Skip to content

Config

Configuration models for experiment runs. All models use Pydantic for validation.

config

Pydantic models for run configuration.

SessionMode

Bases: str, Enum

SessionConfig

Bases: BaseModel

Configuration for a single session within a run.

AgentConfig

Bases: BaseModel

Definition of a subagent available to the main agent.

JudgeConfig

Bases: BaseModel

Auto-judge: an LLM that evaluates the running trajectory against a rubric.

The judge runs every every_n_turns agent turns, sees the trajectory so far, and returns a structured verdict (flagged + reason + confidence). If a verdict is flagged and early_exit is set, the session stops after the current turn. The judge runs independently of the agent engine, so it can judge both Claude Code and Codex runs.

HookCommandConfig

Bases: BaseModel

Shell command run before or after an experiment.

RunConfig

Bases: BaseModel

Top-level run configuration.

load_config(path)

Load a RunConfig from a YAML file.

build_provider_env(config)

Build environment variable dict passed to the engine.

Returns a dict — does NOT mutate os.environ.