Skip to content

v1: apply agentic-judge task weights when an isolated solver fails #2197

Description

@hallerite

Context

In AgenticJudgeEnv(share_runtime=False), a failed solver trace returns before a judge trace is created. finalize() then returns early because no judge exists, so ScoreConfig.task_weight is never applied. With the default task_weight = 0, partial rewards on a failed solver keep their original weights even though the environment is configured to suppress them.

This was found in post-merge review of #2144 and appeared on #2152 only during its automatic retarget to main; agentic-judge is not part of #2152's current diff. Related isolated-mode transport/image fixes are tracked in #2195.

Done when

  • Failed isolated solver traces receive the configured task-weight scaling even when no judge runs.
  • The no-judge infrastructure-failure path remains distinguishable from an actual zero verdict.
  • Shared and successful isolated paths retain their existing reward composition.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions