Runtime types

​

Definition

A runtime is the adapter that executes an Agent Run. The Agent is the durable team member, the Run is one bounded attempt, and the runtime supplies the local or external execution behavior for that attempt.

​

States

The standard Agent configuration UI currently enables these local choices: Availability is setup guidance: Always use Test runtime chain after configuration. A successful probe is stronger evidence than the availability group.

​

Constraints

Runtime capabilities differ. Codex, Claude Code, OpenCode, and Pi can receive Rudder-managed tools through their supported MCP or native bridge. Cursor and Gemini have their own adapter behavior. Remote or unsupported adapters must not claim a managed tool path that they do not have. Model discovery, session reuse, Skill synchronization, transcript parsing, cost metadata, and mid-turn controls are adapter-specific. A fallback runtime is another explicit attempt, not proof that the primary runtime has the same capabilities.

​

Boundaries

Generic process and http adapters are internal or advanced plumbing and do not appear as ordinary runtime picker choices. Registry entries shown as unavailable or coming soon are not promises of an executable setup. Rudder prepares managed runtime state and the selected work context, then the runtime performs the work. Runtime configuration does not grant extra Issue, organization, filesystem, credential, or Browser authority.

​

Examples

• A local Codex Agent needs repository access and Rudder tools. Configure codex_local, select the model and permissions, then run Test runtime chain before assigning work.
• An OpenCode command exists but its provider login is missing. The command may appear ready while the runtime test reports an authentication failure. Fix the provider setup and test again.
• A runtime cannot provide native mid-turn steering. Rudder should report its interrupt-and-continue fallback instead of presenting a native Steer result.