Automated nightly build from main (2026-06-10)

This release brings first-class support for the Claude Fable and Mythos models.

Additional improvements include:

• Improved Immersive Review performance
• /spawn built-in skill for indicating intent to run task in a sub-agent
• backslash symbols like \Pi for writing Π into the input box
• More ergonomic thinking effort keybinds: decrease with Ctrl+Shift+[ and increase with Ctrl+Shift+]

This release brings some minor bug-fixes and usability improvements.

• Solved some long-standing layout flashes on chat send
• Added a per-model minimum thinking level (default: medium) so that cycling thinking is more ergonomic
• Refined plan mode instructions to relax guard on mutations, permitting the agent to delete its own plan file, use local FS for analysis work, create GitHub issues, etc.
  • Opus 4.8 is much more stringent about following the previous mutation guidelines, which lead to long (wasted) reasoning traces about how to execute a request w/o mutation.
• Refined ask_user_question tool call instructions to avoid presenting "recommended" option and focus on genuinely open questions with no clear path forward

This release brings support for Opus 4.8 as well as a few minor improvements.

On the visual side, long plans are now more reviewable with a sticky table of contents that appears in the gutter:

• Cleaned up the Custom Agents system, removed unnecessary configuration options
• Fixed SSH runtime worktree bug
• Bounded syntax highlighting runtime to avoid long stalls
• Many bug-fixes and tweaks to the Goal system

Happy to share the first substantial release in a while.

Key features:

• /goal for Codex-like auto-self-prompting towards a well-defined objective, early testing shows this approach is very useful
• /btw for Claude-Code style one-off Q&A while the agent is running, without disrupting main context
• Agent will now highlight code chunks in the Review Pane it wants you to focus on, further aiding in the transition to fully autonomous agents
• Agent status in sidebar now generated from small model at pre-defined intervals, so they're more consistently up-to-date

Honorable mentions:

• Orchestrator restored in part as /orchestrate built-in skill
• /truncate removed due to dubious value
• SSH workspace spawn times improved to ~2x RTT

Major changes:

• First class support for DeepSeek V4 models
• Sections replaced with sub-projects, which makes context segregation in monorepos more manageable
  • Sections had no consistent adoption pattern and lacked analogues in Claude Code / Codex

Pleasantries:

• Sped up SSH runtime setup significantly
• Solved a variety of layout flashes and performance issues in transcript streaming

Removals:

• Removed the Orchestrator built-in agent
• Removed the file explorer

This release brings GPT 5.5 and GPT-5.5-Pro, supporting both standard API billing and Codex authentication.

This release brings Opus 4.7, which has the same UX as 4.6 with the exception of a new "xhigh" thinking in between "high" and "max".

Features

• Archive Improvements — snapshot-based worktree cleanup on archive, configurable Coder workspace action (keep/stop/delete), and a destructive confirmation for untracked-file failures
• Left Sidebar Redesign — tighter indentation, aligned status dots + sub-agent connectors, hover-only project row actions, subordinate workspace counts

Performance

• Optimized Immersive Review performance on large diffs
• Optimized SSH workspace creation on slow local connections
  • Replaced git bundle with native git push + delta negotiation
  • Sharded OpenSSH exec paths for parallel sync
  • Prefetches origin on SSH host before local→remote push
• Sorted tool definitions deterministically for prompt cache stability
• Avoided extra Anthropic cache breakpoints with explicit TTL

Fixes

• Fixed multiple cases where the chat transcript would vertically tear on chat switches
• Preserves all init logs across workspace switches
• Restores last page and open chat on hard refresh
• Restored Cmd+. agent cycling on macOS
• Reattaches loaded skills after compaction
• Fixed attach_file on macOS (unpacked sharp runtime assets) and added raster image scaling
• Restored archived workspaces with persisted branch mappings
• Stripped leading/trailing newlines from quote input text

Cleanup

• Removed the built-in "Chat with Mux"
  • Our goal is to replicate the capabilities of this chat in regular chat sessions
• Removed Auto mode

Browser & desktop workflows

Enabled in Experiments -> Agent Browser

• New Browser tab support for live agent-browser sessions, including an address bar, reload control, ready-state display, and streaming updates.
• Better Browser tab signal with clearer Recent Actions and less noise.
• Improved reliability for keyboard delivery, shell-launched browser sessions, and server-mode browser sessions.
• Added workspace-scoped virtual desktops via Portable Desktop.

Agents, SSH, and models

• Added grouped task variants and made task variants easier to discover.
• agent-browser is now available as a built-in skill.
• SSH workspaces can now use host-global skills and agents.
• Added support for GPT-5.4 Mini and GPT-5.4 Nano.

Quality of life

• Consistent recovery of interrupted chats on restart.
• Improved queued-message handling and chat layout stability while sending or resizing the composer.
• Polished immersive review with LoC progress, cross-file navigation, and undo.
• Reduced idle compaction / notification noise and improved startup diagnostics.