verifiers 0.3.2.dev130__tar.gz → 0.3.2.dev132__tar.gz

{verifiers-0.3.2.dev130 → verifiers-0.3.2.dev132}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev130
+Version: 0.3.2.dev132
 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.dev130 → verifiers-0.3.2.dev132}/verifiers/v1/harnesses/utils/core.py

@@ -240,7 +240,9 @@ async def chat(
     *,
     tool_choice: str | None = None,
 ):
-    kwargs = {"model": model, "messages": messages, "tools": tools or None}
+    kwargs = {"model": model, "messages": messages}
+    if tools:
+        kwargs["tools"] = tools
     if tools and tool_choice is not None:
         kwargs["tool_choice"] = tool_choice
     raw_stream = await client.chat.completions.create(

{verifiers-0.3.2.dev130 → verifiers-0.3.2.dev132}/verifiers/v1/tasksets/harbor/env.py

@@ -7,7 +7,9 @@ box it worked in. A separate-verifier task is graded by `finalize` instead: the
 solver's declared artifacts travel (collected after its task `finalize` while its
 box is alive), a fresh box is provisioned from the task's verifier declaration,
 `tests/` is staged there, and the verifier's rewards land on the solver's trace.
-No second agent is involved — the verifier is the task's own `tests/test.sh`.
+No second agent is involved — the verifier is the task's own `tests/test.sh`,
+staged by the task's own class, so taskset subclasses customize grading through
+their task hooks rather than a custom env.
 """
 
 import verifiers.v1 as vf
@@ -56,14 +58,16 @@ class HarborEnv(IsolatedVerifierEnv, vf.Env[HarborEnvConfig]):
         Provision a fresh box from the task's verifier declaration, restore the
         solver's collected artifacts, stage `tests/`, run the verifier, and record
         its rewards (and any extra reward.json keys as metrics) on the solver's
-        trace. Setup, restoration, staging, and scoring failures retry per
-        `verifier.retries`; the last one fails the episode."""
+        trace. The grader is the task's own class, so a `HarborTask` subclass's
+        `setup` and `stage_verifier` run in the verifier box too. Setup,
+        restoration, staging, and scoring failures retry per `verifier.retries`;
+        the last one fails the episode."""
         if not isinstance(task, HarborTask) or task.data.verifier is None:
             return
         solution = episode.traces[0]
         if not solution.ok:
             return
-        grader = HarborTask(verifier_box_data(task.data))
+        grader = type(task)(verifier_box_data(task.data), task.config)
         scores, solution = await self.grade(
             self.verifier_config(task),
             grader,