verifiers 0.2.2.dev85__py3-none-any.whl → 0.2.2.dev87__py3-none-any.whl

verifiers/v1/configs/harness.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 import os
 from pathlib import Path
 
-from pydantic import ConfigDict, Field
+from pydantic import ConfigDict, Field, FiniteFloat
 from pydantic_config import BaseConfig
 
 from verifiers.v1.types import ID
@@ -20,7 +20,7 @@ class HarnessConfig(BaseConfig):
     """Extra program variables; harness-owned variables take precedence."""
     forward_env: list[str] = Field(default_factory=list)
     """Host variables to forward without writing secrets into config; explicit `env` wins."""
-    tool_timeout: float = Field(600.0, gt=0)
+    tool_timeout: FiniteFloat = Field(600.0, gt=0)
     """Seconds a single MCP tool call may take; raise it for tools that boot a VM."""
     disabled_tools: list[str] | None = None
     skills: list[Path] = Field(default_factory=list)

verifiers/v1/envs/agentic_judge/__init__.py

@@ -1,16 +1,16 @@
 from verifiers.v1.envs.agentic_judge.env import (
-    AgenticJudgeEnv,
     AgenticJudgeEnvConfig,
     Criterion,
+    IsolatedAgenticJudgeEnv,
     JudgeTaskConfig,
     ScoreConfig,
     TextFile,
 )
 
 __all__ = [
-    "AgenticJudgeEnv",
     "AgenticJudgeEnvConfig",
     "Criterion",
+    "IsolatedAgenticJudgeEnv",
     "JudgeTaskConfig",
     "ScoreConfig",
     "TextFile",

verifiers/v1/envs/agentic_judge/env.py

@@ -1,16 +1,17 @@
-"""agentic-judge: a solver plays the task, then a judge verifies the work.
-
-A reusable env (`--env.id agentic-judge` over any taskset). The solver plays the
-task in a container provisioned from its runtime policy; the judge then grades
-rubric criteria (`[env.task]`: policy prompt, criteria file) and writes its
-verdicts to `/tmp/verdict.json`, with the solver's full trace record uploaded at
-`/tmp/trace.json`. `finalize()` validates them strictly onto the solver's trace —
-`judge/<name>` metrics plus a weighted-mean `judge` reward, composed with the
-taskset's own rewards via `[env.score]` (judge-only by default).
-
-`--env.share-runtime` controls whether the judge uses the solver's runtime. It is
-enabled by default. When disabled, the judge gets a fresh runtime containing the
-task's collected artifacts.
+"""Agentic judging: a solver plays the task, then a judge verifies the work.
+
+Two reusable envs share the grading protocol. `--env.id agentic-judge` provisions
+a fresh box from the solver's runtime policy and restores only the task's collected
+artifacts; `--env.id shared-agentic-judge` explicitly runs the judge in the
+solver's box. The judge grades rubric criteria (`[env.task]`: policy prompt,
+criteria file) and writes its verdicts to `/tmp/verdict.json`, with the solver's
+full trace record uploaded at `/tmp/trace.json`. `finalize()` validates them
+strictly onto the solver's trace — `judge/<name>` metrics plus a weighted-mean
+`judge` reward, composed with the taskset's own rewards via `[env.score]`
+(judge-only by default).
+
+The environment id selects the runtime boundary; there is no mode boolean whose
+value can disagree with the environment's security and artifact semantics.
 """
 
 import json
@@ -260,20 +261,22 @@ class AgenticJudgeEnvConfig(vf.EnvConfig):
     """The solver agent. Its runtime must be a container:
     `--env.solver.runtime.type docker|prime`."""
     judge: vf.AgentConfig = vf.AgentConfig()
-    """The judge agent. Its runtime is ignored when `share_runtime` is enabled;
-    otherwise it must be a container."""
-    share_runtime: bool = True
-    """Whether the judge grades in the solver's runtime."""
+    """The judge agent. Its runtime setting is ignored: both judging modes use the
+    solver's resolved runtime policy, either by borrowing its box or provisioning a
+    fresh equivalent one."""
     task: JudgeTaskConfig = JudgeTaskConfig()
     score: ScoreConfig = ScoreConfig()
 
 
 class AgenticJudgeEnv(vf.Env[AgenticJudgeEnvConfig]):
+    """Common agentic-judge protocol; subclasses choose the runtime boundary."""
+
     def __init__(self, config: AgenticJudgeEnvConfig) -> None:
-        if config.share_runtime:
-            config.judge = config.judge.model_copy(
-                update={"runtime": config.solver.runtime}
-            )
+        # Both modes use the solver's policy. Shared judging borrows that exact box;
+        # isolated judging resolves the mirrored JudgeTask into a fresh equivalent.
+        config.judge = config.judge.model_copy(
+            update={"runtime": config.solver.runtime}
+        )
         super().__init__(config)
         self._check_agents()
         # A missing policy file or a malformed rubric fails here, not mid-episode.
@@ -304,21 +307,6 @@ class AgenticJudgeEnv(vf.Env[AgenticJudgeEnvConfig]):
         # The judge grades the policy; its tokens are never training data.
         agents.judge.trainable = False
 
-    async def run(self, task: vf.Task, agents: vf.Agents) -> None:
-        if self.config.share_runtime:
-            async with agents.solver.provision(task) as box:
-                solution = await agents.solver.run(task, runtime=box)
-                judge_task = JudgeTask.from_trace(solution, self.config.task)
-                await agents.judge.run(judge_task, runtime=box)
-            return
-
-        solution = await agents.solver.run(task)
-        if not solution.ok:
-            return
-        await agents.judge.run(
-            JudgeTask.from_trace(solution, self.config.task, share_runtime=False)
-        )
-
     async def finalize(self, task: vf.Task, episode: vf.Episode) -> None:
         by_agent = {t.agent.name: t for t in episode.traces}
         if "judge" not in by_agent:
@@ -336,3 +324,25 @@ class AgenticJudgeEnv(vf.Env[AgenticJudgeEnvConfig]):
         total = sum(criterion.weight for criterion in criteria)
         reward = sum(c.weight * scores[c.name] for c in criteria) / total
         solution.record_reward("judge", reward, weight=self.config.score.judge_weight)
+
+
+class SharedAgenticJudgeEnv(AgenticJudgeEnv):
+    """Judge the solver in its runtime, preserving the complete mutable workspace."""
+
+    async def run(self, task: vf.Task, agents: vf.Agents) -> None:
+        async with agents.solver.provision(task) as box:
+            solution = await agents.solver.run(task, runtime=box)
+            judge_task = JudgeTask.from_trace(solution, self.config.task)
+            await agents.judge.run(judge_task, runtime=box)
+
+
+class IsolatedAgenticJudgeEnv(AgenticJudgeEnv):
+    """Judge only collected artifacts in a fresh box with the solver's policy."""
+
+    async def run(self, task: vf.Task, agents: vf.Agents) -> None:
+        solution = await agents.solver.run(task)
+        if not solution.ok:
+            return
+        await agents.judge.run(
+            JudgeTask.from_trace(solution, self.config.task, share_runtime=False)
+        )

verifiers/v1/envs/shared_agentic_judge/__init__.py

@@ -0,0 +1,6 @@
+from verifiers.v1.envs.agentic_judge.env import (
+    AgenticJudgeEnvConfig,
+    SharedAgenticJudgeEnv,
+)
+
+__all__ = ["AgenticJudgeEnvConfig", "SharedAgenticJudgeEnv"]

{verifiers-0.2.2.dev85.dist-info → verifiers-0.2.2.dev87.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev85
+Version: 0.2.2.dev87
 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.dev85.dist-info → verifiers-0.2.2.dev87.dist-info}/RECORD

@@ -211,7 +211,7 @@ verifiers/v1/configs/__init__.py,sha256=X7u6X7B3ieD1HZl79Tv1kg-yPHbT0QLvAebaGqxV
 verifiers/v1/configs/agent.py,sha256=V_2FDvPHRB3_bHN9QnNnDYOjlmTP7NtG22SaopU-X64,3261
 verifiers/v1/configs/client.py,sha256=dCi2aC_wVHCrM0xRi9foFcbqp9pN2GR7yxEs1z_ChJM,4371
 verifiers/v1/configs/env.py,sha256=7ESsWZRWJpKqR5EEV1rmj0R1Dkn2Pn8MVSXPJOzRiF8,7885
-verifiers/v1/configs/harness.py,sha256=uDBb3DoX3RGbH2fOAvVVMaluP3VivtMfurGiBbbpuAs,1695
+verifiers/v1/configs/harness.py,sha256=GuXpAEPTwP2i2hp0bM4CYzC1yAMz-cp_j2_4-Qp5eM4,1714
 verifiers/v1/configs/judge.py,sha256=jc9w8eYsqSEyMMY7n12EWyYcwqQl5s-47xV3Me7WA0c,2242
 verifiers/v1/configs/legacy.py,sha256=YOJM6L5u6-jYpZ1CLpPInufKdmz1lEsMNKIRdT2JaaY,1080
 verifiers/v1/configs/retries.py,sha256=-nPlgmz7J_NZ4GOYG6oCc3thXo4La2SufBmvQGvvPGo,821
@@ -231,10 +231,11 @@ verifiers/v1/dialects/base.py,sha256=YZQDnasQseDXK-GHLfJLZgyHIIygS80YhD_sTjmUbfA
 verifiers/v1/dialects/chat.py,sha256=R_otqi5N0snB1QVKIYloNeLHR9kc7OmI06HYg1ir_Lw,14021
 verifiers/v1/dialects/responses.py,sha256=ismnxUikPcPUIm80FUTPBbj-Cn1e0yGAsISMEseAWRc,13679
 verifiers/v1/envs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-verifiers/v1/envs/agentic_judge/__init__.py,sha256=6vwtMCQ_cuvuubCF5-nrE5W7gGgzE-LIPKorkM5DXdw,309
-verifiers/v1/envs/agentic_judge/env.py,sha256=ZgNDanhSQYpP3K5Vz2rUW0UjWGApdqz86ng9BXhj6PA,13914
+verifiers/v1/envs/agentic_judge/__init__.py,sha256=LfEcRHLG_siKzhzWmODDXWSLpLgfum3UFdQDK9XG0gk,325
+verifiers/v1/envs/agentic_judge/env.py,sha256=9HdI8dm6m5QTNcHhpMvu6JJ2GlqH6e6xrc0inrEEZMk,14443
 verifiers/v1/envs/best_of_n/__init__.py,sha256=mubASiwXMNIhnQFDccNAc7yZ0OFsOVK8hH8WIx9b2A4,119
 verifiers/v1/envs/best_of_n/env.py,sha256=KEUJBCVqblnXP_lEwm3mjGdaikBMkecz8UjOFPBhxc0,1987
+verifiers/v1/envs/shared_agentic_judge/__init__.py,sha256=9T4vyYAuqgZSQjqTIceghxnpauvqeW6XGhZSQDCttMA,168
 verifiers/v1/envs/single_agent/__init__.py,sha256=r0edAc_6gtHBhPzeBGVcdHZZqSdCPZnLMje6eVgyz5U,138
 verifiers/v1/envs/single_agent/env.py,sha256=lPs_a6VA7KoH1xqA3T8S32Jzb_d4zrx0_54LrqT5z1Q,1187
 verifiers/v1/envs/user_sim/__init__.py,sha256=nnT695HdqjAEouYL5wtVu6MhIp2wf5yuK-EYgxOVx0s,118
@@ -337,8 +338,8 @@ verifiers/v1/utils/platform.py,sha256=56Ixmk1SER6q5LvzyYcA0hgzGpzhxKZL8Hiqp4c_XV
 verifiers/v1/utils/retries.py,sha256=Y2ZgrAjn-qNkRKZP_RVNL_7EK0iaRcZeCa404lpGYi4,5417
 verifiers/v1/utils/score.py,sha256=-R5Cog14r_tJ6Q_oOfGr5VPAm2CCUhofYZB6B9lm4wM,5787
 verifiers/v1/utils/version.py,sha256=-obEo_-l9-D8FLef4hYxncOe-uJpxrM1g2Hig_37Sgs,1607
-verifiers-0.2.2.dev85.dist-info/METADATA,sha256=SXrfQrFds4Ppx3ye69DlhIspd3t9JAz7gx4EfYB7CA0,4545
-verifiers-0.2.2.dev85.dist-info/WHEEL,sha256=lCkmxWfQsSc9CfIClYeavTdQeEX2toPqufh9gI35EQA,87
-verifiers-0.2.2.dev85.dist-info/entry_points.txt,sha256=v6v0QT9vVExnfn4br42MostI_o3f3ZbWYlTormz-U3g,515
-verifiers-0.2.2.dev85.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.2.2.dev85.dist-info/RECORD,,
+verifiers-0.2.2.dev87.dist-info/METADATA,sha256=U5lglk-lfizY1P98ZGz92ItcVHUBniDOdYTDLrExpaE,4545
+verifiers-0.2.2.dev87.dist-info/WHEEL,sha256=lCkmxWfQsSc9CfIClYeavTdQeEX2toPqufh9gI35EQA,87
+verifiers-0.2.2.dev87.dist-info/entry_points.txt,sha256=v6v0QT9vVExnfn4br42MostI_o3f3ZbWYlTormz-U3g,515
+verifiers-0.2.2.dev87.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.2.2.dev87.dist-info/RECORD,,