September 7–13 reshaped Desktop around the conversation with your agent. Agent chat is now the main workspace, tool activity is easier to follow, and the chat list cleans up after itself. Runtime administrators also get usage charts and control over runtime images, and Claude Code and Codex runtimes are more resilient under heavy work.
A workspace built around agent chat
Agent chat is now the primary view. Other pages open as tabs in a resizable right-hand dock that you can expand when you need more room. A new New Tab launcher gets you to any page quickly, and the left sidebar lists your sessions first when you work in split view. Plans appear inline in the conversation, and long transcripts scroll and render faster.
The runtime for a conversation now sits as a chip in the composer, next to the model selector. Its menu opens Move runtime and Manage runtimes.
Tool calls in one activity line
Consecutive tool calls in a turn now fold into a single line that shows the latest call, elapsed time, the number of calls, and any failures. Open the line to see every call with its details. A call that needs your approval still opens on its own.
Chat activity and archiving
- Chats in the sidebar show a spinner while the agent is working on them.
- Archive a chat from the hover button or by right-clicking Archive chat in All chats. Archived chats are hidden from the list and their open tabs close.
- Conversations idle for seven days are archived automatically. Restoring one keeps it out of the next sweep.
- Choosing a chat that is already open in another tab switches to that tab.
- If a conversation finishes after you close its tab, the notification now includes the final answer or the pending approval.
Runtime settings and images
The agent runtime settings in Settings → Agent are rebuilt as a list with one row per runtime, showing status, image, and account. Expand a row to see CPU, memory, and concurrent session charts for the last 1, 6, or 24 hours, along with conversation defaults. Administrators can also choose and update the container image each Claude Code or Codex runtime uses.
More improvements
- Pod details now let you pick a container to inspect its environment and mounts, and workload templates are labeled as templates rather than live state. Logs and topology include init and ephemeral containers, and Job, ReplicaSet, Secret, RBAC, network, and Helm details are more accurate.
Fixes
- Runtimes have more memory, so one heavy command, such as a large type check, no longer disconnects active conversations.
- In Slack, replies the agent sends after you approve a tool now reach the thread.
- When a runtime fails, the transcript keeps the partial answer and the actual error, including a clear message when the runtime runs out of disk. Desktop no longer reconnects to a failed stream in a loop.
- If a runtime's sign-in is revoked or expired, Nuphos stops retrying and shows a card that links to Agent settings. Members who cannot manage runtimes are pointed to an administrator.
- In Full Access conversations, tool calls in turns the agent starts on its own, such as scheduled wakeups, are no longer cancelled.
- Conversations you follow stay scrolled to the bottom when you switch back to their tab.
Available in Desktop v0.49.0 and later.