Claude docs changes for August 5th, 2026 [diff]

Executive Summary

• New native PowerShell tool and hook support brings Windows shell handling to parity with Bash: hooks, permission checks, and tool-search coverage now recognize Bash|PowerShell uniformly.
• Sandbox credential masking (mode: "mask") is now documented for sandbox.credentials.files on Linux/WSL, letting sandboxed commands use real secrets through a proxy without the container ever seeing the plaintext value.
• Ultraplan has been removed: the /ultraplan command, its keyword trigger, and the plan-approval option are gone, with Plan mode taking its place across the docs.
• The auto-compact window is now configurable via /autocompact, --autocompact, or CLAUDE_CODE_AUTO_COMPACT_WINDOW, giving explicit control over when a session starts compacting.
• Claude Code 2.1.222 closes two security-relevant bugs: worktree-isolated sessions could run destructive git commands against the main checkout, and PreToolUse auto-allow hooks could bypass tool restrictions inside background agent tasks.

New Claude Code versions

2.1.222

Existing feature improvements

• Improved auto mode safety: messages sent to other agent sessions via SendMessage are now evaluated by the permission classifier before dispatch
• Improved the refusal when Claude tries to invoke a skill with disable-model-invocation: Claude is now told to ask you to run the skill instead of replicating its workflow
• Improved the /diff view, the Remote Control workspace diff, and file-edit diffs in Claude Code on the web sessions to use raw git blob content, ignoring workspace-configured diff drivers and textconv
• Changed Remote Control auto-start so repo-local settings (.claude/settings.json or .claude/settings.local.json) can no longer turn it on (they can still turn it off); enable it at user scope via /config
• Removed the Ultraplan feature

Major bug fixes

• Fixed worktree-isolated sessions and their subagents being able to run destructive git commands against the main checkout; isolation now applies to file edits and Bash in every session type
• Fixed PreToolUse auto-allow hooks bypassing tool restrictions in background agent tasks (summaries, compaction, renames)
• Fixed /usage-credits on Team and Enterprise showing "you've already sent a usage credit request" for members whose earlier request was dismissed, blocking them from sending a new one
• Fixed the startup connectivity check hanging and then failing behind an HTTPS proxy; it now uses the same proxy-aware transport as API requests and times out with a clear message
• Fixed "Connection closed mid-response" errors being reported on responses that had actually completed
• Fixed /usage overattributing usage to MCP servers: a server's share now reflects only the requests that actually consumed its tool results, instead of every turn after any call to it
• Fixed sessions not linking to pull requests created after the branch was pushed, including through the GitHub REST API
• Fixed org-restricted model: opus-style subagent and teammate family aliases dropping to the parent model instead of stepping down to the newest org-allowed model in the family
• Fixed stream idle timeout firing on custom ANTHROPIC_BASE_URL gateways despite server keep-alive pings arriving on the wire
• Fixed claude.ai connectors being falsely marked as needing authorization when the session token is invalid — they now show a /login hint instead
• Fixed tool errors not being displayed for tools no longer available locally, for example after an MCP server is removed
• Fixed screen readers re-reading the whole input line on every backspace in --ax-screen-reader mode — end-of-line deletions now echo just the deleted characters

Claude Code changes

New Documents

agent-sdk/troubleshooting [Source]

New Agent SDK troubleshooting reference, keyed to the error you see. Covers CLINotFoundError: Claude Code not found (the Python SDK can't find a claude executable), CLIConnectionError: Refusing to execute batch script (the SDK refuses to spawn claude.cmd/other batch scripts on Windows because cmd.exe re-parsing makes them injectable, with guidance to install a native claude.exe instead), and structured_output coming back None/undefined on an otherwise-successful run (e.g. from an unsatisfiable schema). Also links to the SDK repos for filing new issues.

Changed documents

accessibility [Source]

• Screen-reader-friendly character-level echo in --ax-screen-reader mode now also covers Backspace deletions at the end of the input line, not just typing. [line 51] [Source]

agent-sdk/agent-loop [Source]

• Clarifies that the TypeScript and Python Agent SDKs bundle a native Claude Code binary for most installs (previously stated unconditionally), pointing to the quickstart for exceptions. [line 4] [Source]
• The tool-definition context-cost table now describes MCP tool-search fallback to upfront loading as affecting "unsupported models and certain platforms" rather than enumerating each platform individually. [line 221] [Source]

agent-sdk/hosting [Source]

• Runtime dependency note updated: the bundled Claude Code binary covers most installs, not all; links to the quickstart for which installs need a separately installed native binary. [line 169] [Source]

agent-sdk/mcp [Source]

• New "Connection timing" table defines exactly which server types delay the first turn and their wait timeouts (stdio/uncached HTTP-SSE block on MCP_TIMEOUT, cached remote servers don't block, in-process SDK servers never block), plus two ways to make startup itself block on connection: MCP_CONNECTION_NONBLOCKING=0 (capped by MCP_CONNECT_TIMEOUT_MS) or alwaysLoad: true per server. [lines 140-153] [Source]
• Clarifies exactly when the init message is emitted relative to the first-turn wait, and what status/tools each server type shows at that point. [line 206] [Source]
• "pending" status is now documented as covering two distinct cases: a settings-file server that didn't get the full startup wait, or a cached-tool-list server not yet actually connected. [lines 725-726] [Source]

agent-sdk/quickstart [Source]

• New documented edge cases where the SDK does not bundle a native Claude Code binary — pip installing the Python SDK's source distribution (e.g. ARM64 Windows) and npm installs that skip optional dependencies (e.g. npm ci --omit=optional) — each with a workaround. [lines 82-85] [Source]

agent-sdk/skills [Source]

• In the TypeScript SDK, the skills option now requires exact Skill names: query() throws before starting Claude Code for empty names, names with parentheses/commas/control characters, padded whitespace, or wildcard forms (a bare * or a :* suffix). Use skills: "all" instead of a wildcard. [lines 85-92] [Source]

agent-sdk/tool-search [Source]

• Tool search is now on by default on Google Cloud's Agent Platform for Claude Opus/Sonnet/Haiku 4.5+ (v2.1.221, same minimums as elsewhere); earlier Agent Platform models still get upfront loading with no ENABLE_TOOL_SEARCH override. Previously tool search was off by default for the whole platform. [lines 17-26] [Source]

agent-sdk/typescript [Source]

• The skills option table now documents that the SDK rejects malformed or wildcard skill names with an error before starting the process. [line 439] [Source]
• Same "bundled for most installs" caveat added for the native Claude Code binary. [line 7] [Source]

agent-view [Source]

• /fork's copy now makes code changes in its own new worktree rather than potentially sharing the original session's checkout (except bare-repo layouts or worktree isolation off); when the copy starts in the main working tree, it bases a new branch on the original session's branch. [lines 312-316] [Source]
• Background-session git workflow reorganized into explicit rules: commit-and-push, open a draft PR when warranted, never push to main/force-push/merge, and defer to your own git-handling instructions in the task, CLAUDE.md, or memory. [lines 412-417] [Source]
• /status now shows a Session kind row — background job (attached/unattended) vs. interactive — as of v2.1.221. [line 520] [Source]

channels [Source]

• Plugin install troubleshooting (Telegram, Discord, iMessage, fakechat) now documents that Claude Code auto-refreshes a stale marketplace catalog and retries before reporting "not found," and that the install summary — not a blanket instruction — determines whether /reload-plugins is needed. [lines 34-39] [Source]

claude-apps-gateway [Source]

• Claude Code now forwards parent-supplied sandbox.credentials in stripped form: deny entries keep only path/name and mode, mode: mask file entries are forwarded sentinel-only (with empty injectHosts, dropping extract/onExtractNoMatch/maskDuplicates), and mode: mask envVars entries aren't forwarded at all. [lines 281-285] [Source]

claude-apps-gateway-config [Source]

• New gateway validation: a non-string model value in a request is now rejected with 400 model must be a string, requiring gateway v2.1.221+. [line 395] [Source]

claude-code-on-the-web [Source]

• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE is now described as moving compaction earlier within the auto-compact window, with the new /autocompact command and CLAUDE_CODE_AUTO_COMPACT_WINDOW for changing the window itself. [line 152] [Source]
• Claude Code now computes web diffs (including per-file diffs shown as Claude edits) from raw git blob content, so repository diff/textconv filters don't apply. [line 160] [Source]
• Ultraplan references removed from this page following the feature's removal. [line 77] [Source]

claude-security [Source]

• Same marketplace auto-refresh-and-retry plus install-summary-driven /reload-plugins guidance added as in channels.md. [lines 26-29] [Source]

cli-reference [Source]

• New --autocompact <auto|tokens> flag sets the auto-compact window for a single session launch, requires v2.1.221+. [line 58] [Source]

code-review [Source]

• Clarifies exactly which run types always get text-based findings (terminal /code-review forked-subagent runs, and -p text/JSON runs) versus host apps (e.g. the desktop app) that get the structured ReportFindings list. [lines 286-291] [Source]

commands [Source]

• New /autocompact [auto|<tokens>] command sets the auto-compact window and saves it to user settings, v2.1.221+. [line 36] [Source]
• /ultraplan is marked removed; use plan mode instead. [line 127] [Source]

context-window [Source]

• New "Set the auto-compact window" section documents three configuration layers (/autocompact command, --autocompact flag, CLAUDE_CODE_AUTO_COMPACT_WINDOW env var) with precedence rules, accepted value forms (100K-1M tokens), and which sessions compact earlier by default (cloud sessions, Sonnet/Opus 4.6 without extended context, Sonnet 5). [lines 93-118] [Source]

costs [Source]

• Clarifies that the context/auto-compact warning refers to the session's (now configurable) auto-compact window rather than the model's raw maximum input size. [line 640] [Source]

desktop [Source]

• Renaming a session in the desktop app now syncs to the underlying CLI session name, visible via claude agents in a terminal; requires v2.1.221+. [line 293] [Source]

discover-plugins [Source]

• Plugin install failure handling now auto-refreshes a stale marketplace catalog and retries before reporting "not found" when auto-update is on; before v2.1.221 this required a manual /plugin marketplace update. [lines 35-38] [Source]
• The install summary now reports one of Plugin is now active., Run /reload-plugins to activate., or a load failure shown in the /plugin Errors tab, instead of every install always requiring a manual reload/restart. [lines 279-285] [Source]
• /reload-plugins now warns and skips instead of applying when the reload would invalidate the prompt cache, requiring --force to proceed. [line 358] [Source]

env-vars [Source]

• BASH_MAX_OUTPUT_LENGTH now documents concrete default (30000) and maximum (150000) values. [line 158] [Source]
• BASH_MAX_TIMEOUT_MS: the effective ceiling is now documented as the larger of this and BASH_DEFAULT_TIMEOUT_MS. [line 159] [Source]
• New CLAUDE_CODE_AUTO_COMPACT_WINDOW var documents its accepted range (100000-1000000), that it only accepts a plain integer, and its precedence over /autocompact/--autocompact/the autoCompactWindow setting. [line 179] [Source]
• CLAUDE_CODE_RESUME_INTERRUPTED_TURN: fixed so setting the value to 0 disables resume (previously only unsetting it worked). [line 292] [Source]
• ENABLE_TOOL_SEARCH updated to reflect the new default-on tool search for Claude 4.5+ models on Google Cloud's Agent Platform. [line 371] [Source]

errors [Source]

• When a server error, dropped connection, or stalled stream hits mid-response after Claude has already completed a text block or tool call, Claude Code now still runs any tool calls Claude already completed and continues the turn from their results, instead of just ending the turn; same updated behavior for "Socket is closed." [line 118] [Source]
• New documented error "Your checkout has no branches": /code-review ultra and claude ultrareview now refuse to review a detached-HEAD checkout with no branches before starting a cloud session, requires v2.1.221+. [lines 1234-1245] [Source]

fast-mode [Source]

• Switching to Opus 4.7 now turns fast mode off like any other unsupported model, instead of leaving it on and having the API reject requests (fixed in v2.1.221). [line 38] [Source]
• New fallback when usage credits run out mid-session: Claude Code retries each rejected fast-mode request at standard speed/pricing (no cooldown), with a notification in interactive sessions and a system/notification message in stream-json/SDK sessions. [lines 142-145] [Source]

feature-availability [Source]

• Tool search availability on Google Cloud's Agent Platform changed from "off by default platform-wide" to "unsupported only on models earlier than the Claude 4.5 generation." [line 27] [Source]
• Ultraplan removed from the list of claude.ai-only features. [line 42] [Source]

google-vertex-ai [Source]

• Tool search on Vertex AI is now on by default for Claude Opus/Sonnet/Haiku 4.5+ (previously required ENABLE_TOOL_SEARCH=true); earlier models still load upfront regardless. New ENABLE_TOOL_SEARCH=false opt-out documented, effective v2.1.221. [lines 123-128] [Source]

hooks [Source]

• Hooks now run consistently everywhere Claude Code runs — sessions in the terminal, IDE extensions, the desktop app, and Claude Code on the web all fire the same hook events. [line 5] [Source]
• New native PowerShell hook support: match Bash|PowerShell to cover shell-command hooks on Windows, with a documented PowerShell tool field table and an example .ps1 hook script. [lines 103-165] [Source]
• Hook deduplication changed: a handler defined in more than one settings file now runs once, but a plugin's or skill's copy of the same handler stays separate — previously dedup was purely by command string/args or URL. [line 370] [Source]
• Cloud sessions on Claude Code on the web don't read your local ~/.claude/settings.json; hooks there come from the repo and your organization's server-managed settings. [line 242] [Source]
• PostToolUseFailure's error field format is now documented in detail: a first line Exit code N for Bash/PowerShell, [Request interrupted by user for tool use] for cancellations, a bare failure message when the shell process itself couldn't start, and middle-truncation past 10,000 characters. [lines 1757-1770] [Source]
• Async hooks: in non-interactive mode (-p), Claude Code now kills any async hook still running at teardown and finalizes it with outcome cancelled; work that must outlive the session should start a fully detached process. [lines 2971-2974] [Source]

hooks-guide [Source]

• Hooks needing full shell-command coverage must now match Bash|PowerShell, not just Bash, since Claude can also run shell commands via the PowerShell tool. [line 625] [Source]
• Shell-form command hooks on Windows now fall back to PowerShell when Git Bash isn't installed (previously always used Git Bash). [line 937] [Source]

large-codebases [Source]

• Plugin-install failure handling now auto-refreshes a stale marketplace catalog and retries before reporting "plugin not found"; manual /plugin marketplace update is only needed if auto-update is disabled. [lines 188-191] [Source]

managed-mcp [Source]

• serverCommand/serverUrl policy entries in allowedMcpServers/deniedMcpServers now expand ${VAR} from a pinned environment rather than the live process environment, closing a bypass where a project/user settings-defined variable could redefine what an allowlist entry matches; deniedMcpServers still widens from settings-file variables. Effective v2.1.219. [lines 167-176] [Source]

mcp [Source]

• Same Vertex AI tool-search default change as google-vertex-ai.md — on by default for Claude 4.5+ generation models on Google Cloud's Agent Platform as of v2.1.221; ENABLE_TOOL_SEARCH behavior table updated accordingly. [lines 1078-1092] [Source]
• Plugin install failures now auto-retry after refreshing a stale marketplace catalog; the install summary, not a blanket instruction, tells you whether /reload-plugins is needed. [lines 37-42] [Source]

permission-modes [Source]

• acceptEdits PowerShell auto-approval now excludes commands with a quoted positional argument containing a quote character (e.g. an apostrophe) — these still prompt; use a named parameter like -Value to avoid it. [line 106] [Source]
• Switching permission modes while an auto-mode classifier check is pending now discards the pending verdict — you're re-prompted (or auto-denied in dontAsk mode) instead of a stale verdict being applied. [line 274] [Source]
• The "No, refine with Ultraplan" option was removed from the plan-approval dialog. [line 130] [Source]

plugin-marketplaces [Source]

• Plugin install no longer always requires a manual /reload-plugins afterward — it can activate automatically, and the install summary now tells you whether to run it. [lines 102-107] [Source]

plugins [Source]

• Same reload-plugins/install-summary change — check the install summary for Run /reload-plugins to activate. instead of always running it. [line 233] [Source]

prompt-caching [Source]

• Falling back to standard speed after running out of fast-mode usage credits mid-session now also preserves the prompt cache, matching the existing rate-limit fallback behavior. [line 69] [Source]
• Installing a plugin from the /plugin interface can now activate it immediately without /reload-plugins; cache-cost timing follows whenever the change actually applies. [lines 84-85] [Source]

remote-control [Source]

• Renaming a session from claude.ai or the Claude app now also updates the session's name in the local CLI (prompt bar and claude agents), not just the claude --resume title. Requires v2.1.221+. [line 125] [Source]
• New remote-only support for /autocompact: pass a window size (e.g. /autocompact 500k) or run with no argument to print the current window size. [line 245] [Source]

sandbox-environments [Source]

• --dangerously-skip-permissions now refuses to start when run as root on Linux/macOS. [line 43] [Source]
• Substantially expanded setup guidance for @anthropic-ai/sandbox-runtime: required Linux/WSL2 packages (bubblewrap, socat, ripgrep), required write grants (project dir, ~/.claude, ~/.claude.json, /tmp), required network domains (provider endpoint plus claude.ai/platform.claude.com for OAuth), and a note that Linux/WSL2 only grants writes to paths that already exist. [lines 61-84] [Source]
• New documentation of built-in deny protections (.git/hooks, .git/config, .mcp.json, .claude/commands, .claude/agents, shell startup files) and platform differences (macOS checks per-write; Linux/WSL2 builds the deny list once at launch), plus guidance to review writable paths after unattended runs. [lines 94-109] [Source]

sandboxing [Source]

• New table details whether a managed credentials.files entry pins filesystem.disabled — only mode: "deny" (or a mask entry degraded to deny by validation) pins it. [lines 188-205] [Source]
• Clarifies which protections keep applying with filesystem isolation off (env-var deny/mask and file mask entries still enforced; only file deny/denyRead and TMPDIR redirection are lifted); notes $TMPDIR can now expand empty on Linux, so Claude is told to use mktemp -d. [lines 207-220] [Source]
• Major new feature: credentials.files entries now support "mode": "mask" (previously deny-only), requiring network.tlsTerminate and v2.1.221+. On Linux/WSL2, sandboxed commands see a sentinel copy of the file with the proxy substituting the real value on egress to injectHosts; on macOS the file is simply unreadable (same as deny). New extract, onExtractNoMatch, and maskDuplicates fields documented, plus fallback-to-deny conditions (directories, globs, files >8MiB, non-UTF-8). [lines 250-325] [Source]
• The protected settings-file deny list now also includes .mcp.json at the project root and at each --add-dir//add-dir root, not just settings.json files. [lines 488-491] [Source]

security-guidance [Source]

• Same plugin-install auto-retry/install-summary changes as the other plugin docs. [lines 377-384] [Source]

sessions [Source]

• Renaming a session from claude.ai/the Claude app, or from the desktop app, now applies the same name in the CLI. Requires v2.1.221+. [lines 72-73] [Source]

settings [Source]

• Invalid sandbox.credentials files/envVars entries that still have a valid path/name and a mask/deny mode are now degraded to mode: "deny" with a warning rather than stripped outright, so the credential stays blocked (v2.1.221; previously every invalid entry was stripped). [line 175] [Source]
• New autoCompactWindow setting configures the context-window fill threshold (100,000-1,000,000 tokens) for auto-compaction. [line 208] [Source]
• disableAutoMode can now also be set under permissions.disableAutoMode. [line 232] [Source]
• credentials.files now supports mode: "mask", plus new extract, onExtractNoMatch, maskDuplicates, and injectHosts fields. [lines 393-397] [Source]

skills [Source]

• In non-interactive sessions, Claude Code no longer reserves the help/feedback command names for terminal-only built-ins, so a plugin skill using those names keeps its bare command (a regression from v2.1.216, fixed in v2.1.221). [line 252] [Source]

terminal-config [Source]

• The ultraplan keyword was removed from the rainbow-gradient-highlighted keywords in the prompt input — only ultrathink remains. [line 118] [Source]

tools-reference [Source]

• In auto mode (and plan mode while the classifier reviews commands), the auto-mode classifier now reviews each SendMessage send before delivery, requires v2.1.222+. [line 132] [Source]
• Reworked "Output limits": Bash output streams to a working file as the command runs (killed past 5GB); how much reaches Claude inline now depends on whether the result counts as valid vs. failure — valid results get ~30,000 chars inline then a file path, failures get a ~10,000-char head/tail excerpt with no file. New table lists which commands' exit code 1 counts as benign (grep, rg, find, diff, test, git diff, git grep, etc.) vs. failure. [lines 140-158] [Source]
• PowerShell tool commands are now visible to PreToolUse hooks via tool_input.command; hooks inspecting shell commands should match Bash|PowerShell. [lines 292-293] [Source]

ultraplan [Source]

• The Ultraplan research-preview feature has been removed entirely — the /ultraplan command, the ultraplan keyword trigger, and the "No, refine with Ultraplan" plan-approval option are all gone. The page now just redirects to Plan mode and Claude Code on the web. [lines 1-15] [Source]

ultrareview [Source]

• The "No merge base" fallback is clarified and cross-referenced to the new "no branches" refusal case added in errors.md; the "Plan complex changes with ultraplan" related-resource link was removed. [line 66] [Source]

vs-code [Source]

• New Focus view collapses tool calls/results/thinking into one expandable row per turn (toggle via Settings, Command Palette, or Ctrl+Option+F/Ctrl+Alt+F); persists across sessions and applies to all open sessions immediately. New focusView setting, requires v2.1.221+. [lines 80-291] [Source]

web-quickstart [Source]

• Section heading/anchor updated to drop the "or ultraplan" reference, reflecting the Ultraplan removal. [line 200] [Source]

API changes

Changed documents

api/admin/analytics [Source]

• AnalyticsUser gains a type field (always "user") alongside id/email_address. [line 147] [Source]
• AnalyticsUserActor fields deleted, email, name, type, user_id are now documented as required (previously all optional). [line 165] [Source]

api/admin/analytics/artifacts/list [Source]

• filter[] now documents its supported dimensions explicitly: artifact_type, is_shared, rbac_group_id, user_id, with value formats for each. [line 28] [Source]
• group_by[] is now typed as an enum of "rbac_group_id" / "user_id" instead of a free-form string array. [line 90] [Source]

api/admin/analytics/chat_projects [Source]

• created_by (AnalyticsUser) gains a type field, always "user". [line 59] [Source]

api/admin/analytics/chat_projects/list [Source]

• filter[] now documents supported dimensions: project_id, rbac_group_id, user_id, with value formats. [line 33] [Source]
• created_by.type field added. [line 121] [Source]

api/admin/analytics/connectors/list [Source]

• filter[] now documents supported dimensions: connector_name, product, rbac_group_id, user_id, with value formats and normalization notes. [line 36] [Source]

api/admin/analytics/cost [Source]

• New slack_channel_id field added to cost result rows and per-user UserCost rows, populated when grouping by Slack channel. [lines 91-93] [Source]
• AnalyticsUserActor fields now required rather than optional. [line 149] [Source]

api/admin/analytics/cost/list [Source]

• group_by[] gains a new "slack_channel_id" dimension. [line 76] [Source]
• New slack_channel_ids filter parameter to scope cost data to specific Slack channels. [lines 116-118] [Source]
• Result rows gain a slack_channel_id field. [line 210] [Source]

api/admin/analytics/cost/list_by_user [Source]

• Same Slack-channel additions: group_by[] value "slack_channel_id", a slack_channel_ids filter, and a slack_channel_id result field. [line 83] [Source]

api/admin/analytics/plugins/list [Source]

• filter[] now documents supported dimensions: plugin_name, product (claude_code/cowork only), rbac_group_id, user_id. [line 39] [Source]

api/admin/analytics/retrieve_summaries [Source]

• New capability: the summaries series can now be scoped to a single RBAC group via filter[]=rbac_group_id:<id>. [line 17] [Source]

api/admin/analytics/skills/list [Source]

• filter[] now documents supported dimensions: product, rbac_group_id, share_status, skill_name, user_id. [line 34] [Source]

api/admin/analytics/usage [Source]

• New slack_channel_id field added to usage result rows and per-user data. [lines 95-97] [Source]

api/admin/analytics/usage/list [Source]

• group_by[] gains "slack_channel_id"; new slack_channel_ids filter; result rows gain a slack_channel_id field. [line 74] [Source]

api/admin/analytics/usage/list_by_user [Source]

• Same Slack-channel additions as usage/list: group_by[] value, slack_channel_ids filter, result field. [line 81] [Source]

api/admin/analytics/users [Source]

• user (AnalyticsUser) gains a type field. [line 457] [Source]

api/admin/analytics/users/list [Source]

• filter[] now documents supported dimensions: project_id, rbac_group_id, user_id, including that project_id scopes to chat activity in a project and can't combine with group_by[]. [line 33] [Source]
• user.type field added. [line 511] [Source]

api/admin/cost_report [Source]

• inference_geo is now a typed enum ("global" / "not_available" / "us") instead of a freeform string. [line 75] [Source]

api/admin/cost_report/retrieve [Source]

• Same inference_geo enum change. [line 120] [Source]
• next_page cursor format changed from an RFC 3339 timestamp string to an opaque page_... token. [line 219] [Source]

api/admin/rate_limits [Source]

• Rate-limit entries gain a new id field: a stable identifier for the rate-limit group. [line 27] [Source]

api/admin/rate_limits/list [Source]

• Same id field addition. [line 57] [Source]

api/admin/usage_report [Source]

• Response actor schemas renamed from ClaudeCodeUserActor/ClaudeCodeAPIActor to generic UserActor/APIActor. [line 39] [Source]
• date field changed from YYYY-MM-DD format to an RFC 3339 timestamp at midnight UTC. [line 103] [Source]
• inference_geo now a typed enum. [line 252] [Source]

api/admin/usage_report/retrieve_claude_code [Source]

• Same actor renames (UserActor/APIActor). [line 50] [Source]
• date field format changed to an RFC 3339 timestamp. [line 114] [Source]

api/admin/usage_report/retrieve_messages [Source]

• inference_geo now a typed enum ("global"/"not_available"/"us"). [line 228] [Source]
• next_page cursor format changed from timestamp string to opaque page_... token. [line 347] [Source]

api/admin/workspaces/create [Source]

• allowed_inference_geos, default_inference_geo, and workspace_geo are now typed enums ("global"/"us", or "us" only for workspace_geo) instead of freeform strings. [line 37] [Source]

api/admin/workspaces/rate_limits [Source]

• Workspace rate-limit entries gain two new fields: rate_limit_id (the id of the org-level rate-limit group being overridden) and workspace_id. [line 67] [Source]

api/admin/workspaces/rate_limits/list [Source]

• Same new fields rate_limit_id and workspace_id. [line 99] [Source]

api/admin/workspaces/update [Source]

• Same enum typing change for allowed_inference_geos/default_inference_geo. [line 27] [Source]

build-with-claude/compaction [Source]

• Compaction is no longer beta: the requirement to include beta header compact-2026-01-12 was removed, along with the old beta-only supported-models list.
• New "Compatibility" table documents supported models (Fable 5, Mythos 5/Preview, Opus 4.6-5, Sonnet 4.6/5) and supported platforms, now including Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry (each marked Beta). [lines 589-594] [Source]

build-with-claude/thinking [Source]

• New guidance: use your SDK's message-accumulation helper (e.g., stream.get_final_message() in Python, stream.finalMessage() in TypeScript) to reassemble complete thinking blocks with signatures after streaming, instead of concatenating deltas manually. [line 225] [Source]
• Clarified Ruby SDK behavior: plain hashes use display:, while only the typed ThinkingConfigAdaptive class requires the trailing-underscore display_ param name (previously stated flatly that display_: was needed). [line 171] [Source]

cli-sdks-libraries/libraries/apple-foundation-models [Source]

• Clarified billing requirement: using Claude via Apple Foundation Models requires the organization to have an available credit balance or active billing method. [line 9] [Source]