Claude Fable 5 model, Cursor cloud agents, and Codex CLI relays for daily builds

What shipped

Across this week from 12 June 2026 to 19 June 2026, three coding platforms released updates focused on agent depth, cloud execution, and CLI controls. Claude Code shipped the largest set of changes including a new Mythos-class model and permission rules. Cursor and OpenAI Codex followed with SDK extensions, design tools, and relay features that affect how teams run agents locally or in the cloud.

Cursor

Cursor delivered twenty updates centered on its SDKs, Design Mode, Bugbot reviews, and cloud agent controls. The releases added custom tools, multi-select editing, and local pre-push reviews that compete directly with VS Code extensions. Builders gain faster iteration on UI tasks and parallel cloud runs without leaving the IDE.

• •TypeScript and Python SDKs Cursor added custom local tools, auto-review classification, JSONL stores, and nested subagents to the SDKs with fixes for production scripts.
• •Design Mode multi-select Cursor now lets users select multiple UI elements at once to match styles and layouts while the agent applies changes.
• •Bugbot with Composer 2.5 Bugbot runs 3x faster and 22% cheaper, finds 10% more bugs, and supports local /review runs before Git pushes.
• •Cloud agents launch Users can spin up cloud environments in under 10 minutes and hand off sessions to isolated subagents via the Agents Window.
• •Cursor Automations The new /automate skill plus GitHub and Slack triggers let always-on agents handle repetitive tasks with computer-use visuals.
• •Canvas Design Mode Canvases now support direct UI annotation inside Design Mode plus token usage reports for optimization.
• •SDK updates for agents The TypeScript and Python SDKs gained MCP custom tools, auto-review routing, and infinitely nested subagents.
• •Voice input in Design Mode Continuous voice narration lets users queue UI edits without waiting for each agent step to finish.
• •Local Bugbot reviews Reviews limited to new PR changes now run locally via /review and sync with GitHub and GitLab.
• •Agents Window management Cloud subagents and environment handoffs between local and parallel VMs are now controlled from one window.
• •Enterprise Organizations Teams can manage multiple Cursor groups under one top-level org with centralized IDP and usage rollups.
• •Canvas context reports Interactive reports now show token distribution and allow full-screen shared canvases in the browser.
• •Nested subagents in SDKs Custom tools and JSONL metadata stores support deeper agent hierarchies in production scripts.
• •Multi-select UI edits Design Mode conveys visual relationships to agents by selecting groups of elements at once.
• •Selective PR reviews Bugbot on Composer 2.5 skips unchanged files and runs security checks locally before push.
• •Multiple team management Enterprise orgs add Groups for granular model access and spend limits across teams.
• •Design Mode on Canvases UI element selection and annotation now work directly inside shared canvases.
• •Advanced agent SDKs Performance fixes accompany custom tool support and auto-review routing in both TypeScript and Python.
• •Voice and multi-select Continuous narration and group element selection speed up UI edits in the browser.
• •Faster Bugbot runs Local /review commands now avoid duplicate GitHub reviews while cutting cost by 22%.

Claude Code

Claude Code released forty-nine updates, the largest set this week, focused on sub-agent nesting, permission rules, and the new Claude Fable 5 model. Changes include five-level sub-agent spawning, Tool(param:value) syntax, and safe-mode troubleshooting that improve reliability for complex agent teams. The updates also add /config syntax and post-session hooks for self-hosted runners.

• •Directory navigation command The new /cd command moves a session to a different working directory without breaking the prompt cache.
• •Safe mode and hooks A --safe-mode flag disables custom plugins for troubleshooting while post-session hooks support self-hosted runners.
• •Claude Fable 5 model The Mythos-class model is now available in version 2.1.170 for advanced reasoning and safe general use.
• •Five-level sub-agents Sub-agents can spawn their own sub-agents up to five levels deep for hierarchical task delegation.
• •enforceAvailableModels setting Managed settings now prevent user overrides from widening the allowed model list.
• •Parameter-level permissions Tool(param:value) syntax lets rules block or allow tool calls based on specific input values.
• •Implicit teammate spawning Users spawn teammates directly with the Agent tool name instead of separate TeamCreate commands.
• •Bun runtime 1.4 The bundled runtime upgrade improves execution performance and compatibility for internal processes.
• •Prompt config syntax /config key=value lets users adjust any setting directly from the interactive prompt.
• •Auto mode safety blocks Destructive git and terraform commands are now blocked in auto mode unless explicitly allowed.
• •Safe mode and /cd added The --safe-mode flag and directory change command help troubleshoot without cache loss.
• •Fable 5 general access Claude Fable 5 exceeds prior models and is accessible after updating to 2.1.170.
• •Multi-level sub-agents Sub-agents reach five levels deep with added plugin marketplace search and Bedrock region fixes.
• •Wheel scroll setting A new option disables mouse-wheel acceleration in fullscreen mode and fixes the model picker display.
• •Managed model enforcement The enforceAvailableModels setting keeps default models inside the allowlist.
• •Localized session titles Titles now generate automatically in the conversation language with improved Bedrock caching.
• •Parameter permission rules Tool(param:value) syntax and nested skill directory support resolve name clashes.
• •Connection and WSL2 fixes Mid-stream drops preserve partial responses and mouse-wheel scrolling works in WSL2.
• •Config syntax and Bun 1.4 Prompt-based settings and the upgraded runtime improve long-paragraph streaming.
• •Version constraints in 2.1.163 Managed minimum and maximum version settings plus /plugin list command appear.
• •Fallback models in 2.1.166 Up to three fallback models and glob-pattern deny rules are now configurable.
• •Safe mode in 2.1.169 The troubleshooting flag, post-session hook, and /cd command ship together.
• •Fable 5 in 2.1.170 The new Mythos-class model becomes available with transcript fixes.
• •Nested sub-agents in 2.1.172 Five-level spawning, plugin search, and automatic Bedrock region resolution land.
• •Scroll and usage in 2.1.174 Wheel acceleration control and VS Code usage attribution for subagents are added.
• •Model enforcement in 2.1.175 The managed setting prevents widening of allowed models.
• •Multilingual titles in 2.1.176 Session titles match conversation language with better Bedrock credential handling.
• •Permission rules in 2.1.178 Tool(param:value) syntax and nested skills improve auto-mode evaluation.
• •Connection fixes in 2.1.179 Partial responses survive drops and WSL2 scrolling is corrected.
• •Version ranges in 2.1.163 Required minimum and maximum version settings block unsupported starts.
• •Fallback and thinking controls Up to three fallback models and the ability to disable thinking are added.
• •Fable 5 support The Mythos-class model ships with transcript saving fixes.
• •Fable 5 name normalization Context suffixes are stripped automatically and Windows sandbox warnings are fixed.
• •Scroll acceleration toggle The wheelScrollAccelerationEnabled setting and detailed VS Code usage appear.
• •Model allowlist enforcement enforceAvailableModels keeps defaults inside the managed list.
• •Language-based titles Session titles now match the conversation language automatically.
• •Tool parameter rules Permission syntax and nested skill loading resolve clashes.
• •WaitingFor status in JSON The claude agents --json output now shows what blocks a waiting session.
• •Plugin list command /plugin list supports filtering by enabled or disabled status.
• •Version range enforcement Managed settings prevent launch outside allowed version bounds.
• •FallbackModel setting Up to three fallback models activate automatically on primary overload.
• •/cd command added Sessions move directories mid-run without cache breaks.
• •Safe-mode flag The --safe-mode flag and environment variable disable all customizations.
• •Claude Fable 5 introduction The Mythos-class model exceeds prior capabilities and is safe for general use.
• •Multi-level sub-agent spawn Five-level nesting supports complex hierarchical workflows.
• •Conversation-language titles Titles generate automatically in the active conversation language.

OpenAI Codex

OpenAI Codex shipped twenty-five updates across CLI, desktop app, and iOS that add encrypted relays, Record & Replay, and migration flows from Claude Code. The releases include rate-limit banking for Plus users and Developer mode with CDP access that compete with existing terminal tools. Builders can now hand off threads between CLI and desktop while preserving image attachments.

• •Codex CLI 0.141.0 Authenticated Noise relay channels enable remote executors with cross-platform directory preservation.
• •Record & Replay in App macOS workflows turn into reusable skills while bulk actions appear in automation history.
• •EEA UK Switzerland rollout Codex app features including Computer Use now reach users in those regions.
• •Encrypted Noise relays CLI 0.141.0 adds end-to-end encrypted channels and stdio MCP server activation per thread.
• •CLI 0.140.0 release /usage views, permanent session deletion, and Claude Code import commands arrive.
• •iOS 1.2026.160 update Workspace file browser, directory picker, and LaTeX rendering ship with MCP approvals.
• •Regional feature expansion Computer Use, Chrome extension, and Memories reach EEA, UK, and Switzerland.
• •Realtime voice removal Experimental /realtime controls and audio dependencies are stripped from the TUI.
• •SQLite auto-recovery Corrupted state databases are backed up and rebuilt from rollout data automatically.
• •Bedrock authentication Managed Amazon Bedrock API keys and encrypted local credential storage are added.
• •Unified mentions menu The @ symbol now opens a single menu for files, plugins, and skills.
• •Claude Code import The /import command brings setup, config, and recent chats from Claude Code.
• •Permanent session delete codex delete and /delete commands remove sessions with subagent cleanup.
• •Oversized /goal support Large text blocks and image attachments survive inside remote sessions.
• •Token usage views /usage shows daily, weekly, and cumulative account activity from the CLI.
• •Desktop handoff in 0.138.0 /app hands CLI threads to Codex Desktop on macOS and Windows with image path fixes.
• •Standalone web search Code mode calls web search directly and preserves oneOf schemas in 0.139.0.
• •iOS branch selection New threads support branch choice, setup scripts, and inline review comments.
• •Migration flows in App Onboarding imports from Claude Code and Claude Cowork with plugin filters.
• •Rate-limit banking Plus and Pro users gain reset credits plus Developer mode with CDP access.
• •CLI to Desktop handoff 0.138.0 lets /app move threads to the desktop app on both platforms.
• •Web search in Code mode Standalone search returns plaintext results in 0.139.0 with richer MCP schemas.
• •Codex profiles on iOS Usage stats, /goal support, and branch selection appear in the mobile app.
• •Claude migration in App Revamped plugins screen and expanded Settings search accompany import flows.
• •Developer mode and banking Rate-limit resets and Browser CDP access expand for Enterprise users.

What this means for you

For Vibe Builders: You can now run Cursor Automations and Design Mode voice edits to ship UI changes without writing code. Claude Code five-level sub-agents and Codex Record & Replay turn demonstrated workflows into reusable skills you trigger from chat. These updates let non-coders move from single prompts to always-on agents that handle GitHub tasks and visual demos.

For Non-techies: For daily business work, Cursor Bugbot now runs local pre-push reviews that catch 10% more issues at lower cost. Claude Fable 5 and Codex rate-limit banking give reliable responses during peak hours while iOS branch selection keeps project files organized. The net result is fewer manual reviews and faster handoffs between desktop and mobile.

For Developers: Production teams gain Cursor SDK nested subagents and MCP custom tools, Claude Code Tool(param:value) permissions, and Codex encrypted Noise relays for remote execution. Evaluate the enforceAvailableModels setting and fallbackModel lists against your current stack before rolling out Fable 5 or Composer 2.5. Watch the new post-session hooks and /import commands for safe migration testing this week.

What to watch next

Track Claude Code version 2.1.180 for any Fable 5 follow-up fixes. Watch Cursor for Agents Window usage analytics and Codex for expanded Record & Replay on Windows. Check whether Bedrock region handling and plugin marketplace filters stabilize across all three tools.