling; the "manual compaction" doc link now points at the merged commands section in skills.md. [line 236] [Source]

skills [Source]

• This page was rewritten and now also serves as the SDK's command documentation (the old slash-commands.md content was folded in): it covers discovering commands via slash_commands in the init message, dispatching commands by name, using /compact, and pre-approving tools for skills. [lines 1-98] [Source]
• New "Commands in Agent SDK sessions" section distinguishes built-in commands, bundled skills, user skills, and legacy command files. [line 98] [Source]
• New "Pre-approve tools for skills" section replaces the old tool-restrictions guidance, clarifying that allowed-tools frontmatter only applies to the CLI, not SDK sessions. [line 324] [Source]
• New "Invalid skill name error" section documents the exact TypeScript/Python errors raised when a skills list entry can't work as an exact name. [line 446] [Source]

subagents [Source]

• Documents the new subagent_type is required error when the Agent tool is called with no subagent_type and no general-purpose fallback is available (e.g. CLAUDE_AGENT_SDK_DISABLE_BUILTIN_AGENTS=1); before v2.1.235 this failed with Agent type 'general-purpose' not found. [line 179] [Source]

agent-teams [Source]

• Clarifies that an in-process teammate's run_in_background: true request now also fails (with an error, or by silently running in the foreground) rather than only being refused when the subagent's frontmatter sets background: true. [line 421] [Source]

amazon-bedrock [Source]

• Clarifies that Sonnet 5 always runs with the 1M token context window on both the Invoke API and the Mantle endpoint on Amazon Bedrock. [line 343] [Source]

commands [Source]

• Updated the /goal description to reflect that the goal can now also clear for other reasons (see the goal doc changes below). [line 71] [Source]

cross-session-messaging [Source]

• New limitation: same-machine cross-session messages are capped at about a million serialized characters; Claude Code now refuses oversized messages in the sending session before they leave, instead of silently dropping them. [line 234] [Source]

debug-your-config [Source]

• The MCP troubleshooting table now also covers .mcp.json files that use a top-level servers key (VS Code's format) instead of mcpServers. [line 91] [Source]

env-vars [Source]

• New CLAUDE_CODE_GOAL_CHECKIN_MINUTES variable controls how long background work can keep an active goal waiting (default 30 minutes) before Claude is asked to check on it; set to 0 to disable check-ins. [line 262] [Source]

errors [Source]

• New "Context limit reached" interactive-mode error message replaces the plain Prompt is too long wording shown in the terminal, and now names when auto-compact is off and how to re-enable it via /config. [lines 82-83] [Source]
• New "subagent_type is required" error entry, matching the subagents.md change above. [line 134] [Source]
• New "Message too large for cross-session delivery" entry documenting the refusal message and remediation steps. [line 139] [Source]
• New "The remote sent a reply this version can't display" entry for /context and /btw version mismatches against an attached cloud session; before v2.1.235 this showed a raw JavaScript error instead. [line 78] [Source]

goal [Source]

• New "Errors you have to fix clear the goal" section: an authentication failure, exhausted credit balance, uncompactable context overflow, or unavailable model now clears an active goal with a warning telling you to fix the cause and rerun /goal. [line 109] [Source]
• New "Background work defers evaluation" section: after 30 minutes (configurable) of background work keeping a goal waiting, Claude Code now asks Claude to check on it at each turn end. [lines 120-123] [Source]

interactive-mode [Source]

• New "Check spelling as you type" section documents the v2.1.235 spellcheck feature end to end: prerequisites (aspell/hunspell/ispell), where the spellcheck setting can be set, per-checker/language/color options, what gets skipped (code-like text, CJK/Thai/Lao/Khmer/Myanmar text), and how to debug a checker that isn't underlining anything. [line 362] [Source]

mcp [Source]

• New "Add a server from setup instructions written for another client" section walks through converting a URL, a launch command, or an mcpServers JSON block (as written for Claude Desktop, Cursor, etc.) into the equivalent claude mcp add / claude mcp add-json command. [line 131] [Source]

settings [Source]

• New spellcheck settings-reference entry, readable only from user settings, --settings, and managed settings (not project settings). [line 303] [Source]

sub-agents [Source]

• New line noting that an Agent tool call omitting subagent_type now fails with subagent_type is required when no general-purpose fallback exists. [line 67] [Source]

API changes

Changed documents

access-transparency [Source]

• "Anthropic Workbench" renamed to "Playground (Claude Console)" in the Access Transparency coverage table. [line 171] [Source]

admin-api [Source]

• Role permissions table now says "Playground" instead of "Workbench" for the user, claude_code_user, developer, and billing roles. [lines 82-85] [Source]

agent-skills/enterprise [Source]

• New "Skill content scanning" section: Claude Enterprise orgs can turn on beta automated security scanning for custom Skills uploaded in claude.ai and Claude Cowork; a Skill that fails scanning is blocked, one that passes with a warning stays usable behind a caution notice. Scanning doesn't cover Skills uploaded via the Skills API or Claude Console. [line 43] [Source]
• Versioning strategy now warns that omitting version means requests use the latest version, so any workspace member's upload immediately changes what production agents run. [line 164] [Source]

agent-skills/overview [Source]

• Links to the new Skill content scanning feature described above. [line 225] [Source]
• New pointer to the Skills API audit logging section (see skills-guide.md below). [line 256] [Source]

api-and-data-retention [Source]

• ZDR and HIPAA "what's not covered" sections now say "Claude Console, including Playground" instead of "Console and Workbench". [lines 33, 54]

api/overview [Source]

• "Workbench" renamed to "Playground" when describing how to get API keys. [line 56] [Source]

build-with-claude/files [Source]

• Total storage limit per organization doubled from 500 GB to 1 TB. [lines 249, 274]
• New "Audit logging" section: Files API uploads, downloads, and deletions now appear in the Compliance API Activity Feed as platform_file_uploaded/platform_file_content_downloaded/platform_file_deleted activities (list/metadata calls are not recorded). [line 260] [Source]
• Files API rate limit raised out of beta from ~100 to ~500 requests per minute. [line 305] [Source]

build-with-claude/skills-guide [Source]

• Production guidance now warns that omitting version or setting it to "latest" means any workspace member's upload changes production behavior immediately. [line 576] [Source]
• New "Audit logging" section: Skill and Skill version creation/deletion now appear in the Compliance API Activity Feed. [line 709] [Source]

build-with-claude/vision [Source]

• "Workbench" renamed to "Playground" as a way to try vision requests. [line 16] [Source]

cmek [Source]

• New guidance: attach a CMEK key to a new workspace before sending it traffic; attaching to a workspace that already receives requests can take up to a day to take effect, during which data is still encrypted with Anthropic-managed keys. [line 23] [Source]
• "Encrypted" / "Not encrypted" sections renamed to "Encrypted with CMEK key" / "Encrypted with Anthropic key" for clarity. [lines 53, 90]
• The personal-preferences disablement bullet was removed from the "disabled under CMEK" list and replaced with a new bullet: the "Instructions for Claude" section of personal preferences is now listed as encrypted with the Anthropic key (account-level, shared across orgs) instead of disabled. [line 106] [Source]

cmek-aws-kms [Source], cmek-azure-key-vault, cmek-google-cloud-kms

• Each provider guide adds the same new-workspace-first / up-to-a-day-to-take-effect guidance as cmek.md above. [cmek-aws-kms line 187] [cmek-azure-key-vault line 221] [cmek-google-cloud-kms line 215]

home [Source]

• "Try the Workbench" link renamed to "Try the API in Playground". [line 82] [Source]

intro [Source]

• Developer Console description now points to Playground instead of the Workbench. [line 66] [Source]

release-notes/overview [Source]

• New August 18, 2026 entry formally announcing the Workbench → Playground rename: Playground supports every Messages API parameter, includes feature-demo templates, and shows the full SDK request/response for each run. [line 9] [Source]

usage-cost-api [Source]

• FAQ heading and answer updated from "Anthropic Workbench" to "Playground", noting usage from the legacy Workbench is handled the same way. [line 319] [Source]

wif-providers/aws [Source], wif-providers/azure, wif-providers/gcp, wif-providers/github-actions, wif-providers/kubernetes, wif-providers/okta, wif-providers/spiffe

• Each provider guide's "Verify the setup" section now describes a failed token exchange as an opaque 401 authentication_error (Authentication failed) and points to the new authentication history page to find the deny reason, replacing the old 400 invalid_grant wording. [aws.md line 178] [azure.md line 207] [gcp.md line 174] [github-actions.md line 150] [kubernetes.md line 155] [okta.md line 125] [spiffe.md line 211]

wif-reference [Source]

• All token-exchange failure causes that previously returned 400 invalid_grant now return a uniform 401 authentication_error with the fixed message Authentication failed; the specific deny reason (e.g. match_subject_prefix, workspace_id_required) is now visible on a new authentication history page instead of being logged server-side only. [lines 217-224] [Source]
• The workspace_id_required case moved from a 400 invalid_request to this same 401 family, and a new 400 invalid_request_error case covers a malformed (rather than missing) workspace_id. [line 222] [Source]
• "Troubleshoot a failed exchange" section updated to match the new 401/authentication-history behavior. [line 238] [Source]

workspaces [Source]

• Workspace User role description now says "Use Playground only" instead of "Use the Anthropic Workbench only". [line 37] [Source]

Note: several docs-md/api/api/compliance/** pages (activities/list, organizations, organizations/list, roles, roles/permissions, roles/permissions/list, roles/retrieve, settings, users, users/list) showed large diffs today, but these were crawl failures — the pages rendered as repeated "Loading" placeholders rather than real content — so they're excluded from this changelog as not reflecting an actual documentation change.