See why a model was chosen, workspace-aware activity, and healthier Slack bridges

This release makes several decisions that used to happen out of sight much easier to understand. You can see which model a specialist is expected to use and where that choice came from, specialists can work with activity that stays inside the right workspace, and Slack connections now tell you when messages are stuck or have failed.

See which model will answer

Specialist settings now show the expected model, the setting that selected it, and the fallback order. The model badge on a reply carries the same explanation, such as “via your default model” or “via the workspace default.”

Workspace defaults now work as advertised instead of losing to a specialist’s bundled preference. An explicit personal choice still takes priority. If a saved workspace default is no longer available, Settings names it and explains why it is not being enforced rather than silently dropping it. Members who cannot change the workspace default see a read-only control instead of an edit that fails after saving.

Settings also stops showing a “Leave site?” warning when you reload a page without unsaved changes.

Workspace-aware activity for specialists

Installed miniapps can now contribute activity sources that specialists query through one permission-checked tool. Built-in sources cover active time and GitHub contributions, while Settings lets you disable individual sources or block a specialist from using them.

Workspace boundaries apply throughout the query. Active time is attributed to the workspace where it happened, and GitHub activity only reads repositories attached to projects in the current workspace. Removing a project or disconnecting GitHub removes that access too.

Slack bridges show their health

Channel settings now show pending and failed outbound Slack deliveries, so a connection that looks active no longer hides a broken message path. Slack also retries messages without optional author decoration when Slack rejects those cosmetic fields.

Private Slack channels are included in the channel picker with the updated connection permissions. Existing Slack connections may need to be reconnected before those channels become available.

Specialists can recover and reach linked repositories

Repeated tool failures no longer corrupt a specialist checkpoint while it is trying to recover. Specialists also regain access to repositories linked into a project workspace, fixing permission errors that could stop Git inspection and publishing work before the Git tools even ran.

Opening a workspace or project from a deep link now activates the workspace before loading its data. That prevents valid links from landing on an empty, missing, or blank screen because the previous workspace was still active.

Keyboard and screen-reader fixes

This release includes a broad focus pass across chat. Focus now moves into action menus, Activity, Later, threads, Channel Details, channel groups, resource menus, and task detail controls when they open, then returns to the right place when they close. The composer attachment menu is included too.

Settings validation errors for workspace names, GitHub tokens, and MCP tools are now tied to their fields for assistive technology. MCP remove buttons and scope controls have distinct accessible names, so keyboard and screen-reader users can tell which row or scope they are changing.

Chat polish and stability

Duplicate Chloe rows no longer appear in mention autocomplete. A single leading + stays plain text instead of turning into a bullet, transcript spacing is tighter, and the emoji and microphone controls now match the rest of the composer. Voice controls also use the same icon in the sidebar and the voice panel.

Authentication listener cleanup no longer produces renderer errors during reloads or account changes, explicit button icon colors are preserved, and desktop error reporting is working again so failures do not disappear into a console warning.