Revert "feat(v1)!: move run info from trace to episode"

[mikasenghaas]

Summary

• Reverts feat(v1)!: move run info from trace to episode #2244: run identity returns to Trace.run / Trace.record_run, and Episode.run / Episode.info / Episode.record_run are removed.
• The eval runner call sites return to stamping each trace and persisting via the single-trace episode wrapper, as before feat(v1)!: move run info from trace to episode #2244.

Moving the run onto the episode forced consumers that keep per-trace records (prime-rl's orchestrator) to reconstruct episode shells around traces just to place them, and the episode-level home only pays off with the rest of the episode-first pipeline (#2252 / prime-rl#3206), which we're not taking. Reverting keeps the trace self-describing and makes the prime-rl pin bump trivial.

Verification

uv run pytest tests/v1/test_trace.py tests/v1/test_graph.py
ruff check verifiers/v1/trace.py verifiers/v1/episode.py verifiers/v1/cli/eval/runner.py

Both pass. Only in-tree users of the episode-level fields were Episode.record_run itself and the three eval-runner call sites restored here.

🤖 Generated with Claude Code

Note

Move run info from episode level to trace level in v1 eval runner

• Removes run and info fields and the record_run method from Episode; run metadata is no longer stored at the episode level.
• Adds a run: RunInfo | None field and record_run method to Trace so each trace carries its own run metadata.
• Updates eval.runner to call trace.record_run(...) on each trace within an episode instead of stamping the episode directly.
• In the group execution path, traces are now persisted individually via append_trace and Episode.of(trace) is called without an env argument, leaving env id empty on returned Episode records.
• Behavioral Change: Episode.record_run is removed; callers must now stamp run info on individual traces.

^{Macroscope summarized eaab120.}

---

Note

Medium Risk
Changes persisted artifact shape and public v1 APIs (Episode/Trace); callers that adopted episode-level run must migrate back to per-trace stamping.

Overview
Reverts the episode-first run stamping from #2244 so eval and downstream consumers (e.g. prime-rl) can keep per-trace records without wrapping traces in episodes just to attach run identity.

Trace again has run: RunInfo | None and record_run, which sets run id/type and merges optional metadata into trace.info. Episode drops run, info, and record_run—episodes are no longer the place for consumer-stamped run metadata.

The v1 eval runner stamps EvalRunInfo on every trace in an episode (on_complete, run_unit, and the legacy run_group path). Group rollouts persist each trace with append_trace (with env=config.env_id) and return Episode.of(trace) shells for in-memory results; single-rollout paths still use append_episode after per-trace stamping.

^{Reviewed by Cursor Bugbot for commit eaab120. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

🟡 Medium eval/runner.py:263

In run_group_unit, each trace is persisted with append_trace(..., env=config.env_id), but the in-memory episode returned to callers is built via Episode.of(trace) without passing env. As a result, every group-scored legacy result returned by run_eval_server has an empty episode.env.id, even though the same episode on disk carries the configured environment ID. The previous code constructed Episode.of(trace, env=config.env_id), so this change silently drops the env metadata from in-memory results. Pass env=config.env_id to Episode.of to preserve the in-memory episode's environment identity.

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file @verifiers/v1/cli/eval/runner.py around line 263:

In `run_group_unit`, each trace is persisted with `append_trace(..., env=config.env_id)`, but the in-memory episode returned to callers is built via `Episode.of(trace)` without passing `env`. As a result, every group-scored legacy result returned by `run_eval_server` has an empty `episode.env.id`, even though the same episode on disk carries the configured environment ID. The previous code constructed `Episode.of(trace, env=config.env_id)`, so this change silently drops the env metadata from in-memory results. Pass `env=config.env_id` to `Episode.of` to preserve the in-memory episode's environment identity.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit eaab120. Configure here.}

[cursor]

Returned episodes diverge from persisted records

Medium Severity

append_trace persists one episode shell, while Episode.of(trace) creates a different returned shell with a fresh id and empty env. Group-run callers therefore receive episode identities that differ from traces.jsonl and lack environment provenance.

 

^{Reviewed by Cursor Bugbot for commit eaab120. Configure here.}