Automated nightly build from main (2026-09-16)

This release lets the desktop app connect to remote Xum servers, lets xum server update itself, and adds first-class support for OpenAI GPT-6 Astra and Astra Pro.

Models & providers

• First-class OpenAI GPT-6 Astra, with Astra Pro and advisor reasoning mode; GPT-6 Astra routes through Codex OAuth and Astra Pro mode also works through Coder (#4064, #4106, #4094, #4124)
• Default-hidden Daybreak models and configurable Intuition reasoning effort (#4113, #4112)
• Codex OAuth cost estimates, preserved prompt-cache routing, and correct cache-write accounting in the gateway (#4099, #4159, #4160)

Server & desktop

• The desktop app can connect to remote Xum servers (#4101)
• xum server updates itself under a restart supervisor, including with recoverable watchers and the newest npm builds (#4083, #4127)
• Live chat streaming resumes after a reconnect (#4123)
• Desktop viewer controls with managed popouts, and delegated desktop agents can share the caller's desktop (#4088, #4081)
• Opt-in Claude Design MCP credential reuse (#4105)

Chat & context

• Token-budget context window rollovers, a continuous compaction experiment, and opt-in memory intuition recall (#4097, #4086, #4078)
• The workspace remembers its model and mode on send (#3968)
• Copy selected chat text as Markdown (#4170)
• Chat history loading shows above the composer without a layout flash (#4125, #4163)
• Message edits no longer hang during an active stream (#4153)

Agents & workflows

• Statically declared workflow phases with a live phase rail, and compact expandable workflow stages (#4095, #4084)
• Exec sub-agents inherit the calling chat's Exec settings; readable sub-agent failure cards and clearer outgoing agent communication cards (#4080, #4082, #4077)
• Sub-agent progress reports no longer arrive late or stale, and sub-agent execution survives parent guidance (#4100, #4098)
• Fixes for deleting parents with inactive descendants, archiving with an idle desktop process, interrupt_active during a PREPARING turn, and workspace MCP server lifecycle during archive (#4171, #4128, #4166, #4168)
• Best-of report publication is serialized with grouped finalization, and task_send_message acquires locks in the right order (#4162, #4161)

UI

• Optional flat chat list for the sidebar with project badges (#3994)
• "Archiving..." shows from every archive entry point while the request is in flight (#4068)
• Thin dividers between General settings groups, aligned composer pill heights, and preserved spacing around nested code execution tools (#4079, #4087, #4150)
• Plugin updates are reviewed in place instead of uninstall/reinstall (#4164)
• The native macOS shortcut label formatting from v0.28.4 was reverted (#4075)

Reliability

• The origin fetch during worktree workspace creation is bounded (#4155)
• Compaction and history publication are fenced against resets, truncation, deletion, and stale owners (#4129, #4133, #4135, #4137, #4148)
• A turn rolls back when shutdown lands in its pre-acceptance awaits; bash-monitor wakes wait for the active turn to end; withdrawn queued wakes are dropped (#4154, #4114, #4067)
• The server listener binds before startup recovery, and per-task config.json reloads are gone (#4058)

Internals

• Effect Wave 4 runtime work: scoped turn lifetimes, centralized turn preparation and lifecycle ownership, awaitable cancellation and disposal, runtime-run startup effects
• Deslop passes 1 through 3 removing dead exports, unused re-exports, alias wrappers, and tautological tests (#4089, #4090, #4092)

This release makes Gemini 3.8 Flash (GA today) the default gemini-flash model, and adds browser-based Login with Coder on remote Xum servers.

Models

• Gemini 3.8 Flash as the new gemini-flash model; older Flash tiers remain usable as custom model strings (google:gemini-3.7-flash, google:gemini-3.6-flash)

Coder & backup

• Login with Coder works from a browser connected to a Xum server, including a remote one, via the server's own OAuth callback route
• Opt-in project bundle for Settings Backup: back up and restore the project list and per-project memory files, with an explicit approval flow for projects at unrecorded paths

Reliability

• A send that arrives while an earlier send is still in preflight is queued instead of aborting the first stream
• Terminal-attention wakes no longer inherit the compaction request's disable-all tool policy
• The connection indicator surfaces a slow backend ("Server is slow to respond") instead of staying green

UI

• Native macOS shortcut labels (⌘P instead of ⌘·P)
• Phone-width sidebar hamburger stays inside its button

Internals

• Effect Phase 11 runtime work: ManagedRuntime skeleton, staged core layers, DesktopLive wiring, TestClock sweep
• The legacy mux compat package is version-locked to the root package again

This release promotes Claude Fable 5.1 and Claude Mythos 5.1 (launched by Anthropic today) as the default fable and mythos models, and adds Z.ai as a native provider.

Models

• Claude Fable 5.1 as the new fable model, with launch-verified pricing including the 75% cheaper cache reads ($0.25/M)
• Claude Mythos 5.1 as the new mythos model (trusted-access)
• Fable 5 and Mythos 5 remain usable as custom model strings, keeping their metadata and tokenizer approximations; existing fallback chains are preserved with one-shot reseeding to the 5.1 chain
• Z.ai as a native provider with GLM 5.3 Flash

Workflows & agents

• Durable workflow runs render as live cards in chat and the workspace UI
• Terminal wakes for kernel-launched background workflow runs, and compact machine events for background workflow-result wakes
• Intra-tree agent peer messaging via task_send_message, workspace archive/unarchive via task_workspace_lifecycle, and agentId on workspace tasks

Reliability

• Self-heal corrupt analytics rows and stop a single corrupt persisted row from bricking workspace fetch
• Exit wakes when monitored background bash settles, no more silent bash monitor retirement, and fixed release preparing after superseded monitor wakes
• Blocked duplicate workflow_run launches of an active script

Performance & UI

• Scoped workspace activity lookups and archived workspaces excluded from project listings
• Mobile timeline dialog, model selector dropdown clamped below the top bar, and iOS 26 PWA header fix

This release makes Grok 4.6 the default Grok model, adds a Coder provider with OAuth login, redesigns the workspace chrome, and introduces a durable per-workspace timeline.

Models

• Grok 4.6 as the new default Grok model, with native xhigh reasoning effort and official pricing
• Grok priority processing (Fast mode) and a combined thinking + fast-mode selector
• Grok Responses default to store=false for zero-data-retention parity
• New Coder provider with "Login with Coder" OAuth
• OpenAI GPT-5.6 Luna and Terra price cuts applied

Chat & workspace

• Redesigned workspace chrome: footer info bar, title header, creation hero, and composer
• Durable per-workspace timeline with machine-authored turn classification and a transcript card
• Copy/download context menu and richer previews for tool result images
• Project switcher on the scratch creation page
• Stack-aware PR indicator with a gh-stack dropdown

Agents & tasks

• Simplified persistent sub-agent lifecycle; parents can update sub-agent guidance mid-flight
• Task kind, spawn intent, model, and thinking level surfaced in task_await summaries and reports
• Sub-agent reports resume parents directly and collapse by default

Platform

• Experimental settings backup to a git repository
• Experimental Agent Plugins 1.0.0 support (skills + MCP)
• MCP client migrated to the official TypeScript SDK v2 with 2026-07-28 spec support
• Analytics timezone mode

Plus roughly 30 fixes and performance improvements across transcripts, streaming, mobile layouts, task management, and idle CPU usage.

Full Changelog: v0.28.1...v0.28.2

This release adds Claude Opus 5, native Kimi K3, and Gemini 3.6 Flash support, plus project-less scratch chats and message queueing.

Models

• Claude Opus 5 support
• Native Kimi K3 support via a new Moonshot AI provider
• Gemini 3.6 Flash as the new default Gemini Flash model
• GPT-5.6: native reasoning controls, explicit prompt cache breakpoints for direct OpenAI, and corrected context limits over Codex OAuth

Chat

• Project-less scratch chats
• Messages sent during special sends queue up (FIFO) instead of erroring
• Stage arbitrary pasted/dropped files into the workspace from chat and the creation composer
• Mid-turn thinking-level changes apply at the next model step

Agents & tasks

• Sub-agent reports presented in chat, with incremental in-progress updates
• Sticky sub-agents that persist after reporting
• Skills refresh: invocation control, $ARGUMENTS, dynamic context, and .claude compatibility

UX

• Per-model cost breakdown in the workspace Costs tab
• Context sections split into a Stats Context sub-tab

Plus roughly 17 fixes across streaming, task management, MCP OAuth, and mobile layouts.

Full Changelog: v0.28.0...v0.28.1

This release introduces support for the GPT-5.6 family of models.

This release focuses on durable workflows, agent memory, and background automation.

Durable workflows

• First-class nested workflows, inline workflow_run source, and streamlined authoring
• Workflows right-sidebar tab with live run streaming
• Wall-clock workflow scheduler, workspace tags, and workspace.* host actions
• /orchestrate routes to durable workflows with a gate/fixup verification loop
• Workflow CLI improvements: worktree trust, mux trust command, and the wf alias

Agent memory (experiment)

• Memory tool with global, project, and workspace scopes
• Curation UI, hot-set preloading, and background dream consolidation

Background automation

• Background bash monitors that wake the agent on matching output
• Workspace heartbeats with trigger and when-busy scheduling modes
• Notifications for terminal background work, including monitors lost to a Mux restart

Models

• Claude Sonnet 5 and Claude Mythos 5 support
• Refusal fallback with a transcript badge (default: Fable 5 falls back to Opus 4.8)

UX

• Pin chats to the top of the sidebar and reorder them
• Continue an interrupted stream by clicking the interrupted splitter
• Workspace-staged ZIP attachments
• Optional sidebar age grouping

Plus roughly 75 fixes and several performance improvements, including sealed chat history rotation at compaction boundaries.

Full Changelog: v0.27.0...v0.27.1

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