Claude docs changes for July 1st, 2026 [diff]

Executive Summary

• Claude Sonnet 5 launches as the new default model in Claude Code (v2.1.197), with a native 1M-token context window and promotional pricing of $2/$10 per Mtok through August 31, 2026 — the rollout touches nearly every API and Claude Code doc (pricing, platform availability, adaptive-thinking-only behavior, a new tokenizer producing ~30% more tokens than prior models)
• New self-serve Tunnels API (/v1/tunnels, beta) lets developers create and manage tunnels directly via a Workload Identity Federation token, without org-admin access; the old admin-only MCP Tunnels API (/v1/organizations/tunnels) is now marked deprecated with a migration window
• Claude Desktop launches on Linux (beta) — installable via apt or a .deb package on Ubuntu 22.04+/Debian 12+, with Chat, Cowork, and Code tabs (Computer Use and voice dictation not yet supported)
• Managed Agents gain session-level configuration overrides (swap model, system, tools, mcp_servers, or skills for a single session without a new agent version) and an opt-in incremental event-delta streaming preview for agent messages/thinking
• Permission SDK docs clarify a common gotcha: tools auto-approved by allow rules, acceptEdits, or bypassPermissions never reach canUseTool or hooks — guidance now recommends a PreToolUse hook for logic that must run on every tool call regardless of approval path

New Claude Code versions

2.1.197

New features

• Claude Sonnet 5 is now the default model in Claude Code, with a native 1M-token context window and promotional pricing of $2/$10 per Mtok through August 31, 2026

Claude Code changes

New Documents

Claude Desktop on Linux (beta) [Source]

Documents the new beta Linux release of the Claude Desktop app (Chat, Cowork, and Code tabs), supported on Ubuntu 22.04+/Debian 12+ (x86_64 or arm64). Covers installation via Anthropic's apt repository (with signing-key verification) or a downloaded .deb package, plus update/uninstall via standard apt commands. Lists what's not yet available: Computer Use, voice dictation, the Quick Entry global hotkey on native Wayland (works on X11), and Fedora/RHEL support.

Changed documents

Advisor [Source]

• Sonnet 5 added to the advisor-eligibility table: it can call and act as an advisor for Fable, Opus, and Sonnet 5, but a Sonnet 4.6 advisor is rejected when the main model is Sonnet 5 [line 64]
• Supported main-model requirement loosened from "Sonnet 4.6" to "Sonnet 4.6 or later" [line 117]

Agent SDK — Agent loop [Source]

• Example model ID updated from claude-sonnet-4-6 to claude-sonnet-5 [line 169]

Agent SDK — Permissions [Source]

• New "ask rules" step added to the permission evaluation flow diagram (now 6 steps: hooks, deny rules, ask rules, permission mode, allow rules, canUseTool) [line 47]
• New documented behavior: auto-approved tools (via allow rules, acceptEdits, or bypassPermissions) never reach canUseTool — coverage depends on rule form, and a PreToolUse hook is recommended to guarantee checks run on every call [lines 65-66]

Agent SDK — Python [Source]

• can_use_tool field clarified: only invoked when the permission flow falls through to a prompt, not for calls auto-approved by allowed_tools, allow rules, or permission_mode; a PreToolUse hook is recommended for logic that must run on every call [lines 795, 1075]
• Code example updated to warn against listing gated tools in both allowed_tools and can_use_tool [lines 634-636]
• 1M-context beta retirement migration guidance now includes Claude Sonnet 5 as a migration target [line 1254]

Agent SDK — TypeScript [Source]

• canUseTool clarified to only fire when the permission flow falls through to a prompt, not for calls auto-approved by allowedTools, allow rules, or permissionMode [lines 391, 766]
• 1M-context beta retirement migration guidance now includes Claude Sonnet 5 as a migration target [line 2755]

Agent SDK — User input [Source]

• New explicit statement: canUseTool never fires for auto-approved tools; a bare allowed_tools entry skips it unless an ask rule or plan mode routes the call back to a prompt — recommends PreToolUse hooks for logic that must run on every call [lines 27-29]

Agent teams [Source]

• New documented security behavior: messages relayed between teammates via SendMessage are marked as coming from another Claude session, not the user; a teammate can't approve permission prompts or relay bypass-consent on another teammate's behalf, and in auto mode a relayed "approval" is treated as untrusted input [lines 220-221]

Agent view [Source]

• @<repo> dispatch clarified: typing @ lists git repositories one level below the launch directory plus any directory that already has a session; directories with a space in the name are not listed [lines 225, 237-240]

Amazon Bedrock [Source]

• Sonnet 5 added to the models supporting the 1M-token context window on Bedrock; runs via the Mantle endpoint and always uses 1M context (no [1m] variant needed) [lines 280-281, 327]

Claude apps gateway [Source]

• Intro rewritten to describe when the gateway is preferred vs. when Claude Enterprise is a better fit, now pointing to a new feature-availability comparison page [line 3]
• Auto mode compatibility note broadened from "only Opus models" to "only the models eligible on third-party providers," reflecting Sonnet 5 eligibility [line 277]

Claude Platform on AWS [Source]

• Example ANTHROPIC_DEFAULT_SONNET_MODEL updated to claude-sonnet-5 [line 78]

CLI reference [Source]

• --model flag example updated from claude-sonnet-4-6 to claude-sonnet-5 [line 85]

Context window [Source]

• Sonnet 5 added to the models supporting the 1M-token context window, with a pointer to its own auto-compaction/LLM-gateway section [line 96]

Desktop application [Source]

• New Linux (beta) download entry added, linking to the new Linux page; Linux is no longer described as unsupported [lines 13-16]
• Auto mode model requirement loosened to "Sonnet 4.6 or later"; Vertex AI Enterprise auto-mode support expanded to include Sonnet 5 [line 72]
• Sonnet 5 added to the models that always use adaptive reasoning (fixed-budget override doesn't apply) [line 467]
• Linux limitation entry updated: desktop app now available (beta), but Computer Use isn't yet supported on Linux [line 643]
• Update-checking troubleshooting note added: Linux updates go through apt, not app auto-update [line 675]

Desktop quickstart [Source]

• New Linux (beta) download entry (apt or .deb) and Linux-specific install/sign-in steps (launcher instead of Applications/Start menu) [lines 13-16, 36]

Environment variables [Source]

• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE / CLAUDE_CODE_AUTO_COMPACT_WINDOW: now note Sonnet 5 has its own default proactive-compaction threshold [lines 130, 142]
• CLAUDE_CODE_DISABLE_1M_CONTEXT: now documents that setting this also forces Sonnet 5 sessions down to a 200K window [line 152]
• CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING: documented as having no effect on Sonnet 5 (in addition to Fable 5 and Opus 4.7+) [line 153]
• CLAUDE_CODE_SHELL: description substantially expanded — only accepts a path to a bash/zsh binary (other shells like fish unsupported), falls back to auto-detection otherwise, and documents the detection algorithm ($SHELL, then first working zsh then bash on PATH) [line 242]

Errors [Source]

• "thinking.type.enabled is not supported" error now also attributed to Claude Code versions older than the Sonnet 5 minimum [line 643]
• New version requirement documented: Sonnet 5 needs Claude Code v2.1.197+ (TypeScript SDK v0.3.197+) [lines 651, 653]

GitHub Actions [Source]

• Example model IDs updated from claude-sonnet-4-6 to claude-sonnet-5 across config examples [lines 93, 106, 603, 610]

Glossary [Source]

• "Effort level" entry: supported model list loosened to "Sonnet 4.6 and later" [line 103]

Google Vertex AI [Source]

• Example ANTHROPIC_DEFAULT_SONNET_MODEL updated to claude-sonnet-5; Sonnet 5 added to the 1M-token context-window model list (always 1M, no [1m] variant needed) [lines 136, 176]

Hooks [Source]

• New documented limitation: PreToolUse does not fire for files referenced with @ in a prompt (inserted directly, no tool call); use a Read deny rule to block specific paths instead [lines 1234-1235]
• New PowerShell guidance: ${CLAUDE_PROJECT_DIR} substitution in shell-form commands only works for plugin hooks; settings.json hooks on PowerShell should use $env:CLAUDE_PROJECT_DIR or exec form, with a new example [lines 2877-2886]

Microsoft Foundry [Source]

• Example ANTHROPIC_DEFAULT_SONNET_MODEL updated to claude-sonnet-5 [line 78]

Model configuration [Source]

• On the Anthropic API, the sonnet alias now resolves to Sonnet 5 (previously Sonnet 4.6); requires Claude Code v2.1.197+ [lines 29, 31]
• Default model table updated: Pro/Team Standard/Enterprise seats now default to Sonnet 5 [line 197]
• Effort-level tables updated: Sonnet 5 added at the low/medium/high/xhigh/max tier, defaulting to high [lines 284, 287, 301]
• New dedicated "Sonnet 5 context window" subsection: always runs with 1M context (no 200K variant, no [1m] suffix, no usage credits needed), auto-compacts around 967K tokens by default; documents two exceptions that budget it at 200K (behind an LLM gateway without explicitly selecting sonnet[1m], or CLAUDE_CODE_DISABLE_1M_CONTEXT=1) [lines 347-349, 370-378]
• Sonnet 5 added to models that always use adaptive reasoning [line 330]

Monitoring usage [Source]

• Example model attribute values updated from claude-sonnet-4-6 to claude-sonnet-5 across telemetry attribute tables [lines 421, 443, 461, 572, 595]

Permission modes [Source]

• Auto mode model requirement loosened to "Sonnet 4.6 or later"; third-party provider (Bedrock/Vertex/Foundry/gateway) support expanded to include Sonnet 5, alongside Opus 4.7/4.8 [lines 143, 151]

Server-managed settings [Source]

• Security framing strengthened: explicitly states server-managed settings are "not a security boundary" — bypassing on unmanaged devices doesn't require admin/sudo access [line 205]
• New rows in the security-considerations table: a modified binary bypasses client-side controls; an old Claude Code version predating server-managed settings never fetches them; intercepted/redirected network traffic can alter received settings [lines 211-212, 216]
• New reference to "Tenant Restrictions" (network-level access control) in the Claude Help Center [line 219]

Settings [Source]

• enforceAvailableModels clarified: fallback to the first allowlisted entry only occurs when the account-type default model is not already allowlisted [line 229]
• Example fallbackModel/model values and default commit-attribution trailer updated to Sonnet 5 [lines 231, 248, 437]

Setup [Source]

• Desktop app download callout now includes a Linux download link alongside macOS and Windows [line 26]

Skills [Source]

• New documented capability: skill entries can be symlinks to directories elsewhere on disk — Claude Code follows the symlink and reads SKILL.md from the target, deduplicating shared targets; plugin skills handle symlinks differently [line 105]

Sub-agents [Source]

• Full model ID example updated from claude-sonnet-4-6 to claude-sonnet-5 [line 269]

Terminal guide [Source]

• Desktop app download callout now includes a Linux download link [line 8]

Troubleshoot installation [Source]

• New troubleshooting entry for A parameter cannot be found that matches parameter name 'fsSL' — covers running the macOS/Linux curl installer in Windows PowerShell (where curl aliases to Invoke-WebRequest), with a link to the correct PowerShell command [lines 21, 40, 436-442]
• Desktop app download callout now includes a Linux download link [line 40]

API changes

New Documents

Tunnels [Source]

New self-serve, non-admin Tunnels API at /v1/tunnels (research preview, requires the anthropic-beta: mcp-tunnels-2026-06-22 header). It supersedes the org-admin-only Tunnels API (/v1/organizations/tunnels): the underlying resources (BetaTunnel, BetaTunnelToken, BetaTunnelCertificate) and operations (create/list/retrieve/archive tunnel, reveal/rotate token, and a certificates sub-resource with create/list/retrieve/archive) are the same, but authorization now uses a Workload Identity Federation token scoped to workspace:manage_tunnels (previously org:manage_tunnels), and POST /v1/tunnels can allocate a brand-new tunnel programmatically instead of only attaching to a Console-created one. Equivalent per-language SDK reference pages were also added under each language's beta/tunnels/ path.

Changed documents

Admin [Source]

• All MCP Tunnels endpoints (tunnel and tunnel-certificate operations) are now marked Deprecated in the endpoint index [lines 366-410]

Admin — Analytics [Source]

• assigned_seat_count, daily_adoption_rate, monthly_adoption_rate, weekly_adoption_rate, and pending_invite_count are now documented as null when a response is scoped to an RBAC group

Admin — Chat Projects usage [Source]

• distinct_conversation_count changed from a required field to optional number, now null on aggregated rows where a distinct count can't be computed [lines 20-60]

Admin — Get Activity Summaries [Source]

• Same RBAC-group null behavior added for assigned_seat_count, daily_adoption_rate, monthly_adoption_rate, weekly_adoption_rate, pending_invite_count [lines 40-92]

Admin — Get Per-User Cost [Source]

• Clarified this endpoint only includes cost attributable to seat users; org-wide totals (including direct API-key/automation traffic) require the new bucketed /v1/organizations/analytics/cost_report endpoint [lines 11-17]

Admin — Get Per-User Token Usage [Source]

• Clarified this endpoint only includes usage attributable to seat users; org-wide totals require the new bucketed /v1/organizations/analytics/usage_report endpoint [lines 11-17]

Admin — Users analytics [Source]

• New bioscience_metrics object added to per-user analytics: delegation_count, distinct_session_count, message_count, remote_compute_job_count, skills_used_count [lines 20-45]
• chat_metrics.connectors_used_count redefined as "Number of MCP connector invocations" (no longer nullable/distinct-based); new chat_metrics.distinct_connectors_used_count field added (nullable on aggregated rows) [lines 48-60]

Admin — External keys [Source]

• AWS provider_config.role_arn is now deprecated and optional (previously required) — Anthropic reaches the KMS key via a managed intermediate role, and the field is ignored if supplied; applies across create/list/retrieve/update

Admin — MCP Tunnels [Source]

• All endpoints (Create/Get/List/Archive Tunnel, Reveal/Rotate Token, and the Tunnel Certificates create/list/retrieve/archive sub-resource) are now marked Deprecated, superseded by the new /v1/tunnels beta API; existing integrations keep working with the old header (mcp-tunnels-2026-05-19) and scope (org:manage_tunnels) during a migration window [line 9 in each endpoint page]

Admin — Approve Spend Limit Increase Request [Source]

• Example response status corrected from "pending" to "approved", reflecting the endpoint's actual post-approval state [lines 372, 438]

Admin — List Effective Spend Limits [Source]

• Terminology clarified from "cap" to "spend limit" throughout (no behavior change)

Beta [Source]

• New Tunnels and Tunnel Certificates endpoint groups indexed (see the new Tunnels page above) [line 785]

MCP tunnels — Architecture and components [Source]

• Required WIF scope renamed from org:manage_tunnels to workspace:manage_tunnels [line 31]

MCP tunnels — Manage tunnels in the Console [Source]

• Same scope rename (org:manage_tunnels → workspace:manage_tunnels) in prerequisites and federation-rule setup steps [lines 18, 42]

MCP tunnels — Deploy with Docker Compose [Source]

• Tunnel creation is no longer required up front: leaving TUNNEL_ID unset now lets the setup component create the tunnel itself [line 17]
• mcp-proxy image digest bumped; install/rotation instructions updated so TUNNEL_ID is optional; setup init reruns now reuse the stored tunnel ID and never create a second tunnel [lines 89, 133, 174, 189, 196, 268]

MCP tunnels — Deploy with Helm [Source]

• Chart now supports creating a new tunnel during install (not only attaching to a pre-existing Console tunnel); Helm chart version bumped from 1.0.0 to 2.0.0 [lines 11, 16, 131, 183, 196, 255, 273]
• values.yaml schema changed: api.wif.tunnelId removed in favor of a new tunnel.id field (empty = auto-create); new guidance on reading the auto-created tunnel's domain from the mcp-tunnel Secret [lines 148-155, 205-215]
• New "Upgrade from chart 1.x" section documenting the api.wif.tunnelId → tunnel.id migration and required scope rename [lines 244-246]

MCP tunnels overview [Source]

• Prerequisites updated: a tunnel can now be created via the API or auto-created by the Helm setup hook, not only via the Console; WIF scope renamed to workspace:manage_tunnels [lines 29, 31]

MCP tunnels quickstart [Source]

• mcp-proxy image digest bumped to match the new build [line 149]

MCP tunnels reference [Source]

• Major rewrite of the "Tunnels API" section: points to the new /v1/tunnels beta API as primary, with the old admin API called out as deprecated with a migration window and explicit migration steps [lines 39-49]
• Required scope updated to workspace:manage_tunnels, required beta header updated to mcp-tunnels-2026-06-22; CA certificate upload path updated to /v1/tunnels/{tunnel_id}/certificates [lines 51, 61, 69]
• setup init / --tunnel-id flag: now optional — omitting it creates a new tunnel, a previously-stored ID is reused on reruns [lines 97, 100]

MCP tunnels security [Source]

• "Archive over the API" link retargeted from the deprecated admin API to the new beta Tunnels API archive endpoint [line 55]

Troubleshoot MCP tunnels [Source]

• WIF scope in the 403 troubleshooting tip renamed to workspace:manage_tunnels [line 85]

Computer use tool [Source]

• Claude Sonnet 5 added to models supporting the computer-use-2025-11-24 beta and enhanced actions (e.g. zoom); added to the higher screenshot resolution tier (up to 2576px long edge) [lines 10-13, 297-300, 568-571]

Web fetch tool [Source]

• Claude Sonnet 5 added to the models supporting dynamic filtering in web_fetch_20260318 [lines 6-8]

Web search tool [Source]

• Claude Sonnet 5 added to the models supporting dynamic filtering in web_search_20260318 [lines 6-8]

API reference (all SDKs) — Messages: create, count_tokens, batches [Source]

• New refusal category enum value military_weapons (alongside cyber, bio, frontier_llm, reasoning_extraction)
• New tool versions web_search_20260318/web_fetch_20260318 adding allowed_callers, defer_loading, use_cache, max_content_tokens, and response_inclusion (full/excluded — controls whether nested server_tool_use/result blocks are dropped when consumed by a same-turn code_execution call)
• New header anthropic-user-profile-id (optional) on count_tokens and batch create to attribute requests to a user profile (requires the user-profiles beta header)

API reference (all SDKs) — Beta Messages: create, count_tokens, batches [Source]

• New header anthropic-user-profile-id (optional) on create/count_tokens/batch-create to attribute a request to a user profile, requiring the user-profiles beta header; removes the previous user_profile_id body field it supersedes
• Same new web_search_20260318/web_fetch_20260318 tool versions and military_weapons refusal category as the non-beta Messages API

Adaptive thinking [Source]

• Claude Sonnet 5 added as a supported model: adaptive thinking is on by default, manual {type: "enabled", budget_tokens: N} is rejected with a 400 error, and {type: "disabled"} turns it off; added to max/xhigh effort levels and to models where thinking.display defaults to "omitted" [lines 11, 28, 79-80, 139-141, 215, 226, 243]

Batch processing [Source]

• New Batch API pricing for Sonnet 5: $1/$5 per MTok through August 31, 2026, then $1.50/$7.50 per MTok from September 1, 2026; added to the output-300k-2026-03-24 beta (300k max_tokens for batch requests) [lines 97-98, 432]

Claude in Amazon Bedrock [Source]

• Sonnet 5 now open to all Bedrock customers (anthropic.claude-sonnet-5), including via the global (non-regional) endpoint [lines 15, 184, 221]

Claude in Microsoft Foundry [Source]

• Claude Sonnet 5 (preview) added to the Foundry model table with 1M-token context window support [lines 266, 311]

Claude on Amazon Bedrock (legacy) [Source]

• Note added: Sonnet 5 is not available on the legacy ARN-based Bedrock surface — use Claude in Amazon Bedrock or Claude Platform on AWS instead [line 88]

Claude on Google Cloud [Source]

• claude-sonnet-5 added to the Vertex AI model table and the 1M-token context-window model list [lines 70, 159]

Claude Platform on AWS (API) [Source]

• claude-sonnet-5 added to the list of models available on Claude Platform on AWS [line 292]

Compaction [Source]

• Claude Sonnet 5 added to the models supporting compaction [line 41]

Context windows [Source]

• Claude Sonnet 5 added to the 1M-token context-window list, context-awareness support (1M budget vs. 200k for Sonnet 4.5/Haiku 4.5), and beta server-side compaction support [lines 37, 108, 120, 142]

Effort [Source]

• Claude Sonnet 5 added as supporting the effort parameter and the max/xhigh tiers; new "Recommended effort levels for Claude Sonnet 5" subsection (defaults to high) [lines 15, 44-45, 54-63, 131-132]
• New note on Sonnet 5's adaptive-thinking/effort interaction: manual thinking unsupported (400 error), {type:"disabled"} turns thinking off [line 166]

Extended thinking [Source]

• Claude Sonnet 5 added to the model-support table: manual extended thinking not supported (400 error) — use adaptive thinking + effort instead; added to the 128k max-output-token group, the interleaved-thinking-2025-05-14 deprecated/ignored group, and the thinking-behavior comparison table [lines 23, 101, 111, 187, 441, 453-454, 675]

Prompt caching [Source]

• New prompt-caching pricing for Sonnet 5: $2/$2.50/$4/$0.20/$10 per MTok through August 31, 2026, then $3/$3.75/$6/$0.30/$15 per MTok from September 1, 2026; added to the 1,024-token minimum cacheable prompt length group [lines 91-92, 280]

Search results [Source]

• claude-sonnet-5 added to the models supporting the search results feature [line 18]

Using Agent Skills with the API [Source]

• Corrected the "delete all skill versions" CLI example: replaced a shell loop piping versions list into versions delete with an explicit two-step (list, then delete each version id) [lines 388-397]

Structured outputs [Source]

• Claude Sonnet 5 added to the Claude API, Google Cloud, and Amazon Bedrock general-availability lists for structured outputs [line 16]

Task budgets [Source]

• Claude Sonnet 5 added to the task-budgets model-support table as "Not supported" [line 256]

Token counting [Source]

• New note: Opus 4.7+ models, Fable 5, Mythos 5/Preview, and Sonnet 5 use a newer tokenizer producing ~30% more tokens for the same input text than earlier models — recommends recounting prompts per target model [lines 31, 33]

Vision [Source]

• Claude Sonnet 5 added to the high-resolution image tier (2576px max long edge, 4784 max visual tokens) [line 230]
• Cost-estimate example switched from Sonnet 4.6 to Haiku 4.5 as the standard-tier example [line 246]

Using the Messages API [Source]

• Claude Sonnet 5 added to the models where response prefilling is unsupported (400 error) [line 127]

Intro to Claude [Source]

• Model lineup updated: Claude Sonnet 5 replaces Claude Sonnet 4.6 as the frontier Sonnet-tier model, with a link to the announcement [line 17]

Create an Admin API key [Source]

• Compliance API scope table merges read:compliance_org_settings into read:compliance_org_data: reading org metadata and effective settings now only requires read:compliance_org_data [line 78]

API and data retention [Source]

• MCP tunnels API endpoint path updated from /v1/organizations/tunnels to /v1/tunnels [line 188]

Get access to the Compliance API [Source]

• read:compliance_org_settings scope removed from the scope-selection table; read:compliance_org_data now also grants read access to effective org settings [lines 77-80]

Handle Compliance API errors [Source]

• read:compliance_org_settings scope retired as of June 30, 2026; the settings endpoint now requires read:compliance_org_data — existing keys with only the retired scope will fail on every call; fix is to create a new key with the new scope [lines 137-160]

List organizations, users, roles, groups, and settings [Source]

• read:compliance_org_settings scope retired in favor of read:compliance_org_data [lines 13, 240-247]
• New api_keys array added to the effective org settings response, listing every Compliance Access Key on the parent org (id, name, scopes, is_active, created_at, expires_at, created_by_id); secret values are never returned, and deactivated/retired-scope keys remain listed for audit [lines 290-300, 310-313]

Usage and Cost API [Source]

• Removed "Priority Tier optimization" as a stated use case of the Usage & Cost API [line 18]

Manage WIF with the Admin API [Source]

• OAuth scope renamed: org:manage_tunnels → workspace:manage_tunnels [lines 220-221]

Use WIF with Microsoft Entra ID [Source]

• New required step: register a dedicated Entra app registration/service principal as the "Claude API audience" (api://<APP_ID>) before token exchange will succeed [lines 9-63]
• New guidance: managed-identity tokens need up to a 24-hour lifetime, requiring max_jwt_lifetime_seconds: 86400 on the federation issuer (the wizard's default 7500 is insufficient) [lines 158-163]
• New documented App Service/Functions/Container Apps token endpoint (IDENTITY_ENDPOINT/IDENTITY_HEADER) as an alternative to IMDS [lines 100-105]
• Full step-by-step AKS Entra Workload Identity path added (issuer/cluster setup, federated credential, service account annotation, two-hop token exchange); new "If your tokens are v1.0" section on legacy claim differences [lines 250-446, 492-502]

WIF reference [Source]

• OAuth scope renamed to workspace:manage_tunnels; its granted actions now include create tunnel (previously list/get/certs/reveal/rotate/archive only) [line 132]

Workload Identity Federation [Source]

• Default MCP tunnels rule scope renamed from org:manage_tunnels to workspace:manage_tunnels [line 44]

Workspaces [Source]

• MCP tunnels are now managed with a workspace-scoped workspace:manage_tunnels OAuth token instead of the org-scoped org:manage_tunnels [lines 212-214]

Define your agent [Source]

• New capability: session-level overrides for model, system, tools, mcp_servers, and skills without creating a new agent version [line 27]

Session event stream [Source]

• New "Event deltas" opt-in incremental streaming preview for agent.message/agent.thinking via a new event_deltas[] query parameter on the events-stream endpoint; new non-persisted event_start/event_delta event types with their own wire format [lines 18, 89-208]
• Documented limitations: best-effort/lossy under load, no replay on reconnect, primary-thread-text-only, thinking previews are start-only, deltas never persisted in event history [lines 202-208]

Multi-agent sessions [Source]

• Session-level configuration overrides now apply to the coordinator and its self copies (not roster entries referenced by ID); session-creation overrides can now replace the coordinator's MCP servers [lines 19, 60, 88]

Managed Agents reference [Source]

• New "Event deltas" event-type category added alongside User/Session/Span/Agent/System events [lines 15, 37-40]

Scheduled deployments [Source]

• New webhook events for deployment lifecycle changes and deployment run outcomes; new endpoint GET /v1/deployment_runs/{deployment_run_id} to retrieve a single run by ID [lines 84, 126]
• Unrecoverable session-creation errors (e.g. archived environment or vault), not just agent_archived_error, now auto-pause the deployment and record a failed run [line 166]

Session operations [Source]

• Clarified: only tools and mcp_servers can be updated on an existing session; model, system, skills require the new session-creation overrides instead [lines 28-29]
• New pagination documentation for GET /v1/sessions: limit, next_page/prev_page cursors, order (default desc); cursors are order-specific [lines 64-68]

Start a session [Source]

• New "Override agent configuration for a session" feature: pass agent as type: agent_with_overrides with an id, optional version, and any of model/system/tools/mcp_servers/skills to override for a single session, with detailed field-clearing semantics [lines 43-77]

Authenticate with vaults [Source]

• New capability: injection_location on environment-variable credentials can now be updated/rotated (previously only secret values and display_name were mutable); updates propagate to running sessions without restart [line 143]

Subscribe to webhooks [Source]

• New webhook event categories: Agent events, Deployment events, Deployment run events, alongside existing Session/Vault events [lines 21-31]