verifiers 0.3.2.dev136__py3-none-any.whl → 0.3.2.dev138__py3-none-any.whl

verifiers/v1/agent.py

@@ -12,13 +12,15 @@ import logging
 from collections.abc import AsyncIterator, Callable, Mapping
 from contextlib import asynccontextmanager, nullcontext
 from dataclasses import dataclass, replace
-from typing import Self
+from typing import Generic, Self, cast
+
+from typing_extensions import TypeVar
 
 from verifiers.v1.clients import (
     EvalClientConfig,
     ModelContext,
 )
-from verifiers.v1.configs.agent import AgentConfig, TimeoutConfig
+from verifiers.v1.configs.agent import AgentConfig, TimeoutConfig, agent_config_fields
 from verifiers.v1.configs.runtime import NetworkPolicyConfig
 from verifiers.v1.dialects import parse_message
 from verifiers.v1.harness import Harness
@@ -42,6 +44,7 @@ from verifiers.v1.types import (
     ToolMessage,
     UserMessage,
 )
+from verifiers.v1.utils.aio import run_shielded
 from verifiers.v1.utils.compile import (
     cap_remote_agent_timeout,
     resolve_runtime_config,
@@ -462,9 +465,8 @@ class Agent:
                     run.trace.stop("agent_completed")
             trace = await run.close()
         except BaseException:
-            # A cancellation mid-run (or a lifetime bug raised to the caller) means
-            # close() never runs — free the run's servers and owned runtime first.
-            await run.abort()
+            # Finish cleanup even if another cancellation arrives during abort().
+            await run_shielded(run.abort())
             raise
         if trace.agent.runtime is not None:
             trace.agent.runtime.borrowed = runtime is not None
@@ -720,27 +722,22 @@ def make_agent(
     return Agent(config, interception=interception)
 
 
-MakeAgent = Callable[[str, AgentConfig], Agent]
+AgentT = TypeVar("AgentT", bound=Agent, default=Agent)
+MakeAgent = Callable[[str, AgentConfig], AgentT]
 """An agent factory keyed by name — what `Agents` calls per scraped config field."""
 
 
-def agent_config_fields(config) -> dict[str, AgentConfig]:
-    """The top-level `AgentConfig` fields declared on a config, in declaration
-    order — the env's agents, keyed by field name (the only naming site)."""
-    return {name: value for name, value in config if isinstance(value, AgentConfig)}
-
-
-class Agents:
+class Agents(Generic[AgentT]):
     """A config's agents, addressed by attribute: every top-level `AgentConfig`
     field becomes an `Agent` under the field's name (`agents.solver`)."""
 
-    def __init__(self, config, make: MakeAgent | None = None) -> None:
-        self._agents: dict[str, Agent] = {
-            name: make_agent(value) if make is None else make(name, value)
+    def __init__(self, config, make: MakeAgent[AgentT] | None = None) -> None:
+        self._agents: dict[str, AgentT] = {
+            name: cast(AgentT, make_agent(value)) if make is None else make(name, value)
             for name, value in agent_config_fields(config).items()
         }
 
-    def __getattr__(self, name: str) -> Agent:
+    def __getattr__(self, name: str) -> AgentT:
         # self.__dict__ directly: attribute lookup re-entering __getattr__ before
         # __init__ ran (copy/unpickle) must raise, not recurse.
         agents = self.__dict__.get("_agents")

verifiers/v1/cli/dashboard/eval.py

@@ -81,9 +81,9 @@ def format_cost_usd(cost_usd: float) -> str:
 def _seat_value(config: EvalConfig, read):
     """A cap as the overview shows it: the declared seats' shared value, the
     string 'per-seat' when they disagree (caps live on the seats)."""
-    from verifiers.v1.configs.env import _declared_agent_configs
+    from verifiers.v1.configs.agent import agent_config_fields
 
-    values = {read(spec) for spec in _declared_agent_configs(config.env).values()}
+    values = {read(spec) for spec in agent_config_fields(config.env).values()}
     return values.pop() if len(values) == 1 else "per-seat"
 
 

verifiers/v1/cli/debug.py

@@ -17,7 +17,7 @@ from uuid import uuid4
 from pydantic_config import cli
 
 import verifiers.v1 as vf
-from verifiers.v1.cli.output import append_trace, save_config
+from verifiers.v1.cli.output import save_config
 from verifiers.v1.cli.resolve import (
     extract_id,
     narrow_taskset_config,
@@ -34,6 +34,7 @@ from verifiers.v1.utils.compile import resolve_runtime_config
 from verifiers.v1.utils.decorators import invoke
 from verifiers.v1.utils.interrupt import install_interrupt
 from verifiers.v1.utils.logging import setup_logging
+from verifiers.v1.utils.trace_store import append_trace
 
 logger = logging.getLogger(__name__)
 

verifiers/v1/cli/eval/main.py

@@ -10,7 +10,6 @@ from pydantic_config import cli
 
 from verifiers.v1.cli.eval.runner import run_eval
 from verifiers.v1.cli.output import (
-    TRACES_FILE,
     create_attempt_log_dir,
     output_path,
     saved_config_path,
@@ -26,6 +25,7 @@ from verifiers.v1.cli.resolve import (
 from verifiers.v1.configs.cli.eval import EvalConfig
 from verifiers.v1.utils.interrupt import install_interrupt
 from verifiers.v1.utils.logging import setup_logging
+from verifiers.v1.utils.trace_store import TRACES_FILE
 
 logger = logging.getLogger(__name__)
 

verifiers/v1/cli/eval/resume.py

@@ -16,9 +16,9 @@ from pathlib import Path
 
 from pydantic_core import from_json
 
-from verifiers.v1.cli.output import TRACES_FILE
 from verifiers.v1.episode import WireEpisode
 from verifiers.v1.task import task_key
+from verifiers.v1.utils.trace_store import TRACES_FILE
 
 
 def load(

verifiers/v1/cli/eval/runner.py

@@ -17,11 +17,7 @@ from typing import TypeVar, cast
 from verifiers.v1.cli.dashboard import dashboard
 from verifiers.v1.cli.eval import resume
 from verifiers.v1.cli.eval.hint import PRIME_RL_HINT
-from verifiers.v1.cli.output import (
-    append_episode,
-    output_path,
-    save_config,
-)
+from verifiers.v1.cli.output import output_path, save_config
 from verifiers.v1.cli.resume import distribute
 from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.cli.eval import EvalConfig
@@ -35,6 +31,7 @@ from verifiers.v1.utils.platform import (
     log_episodes,
     open_run,
 )
+from verifiers.v1.utils.trace_store import append_episode
 
 logger = logging.getLogger(__name__)
 

verifiers/v1/cli/gepa.py

@@ -16,7 +16,7 @@ import uuid
 from pydantic_config import cli
 
 import verifiers.v1 as vf
-from verifiers.v1.cli.output import TRACES_FILE, output_path, write_config
+from verifiers.v1.cli.output import output_path, write_config
 from verifiers.v1.cli.resolve import (
     extract_id,
     narrow_config,
@@ -27,6 +27,7 @@ from verifiers.v1.cli.resolve import (
 from verifiers.v1.gepa import GEPAConfig, run_gepa
 from verifiers.v1.utils.interrupt import install_interrupt
 from verifiers.v1.utils.logging import setup_logging
+from verifiers.v1.utils.trace_store import TRACES_FILE
 
 logger = logging.getLogger(__name__)
 

verifiers/v1/cli/output.py

@@ -1,31 +1,13 @@
-"""On-disk output: traces.jsonl (one rollout episode per line) + configs/<cli>.json.
-
-Each line is an `Episode` — the episode's standing (`id`/`env`/`errors`) inlined
-next to its flat, self-contained traces — so an episode persists whole or not at all: a torn line is the
-whole episode owed on resume, and a failure before any trace minted still leaves
-its errors on disk. The JSON file is the run's resolved config in the format the
-CLI reads (`@ configs/<cli>.json`), so a run is re-runnable from its own output. Lines
-append as episodes complete, so results are durable mid-run. Files written
-by this surface contain episodes only.
-"""
-
-import asyncio
+"""CLI output directories, saved launch/config files, and per-attempt log paths."""
+
 import json
 import os
-from functools import cache
 from pathlib import Path
 
-from pydantic import BaseModel, TypeAdapter
+from pydantic import BaseModel
 
 from verifiers.v1.configs.cli.eval import EvalConfig
-from verifiers.v1.episode import EnvInfo, Episode, WireEpisode
-from verifiers.v1.state import StateT
-from verifiers.v1.task import DataT
-from verifiers.v1.trace import AgentConfigT, Trace
-from verifiers.v1.utils.aio import run_shielded
-
-TRACES_FILE = "traces.jsonl"
-"""Filename a run's rollout episodes are written to (one JSON episode per line)."""
+from verifiers.v1.utils.trace_store import TRACES_FILE
 
 CONFIG_DIR = "configs"
 """Directory inside a run dir holding its configs: the launch TOML copied verbatim to
@@ -78,10 +60,6 @@ def saved_config_path(run_dir: Path) -> Path | None:
     return candidates[0] if candidates else None
 
 
-# Compiling an adapter is the expensive part; run output reuses only a few model classes.
-type_adapter = cache(TypeAdapter)
-
-
 def output_path(config: EvalConfig) -> Path:
     """Where this run writes: `output_dir / run.dir` — the same grouping convention as
     training. The run directory defaults to the auto-generated run name
@@ -142,64 +120,3 @@ def save_config(
     (results_dir / TRACES_FILE).write_text(
         ""
     )  # fresh; appended to as rollouts complete
-
-
-def write_episode(
-    results_dir: Path, episode: Episode[DataT, StateT, AgentConfigT]
-) -> None:
-    """Serialize and append one rollout episode in the worker thread."""
-    # Preserve fields declared by typed Trace subclasses nested in the episode.
-    data = type_adapter(type(episode)).dump_json(episode, exclude_none=True)
-    with (results_dir / TRACES_FILE).open("ab") as f:
-        f.write(data + b"\n")
-
-
-def read_episodes(results_dir: Path, trace_type: type) -> list[WireEpisode]:
-    """Load a run's saved rollouts from `traces.jsonl` with traces typed as
-    `trace_type` (`Trace[WireTaskData, ...]` reads any taskset's file without
-    importing it)."""
-    trace_adapter = type_adapter(trace_type)
-    episodes: list[WireEpisode] = []
-    with (results_dir / TRACES_FILE).open(encoding="utf-8") as f:
-        for line in f:
-            if not line.strip():
-                continue
-            row = json.loads(line)
-            record = WireEpisode.model_validate({**row, "traces": []})
-            record.traces = [
-                trace_adapter.validate_python(trace) for trace in row["traces"]
-            ]
-            episodes.append(record)
-    return episodes
-
-
-async def append_episode(
-    results_dir: Path,
-    episode: Episode[DataT, StateT, AgentConfigT],
-    lock: asyncio.Lock,
-) -> None:
-    """Append one finished rollout episode without blocking the event loop. The run's
-    shared lock preserves whole-line ordering, and awaiting the worker preserves
-    per-episode durability."""
-
-    async def persist() -> None:
-        async with lock:
-            await asyncio.to_thread(write_episode, results_dir, episode)
-
-    # Run lock acquisition and the worker to completion even under cancellation, so
-    # finalized episodes are never lost mid-write (`run_shielded` re-raises the cancellation).
-    await run_shielded(persist())
-
-
-async def append_trace(
-    results_dir: Path, trace: Trace, lock: asyncio.Lock, env: str = ""
-) -> None:
-    """Append one finished trace as a single-agent rollout episode — debug and replay,
-    which complete trace-at-a-time, both go through here."""
-    episode = Episode(
-        env=EnvInfo(id=env),
-        task=trace.task,
-        traces=[trace],
-        ok=trace.ok,
-    )
-    await append_episode(results_dir, episode, lock)

verifiers/v1/cli/replay.py

@@ -22,13 +22,7 @@ from pydantic_config import cli
 
 import verifiers.v1 as vf
 from verifiers.v1.cli.dashboard.replay import ReplayProgress, replay_dashboard
-from verifiers.v1.cli.output import (
-    append_trace,
-    read_episodes,
-    save_config,
-    saved_config_path,
-    write_config,
-)
+from verifiers.v1.cli.output import save_config, saved_config_path, write_config
 from verifiers.v1.cli.resolve import narrow_taskset_config
 from verifiers.v1.configs.agent import WireAgentConfig
 from verifiers.v1.configs.cli.replay import ReplayConfig
@@ -37,6 +31,7 @@ from verifiers.v1.task import Task, WireTaskData
 from verifiers.v1.trace import Trace
 from verifiers.v1.utils.interrupt import install_interrupt
 from verifiers.v1.utils.logging import setup_logging
+from verifiers.v1.utils.trace_store import append_trace, read_episodes
 
 logger = logging.getLogger(__name__)
 

verifiers/v1/configs/agent.py

@@ -1,6 +1,8 @@
 """One env agent's config: who plays the seat, and its per-run caps."""
 
-from pydantic import SerializeAsAny, model_validator
+from typing import Any
+
+from pydantic import BaseModel, SerializeAsAny, model_validator
 from pydantic_config import BaseConfig
 
 from verifiers.v1.clients import ClientConfig
@@ -8,6 +10,7 @@ from verifiers.v1.configs.harness import HarnessConfig, WireHarnessConfig
 from verifiers.v1.configs.retries import RetryConfig
 from verifiers.v1.runtimes import PrimeConfig, RuntimeConfig
 from verifiers.v1.types import SamplingConfig
+from verifiers.v1.utils.generic import deep_merge
 
 
 class TimeoutConfig(BaseConfig):
@@ -78,3 +81,50 @@ class WireAgentConfig(AgentConfig):
     def _resolve_harness(cls, data):
         """Override: a record read resolves no plugins."""
         return data
+
+
+def agent_config_fields(config: BaseModel) -> dict[str, AgentConfig]:
+    """Top-level agent configs, in declaration order, keyed by their field names."""
+    return {name: value for name, value in config if isinstance(value, AgentConfig)}
+
+
+def merge_agent_defaults(config: type[BaseModel], data: Any) -> Any:
+    """Merge partial agent overrides onto their declared defaults."""
+    if isinstance(data, dict):
+        for name, field in config.model_fields.items():
+            if isinstance(field.default, AgentConfig) and isinstance(
+                data.get(name), dict
+            ):
+                data[name] = deep_merge(
+                    field.default.model_dump(exclude_none=True), data[name]
+                )
+    return data
+
+
+def resolve_agent(
+    spec: AgentConfig,
+    *,
+    model: str | None = None,
+    client: ClientConfig | None = None,
+    sampling: SamplingConfig | None = None,
+    harness: HarnessConfig | None = None,
+) -> AgentConfig:
+    """`spec` with what it leaves unset filled from the run's defaults; its own
+    sampling values merge over the run's. The one place a seat's identity resolves,
+    for configured agent roles."""
+    merged = spec.sampling if sampling is None else sampling
+    if sampling is not None and spec.sampling is not None:
+        merged = sampling.model_copy(
+            update=deep_merge(
+                sampling.model_dump(exclude_unset=True),
+                spec.sampling.model_dump(exclude_unset=True),
+            )
+        )
+    return spec.model_copy(
+        update={
+            "harness": spec.harness if spec.harness is not None else harness,
+            "model": spec.model if spec.model is not None else model,
+            "client": spec.client if spec.client is not None else client,
+            "sampling": merged,
+        }
+    )

verifiers/v1/configs/env.py

@@ -6,13 +6,16 @@ from typing import get_args
 from pydantic import Field, SerializeAsAny, model_validator
 from pydantic_config import BaseConfig
 
-from verifiers.v1.configs.agent import AgentConfig
+from verifiers.v1.configs.agent import (
+    AgentConfig,
+    agent_config_fields,
+    merge_agent_defaults,
+)
 from verifiers.v1.configs.harness import HarnessConfig
 from verifiers.v1.configs.retries import RetryConfig
 from verifiers.v1.configs.taskset import TasksetConfig
 from verifiers.v1.interception import ElasticInterceptionPoolConfig, InterceptionConfig
 from verifiers.v1.types import ID
-from verifiers.v1.utils.generic import deep_merge
 
 
 class TimeoutConfig(BaseConfig):
@@ -73,7 +76,7 @@ class EnvConfig(BaseConfig):
         default = default_agent_harness(self.taskset.id)
         return {
             name: cfg.harness if cfg.harness is not None else default
-            for name, cfg in _declared_agent_configs(self).items()
+            for name, cfg in agent_config_fields(self).items()
         }
 
     @model_validator(mode="before")
@@ -110,17 +113,7 @@ class EnvConfig(BaseConfig):
     @model_validator(mode="before")
     @classmethod
     def _merge_role_defaults(cls, data):
-        """Deep-merge partial role data over the field's declared default — plain
-        validation would replace the instance wholesale, resetting its other pins."""
-        if isinstance(data, dict):
-            for name, field in cls.model_fields.items():
-                if isinstance(field.default, AgentConfig) and isinstance(
-                    data.get(name), dict
-                ):
-                    data[name] = deep_merge(
-                        field.default.model_dump(exclude_none=True), data[name]
-                    )
-        return data
+        return merge_agent_defaults(cls, data)
 
     @classmethod
     def __pydantic_init_subclass__(cls, **kwargs):
@@ -146,16 +139,6 @@ class EnvConfig(BaseConfig):
                 )
 
 
-def _declared_agent_configs(config: EnvConfig) -> dict[str, AgentConfig]:
-    """The `AgentConfig` fields declared on an env's config, in declaration order —
-    the env's roles, each seat keyed by its field name (the only naming site)."""
-    return {
-        name: getattr(config, name)
-        for name, field in type(config).model_fields.items()
-        if isinstance(field.default, AgentConfig)
-    }
-
-
 def default_agent_harness(taskset_id: str) -> HarnessConfig:
     """What an unpinned role's `harness=None` resolves to: the taskset's bundled
     harness when it ships one, else the built-in `bash`."""

verifiers/v1/env.py

@@ -14,12 +14,12 @@ from typing import (
 
 from verifiers.v1.agent import Agent, Agents, _EpisodeAgent
 from verifiers.v1.clients import ModelContext
-from verifiers.v1.configs.agent import AgentConfig
-from verifiers.v1.configs.env import (
-    EnvConfig,
-    _declared_agent_configs,
-    default_agent_harness,
+from verifiers.v1.configs.agent import (
+    AgentConfig,
+    agent_config_fields,
+    resolve_agent,
 )
+from verifiers.v1.configs.env import EnvConfig, default_agent_harness
 from verifiers.v1.episode import EnvInfo, Episode
 from verifiers.v1.errors import EnvError, boundary
 from verifiers.v1.harness import Harness, HarnessConfig
@@ -32,7 +32,7 @@ from verifiers.v1.mcp import SharedToolServer, serve_shared
 from verifiers.v1.runtimes import SubprocessConfig, runtime_is_local
 from verifiers.v1.task import Task
 from verifiers.v1.trace import Error, Trace, TraceTask
-from verifiers.v1.utils.generic import concrete_type, deep_merge
+from verifiers.v1.utils.generic import concrete_type
 from verifiers.v1.utils.memory import trim_memory_periodically
 from verifiers.v1.utils.retries import run_episode_with_retry
 
@@ -105,7 +105,7 @@ class Env(ABC, Generic[ConfigT]):
         self._default_harness = default_agent_harness(config.taskset.id)
         task_cls = type(self.taskset).task_type()
         self._task_cls: type[Task] = task_cls
-        self._agent_specs: dict[str, AgentConfig] = _declared_agent_configs(self.config)
+        self._agent_specs: dict[str, AgentConfig] = agent_config_fields(self.config)
         if not self._agent_specs:
             raise ValueError(
                 f"{type(self).__name__} declares no agents; declare each as an "
@@ -205,23 +205,12 @@ class Env(ABC, Generic[ConfigT]):
 
         def make(name: str, spec: AgentConfig) -> Agent:
             # Unpinned fields fall back to the run's ctx / the taskset's harness.
-            sampling = ctx.sampling
-            if spec.sampling is not None:
-                sampling = sampling.model_copy(
-                    update=deep_merge(
-                        sampling.model_dump(exclude_unset=True),
-                        spec.sampling.model_dump(exclude_unset=True),
-                    )
-                )
-            resolved = spec.model_copy(
-                update={
-                    "harness": spec.harness
-                    if spec.harness is not None
-                    else self._default_harness,
-                    "model": spec.model if spec.model is not None else ctx.model,
-                    "sampling": sampling,
-                    "client": spec.client if spec.client is not None else ctx.client,
-                }
+            resolved = resolve_agent(
+                spec,
+                model=ctx.model,
+                client=ctx.client,
+                sampling=ctx.sampling,
+                harness=self._default_harness,
             )
             return _EpisodeAgent(
                 resolved,

verifiers/v1/gepa/runner.py

@@ -11,7 +11,7 @@ import logging
 from gepa.api import optimize
 from gepa.core.result import GEPAResult
 
-from verifiers.v1.cli.output import append_episode, output_path, save_config
+from verifiers.v1.cli.output import output_path, save_config
 from verifiers.v1.clients import ModelContext
 from verifiers.v1.env import Env
 from verifiers.v1.episode import Episode
@@ -22,6 +22,7 @@ from verifiers.v1.gepa.dataset import (
     split_tasks,
 )
 from verifiers.v1.gepa.reflection import build_reflection_lm
+from verifiers.v1.utils.trace_store import append_episode
 
 logger = logging.getLogger(__name__)
 

verifiers/v1/harnesses/utils/core.py

@@ -4,7 +4,6 @@ import argparse
 import asyncio
 import json
 import logging
-import random
 import subprocess
 from contextlib import AsyncExitStack
 from pathlib import Path
@@ -13,6 +12,13 @@ from typing import TYPE_CHECKING
 import httpx
 from openai import APIConnectionError, APIStatusError, AsyncOpenAI, omit
 from openai.lib.streaming.chat import AsyncChatCompletionStream
+from tenacity import (
+    AsyncRetrying,
+    before_sleep_log,
+    retry_if_exception_type,
+    stop_after_attempt,
+    wait_random_exponential,
+)
 
 if TYPE_CHECKING:
     # The harness bundles this module into the generated script before execution.
@@ -247,9 +253,16 @@ async def chat(
         kwargs["tools"] = tools
     if tools and tool_choice is not None:
         kwargs["tool_choice"] = tool_choice
-    for attempt in range(client.max_retries + 1):
+    async for attempt in AsyncRetrying(
+        retry=retry_if_exception_type((APIConnectionError, httpx.TransportError)),
+        stop=stop_after_attempt(client.max_retries + 1),
+        wait=wait_random_exponential(multiplier=0.5, max=8.0),
+        before_sleep=before_sleep_log(logging.getLogger(__name__), logging.WARNING),
+        reraise=True,
+    ):
         # Reuse the interception server's body-digest replay guard on stream retries.
-        headers = {"x-stainless-retry-count": str(attempt)} if attempt else omit
+        retry_count = attempt.retry_state.attempt_number - 1
+        headers = {"x-stainless-retry-count": str(retry_count)} if retry_count else omit
         raw_stream = await client.chat.completions.create(
             **kwargs,
             stream=True,
@@ -257,19 +270,8 @@ async def chat(
             extra_headers=headers,
         )
         # The SDK retries request setup; only stream consumption is retried here.
-        try:
+        with attempt:
             return await _read_chat_completion(raw_stream)
-        except (APIConnectionError, httpx.TransportError) as error:
-            cause = error.__cause__ or error
-            if attempt == client.max_retries:
-                raise
-            logging.getLogger(__name__).warning(
-                "Retrying interrupted model stream (%s/%s): %r",
-                attempt + 1,
-                client.max_retries,
-                cause,
-            )
-            await asyncio.sleep(min(0.5 * 2**attempt, 8.0) * random.uniform(0.75, 1.0))
 
 
 async def _read_chat_completion(raw_stream):

verifiers/v1/runtimes/docker/__init__.py

@@ -17,13 +17,18 @@ from urllib.parse import urlsplit
 
 from verifiers.v1.configs.runtime import NetworkPolicyConfig
 from verifiers.v1.errors import SandboxError
-from verifiers.v1.runtimes.base import SERVICE_PORT, BaseRuntimeInfo, parse_gpu
+from verifiers.v1.runtimes.base import (
+    SERVICE_PORT,
+    BaseRuntimeInfo,
+    parse_gpu,
+)
 from verifiers.v1.runtimes.container import ContainerConfig, ContainerRuntime, cli
 from verifiers.v1.runtimes.docker.egress import (
     EgressProxy,
     NetworkPolicy,
     is_loopback_host,
 )
+from verifiers.v1.utils.scope import run_scope
 
 logger = logging.getLogger(__name__)
 
@@ -157,9 +162,11 @@ class DockerRuntime(ContainerRuntime):
             for key, value in self.env.items()
             for arg in ("--env", f"{key}={value}")
         ]
+        self._label_args = ["--label", f"verifiers.run={run_scope()}"]
         run = await cli(
             self.engine,
             "run",
+            *self._label_args,
             "--detach",
             *options,
             *env_args,
@@ -296,6 +303,7 @@ class DockerRuntime(ContainerRuntime):
             helper = await cli(
                 self.engine,
                 "run",
+                *self._label_args,
                 "--rm",
                 "--user",
                 "0",
@@ -385,6 +393,7 @@ class DockerRuntime(ContainerRuntime):
         cut = await cli(
             self.engine,
             "run",
+            *self._label_args,
             "--rm",
             "--network",
             f"container:{self._container}",

verifiers/v1/runtimes/prime.py

@@ -185,12 +185,14 @@ class PrimeRuntime(Runtime):
             "gpu_type": gpu_type,
             "region": self.config.region,
         }
+        scope = run_scope()
+        labels = [*BASE_LABELS, *self.config.labels, scope]
         try:
             async with (
                 creation_limiter(
                     (self.config.creates_per_min or 0) / 60,
                     "prime-sandbox",
-                    run_scope(),
+                    scope,
                 )
                 or contextlib.nullcontext()
             ):
@@ -201,9 +203,7 @@ class PrimeRuntime(Runtime):
                     sandbox = await self._client.create(
                         CreateSandboxRequest(
                             name=self.name,
-                            labels=list(
-                                dict.fromkeys([*BASE_LABELS, *self.config.labels])
-                            ),
+                            labels=list(dict.fromkeys(labels)),
                             docker_image=self.config.image,
                             environment_vars=self.env,
                             **{k: v for k, v in options.items() if v is not None},

verifiers/v1/serve/delta.py

@@ -1,4 +1,4 @@
-"""Turn-by-turn episode deltas over the env-serve wire.
+"""Incremental trace deltas for episode streams and file writers.
 
 The worker streams a served episode as it grows. Each trace announces its own changes
 (`Trace.notify`, fired by the rollout at every phase change and by the interception proxy
@@ -25,7 +25,7 @@ from __future__ import annotations
 import asyncio
 import copy
 import logging
-from collections.abc import Awaitable, Callable
+from collections.abc import Awaitable, Callable, Iterable
 from typing import TYPE_CHECKING, Any, Self
 
 import msgpack
@@ -36,7 +36,6 @@ from verifiers.v1.graph import _decode_ndarray, _encode_ndarray
 from verifiers.v1.serve.encoding import msgpack_encoder
 
 if TYPE_CHECKING:
-    from verifiers.v1.env import RunSlot
     from verifiers.v1.trace import Trace
 
 logger = logging.getLogger(__name__)
@@ -102,16 +101,19 @@ class TraceCursor:
 
 
 class DeltaStreamer:
-    """Sends a `RunSlot`'s trace changes as deltas, one flush per burst of changes.
+    """Sends the current traces' changes as deltas, one flush per burst of changes.
 
-    `watch` subscribes a minted trace (pass it as `run_slot`'s `on_trace`); each
+    `watch` subscribes a minted trace (pass it as the runner's `on_trace`); each
     `Trace.notify` schedules a flush, and changes landing in the same loop iteration
     ride one flush. Use as an async context manager around the rollout: a clean exit
-    flushes the final state (the slot then holds the finished episode's traces), a
-    cancelled rollout flushes nothing — its client has already gone."""
+    flushes the final state; a cancelled rollout settles pending sends without a final flush."""
 
-    def __init__(self, slot: RunSlot, send: Callable[[bytes], Awaitable[None]]) -> None:
-        self.slot = slot
+    def __init__(
+        self,
+        traces: Callable[[], Iterable[Trace]],
+        send: Callable[[dict], Awaitable[None]],
+    ) -> None:
+        self.traces = traces
         self.send = send
         self.cursors: dict[str, TraceCursor] = {}
         self._scheduled: asyncio.Handle | None = None
@@ -132,6 +134,7 @@ class DeltaStreamer:
         else:
             for task in self._flushes:
                 task.cancel()
+            await asyncio.gather(*self._flushes, return_exceptions=True)
 
     def watch(self, trace: Trace) -> None:
         trace.watch(self._changed)
@@ -153,7 +156,7 @@ class DeltaStreamer:
         async with self._lock:
             for trace_id, delta, cursor in self.diff():
                 try:
-                    await self.send(pack(delta))
+                    await self.send(delta)
                 except Exception:  # the cursor stays put: the next flush diffs it again
                     logger.warning(
                         "failed to send delta for %s", trace_id, exc_info=True
@@ -193,7 +196,7 @@ class DeltaStreamer:
         """Each trace's delta against its sent cursor, with the cursor as it stands once
         that delta is sent (None for a discard). Nothing here is committed: `flush`
         stores a cursor only after its delta left."""
-        traces = list(self.slot.traces)
+        traces = list(self.traces())
         live = {trace.id for trace in traces}
         deltas: list[tuple[str, dict, TraceCursor | None]] = []
         for trace_id in [trace_id for trace_id in self.cursors if trace_id not in live]:

verifiers/v1/serve/server.py

@@ -8,7 +8,7 @@ import zmq.asyncio
 from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.client import ClientConfig
 from verifiers.v1.configs.env import EnvConfig
-from verifiers.v1.serve.delta import DeltaStreamer, TraceSummary, dump
+from verifiers.v1.serve.delta import DeltaStreamer, TraceSummary, dump, pack
 from verifiers.v1.serve.encoding import msgpack_encoder
 from verifiers.v1.serve.types import (
     BaseResponse,
@@ -100,15 +100,15 @@ class EnvServer:
         ctx = self._context(req.client, req.model, req.sampling)
         (slot,) = self.env.slots(self._build_task(req.task_data))
 
-        async def send_delta(data: bytes) -> None:
+        async def send_delta(delta: dict) -> None:
             await self.frontend.send_multipart(
-                [client_id, request_id, b"delta", data], copy=False
+                [client_id, request_id, b"delta", pack(delta)], copy=False
             )
 
         # The gate spans requests: `--max-concurrent` bounds this worker's episodes
         # in flight the same way the in-process eval's semaphore does. The streamer
         # ships each trace as it changes; the reply below carries only the rest.
-        async with DeltaStreamer(slot, send_delta) as streamer:
+        async with DeltaStreamer(lambda: slot.traces, send_delta) as streamer:
             episode = await self.env.run_slot(
                 slot, ctx, self._gate, on_trace=streamer.watch
             )

verifiers/v1/utils/trace_store.py

@@ -0,0 +1,80 @@
+"""Read and write standard saved traces."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+from functools import cache
+from pathlib import Path
+
+from pydantic import TypeAdapter
+
+from verifiers.v1.episode import EnvInfo, Episode, WireEpisode
+from verifiers.v1.state import StateT
+from verifiers.v1.task import DataT
+from verifiers.v1.trace import AgentConfigT, Trace
+from verifiers.v1.utils.aio import run_shielded
+
+TRACES_FILE = "traces.jsonl"
+type_adapter = cache(TypeAdapter)
+
+
+def write_episode(
+    results_dir: Path, episode: Episode[DataT, StateT, AgentConfigT]
+) -> None:
+    """Serialize and append one rollout episode in the worker thread."""
+    # Preserve fields declared by typed Trace subclasses nested in the episode.
+    data = type_adapter(type(episode)).dump_json(episode, exclude_none=True)
+    with (results_dir / TRACES_FILE).open("ab") as f:
+        f.write(data + b"\n")
+
+
+def read_episodes(results_dir: Path, trace_type: type) -> list[WireEpisode]:
+    """Load a run's saved rollouts from `traces.jsonl` with traces typed as
+    `trace_type` (`Trace[WireTaskData, ...]` reads any taskset's file without
+    importing it)."""
+    trace_adapter = type_adapter(trace_type)
+    episodes: list[WireEpisode] = []
+    with (results_dir / TRACES_FILE).open(encoding="utf-8") as f:
+        for line in f:
+            if not line.strip():
+                continue
+            row = json.loads(line)
+            record = WireEpisode.model_validate({**row, "traces": []})
+            record.traces = [
+                trace_adapter.validate_python(trace) for trace in row["traces"]
+            ]
+            episodes.append(record)
+    return episodes
+
+
+async def append_episode(
+    results_dir: Path,
+    episode: Episode[DataT, StateT, AgentConfigT],
+    lock: asyncio.Lock,
+) -> None:
+    """Append one finished rollout episode without blocking the event loop. The run's
+    shared lock preserves whole-line ordering, and awaiting the worker preserves
+    per-episode durability."""
+
+    async def persist() -> None:
+        async with lock:
+            await asyncio.to_thread(write_episode, results_dir, episode)
+
+    # Run lock acquisition and the worker to completion even under cancellation, so
+    # finalized episodes are never lost mid-write (`run_shielded` re-raises the cancellation).
+    await run_shielded(persist())
+
+
+async def append_trace(
+    results_dir: Path, trace: Trace, lock: asyncio.Lock, env: str = ""
+) -> None:
+    """Append one finished trace as a single-agent rollout episode — debug and replay,
+    which complete trace-at-a-time, both go through here."""
+    episode = Episode(
+        env=EnvInfo(id=env),
+        task=trace.task,
+        traces=[trace],
+        ok=trace.ok,
+    )
+    await append_episode(results_dir, episode, lock)

{verifiers-0.3.2.dev136.dist-info → verifiers-0.3.2.dev138.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev136
+Version: 0.3.2.dev138
 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-0.3.2.dev136.dist-info → verifiers-0.3.2.dev138.dist-info}/RECORD

@@ -1,7 +1,7 @@
 verifiers/__init__.py,sha256=F5IVMw6mhGaPBMCw7QKk0la_r5mMwydP2hdTTdQucik,352
 verifiers/v1/__init__.py,sha256=zzXyzUkCJCXfPRajYNNQlgIAz6_-uu9oYNb-XsbIjeU,8434
-verifiers/v1/agent.py,sha256=Lc4ySCWBCZRqED-M2LwpPcewAH9k4AuS_ubq-E6-wDQ,32808
-verifiers/v1/env.py,sha256=Hon7gcepciMYgki_jN4-l4XCcKRZRYuYm7c17nfeiFE,18520
+verifiers/v1/agent.py,sha256=10hXkda0sC5FxT2a32t9En5QlExqcxgnBYtGYk7oins,32651
+verifiers/v1/env.py,sha256=5xB51-IwVEMoEzQ44vV2HD1O2VFcX3H3-vPBCPl_49o,17956
 verifiers/v1/episode.py,sha256=DKUkG9YWBRbaqAP7VCg9DcOB95WSH_EM_LY4jkJ63hA,5817
 verifiers/v1/errors.py,sha256=uQySmCUrjDYFkBIii1vBFPiEcPFPQZ9EGpas_6HvcQI,5279
 verifiers/v1/graph.py,sha256=3HtCYL36MDAoiBk7u8yi6FxUBkIqQ7VYn9_IL0NEOog,41559
@@ -18,33 +18,33 @@ verifiers/v1/types.py,sha256=xPetvJxksKPYKic389RFnVvy2vN-gD2bPZFZoEyzmp4,10191
 verifiers/v1/acp/__init__.py,sha256=9RySmxFeEMT2XSJy5wYevqEhQdul2jHl9f5XAribG3A,13631
 verifiers/v1/acp/runner.py,sha256=zPo-2ZXmFMmQchhD7nNzlZUrabG09qiBvpB67KVGm3M,14095
 verifiers/v1/cli/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-verifiers/v1/cli/debug.py,sha256=sgXi2y17qr2yd5kbkL_IzwWN1nIFBOa1adu7bcFzaVk,11780
-verifiers/v1/cli/gepa.py,sha256=F0Q1ApfUL_mInFnmaJ5-yBDRYBtBuH72OkWelPoGCp4,4332
+verifiers/v1/cli/debug.py,sha256=G05-k9mqsMNIf7L-Vlxgdmw3QFwFF94-FodKXV7GYbQ,11822
+verifiers/v1/cli/gepa.py,sha256=IyS3YZCCHVydb-OyPeGdtCYsz6OYqXXddXqYq8RLucM,4374
 verifiers/v1/cli/init.py,sha256=6wr14011_Dygv10R6_YWg5A3oSrz187E2WspGRHIfZw,7608
-verifiers/v1/cli/output.py,sha256=lAl6vwKf9mVuNZb8PzOexv6Q2h-H6t0oUPcapffY3uY,8351
-verifiers/v1/cli/replay.py,sha256=JjhaGczQ41zN6wz1P1nbGGvv0xdfQlX4OmZItGPDpZc,10395
+verifiers/v1/cli/output.py,sha256=4FRuQ6UIZCigMfUUO2nCFFJKxls3DtWDXswMCVTxjIA,5010
+verifiers/v1/cli/replay.py,sha256=ukCmu4-SGYZa2kB9TpNERUYT6mAVEjNJxRNqwIH5eYk,10412
 verifiers/v1/cli/resolve.py,sha256=Q1EHM7wWQo0YkwJA898qtZbYLaIkjIVq3Q7kUp2YIxs,4346
 verifiers/v1/cli/resume.py,sha256=hqV5AxxqeIInJeEZQerQznuijhMNijFEFcHZIltfLdE,470
 verifiers/v1/cli/validate.py,sha256=Z_rUldl8jzXz52oy3FBIRGr4HYhWFXrq43WmU0sJ-4M,17325
 verifiers/v1/cli/dashboard/__init__.py,sha256=v-baMxQuWxOCsbU7-p_jj2Q9BUnTN-TWi1q2hK6rU2s,198
 verifiers/v1/cli/dashboard/base.py,sha256=kUP93zJSIVLptSbnWX6MOy8kGg63fpeAzPqakCpPyDc,3547
-verifiers/v1/cli/dashboard/eval.py,sha256=HEXPEegNXdgU9zAIuHzsWsoSIHyiOcDZh4tCGFNxKAk,38353
+verifiers/v1/cli/dashboard/eval.py,sha256=69vKmLaEY1ll1wahjRGvO1_7QsUQ1URtZOG7ZawvRCw,38347
 verifiers/v1/cli/dashboard/replay.py,sha256=_X5up9MJsRbd_sWe4LgNTPNT8k1ip4xNpo50hxI3Ljk,2697
 verifiers/v1/cli/dashboard/validate.py,sha256=xrpK3Y90JsoDCQYLLvvsnJ4CPRMvSSJIghfHK62E6FQ,3687
 verifiers/v1/cli/eval/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 verifiers/v1/cli/eval/hint.py,sha256=NXP2_JYrQHguqr68LxOJG3lYAsue_VNoepyTVPn-N4E,256
-verifiers/v1/cli/eval/main.py,sha256=L_vKnMJEbpo1IKjbm0TXjXv_1x9we7mFv-jH_xY211Q,5946
-verifiers/v1/cli/eval/resume.py,sha256=fYkWa0Fudq--IWfk25j_McLm67bXYWvksWmJR4u4QAU,3811
-verifiers/v1/cli/eval/runner.py,sha256=z1yU2_rAA6tHz3PL7NQuzm9v3-4Ri18DXD5SRBKXqVY,6962
+verifiers/v1/cli/eval/main.py,sha256=46aHpha6yhFzbqUrRogC5wC2XAwdZcpp-4siqLpUExY,5984
+verifiers/v1/cli/eval/resume.py,sha256=cmFdxLwyU66lyAuXb7U8mNBB-8NLX41JdCg1EuU4w9g,3818
+verifiers/v1/cli/eval/runner.py,sha256=JLitCJajj6h34L0Kq5fWklS7FAsiGOrdGp_c0qVN7Xc,6987
 verifiers/v1/clients/__init__.py,sha256=Ysig0tE_0E4Jsfgfes1XHN-fK1s_RCXdqZD6E7lK4PU,507
 verifiers/v1/clients/base.py,sha256=PoDw4GMqrfuPTFVK6n2jDmlD0_lJ5zTrYcNio5yq6K8,1691
 verifiers/v1/clients/client.py,sha256=zqC_AkiD9pl0kxxIbupNiehbS-YnVdp3EMZ0LOHTd8U,3125
 verifiers/v1/clients/eval.py,sha256=cJucyUFsbF1JlG8nF0Tlw2EBMSoAF0YMklgYiOSoNpg,8322
 verifiers/v1/clients/train.py,sha256=jm_z_Gkwqo4IKEVN8WrDnEyJykXxmswFHzlsnFop7_I,18945
 verifiers/v1/configs/__init__.py,sha256=X7u6X7B3ieD1HZl79Tv1kg-yPHbT0QLvAebaGqxVtvA,317
-verifiers/v1/configs/agent.py,sha256=PiWMNueum67vnyLLpIgxXZDb3FTQEUFqo2Z1siOOJqs,3342
+verifiers/v1/configs/agent.py,sha256=76ZvO0EE_JCEep96vOYfvAmC4fsLHtJn_cKbJhve6gU,5241
 verifiers/v1/configs/client.py,sha256=fPWtBJv5lPkqttwk0AF2ubfLdzxX2XCO0Zr-A9V4mtk,4440
-verifiers/v1/configs/env.py,sha256=-Bcdd4MEPfekVc_fnpiwsb9UDwCNdNCKrAHGoQuuXu8,7784
+verifiers/v1/configs/env.py,sha256=xDcfoBcor6loLrK79YdwEdui5OGVLM_InVA8HnLcjPQ,6864
 verifiers/v1/configs/harness.py,sha256=ZrR-Egzfft95oMoqQ78bAq9TqP882Q68M_47Ajw9lE8,2918
 verifiers/v1/configs/judge.py,sha256=WEwA8D4X5JSaIP3c-E2wPnvgByWeT6JRAn1d8gfiQ40,2142
 verifiers/v1/configs/retries.py,sha256=-nPlgmz7J_NZ4GOYG6oCc3thXo4La2SufBmvQGvvPGo,821
@@ -81,7 +81,7 @@ verifiers/v1/gepa/adapter.py,sha256=YNvHMR2L5Utl-vtPfjZxmDCd2tTcoAPWG6aVGnV1icM,
 verifiers/v1/gepa/config.py,sha256=PhZRLkhaPciNm_IhwVX3e9DM8Y2IdCe4Fbi5WjH5aro,4883
 verifiers/v1/gepa/dataset.py,sha256=SPGG55BVbJ9HtCEvm-xyl52QSbldFwGqWvRIcaLelsM,2171
 verifiers/v1/gepa/reflection.py,sha256=ptHhx0lcDLcWhWhXuslh9T8gPVNpnX5XLPtXoneW1yA,1054
-verifiers/v1/gepa/runner.py,sha256=fGEyFzejdQKBzBbvdiTDHV3JtRhXMXMGSBU5y8C8i3w,6054
+verifiers/v1/gepa/runner.py,sha256=s2eUV0T0mUGy-7XxupoX757z9kZoKAR73NN5M6rUi6E,6096
 verifiers/v1/harnesses/__init__.py,sha256=2JPwrRoYoigH-HfBtiFnatRdFEVHLv4rwlKfsFEibyE,1803
 verifiers/v1/harnesses/node.py,sha256=-lDPh30p65IETPVLkKn4yhEK49rtEGVQf9TNs7MVQYE,2133
 verifiers/v1/harnesses/bash/__init__.py,sha256=IAV4eMMqDHSl-9ANb0Q3kxmBZ22SeHZFBM448X4gQio,140
@@ -118,7 +118,7 @@ verifiers/v1/harnesses/terminus_2/harness.py,sha256=cKQJluW01mkxEccTLFwSWsVMavyi
 verifiers/v1/harnesses/terminus_2/program.py,sha256=X8MZdyItU3gbXGXnGOlJ9LWcUTnqIDaxlhkZvoRiqKY,2703
 verifiers/v1/harnesses/utils/__init__.py,sha256=eTS-Ft2HwoLiYn5u1t_KueLmaZB_W12cdYmcVsE-G5A,364
 verifiers/v1/harnesses/utils/compaction.py,sha256=DeY6CYLEsv7-vJm26sT2mRaPHu3W2ywiZhvFr_pi6Hg,10495
-verifiers/v1/harnesses/utils/core.py,sha256=Kqp0npmiJlGtT5vt7r2sX70R29W9d_ODR7gN-ufKrug,19642
+verifiers/v1/harnesses/utils/core.py,sha256=A4B_TbZBQVd6IYQyytrO8gbmrO-aVqjRl0-w8GClWF4,19676
 verifiers/v1/harnesses/utils/install.py,sha256=BDkkUVZ1brSzCrbQOvMO5mVIE2DsV2pE7rTC5yROYUw,2992
 verifiers/v1/harnesses/utils/launch.py,sha256=RqPuUY-kT2XVuLSGtXW7kRQEaP-o2KzkiJ7vcoo2neM,2883
 verifiers/v1/harnesses/utils/mcp.py,sha256=oNhKfLjAW-wScJcDIsW_HfuvKY-JkaGBBZrRLqgRK1g,7355
@@ -145,16 +145,16 @@ verifiers/v1/runtimes/base.py,sha256=IKmrJSJltWX9M1nPUZExKluPJM4N82VsnqQGFz0ECV8
 verifiers/v1/runtimes/container.py,sha256=puQD8X9acP6GefoCNEnq40POi-nXPD-729bUjTVShfM,9554
 verifiers/v1/runtimes/limiters.py,sha256=rOfQJcMdYcLSKq8AL9O09BfGOLqpYmNgrCxXXNcHLmU,2849
 verifiers/v1/runtimes/modal.py,sha256=6ccitC0HVsAP_l8soJ_RQYNIVKrJjDsz2vAJBtxUZjw,17476
-verifiers/v1/runtimes/prime.py,sha256=Ly7eTEJY06-kjBBZhpJVALtsW23M-ZLJ1mlq27uSS9U,19158
+verifiers/v1/runtimes/prime.py,sha256=vuRlBxuYLMLLy6J4cL2GhifnNhsLnzaN0e9FUN-yinc,19149
 verifiers/v1/runtimes/subprocess.py,sha256=QmpQ235_xrX6gowrr8557VgbydOXpLbfFq0KKGfJThk,8774
-verifiers/v1/runtimes/docker/__init__.py,sha256=2wceP-8V-w0dQBEeLtoJFfCZ135DxP2rcdcUBmSWiNU,18543
+verifiers/v1/runtimes/docker/__init__.py,sha256=KaUbd87exTIT5JOoyNmEYFJ-JmnXf-p2MriBn2BdQtQ,18775
 verifiers/v1/runtimes/docker/egress.py,sha256=wtKWTL2_zFidxMR7lDRmfevQlOsl2Iyk88ebdO4Kqqw,24566
 verifiers/v1/serve/__init__.py,sha256=wOKuwfugzmDT4yIniwzCXlf4G6tv6Cg8Kc--WMa38Bs,553
 verifiers/v1/serve/client.py,sha256=E2y2x4qqYveWRWC9VZYQEu19aMEryPWTdXnYzbq5uTA,9829
-verifiers/v1/serve/delta.py,sha256=vbwFy9vGcSIL1gJad5gvLHr2vEBiPtl3-sO_HoHeqSM,13574
+verifiers/v1/serve/delta.py,sha256=I3sJxS39rksOCeX9EO5FNWAR33Z-9tccxeEkU1txgLw,13615
 verifiers/v1/serve/encoding.py,sha256=hBZFucAZK9riXOV3DaHcskq9zHTT8gVGkoFVOPfnr30,2499
 verifiers/v1/serve/pool.py,sha256=bT-FOlIuOiBtOasuPB-p9887tkpEqvQUJBgn1Q45ph0,15912
-verifiers/v1/serve/server.py,sha256=i6XrY_PMAapkrP54tYX8-K6iGmOmrR7BbQ-3Qhn3w4s,9885
+verifiers/v1/serve/server.py,sha256=kaQFh_tVRcmRaQoB5o6CU4y8Fh9X-zFg-cmcvU2GExg,9913
 verifiers/v1/serve/types.py,sha256=Qw5IjiJhsZz4giSXUBi9dfInBnQZ5_xiUmgH2qp8jPI,1741
 verifiers/v1/tasksets/__init__.py,sha256=2ijE3qIlLoUA02g6iysOwWQylyBBQL1GhSpjv8_bp0w,521
 verifiers/v1/tasksets/harbor/__init__.py,sha256=JxThzHFYEkuSfQbVf3pEnjordgdfe4MU-QOsSSbyovU,326
@@ -189,9 +189,10 @@ verifiers/v1/utils/prime.py,sha256=UTYRjp9cbjNb6CVmBHda-1wWZAIIfxNOmTyNuT7_wL4,9
 verifiers/v1/utils/retries.py,sha256=Y2ZgrAjn-qNkRKZP_RVNL_7EK0iaRcZeCa404lpGYi4,5417
 verifiers/v1/utils/scope.py,sha256=bzUEiWDOVdbj7dXOwGOlzsQ_-Xu0j9NR5TalQB-8TS8,838
 verifiers/v1/utils/score.py,sha256=493yJVMw8teCu9JxapxMFPFzI0hNUqdo0Y2nGW4kckk,6200
+verifiers/v1/utils/trace_store.py,sha256=RsCDonGR-bs-tvEstcppiJm3jFmAdIiG83fA5KtCbAw,2801
 verifiers/v1/utils/version.py,sha256=-obEo_-l9-D8FLef4hYxncOe-uJpxrM1g2Hig_37Sgs,1607
-verifiers-0.3.2.dev136.dist-info/METADATA,sha256=Dody0KdHsobWBNx2Mc1KbvuGAH152EZaPxEonMGcoJ4,4161
-verifiers-0.3.2.dev136.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
-verifiers-0.3.2.dev136.dist-info/entry_points.txt,sha256=iugElcdWPKbQM7uFF0lZ8iUpHsNr17-BwEAAjJWxV3U,259
-verifiers-0.3.2.dev136.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.3.2.dev136.dist-info/RECORD,,
+verifiers-0.3.2.dev138.dist-info/METADATA,sha256=5cCtrF7ceNh2ee4zKBCZH_VkAuhIRxR2TOO6Nev91jM,4161
+verifiers-0.3.2.dev138.dist-info/WHEEL,sha256=THafob7ofN-NsuMN7Mg4qZyHaQI7KkD-QlcQatYhXPo,87
+verifiers-0.3.2.dev138.dist-info/entry_points.txt,sha256=iugElcdWPKbQM7uFF0lZ8iUpHsNr17-BwEAAjJWxV3U,259
+verifiers-0.3.2.dev138.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.3.2.dev138.dist-info/RECORD,,

{verifiers-0.3.2.dev136.dist-info → verifiers-0.3.2.dev138.dist-info}/WHEEL

@@ -1,4 +1,4 @@
 Wheel-Version: 1.0
-Generator: hatchling 1.32.0
+Generator: hatchling 1.32.3
 Root-Is-Purelib: true
 Tag: py3-none-any