| """Shared types for the solver taskset and its toolset (kept here to avoid an import cycle).""" | |
| from __future__ import annotations | |
| import verifiers.v1 as vf | |
| class GeneralAgentState(vf.State): | |
| """Per-rollout state shared between the toolset and scoring. The toolset writes the agent's | |
| live task DB here after each tool call (as a plain dict); the task's reward reads it back | |
| off `trace.state` to hash against the gold solution.""" | |
| db: dict | None = None | |
| class GeneralAgentToolsetConfig(vf.ToolsetConfig): | |
| """Placement for the per-task tool server (defaults to its own host runtime, where the task | |
| cache lives). The task to serve is fetched per rollout over the `/task` channel.""" | |