fix: honor group scoring in the v0 legacy env server

[mikasenghaas]

Summary

Fixes group scoring in the v0 legacy env server (LegacyEnvServer).

• requires_group_scoring was hardcoded False, so a v0 env whose rubric defines group/preference reward funcs was advertised (via the info response) as single-rollout, and the orchestrator routed it down the per-rollout path.
• _run_group looped env.run_rollout (→ rubric.score_rollout) n times. score_rollout asserts there are no group-level reward funcs, so every rollout of such an env errored.

Now:

• requires_group_scoring = self.env.requires_group_rollouts — reflects the env's actual rubric (rubric.has_group_rewards). A non-group env still reports False and keeps the single-rollout path.
• _run_group calls env.run_group(...) once, so the rubric scores the rollouts together (score_group) and cross-rollout (group/preference) rewards apply. score_group also handles individual-level reward funcs, so it's correct whether or not the env defines group rewards.

Verification

• ruff check --isolated / ruff format --isolated --check clean.
• LegacyEnvServer imports and the new _run_group / requires_group_scoring resolve against the v0 Environment.run_group / requires_group_rollouts API.

---

Note

Medium Risk
Changes orchestration routing and the group rollout scoring path for legacy v0 envs; behavior fix but affects reward computation for group-rubric environments.

Overview
Fixes group scoring for v0 environments served through LegacyEnvServer.

requires_group_scoring now comes from self.env.requires_group_rollouts instead of always being False, so the server info response and orchestrator correctly use the group rollout path when the rubric has group/preference rewards.

_run_group no longer runs n independent run_rollout calls (which hit score_rollout and fail on group-level reward funcs). It calls env.run_group once with n copies of the task input, so scoring goes through score_group and cross-rollout rewards apply.

^{Reviewed by Cursor Bugbot for commit 3112158. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix group scoring in LegacyEnvServer to delegate to the v0 env's run_group API

• LegacyEnvServer.__init__ now reads requires_group_scoring from the v0 env's requires_group_rollouts flag instead of hardcoding False.
• LegacyEnvServer._run_group replaces n independent _run_v0 calls with a single self.env.run_group call, passing duplicated dataset row inputs for the task.
• Behavioral Change: group rollouts are now executed as a true group rather than as independent single rollouts, which may affect scoring for environments that depend on cross-rollout group context.

^{Macroscope summarized 3112158.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This change enables group scoring in the legacy env server by propagating the environment's requires_group_rollouts config (previously hardcoded to False) and switching from individual rollout execution to batch processing via run_group. This alters how rewards are calculated for preference-based reward functions - a significant runtime behavior change that warrants human review.

^{You can customize Macroscope's approvability policy. Learn more.}