verifiers 0.2.2.dev80__tar.gz → 0.2.2.dev82__tar.gz

{verifiers-0.2.2.dev80 → verifiers-0.2.2.dev82}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev80
+Version: 0.2.2.dev82
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.2.2.dev80 → verifiers-0.2.2.dev82}/verifiers/v1/cli/eval/runner.py

@@ -10,7 +10,6 @@ from verifiers.v1.cli.dashboard import dashboard
 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_eval_server(config: EvalConfig) -> 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_eval_server(config: EvalConfig) -> 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-0.2.2.dev80 → verifiers-0.2.2.dev82}/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 @@ class Episode(BaseModel, Generic[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-0.2.2.dev80 → verifiers-0.2.2.dev82}/verifiers/v1/tasksets/harbor/taskset.py

@@ -140,6 +140,8 @@ class HarborData(TaskData):
     tags: list[str] = Field(default_factory=list)
     task_dir: str = ""
     """Host path to the task dir; used to stage tests/ to verify."""
+    upload_environment: bool = False
+    """Whether to stage environment/ into the workdir."""
     verifier_env: dict[str, str] = Field(default_factory=dict)
     """Raw [verifier.env] entries (literals or `${VAR}`/`${VAR:-default}` templates).
     Resolved against the host environment at scoring time, like `harbor run` — so a
@@ -155,6 +157,27 @@ class HarborData(TaskData):
 class HarborTask(Task[HarborData]):
     """Stage and run Harbor's verifier inside the task's live runtime."""
 
+    async def setup(self, runtime: Runtime) -> None:
+        if not self.data.upload_environment:
+            return
+        await runtime.write(
+            "/tmp/environment.tgz",
+            make_tar(Path(self.data.task_dir) / "environment"),
+        )
+        result = await runtime.run(
+            [
+                "sh",
+                "-c",
+                "tar --no-same-owner -xzf /tmp/environment.tgz && rm /tmp/environment.tgz",
+            ],
+            {},
+        )
+        if result.exit_code:
+            raise RuntimeError(
+                f"environment setup failed (exit {result.exit_code}): "
+                f"{(result.stderr or result.stdout).strip()[-500:]}"
+            )
+
     async def finalize(self, trace: Trace, runtime: Runtime) -> None:
         """Run Harbor's collect hooks while the agent's box is still alive.
 
@@ -437,6 +460,7 @@ def resolve_image(
 
 def parse_task(task_dir: Path, idx: int, harbor_config: HarborConfig) -> HarborData:
     # Harbor is optional, so imports stay deferred until a Harbor task loads.
+    from harbor.environments.definition import should_upload_environment_dir
     from harbor.models.task.config import NetworkMode
     from harbor.models.task.task import Task as HarborModelTask
 
@@ -444,6 +468,11 @@ def parse_task(task_dir: Path, idx: int, harbor_config: HarborConfig) -> HarborD
     parsed = harbor_task.config
     artifacts, hooks, verifier = parse_verifier_extras(task_dir, parsed, harbor_config)
     environment = parsed.environment
+    environment_dir = task_dir / "environment"
+    upload_environment = should_upload_environment_dir(
+        environment_dir,
+        docker_image=environment.docker_image,
+    )
     network = parsed.agent.explicit_phase_policy() or environment.resolve_baseline()
     task, meta = parsed.task, parsed.metadata
     authors = (
@@ -499,6 +528,7 @@ def parse_task(task_dir: Path, idx: int, harbor_config: HarborConfig) -> HarborD
         category=meta.get("category"),
         tags=meta.get("tags", []),
         task_dir=str(task_dir),
+        upload_environment=upload_environment,
         verifier_env=parsed.verifier.env,
         artifacts=artifacts,
         collect=hooks,
@@ -658,9 +688,9 @@ def verifier_env(task: HarborData) -> dict[str, str]:
     return resolve_env_vars(task.verifier_env)
 
 
-# Downloaded test directories are immutable. Cache only the latest archive to
-# bound memory while reusing it across rollouts of the current task.
-@lru_cache(maxsize=1)
+# Downloaded task directories are immutable. Cache the current task's environment
+# and tests to bound memory while reusing both archives across rollouts.
+@lru_cache(maxsize=2)
 def make_tar(directory: Path) -> bytes:
     buffer = io.BytesIO()
     with tarfile.open(fileobj=buffer, mode="w:gz") as tar:

{verifiers-0.2.2.dev80 → verifiers-0.2.2.dev82}/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 @@ class Trace(BaseModel, Generic[DataT, StateT, AgentConfigT]):
         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