Claude docs changes for August 30th, 2026 [diff]
Executive Summary
- New
PreModelSwitchandPostModelSwitchhooks let you block or confirm model switches before they happen, see their cache/cost impact, and inject model-specific guidance after a switch — documented across the hooks reference, hooks guide, and Agent SDK types. - Cloud environments now support API credentials: keys or tokens an org admin stores on an environment that Anthropic's agent proxy attaches to matching requests without ever exposing them to Claude or the session's environment variables.
- A new
desktopSessionCleanupPeriodDayssetting lets you cap how long Claude Desktop and Cowork session transcripts are kept; previously they were retained indefinitely. - Managed settings gained a documented recipe for turning off Anthropic telemetry org-wide via
DISABLE_TELEMETRY, including why it's needed even under customer-managed encryption keys. - Auto mode now turns off immediately in a running session when an admin-deployed policy disables it, instead of waiting until the session ends.
Claude Code changes
Changed documents
admin-setup [Source]
- Server-managed settings are now fetched at session startup, not just at authentication time, in addition to the hourly refresh. [line 43] [Source]
- Clarified what the WSL
/status"Setting sources" line actually tells you: it always shows "Enterprise managed settings (remote)" once any server-managed keys arrive, so it doesn't confirm a Windows registry/file deployment in that case — only in sessions that aren't fetching server-managed settings. [lines 62-65] [Source]
agent-sdk/hooks [Source]
- Added
PreModelSwitchandPostModelSwitchto the SDK hooks table, along with their timeout defaults (30 seconds) and blocking behavior. [line 163] [Source]
agent-sdk/modifying-system-prompts [Source]
- Clarified that
keep-coding-instructionsin an output style's frontmatter only has an effect on Claude Code's full system prompt — it does nothing when a session uses the shorter system prompt viaCLAUDE_CODE_SIMPLE_SYSTEM_PROMPT. [line 106] [Source]
agent-sdk/typescript [Source]
- Added
PreModelSwitchHookInputandPostModelSwitchHookInputtypes, plus the new"PreModelSwitch"/"PostModelSwitch"hook event names and theirSyncHookJSONOutputvariants (permissionDecisionfor pre-switch,additionalContextfor post-switch). [lines 1645-1966] [Source]
claude-code-on-the-web [Source]
- Split the "Credential protection" bullet into two: one for git/signing credentials, and a new one describing the cloud environment API credentials feature, which keeps API keys outside the sandbox and attaches them to matching requests via a proxy. [lines 265-266] [Source]
claude-directory [Source]
- Documented that Claude Desktop and Cowork session transcripts are kept indefinitely by default, with a new
desktopSessionCleanupPeriodDayssetting to give them an age limit (requires v2.1.248+). Previously these transcripts had no special-cased retention rule described here. [lines 205-209] [Source] - The retention-sweep pause condition now also triggers when
desktopSessionCleanupPeriodDaysis explicitly set and settings have validation errors. [line 234] [Source] - Added
desktopSessionCleanupPeriodDaysto the list of ways to reduce credential-exposure risk from unencrypted transcripts. [line 235] [Source]
cloud-environments [Source]
- New Add API credentials section: org admins can store an API key/token (Bearer or other auth type) on a cloud environment, scoped to specific hosts; Anthropic's agent proxy attaches it to matching requests after they leave the session's VM, so Claude and its commands never see the value. Covers requirements (admin role, Anthropic-hosted environment, no customer-managed encryption keys), how to add one, and the requests that never get a credential (GitHub, the Anthropic API, and public package registries). [lines 36-91] [Source]
- API-credential hosts are now reachable regardless of the environment's network access level, alongside GitHub and MCP connector traffic. [lines 169-174] [Source]
- The "what carries over" table now says static API keys/tokens are supported, via API credentials, reversing the previous "no dedicated secrets store" guidance. [line 235] [Source]
- Shared (org-wide) environments can now also be edited from the environment selector at claude.ai/code by an Admin or Owner, not only from the admin Cloud environments page. [line 108] [Source]
commands [Source]
- Expanded the
/terminal-setupdescription: it now also covers enabling Option+Enter and disabling the audible bell in Apple Terminal, and enabling clipboard access for/copyin iTerm2. [line 130] [Source]
data-usage [Source]
- Noted that Claude Desktop/Cowork transcripts are exempt from the 30-day local caching limit by default. [line 44] [Source]
discover-plugins [Source]
- Clarified that in cloud sessions, Claude Code doesn't start plugin language servers, so the LSP tool isn't available there even with a code-intelligence plugin installed. [line 44] [Source]
env-vars [Source]
DISABLE_TELEMETRYentry now links to the new "Turn telemetry off for your organization" section in managed settings. [line 397] [Source]
glossary [Source]
- Trimmed the "Output style" entry to drop the detail about which built-in styles keep coding instructions by default, now covered on the dedicated output-styles/system-prompts pages. [line 163] [Source]
hooks [Source]
- Full reference sections added for the new
PreModelSwitchandPostModelSwitchhook events (requires v2.1.251+): what triggers each, matcher rules against the target model's canonical name, input schema (from_model,to_model,requested_model,source,context_tokens,prompt_cache_warm,cache_ttl,estimated_cache_write_usd,pricing), and decision control —PreModelSwitchcanallow/deny/ask(withaskonly honored by an interactive/model), and a timed-outPreModelSwitchhook blocks the switch by default (unlike most other events).PostModelSwitchcan only addadditionalContextfor Claude's next request. [lines 173-363] [Source] - Added
PreModelSwitch/PostModelSwitchto the lifecycle diagram, event summary table, matcher-value table, timeout-default text, exit-code-0 exceptions, exit-code-2 table, and the command/http/mcp_tool-only hook-type list. [lines 15-134] [Source] SessionStarthooks resuming or forking a transcript with at least one prior response now also receiveseconds_since_last_response,context_tokens,prompt_cache_likely_expired, andestimated_cache_write_usd, letting a hook report the cost of resuming a stale conversation (requires v2.1.251+). [lines 141-166] [Source]
hooks-guide [Source]
- Added
PreModelSwitch/PostModelSwitchto the event list, the exit-code-0 context-injection list, the matcher-field table, and the timeout-defaults text; also summarizedPreModelSwitch'spermissionDecisionsemantics. [lines 479-889] [Source]
managed-settings [Source]
- New "Turn telemetry off for your organization" section: deliver
DISABLE_TELEMETRY: "1"through the managedenvblock to disable telemetry org-wide without relying on each developer's shell; explains it's applied without the approval dialog and also disables feature-flag fetching (Remote Control, default auto mode, etc.). [lines 298-311] [Source]
monitoring-usage [Source]
- Added
transcripts_exempted_desktopto theretention_sweepevent fields, counting transcripts kept under the new Claude Desktop/Cowork retention rule (requires v2.1.248+). [line 1051] [Source]
permission-modes [Source]
- A running session now exits auto mode immediately once an admin-deployed
disableAutoModepolicy reaches it, showing "auto mode disabled by settings"; before v2.1.251, a running session kept auto mode until it ended. [line 233] [Source]
plugins-reference [Source]
prompt-caching [Source]
- Noted that a
PreModelSwitchhook can require or skip the/modelcache-warm confirmation prompt. [line 66] [Source] - Removed the description of the LSP tool being "sticky" for the rest of a session once a language server becomes available (the tools-reference page now describes simpler behavior). [lines 103-107] [Source]
routines [Source]
- Routines no longer require "Claude Code on the web" to be separately enabled — available to all Pro/Max/Team/Enterprise users. [line 13] [Source]
- Routine environment variables guidance now points to the new API credentials feature instead of just warning not to add secrets. [line 61] [Source]
server-managed-settings [Source]
- Invalid
cleanupPeriodDaysanddesktopSessionCleanupPeriodDaysvalues are now kept in the cached remote-settings copy (but never applied) rather than stripped, and a payload consisting only of those invalid retention keys no longer causes a full payload rejection. [line 168] [Source]
sessions [Source]
/clearcan now take a name argument (e.g./clear release-prep) to name the conversation you're leaving instead of carrying its name to the new one. [line 171] [Source]- Added
desktopSessionCleanupPeriodDaysto the storage-configuration table. [line 207] [Source]
settings-reference [Source]
- Documented the new
desktopSessionCleanupPeriodDayssetting (user or managed scope, default0= no age limit; ignored when managedcleanupPeriodDaysis set; requires v2.1.248+). [lines 5021-5036] [Source] - Page now describes an interactive filter/sort UI (topic and scope dropdowns, sort-by control, settings count) above the settings table. [lines 3-11] [Source]
skills [Source]
third-party-integrations [Source]
- New note: customer-managed encryption keys (CMEK) don't cover Claude Code's operational telemetry when routing through an LLM gateway or custom
ANTHROPIC_BASE_URL, so those orgs should useDISABLE_TELEMETRYvia managed settings if they want it off. [line 50] [Source]
tools-reference [Source]
- Simplified the LSP tool's activation behavior description and added that it stays inactive in cloud sessions, since plugin language servers don't start there. [lines 268-269] [Source]
web-quickstart [Source]
- New troubleshooting entry: if
/web-setupwarns your GitHub CLI token lacks theworkflowscope, some pushes (e.g. to GitHub Actions workflow files) may be rejected; fix withgh auth refresh -s workflow. [lines 203-205] [Source]
API changes
No changes to the Claude API documentation today.