Kimi Code 2.5.0 adds session forks and makes model switches take effect on the next turn

Moonshot released Kimi Code 2.5.0 with client changes separate from model weights availability.

Moonshot AI released Kimi Code 2.5.0 on November 4. The coding client can now fork running sessions or branch from a selected earlier turn. The release also lets a developer change the model and effort settings while an agent is working. When a turn has already started calling a model, that turn keeps its existing settings. The switch applies from the next turn. [Moonshot AI release notes](https://github.com/MoonshotAI/kimi-code/releases/tag/@moonshot-ai/[email protected])

What branching changes during coding work

A developer investigating a problem can branch from the point where the session identified the likely source of a failure, then explore a different fix in the new session. Keeping a branch at a specific conversation turn makes it easier to compare the assumptions behind two proposals. Teams should still distinguish the conversation branch from their repository state. If two sessions can edit the same checkout, the team needs a way to see which session wrote each change and reconcile overlapping edits.

The release also documents how a pending model switch behaves. Changing a setting mid-task is useful if the first approach needs more effort on the next turn. Because a turn already calling a model keeps its settings, a reviewer can check the setting associated with the actual turn rather than assuming a change took immediate effect. [Model-switch behavior](https://github.com/MoonshotAI/kimi-code/releases/tag/@moonshot-ai/[email protected])

Where this upgrade affects integrations

Moonshot also documents creation-time environments for SSH, Docker and custom commands. Operators need to know where each session will run before handing it files or credentials. The release separately changes WebSocket subscriptions and removes some REST endpoints for task and session state. Integrations using those surfaces should check the documented replacement events and verify how they recover after reconnecting. [Environment and API changes](https://github.com/MoonshotAI/kimi-code/releases/tag/@moonshot-ai/[email protected])

This release concerns the Kimi Code client and its interfaces. It is a separate development from the earlier hosted Kimi K3.1 Pro model release. The client changelog provides evidence for the session changes. It does not establish delivery of model weights.