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

verifiers/v1/cli/eval/runner.py

@@ -10,7 +10,6 @@
 from verifiers.v1.cli.eval import resume
 from verifiers.v1.cli.output import (
     append_episode,
-    append_trace,
     output_path,
     save_config,
 )
@@ -77,8 +76,7 @@ async def run_eval(env: Env, config: EvalConfig) -> list[Episode]:
     write_lock = asyncio.Lock()
 
     async def on_complete(episode: Episode) -> None:
-        for trace in episode.traces:
-            trace.record_run(EvalRunInfo(id=config.uuid))
+        episode.record_run(EvalRunInfo(id=config.uuid))
         await append_episode(out, episode, write_lock)
 
     # Serving resources (shared tool servers, interception) come up once for the
@@ -258,9 +256,10 @@ async def run_group_unit(idx: int) -> list[Episode]:
                 )
             records = []
             for trace in traces:
-                trace.record_run(EvalRunInfo(id=config.uuid))
-                await append_trace(out, trace, write_lock, env=config.env_id)
-                records.append(Episode.of(trace))
+                record = Episode.of(trace, env=config.env_id)
+                record.record_run(EvalRunInfo(id=config.uuid))
+                await append_episode(out, record, write_lock)
+                records.append(record)
             return records
 
         async def run_unit(payload: dict) -> list[Episode]:
@@ -271,8 +270,7 @@ async def run_unit(payload: dict) -> list[Episode]:
                     sampling=config.sampling,
                     **payload,
                 )
-            for trace in episode.traces:
-                trace.record_run(EvalRunInfo(id=config.uuid))
+            episode.record_run(EvalRunInfo(id=config.uuid))
             await append_episode(out, episode, write_lock)
             return [episode]
 

verifiers/v1/episode.py

@@ -1,14 +1,14 @@
 """The episode — one run's traces plus their shared standing, whole."""
 
 import uuid
-from typing import Generic
+from typing import Any, Generic
 
 from pydantic import BaseModel, Field
 
 from verifiers.v1.configs.agent import WireAgentConfig
 from verifiers.v1.state import State, StateT
 from verifiers.v1.task import DataT, WireTaskData
-from verifiers.v1.trace import AgentConfigT, Error, Trace
+from verifiers.v1.trace import AgentConfigT, Error, RunInfo, Trace
 from verifiers.v1.types import Usage
 
 
@@ -26,12 +26,19 @@ class Episode(BaseModel, Generic[DataT, StateT, AgentConfigT]):
 
     env: EnvInfo = Field(default_factory=EnvInfo)
     """The env that produced this episode."""
+    run: RunInfo | None = None
+    """The run this episode belongs to (eval or train), consumer-stamped. It lives here rather than
+    on each trace because the episode is what a consumer dispatches, and an episode that produced
+    no traces would otherwise have nowhere to say which run it was."""
     ok: bool = False
     """Whether the episode completed successfully."""
     errors: list[Error] = Field(default_factory=list)
     """Every error captured across attempts, oldest to newest."""
     traces: list[Trace[DataT, StateT, AgentConfigT]] = Field(default_factory=list)
     """Every agent's trace, in completion order."""
+    info: dict[str, Any] = Field(default_factory=dict)
+    """Scratch space for episode-level metadata, the counterpart to `Trace.info`. What describes
+    the whole episode belongs here rather than repeated on each of its traces."""
 
     @property
     def last_error(self) -> Error | None:
@@ -72,6 +79,13 @@ def by_agent(self) -> dict[str, list[Trace[DataT, StateT, AgentConfigT]]]:
             grouped.setdefault(trace.agent.name, []).append(trace)
         return grouped
 
+    def record_run(self, run: RunInfo | None = None, **info: Any) -> None:
+        """Record the run identity and any extra metadata about this episode. Both describe the
+        episode as a whole, so they are recorded once here rather than repeated on every trace."""
+        if run is not None:
+            self.run = run
+        self.info.update(info)
+
     @classmethod
     def of(cls, trace: Trace, env: str = "") -> "Episode":
         """The single-agent record: one trace as its own episode."""

verifiers/v1/trace.py

@@ -304,9 +304,6 @@ class Trace(BaseModel, Generic[DataT, StateT, AgentConfigT]):
     """Unique ID for this trace, auto-generated."""
     verifiers: VersionInfo = Field(default_factory=_current_build)
     """The verifiers version that produced this trace."""
-    run: RunInfo | None = None
-    """The run this trace belongs to (eval or train), consumer-stamped."""
-
     task: TraceTask[DataT]
     """The task data that seeded this trace."""
     agent: AgentInfo[AgentConfigT]
@@ -480,12 +477,6 @@ def record_judge(self, response: JudgeResponse) -> None:
         if response.usage is not None:
             self.extra_usage.append(response.usage)
 
-    def record_run(self, run: RunInfo | None = None, **info: Any) -> None:
-        """Record the run identity (eval / train), and optional extra info."""
-        if run is not None:
-            self.run = run
-        self.info.update(info)
-
     def stop(self, condition: str) -> None:
         """Stop the trace, optionally with a stop condition."""
         self.is_completed = True