EvSys

harbor_agents

Harbor glue classes - agent + environment (harbor 0.13.2).

These subclass harbor base classes, so harbor is imported at module top. Our own code never imports this module directly - harbor loads these by the string import_path recorded in the job/agent config at runtime (see :mod:evsys_sdk.training.harbor_engine). That keeps the optional [harbor] extra out of the base import path.

  • :class:NoOpEnvironment - sandbox-free BaseEnvironment (no container).
  • :class:BasicLoopAgent - drives Chat(\<llm>) on-policy, records token-level rollout_details onto the AgentContext, and writes the completion to the agent dir so the verifier can read it. The default agent. model_client picks the sampler: "tinker" (on-policy TinkerLLM) or "litellm" (any provider litellm supports, for benchmarking closed/API models through the same path). Users can also plug any BaseAgent via agent_import_path.
  • :class:EvsysVerifier - a harbor BaseVerifier that wraps our registered verifier fns. It runs host-side (no container exec): reads the completion the agent wrote + the per-task spec (fn_name/expected/params) materialized into the task dir, and returns the reward as a harbor VerifierResult. Used in SHARED mode (a dummy tests/test.sh, written by the materializer, satisfies harbor's task-load check and is never executed).
attribute__all__
= ['BasicLoopAgent', 'EvsysVerifier', 'NoOpEnvironment']
func_read_text(path) -> str
parampathPath

Returns

str
func_read_json(path) -> dict[str, Any]
parampathPath

Returns

dict[str, typing.Any]