Claude docs changes for August 8th, 2026 [diff]

Executive Summary

• New cross-session messaging lets a Claude Code session discover and message your other independent sessions — directly on the same machine, or reply-only to sessions on other machines and Claude Code on the web via Remote Control — via new ListAgents/SendMessage tools, the /list-agents command, and settings (crossSessionInbound, isolatePeerMachines) to control it.
• Auto mode becomes the default permission mode for new Claude Code sessions on Pro, Max, and Team plans starting August 14, 2026 (existing user- or org-set defaults are preserved).
• Claude Managed Agents gains session budgets (hard spend caps that pause a session with budget_reached), session advisors (a mid-turn consultation model added to a session's roster), inference_geo pinning at the agent level, and skill discovery from a mounted GitHub repository's .claude/skills directory.
• New claude import / /import [codex|gemini] command migrates configuration — CLAUDE.md content, MCP servers, commands, subagents, and skills — from Codex or Gemini into Claude Code.
• Claude Code 2.1.225 fixes several reliability issues, including a transient 401 that could replace a long-lived CLAUDE_CODE_OAUTH_TOKEN and break headless sessions until restart, and intermittent MCP OAuth failures on macOS after a keychain read timeout.

New Claude Code versions

2.1.225

New features

• Added gateway spend-limit support to Claude Code's usage warning; the limit-reached message now names the cap, its reset time, and the operator's message (requires the gateway on 2.1.225)
• Added a workspace trust prompt to claude agents for untrusted directories, matching the behavior of claude
• SendMessage can now start a conversation with your Remote Control sessions on other machines by name (ListAgents shows them as name [ref]), instead of only replying after they message you first

Existing feature improvements

• Improved Remote Control: photos attached from the Claude app are now shown to Claude directly instead of being read from disk with a separate tool call
• SendMessage: a Remote Control recipient you already confirmed is never swapped for a same-named session on this machine when its own list couldn't be checked

Major bug fixes

• Fixed a transient 401 replacing a long-lived CLAUDE_CODE_OAUTH_TOKEN with a stored login's short-lived token, breaking headless sessions until restart
• Fixed MCP OAuth servers on macOS intermittently failing with a burst of 401 errors, as if never authenticated, after a keychain read timed out
• Fixed auto mode counting a safety-filter refusal of its own permission check toward the consecutive-block limit; the action is still denied, but the model is now told to move on rather than retry
• Fixed cross-session messages staying parked without a notice or expiry in headless sessions and during startup
• Fixed conversation history breaking on Remote Control session resume after very large conversations were compacted
• Fixed hovering over a session in another project in the agents list changing the directory the next agent starts in
• Fixed claude self-hosted-runner registering and then failing every session when --base-dir cannot be created or written; it now exits at startup with a clear error
• Fixed Claude Code on the web sessions being misreported as stuck, re-sending a growing event backlog on every reconnect
• [VSCode] Fixed Focus view folding away the latest to-do list, a pending question's context, and settled answers; thinking-only folds show "Thought for Ns" and re-collapse when their turn completes

2.1.226

Major bug fixes

• Bug fixes and reliability improvements (no further detail provided)

Claude Code changes

New Documents

agent-sdk/examples [Source]

New page routing to complete, runnable Agent SDK example projects and guided Claude Cookbook recipes. Links the claude-agent-sdk-demos GitHub repo for TypeScript demos (from an email client to a multi-agent research system), the Claude Cookbook's Agent SDK Python notebook series (a progression from a one-liner research agent to sophisticated multi-agent systems), and the Agent SDK quickstart / minimal "Hello World" project for starting from scratch.

cross-session-messaging [Source]

New page documenting cross-session messaging (requires Claude Code v2.1.224+, macOS/Linux only), which lets Claude use new ListAgents and SendMessage tools to discover and message a user's other independent Claude Code sessions — directly via a per-session inbox socket on the same machine, or reply-only to sessions on other machines/Claude Code on the web via Remote Control. It covers message delivery semantics (delivered/held/refused), the crossSessionInbound and isolatePeerMachines settings for controlling inbound/outbound messages, safety guarantees (a message never grants consent or changes config), and platform/provider availability limits.

Changed documents

agent-sdk/agent-loop [Source]

• bypassPermissions mode description now notes the cross-session messaging safeguards still apply even in this mode. [line 198] [Source]

agent-sdk/cost-tracking [Source]

• total_cost_usd/usage fields in the Python SDK are now documented as optionally None on some error paths. [lines 43, 251]
• The 5-minute default prompt-cache TTL now also explicitly applies when authenticating via Claude Platform on AWS. [line 264] [Source]
• New clarification: subscription users only get the automatic 1-hour cache TTL "within included usage"; once drawing on usage credits the SDK drops to a 5-minute TTL unless ENABLE_PROMPT_CACHING_1H is set. [line 306] [Source]

agent-sdk/permissions [Source]

• bypassPermissions mode section updated: cross-session messaging safeguards still apply; subagent inheritance note now mentions the isolatePeerMachines approval requirement for cross-machine messages. [lines 101, 230-238]

agent-sdk/sessions [Source]

• persistSession: false is now clarified as TypeScript-only; Python has no equivalent (use CLAUDE_CODE_SKIP_PROMPT_HISTORY instead). [line 18] [Source]
• Session resume cross-directory lookup behavior expanded into explicit bullets ("Cross-directory lookup", "Same machine only"); notes that before v2.1.223 lookup was scoped only to the current project directory and its worktrees, and older SDK-bundled CLIs still behave that way. [lines 254-264, 375-378]

agent-sdk/typescript [Source]

• peer origin type documentation expanded: a cross-session peer can now be on another of the user's machines or on Claude Code on the web (via Remote Control), not just a local process; a one-way cross-machine reply has no reply address and from resolves to "unknown". [lines 1465-1476] [Source]

agent-teams [Source]

• Cross-references cross-session messaging as the tool for separate sessions passing messages without forming a team. [line 20] [Source]
• "Messages between agents" section: the no-consent/no-relay rules now explicitly extend to cross-session messages, not just teammates; auto-mode classifier behavior restated as two distinct checks. [lines 234-241] [Source]
• Removed the specific "5-6 tasks per teammate" sizing guidance sentence. [line 309] [Source]
• Troubleshooting section restructured: "Teammates stopping on errors" renamed to "Agents stopping early" and merged with the former separate "Lead shuts down before work is done" section. [lines 360-368] [Source]

agent-view [Source]

• /fork behavior clarified: after forking, the two conversations are independent by default, but in sessions with cross-session messaging enabled, either side's Claude can now explicitly message the other. [lines 305-309] [Source]

agents [Source]

• Now lists three (previously two) coordination-support tools, adding cross-session messaging alongside worktrees and /batch. [lines 10-15] [Source]
• "Do the workers need to talk to each other?" guidance updated to mention cross-session messaging as an option for separate sessions. [line 33] [Source]

amazon-bedrock [Source]

• New ANTHROPIC_BEDROCK_REGION_PREFIX environment variable introduced as an alternative to pinning model IDs, letting users set a preferred cross-region inference-profile prefix while keeping the built-in default models. [lines 184-192] [Source]
• New "Cross-region inference profile prefixes" section (requires v2.1.224+): documents the default prefix per AWS region (us-gov., us., eu., apac., global.), valid override values, the resolution order Claude Code follows when it can vs. can't list inference profiles in the account, and that GovCloud always forces us-gov. regardless of the variable. [lines 257-296] [Source]

auto-mode-config [Source]

• New announcement: starting August 14, 2026, auto mode becomes the default permission mode for new sessions on Pro, Max, and Team plans (user- or org-set defaults are preserved). [line 2] [Source]

channels [Source]

• The list of checks that still prompt under --dangerously-skip-permissions in a channels session expands from 3 to 5 items, adding "removals targeting / or your home directory" and the cross-session messaging safeguards. [lines 281-291] [Source]

claude-apps-gateway-config [Source]

• ${file:/path} reference must now be the entire field value, not embedded in a longer string like postgres_url. [line 32] [Source]
• public_url requirement changed from "required behind a proxy" to "required unless host is loopback," including when TLS terminates at the gateway itself; boot now fails without it in that case. [line 44] [Source]
• allowed_groups matching clarified as exact, case-sensitive string comparison with no nested-group expansion. [line 58] [Source]
• ca_cert_pem clarified to require the literal PEM content, not a file path (use ${file:...} to load from a mounted file). [line 72] [Source]
• postgres_url: schema migrations now explicitly run "at boot and on upgrade," not just boot. [line 89] [Source]
• Bedrock "Model access" row rewritten: model access is now enabled by default in commercial regions; the remaining gate is Anthropic's one-time AWS Organizations use-case form. [line 162] [Source]
• upstream_model map behavior clarified: for built-in Claude models, an upstream omitted from the map still serves the model with the provider's default ID; only a fully custom model id skips upstreams missing from its map. [lines 286, 307]
• fail_closed_on_error now requires an admin: block configured; the gateway refuses to start if set true without one. [line 351] [Source]
• New note: each upstream_model key must match a configured upstream's name, or boot fails. [line 369] [Source]
• New requirement: telemetry.forward_to URLs must use https://, with a narrow loopback exception gated by CLAUDE_GATEWAY_ALLOW_LOOPBACK=1. [line 576] [Source]
• New telemetry detail: a signed-in CLI with OTLP export enabled sends exports to the gateway even without a configured forward_to destination for that signal, which the gateway then silently discards. [line 591] [Source]
• CLAUDE_GATEWAY_LOG_LEVEL gains a new debug level that also logs id_token claim names, for diagnosing groups_claim/email_claim config. [line 620] [Source]

claude-apps-gateway-deploy [Source]

• Removed documentation of the "Anthropic-operated public gateway endpoints" exemption from the private-network deployment requirement. [line 10] [Source]
• Audit event list expanded: new device.callback event; new admin.limit.upsert and admin.limit.delete events; admin.denied now distinguishes invalid_key, bearer_rejected, and no_credentials reasons. [line 95] [Source]
• Database role requirement simplified to "rights to create and alter tables" (removed the old pre-create-schema-for-DDL-restricted-roles guidance); concurrent gateway replicas now serialize schema migrations via a Postgres advisory lock during rolling upgrades. [line 142] [Source]
• SSRF loopback guard clarified: CLAUDE_GATEWAY_ALLOW_LOOPBACK=1 also skips a boot-time check of cloud-metadata-endpoint reachability. [line 171] [Source]
• Troubleshooting table updated to match the removed public-endpoint exemption and the new DDL-rights guidance. [lines 211-234] [Source]

claude-apps-gateway-on-aws [Source]

• public_url requirement updated to "required for any non-loopback bind." [line 251] [Source]
• SSRF guard clarified: the gateway now rejects an IP-literal loopback telemetry URL (e.g. http://127.0.0.1:4318) at boot, not just at send time. [line 259] [Source]

claude-apps-gateway-on-gcp [Source]

• Reworded architecture description: the internal Application Load Balancer is now something the operator provides; the walkthrough configures the gateway for it but doesn't create it. [line 273] [Source]
• Deploy command changed: added --max-instances=8; changed --ingress=internal-and-cloud-load-balancing to --ingress=internal; new guidance to keep max-instances × store.max_connections below the Cloud SQL tier's connection limit. [lines 300-311] [Source]
• Clarifies internal Cloud Run ingress works with or without an internal ALB in front, so internal-and-cloud-load-balancing is no longer needed for either topology on this page. [lines 318-321] [Source]

claude-apps-gateway-spend-limits [Source]

• New documented behavior: spend caps reset on UTC calendar boundaries (daily 00:00 UTC, weekly Monday 00:00 UTC, monthly 1st 00:00 UTC). [line 343] [Source]
• GET /v1/organizations/spend_limits gains a new scope_type filter query param. [line 352] [Source]
• /audit endpoint gains after_id pagination, default limit 100. [lines 358, 381]

claude-apps-gateway [Source]

• Removed the "Anthropic-operated public gateway endpoints" exemption text (matches claude-apps-gateway-deploy.md). [line 393] [Source]
• HTTPS prerequisite clarified: plain http:// is accepted only when the host is literally localhost/127.0.0.1/::1. [line 403] [Source]
• Boot sequence log example updated: now shows a "waiting for migration lock" line and one "migration N applied" line per schema migration on a fresh database. [lines 431-440] [Source]
• Claude Desktop policy delivery clarified: Desktop builds its own model/disabled-tools list from /user/bootstrap; only the egress allowlist requires the parentSettingsBehavior: "merge" opt-in, while hooks/env/scoped permission rules only reach /login clients. [lines 457-463] [Source]
• sandbox.credentials forwarding rules gain a new bullet: awsPairs and sigv4 blocks are forwarded restriction-only (only deny values kept; a configured sigv4 block pins all three AWS request forms to deny). [lines 470-475] [Source]
• Telemetry destination behavior changed: the CLI now always sends OTLP/HTTP exports to the gateway (ignoring any local OTEL_EXPORTER_OTLP_ENDPOINT), and the gateway relays to forward_to destinations, discarding signals with none configured. [line 484] [Source]

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

• New teleport method: from inside a cloud session, typing /teleport now replies with the exact claude --teleport <session-id> command (requires v2.1.223+ in the session's environment). [line 149] [Source]
• "Correct repository" teleport requirement rewritten with clearer error messaging when the checkout doesn't match the session's repo. [lines 23-24] [Source]

claude-directory [Source]

• cleanupPeriodDays minimum clarified as 1 (0 fails validation); the same cutoff now applies to automatic orphaned-worktree removal. [line 55] [Source]
• New behavior: if Claude Code can't read/parse a settings file, it pauses the retention cleanup sweep and warns in /status until fixed, unless managed settings supply cleanupPeriodDays (before v2.1.203 cleanup ran at the 30-day default and could delete transcripts meant to be kept longer). [line 63] [Source]
• New documented path ~/.claude/paste-cache/, holding pasted text for recalled prompts. [line 80] [Source]

cli-reference [Source]

• New claude import [codex|gemini] command documented, importing configuration from other coding agents (--dry-run/--yes supported); requires v2.1.213+. [line 99] [Source]

cloud-environments [Source]

• New note: self-hosted environment IDs (ccpool_...) follow a stricter settings-source rule for remote.defaultEnvironmentId than Anthropic-hosted IDs. [line 111] [Source]
• New "GraphQL restrictions" bullet: the GitHub proxy in Anthropic-hosted cloud sessions now serves only a pinned set of GraphQL operations, rejecting everything else with 403 regardless of supplied credentials (Projects v2 unreachable through the proxy). [line 119] [Source]
• VM now documented explicitly as Ubuntu 24.04 on x86_64, with guidance to use x86_64 Linux precompiled binaries; sessions routed to a self-hosted environment run on the org's own runners instead. [lines 128-130] [Source]
• Node.js version install paths documented (/opt/node20/21/22), with 22 on PATH by default. [line 147] [Source]

code-review [Source]

• Effort-level "reuse last" rule clarified: it reuses the last low through max level typed, and ultra neither updates nor consumes the remembered level. [lines 168-169] [Source]

commands [Source]

• New /import [codex|gemini] [--dry-run] [--yes] command documented; /init now offers to run /import when it detects importable config from another coding agent. [lines 197-199] [Source]
• New /list-agents command (alias /peers) documented — lists messageable subagents/other sessions; requires v2.1.224+ and only available where cross-session messaging is enabled. [line 204] [Source]

costs [Source]

• New note: usage-credit users can set ENABLE_PROMPT_CACHING_1H=1 to keep the 1-hour cache TTL. [line 226] [Source]
• New bullet: "Cross-session messages" listed as a cause of unexpected token usage — delivering a message from another session as a new turn resends full context; can be mitigated by setting crossSessionInbound to hold. [line 228] [Source]

data-usage [Source]

• Cloud sessions clarified to run in Anthropic-managed VMs "by default"; sessions routed to a self-hosted environment run on customer infrastructure instead. [line 241] [Source]
• WebFetch domain safety check caching behavior changed: only a hostname that passes the check is cached for 5 minutes; a blocked or failed hostname is now re-checked on the next request. [line 250] [Source]

desktop-scheduled-tasks [Source]

• Clarified that scheduled-task sessions can do everything a normal session can (edit, run commands, commit, open PRs) but can't send/receive messages through the desktop app's session-to-session messaging surface. [line 15] [Source]

desktop [Source]

• bypassPermissions mode's exception list expands to include removals targeting //home directory and the cross-session messaging safeguards. [line 29] [Source]
• "Work across sessions" scoped explicitly to the desktop app's own Code-tab surface; new note that in sessions with cross-session messaging enabled, Claude can separately list/message other Claude Code sessions on the machine, including terminal sessions. [lines 37-45] [Source]

devcontainer [Source]

• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC now also disables cross-session messaging in containers (previously only Remote Control). [line 57] [Source]

env-vars [Source]

• New ANTHROPIC_BEDROCK_REGION_PREFIX variable documented. [line 70] [Source]
• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, DISABLE_GROWTHBOOK, DISABLE_TELEMETRY, and DO_NOT_TRACK all updated to note they also disable cross-session messaging. [lines 79, 112, 121, 125]
• New CLAUDE_CODE_MESSAGING_SOCKET variable documented (auto-set by Claude Code, not user-settable). [line 87] [Source]
• New CLAUDE_CODE_USER_DIALOG_TIMEOUT_MS variable documented. [line 95] [Source]
• New CLAUDE_SUBAGENT_BG_SHELL_MAX_MS variable documented — caps subagent background-shell command lifetime at 60 minutes by default (requires v2.1.133+). [line 103] [Source]
• ENABLE_PROMPT_CACHING_1H updated: usage-credit-drawing subscription users can now set it to keep the 1-hour TTL. [line 128] [Source]

errors [Source]

• New error row/section: 403 with "This GraphQL query is not enabled for this session" in cloud sessions. [line 140] [Source]
• New error row/section: "Could not locate the Claude CLI on PATH" — a VS Code extension error on Windows/PowerShell terminals, with full cause and fix steps. [lines 148, 165-177]

feature-availability [Source]

• New "Cross-session messaging" row added to the feature availability table (macOS/Linux only), gated to v2.1.224+, with WSL2 counting as Linux, native Windows unsupported, and cross-machine messaging requiring the same-machine or Remote Control conditions to be met (API key auth doesn't support cross-machine). [lines 199-209] [Source]
• /list-agents//peers commands' availability gated on cross-session messaging noted. [line 191] [Source]
• Provider "Not available" feature lists (Vertex/Bedrock/Foundry/Claude Platform on AWS tabs) all updated to add cross-session messaging as unavailable. [lines 217-243] [Source]

features-overview [Source]

• Cross-references cross-session messaging for separate sessions passing messages without forming a team. [line 279] [Source]

fullscreen [Source]

• New behavior: users can scroll back to the start of a session even after compaction — Claude Code retains every earlier message in fullscreen scrollback across repeated compactions. [line 291] [Source]

gateways [Source]

• Telemetry description updated: gateway pins OTLP export to itself and relays it to configured forward_to destinations. [line 304] [Source]

goal [Source]

• Removed documented behavior: the description of running /goal with no argument to see turns/tokens spent, and that a goal runs until met or /goal clear, was deleted from the page. [line 316] [Source]

headless [Source]

• New clarification: total_cost_usd and the per-model cost breakdown in JSON output are client-side estimates and can differ from the actual bill. [line 329] [Source]

hooks-guide [Source]

• New DirectoryAdded row added to the hook event field-matching table (slash_command, register_repo_root match values). [line 638] [Source]

hooks [Source]

• DirectoryAdded hook documentation expanded: now explicitly lists three cases where it does not fire (--add-dir startup flag, adding a directory via the /permissions Workspace tab, or re-adding an already-present directory, which errors); also clarifies the hook runs asynchronously in the background with the 600-second default timeout and doesn't block the add. [lines 2402-2410] [Source]

interactive-mode [Source]

• Command history section substantially rewritten: Up-arrow recall now reaches prompts from past sessions of the same project, not just the current session; running /clear no longer "resets" history — it lists the new session's prompts first, then earlier sessions' prompts; new note that recalling a prompt with pasted text resends the full paste content, referencing the paste-cache expiry behavior. [lines 221-231] [Source]

llm-gateway-protocol [Source]

• Amazon Bedrock InvokeModel format gains an optional /model/{model}/count-tokens endpoint. [line 31] [Source]
• Startup traffic changed: Anthropic Messages-format gateways now receive a HEAD /api/hello connection-warming probe (replacing the old HEAD /), skipped when a proxy or client cert is configured; Bedrock-format gateways also receive GET /inference-profiles/{profile} lookups when the configured model is an inference profile. [line 41] [Source]

mcp-quickstart [Source]

• "Failed to connect"/"Connection error" troubleshooting expanded: claude mcp list/get now surface the HTTP status/error code and server-reported error text for "Failed to connect" (as of v2.1.219); "Connection error" still shows no detail; new guidance to check claude mcp list for hidden-whitespace warnings. [lines 288-298] [Source]

mcp [Source]

• New behavior: in --output-format stream-json runs, a skipped --mcp-config entry is now reported in the system/init event's mcp_server_errors field (requires v2.1.219+). [line 15] [Source]
• New "Configuration warnings" subsection: Claude Code now detects and warns about hidden leading/trailing whitespace in MCP config values (command, url, args, env/headers keys and values) in claude mcp list and /mcp, without echoing the values. [line 49] [Source]
• New capability: with tool search enabled, when a server finishes connecting mid-turn, Claude Code surfaces its tool names to Claude on the next request within the same turn instead of waiting for the user's next message. [line 62] [Source]
• GitHub PAT connection-failure example now shows the HTTP status returned (e.g. 401). [line 113] [Source]

memory [Source]

• New cross-reference: /init//import can bring another coding agent's configuration into Claude Code, appending instruction files to CLAUDE.md and importing MCP servers, commands, subagents, and skills (requires v2.1.213+). [line 168] [Source]

model-config [Source]

• Clarified that a skill/command running in a forked subagent (context: fork) follows the subagent model-substitution rules instead of being ignored when its model override is blocked. [line 181] [Source]

permission-modes [Source]

• New announcement: auto mode becomes the default permission mode for new sessions on Pro/Max/Team starting August 14, 2026. [line 208] [Source]
• New note: the Cowork tab has its own separate, independently-enabled permission modes and doesn't use the CLI's mode selector. [line 200] [Source]
• bypassPermissions mode gains two new documented cross-session messaging safeguards: the isolatePeerMachines approval prompt for cross-machine sends still fires, and inbound messages follow crossSessionInbound rules, delivered without asking only when the sender also identifies as bypassing. [lines 223-230] [Source]
• Classifier token-usage note narrowed: classifier calls count toward usage specifically on Enterprise plans and API-key/Bedrock/Vertex/Foundry/Claude-Platform-on-AWS accounts. [line 218] [Source]

permissions [Source]

• bypassPermissions mode row/prose updated to note the cross-session messaging safeguards still apply. [lines 243, 246]

prompt-caching [Source]

• New note: set ENABLE_PROMPT_CACHING_1H=1 to keep the 1-hour TTL while drawing on usage credits. [line 51] [Source]

remote-control [Source]

• New paragraph: the Remote Control connection now also carries cross-session messages between a user's machines and from Claude Code on the web sessions (requires v2.1.224+). [line 63] [Source]
• New limitation documented: non-permission/non-AskUserQuestion forwarded dialogs now expire after a default 5-minute deadline, configurable via dialogExpiry (requires v2.1.224+). [line 71] [Source]

sandbox-environments [Source]

• --dangerously-skip-permissions exception list reformatted into bullets, adding the cross-session messaging safeguards as a new exception. [lines 150-159] [Source]

sandboxing [Source]

• WSL2 Windows-binary-launch behavior changed: now governed by the sandbox's Unix-socket settings (allowAllUnixSockets/allowUnixSockets) and requires the optional seccomp filter to be installed to block it in the first place. [line 92] [Source]
• New optional env-var masking fields: extract (regex-based structured masking), onExtractNoMatch (warn/deny/error), and decode: "jwt" with maskClaims for JWT-aware masking (requires v2.1.224+). [lines 189-198] [Source]
• New "Re-sign AWS requests" section: documents credentials.awsPairs (group non-standard AWS variable names for SigV4 re-signing) and credentials.sigv4 (per-request-form policy — streaming/presigned/sigv4a — deny vs passthrough), requires v2.1.224+. [lines 200-235] [Source]
• Credential-file masking gains the same decode: "jwt"/maskClaims support as env vars. [lines 244-253] [Source]

security [Source]

• Cloud execution security section now clarifies self-hosted environment sessions run on customer infrastructure with isolation/network/git-credential responsibility on the deploying org. [line 29] [Source]

sessions [Source]

• Cross-project session ID resolution now requires the other project to hold "a transcript with messages" (stricter than just any transcript) to disambiguate. [line 55] [Source]

settings [Source]

• New crossSessionInbound setting fully documented (accept/hold/refuse, with precedence rules), requires v2.1.224+. [line 70] [Source]
• New dialogExpiry setting documented — default 5m, governs forwarded-dialog and held-cross-session-message-approval deadlines. [line 73] [Source]
• env setting: CLAUDE_CODE_MESSAGING_SOCKET is now explicitly ignored if set there, since Claude Code exports its own per-session value (requires v2.1.224+). [line 82] [Source]
• New isolatePeerMachines setting documented — requires approval before a SendMessage reply reaches a session beyond the local machine; a true from any scope applies (requires v2.1.224+). [line 90] [Source]
• remote.defaultEnvironmentId: self-hosted environment IDs (ccpool_...) are now honored only from user/managed/--settings sources, ignored with a warning from project/local settings. [line 99] [Source]
• Major new sandbox-credentials settings documented: credentials.files[].decode, credentials.files[].maskClaims, credentials.envVars[].extract, credentials.envVars[].onExtractNoMatch, credentials.envVars[].decode, credentials.envVars[].maskClaims, credentials.awsPairs, credentials.sigv4 (all requiring v2.1.224+). [lines 112-133] [Source]
• network.allowAllUnixSockets behavior clarified as tied to seccomp filter installation on Linux/WSL2; on WSL2 it also reopens the Windows-binary interop socket. [line 136] [Source]
• Settings precedence section: new documented exceptions for isolatePeerMachines (a true from any scope always wins) and crossSessionInbound (stricter project/local value overrides trusted sources). [lines 144-145] [Source]
• Plugin marketplace source gains a new file type (local path to a marketplace.json file, distinct from the existing directory type). [line 153] [Source]

skills [Source]

• model frontmatter field: with context: fork, the model override now applies to the forked subagent's model rather than the main session, following subagent model-substitution rules. [line 33] [Source]

sub-agents [Source]

• New cross-reference distinguishing cross-session messaging (separate sessions) from agent teams (Claude-supervised). [line 76] [Source]
• ListAgents tool now follows the same foreground/background filters as other built-in tools — inherited by foreground subagents when cross-session messaging is enabled, removed from background subagents. [line 93] [Source]
• bypassPermissions subagent mode's still-prompting list gains the isolatePeerMachines cross-machine-message approval. [lines 101-110] [Source]
• SendMessage tool description expanded: in cross-session-messaging-enabled sessions, subagents can message other Claude Code sessions on the machine, or reply to sessions beyond it. [line 119] [Source]

terminal-config [Source]

• New paste-cache behavior documented: pasted content is cached under ~/.claude/paste-cache/ so recalled prompts resend full content (even across sessions) until the cache file ages out per cleanupPeriodDays; documents distinct handling when the cache has expired (plain prompt vs. shell-mode/slash-command vs. empty-after-removal cases). [lines 131-136] [Source]

tools-reference [Source]

• New ListAgents tool row added — lists messageable agents (subagents, local sessions, reply-only Remote Control sessions); backs /list-agents; requires v2.1.224+, only in cross-session-messaging-enabled sessions. [line 26] [Source]
• SendMessage tool description expanded to cover messaging other Claude Code sessions (same-machine or cross-machine reply via Remote Control). [line 157] [Source]

troubleshoot-install [Source]

• New error row and full new section: npm error code ENOTEMPTY during update/reinstall, with diagnosis and fix steps (identify and remove leftover package/temp directories, reinstall, verify). [lines 175, 241-279]
• New error row and expanded section: Windows install command printing script text instead of installing (missing | iex or -o), with fixed commands for PowerShell and CMD. [lines 176, 184-205]
• dyld: cannot load on macOS section reorganized to clearly separate two distinct causes (old macOS/libicucore Symbol not found vs. loader rejecting the binary's load commands / Abort trap), each with its own example. [lines 211-234] [Source]

whats-new/2026-w24 [Source]

• Retroactive terminology correction: an existing changelog entry originally reading "Cross-session messaging is hardened" was changed to "Agent messaging is hardened," disambiguating it from the new cross-session-messaging feature (that entry was actually about intra-team agent message hardening). [line 292] [Source]

workflows [Source]

• New runtime constraint: a dynamic workflow script fails before the run starts if it contains import() — module loading is disallowed; the script body is plain JavaScript. [line 305] [Source]
• New behavior: when an org's availableModels allowlist blocks a model a workflow script requests for an agent, that agent now runs on a substituted model per the standard subagent substitution rules, with a warning shown in the /workflows progress view. [line 313] [Source]

API changes

Changed documents

about-claude/pricing [Source]

• The data residency multiplier (1.1x) now explicitly applies to Claude Managed Agents sessions when model.inference_geo is pinned to "us"; the pricing table row previously stating this multiplier didn't apply to Managed Agents was removed. [line 391] [Source]

agents-and-tools/tool-use/advisor-tool [Source]

• New section documenting the advisor tool's availability within Claude Managed Agents sessions, configured as a roster entry ({"type": "advisor", "model": ...}) rather than a tool definition; no max_uses/max_tokens/caching, and advice arrives as thread events instead of advisor_tool_result blocks. [lines 654-657] [Source]

api/overview [Source]

• Rate limits and spend limits are now viewed/managed on separate Console pages (Rate limits page vs. Billing page), replacing the old single "Limits" page reference. [line 162] [Source]

manage-claude/data-residency [Source]

• Claude Managed Agents now supports geographic inference pinning at the agent level via inference_geo in the agent's model configuration (with per-session overrides), reversing prior documentation stating this wasn't supported. [line 14] [Source]
• The 1.1x data-residency pricing multiplier now explicitly applies to Managed Agents sessions running an agent pinned to "us". [line 113] [Source]

manage-claude/inference-hooks [Source]

• Added an explanatory flow diagram (with new image asset) walking through a Cowork example, clarifying that hooks fire at two points: prompt arrival and tool-result return. [lines 26-28] [Source]

manage-claude/workspaces [Source]

• Workspace limits configuration is now split across two separate Console tabs, "Rate limits" and "Spend limits" (previously a single "Limits" tab combining rate limits and spend notifications). [lines 121-124] [Source]
• Corresponding update to the workspace limits explanation, pointing to the specific "Spend limits" and "Rate limits" settings tabs. [lines 233-234] [Source]

managed-agents/agent-setup [Source]

• The model object now accepts an inference_geo field in addition to speed/effort. [line 20] [Source]
• New documentation: how to pin inference_geo ("us" or "global") on an agent's model config, validation against workspace allowed_inference_geos at save/session-create/each turn, a 400 error if set on an unsupported model, and the requirement that in a multiagent config the coordinator and all roster members must share the same pin (or all be unset). [lines 63-67] [Source]
• Clarified update semantics: supplying a model object without inference_geo clears any existing pin, in addition to the existing effort reset behavior. [line 143] [Source]

managed-agents/events-and-streaming [Source]

• New "Reaching a session budget" section: describes the budget_reached stop reason, the pause sequence (three ordered stream events), how a thread that both crosses the cap and finishes reports end_turn while the session reports budget_reached, which events are still accepted while paused, and that resuming requires a budget update. [lines 396-408] [Source]
• The session usage object now documents new fields: list_cost, active_seconds, and server_tool_use (web_search_requests/web_fetch_requests), alongside existing token counts, plus per-thread vs. session-level usage reconciliation. [lines 450-471] [Source]
• New guidance that the session.usage stream event carries the same cumulative snapshot plus budget, emitted before idle transitions and when a thread pauses at budget, with a recommendation to use session budgets instead of manually polling usage to enforce spend limits. [lines 469-471] [Source]

managed-agents/github [Source]

• Mounting a GitHub repository now also loads skills from its root .claude/skills directory automatically, discovered once at session start. [line 57] [Source]

managed-agents/multiagent-orchestration [Source]

• New roster entry type {"type": "advisor", "model": "<model id>"} giving the session's primary thread a mid-turn advisor (max one per roster). [line 63] [Source]
• New "Give the session an advisor" section: entry schema, reserved roster name anthropic.advisor, model-capability pairing rules, differences from the Messages API advisor tool, consultation thread lifecycle/event ordering, redacted vs. plaintext advice delivery per advisor model policy, interrupt behavior, exemption from concurrent-thread limits, prompt caching/billing, and how to remove the advisor. [lines 71-130] [Source]
• New rule: when agents pin inference_geo, the coordinator's and all roster members' pins must match (all same value or all unset), enforced with a 400 on save or session-create override. [line 69] [Source]
• New note: session budgets are a single shared cap across all threads, with each thread priced at its own served model; advisor consultation threads are exempt from the 25-concurrent-thread limit. [lines 202-206] [Source]
• The multiagent thread events table now notes advisor consultations emit the same events under the reserved name anthropic.advisor. [line 246] [Source]

managed-agents/scheduled-deployments [Source]

• Deployments can now take an optional budget object (same shape as a session budget) that's copied onto every session the deployment starts, bounding each run independently rather than acting as a cumulative cap; example shows updating an existing deployment's budget via PATCH, and budgets can be cleared with "budget": null. [lines 84-110] [Source]

managed-agents/self-hosted-sandboxes [Source]

• Added Fly.io to the list of platform-specific self-hosted sandbox integration guides. [line 36] [Source]

managed-agents/session-operations [Source]

• Clarified that an agent's model configuration, including its inference_geo pin, cannot change mid-session — it must be set at agent save time or via a session-create model override. [line 28] [Source]
• New "Updating the session budget" section: how to replace or remove (null) a session's budget, both of which auto-resume paused work; a replacement cap must exceed consumed list cost, and removal is one-way (can't re-add). [lines 52-54] [Source]

managed-agents/sessions [Source]

• New behavior: a session-create model override now also sets/clears the inference_geo pin for that session, validated against workspace allowed_inference_geos. [line 103] [Source]
• New "Set a session budget" section: budget object schema (type: "limit", max_list_cost with amount in cents as a string + currency, USD only), pause behavior at budget_reached, budget can only be set at creation, and can be changed or removed later (not added after the fact), with a full cURL example. [lines 135-163] [Source]

managed-agents/skills [Source]

• Skills can now reach an agent two ways: attached via the agent's skills array, or newly, loaded from a GitHub repository mounted on the session. [line 9] [Source]
• New major section "Load skills from a GitHub repository": discovery at exactly .claude/skills/<skill-name>/SKILL.md (one level deep at repo root), requires the read tool, a security warning about the trust boundary (repo contributors can inject instructions), cloud-sandbox-only (not supported for self-hosted sandboxes), discovery runs once at session start against the checked-out ref, and repository skills can coexist with attached skills. [lines 73-142] [Source]

managed-agents/webhooks [Source]

• New webhook event session.budget_reached: fires when a session hits its budget and pauses; fires at most once per budget value, re-armed by changing the budget. [line 45] [Source]
• session.thread_created description updated to also cover advisor consultations starting. [line 48] [Source]

release-notes/overview [Source]

• New "August 7, 2026" release notes entry covering four features: session budgets, session advisors, Managed Agents inference_geo pinning, and loading skills from a GitHub repository. [lines 15-20] [Source]