August 31–September 6 brought the second agent to Nuphos: Codex now runs alongside Claude Code, and a workspace can keep several agent runtimes at once. The week also added per-runtime model defaults, a way to move conversations between runtimes, an in-app pull request view, and fixes for long-running turns.
Codex and multiple runtimes
Workspace administrators can now add Codex as an agent runtime, next to Claude Code. Codex signs in with a device code: Nuphos shows the code in Desktop, you approve it in the browser, and the runtime starts. There is no auth.json to find or upload.
A workspace is no longer limited to one runtime. You can run several named Claude Code and Codex instances at the same time, including more than one of the same agent, for example on different accounts. Manage each instance's name, account, and status in Settings → Agent.
When you start a conversation, pick the runtime from the selector next to Send. Existing conversations stay on the runtime they started with, even if other instances are added, disabled, or removed. In an existing conversation, the same spot shows a Model selector with the model, reasoning effort, and Fast options the runtime supports. Codex conversations get the same streaming, mid-turn steering, approvals, Skills, and Memory as Claude Code.
Model defaults per runtime
Administrators can set a default model, effort, and Fast mode for each runtime in Settings → Agent. New conversations take the defaults when they start; existing conversations keep their settings. Choose Runtime default to clear an override.
Move conversations between runtimes
Deleting a runtime no longer strands its conversations. Nuphos saves each conversation's workspace files before removing the runtime, and conversation owners can use Move runtime in the chat to continue on another runtime, either with history and workspace files restored or with history only. The new runtime rebuilds the agent's context from the Nuphos transcript.
Pull requests and Kubernetes resources, in the app
- GitHub pull request links from agent chat now open in Nuphos, on a page with the conversation, checks, reviews, and diff, laid out like GitHub. You can search the PR list by number, label, or workflow run.
- Kubernetes tables are more compact, sort CPU and memory by actual usage, and show clearer health for Pods and workloads. Pod and workload details now cover containers, environment sources, volumes, ownership, scheduling, and a navigable resource topology, with explicit links between related resources.
- Going back from a detail page keeps the list you had loaded.
More improvements
- Long conversations load faster: the backend now compresses API responses.
- Desktop uses much less bandwidth while a conversation is streaming.
- Agent messages and the composer now share the same spacing, so messages no longer extend past the composer while scrolling.
Fixes
- Tool calls that run longer than three minutes without output are no longer interrupted.
- Long agent turns that keep producing output are no longer cut off after 15 minutes.
- After a turn fails, the next message starts a fresh session instead of failing again.
- Choosing Full Access on the home composer no longer reverts to Auto Mode after the first message.
Available in Desktop v0.46.0 and later.