verifiers 0.2.2.dev31__tar.gz → 0.2.2.dev33__tar.gz

{verifiers-0.2.2.dev31 → verifiers-0.2.2.dev33}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev31
+Version: 0.2.2.dev33
 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.dev31 → verifiers-0.2.2.dev33}/verifiers/v1/__init__.py

@@ -37,7 +37,7 @@ from verifiers.v1.errors import (
 )
 from verifiers.v1.configs.harness import HarnessConfig
 from verifiers.v1.harness import Harness
-from verifiers.v1.configs.judge import JudgeConfig, JudgeSamplingConfig, Judges
+from verifiers.v1.configs.judge import JudgeConfig, Judges
 from verifiers.v1.judge import Judge, JudgeResponse, JudgeView
 from verifiers.v1.judges import (
     ReferenceJudge,
@@ -267,7 +267,6 @@ __all__ = [
     "Judge",
     "JudgeConfig",
     "Judges",
-    "JudgeSamplingConfig",
     "JudgeResponse",
     "JudgeView",
     "ReferenceJudge",

{verifiers-0.2.2.dev31 → verifiers-0.2.2.dev33}/verifiers/v1/configs/judge.py

@@ -12,10 +12,6 @@ from verifiers.v1.types import ID, SamplingConfig
 from verifiers.v1.utils.install import env_name
 
 
-class JudgeSamplingConfig(SamplingConfig):
-    pass
-
-
 class JudgeConfig(BaseClientConfig):
     id: ID = ""
     """Plugin id; empty for a judge called directly by task code."""
@@ -23,7 +19,7 @@ class JudgeConfig(BaseClientConfig):
     """Reward key override for a plugged judge."""
     weight: float = 1.0
     model: str = "openai/gpt-5.4-nano"
-    sampling: JudgeSamplingConfig = JudgeSamplingConfig()
+    sampling: SamplingConfig = SamplingConfig()
     prompt: str | None = None
     prompt_file: Path | None = None
     """Prompt file override, mutually exclusive with `prompt`."""

{verifiers-0.2.2.dev31 → verifiers-0.2.2.dev33}/verifiers/v1/dialects/responses.py

@@ -1,9 +1,9 @@
 """The OpenAI Responses dialect (codex and friends).
 
 Request parsing walks the `input` items, folding each run of assistant-side items (reasoning /
-assistant message / function_call) into one typed assistant message; response parsing reads the
-`output` items. Relay-only: the eval client forwards the program's bytes to a `/responses`
-endpoint and this dialect parses a copy for the trace. Server-side statefulness
+assistant message / function or custom tool call) into one typed assistant message; response
+parsing reads the `output` items. Relay-only: the eval client forwards the program's bytes to a
+`/responses` endpoint and this dialect parses a copy for the trace. Server-side statefulness
 (`previous_response_id`) is not emulated — the endpoint owns it.
 """
 
@@ -97,8 +97,7 @@ def parse_content(content) -> str | list[ContentPart]:
 
 
 def fold_assistant(items: list[dict]) -> AssistantMessage:
-    """One run of assistant-side items (reasoning / message / function_call) -> one typed
-    assistant message."""
+    """One run of assistant-side items -> one typed assistant message."""
     content = ""
     reasoning: list[str] = []
     calls: list[ToolCall] = []
@@ -106,12 +105,12 @@ def fold_assistant(items: list[dict]) -> AssistantMessage:
         if item.get("type") == "reasoning":
             reasoning += [s.get("text", "") for s in item.get("summary") or []]
             reasoning += [c.get("text", "") for c in item.get("content") or []]
-        elif item.get("type") == "function_call":
+        elif item.get("type") in ("function_call", "custom_tool_call"):
             calls.append(
                 ToolCall(
                     id=item.get("call_id", ""),
                     name=item.get("name", ""),
-                    arguments=item.get("arguments", ""),
+                    arguments=item.get("arguments", item.get("input", "")),
                 )
             )
         else:  # an assistant message item
@@ -151,12 +150,12 @@ def response_from_wire(response: OpenAIResponse) -> Response:
         elif kind == "reasoning":
             reasoning += [s.get("text", "") for s in item.get("summary") or []]
             reasoning += [c.get("text", "") for c in item.get("content") or []]
-        elif kind == "function_call":
+        elif kind in ("function_call", "custom_tool_call"):
             calls.append(
                 ToolCall(
                     id=item.get("call_id", ""),
                     name=item.get("name", ""),
-                    arguments=item.get("arguments", ""),
+                    arguments=item.get("arguments", item.get("input", "")),
                 )
             )
     tool_calls = calls or None
@@ -279,7 +278,10 @@ class ResponsesDialect(Dialect[dict, OpenAIResponse]):
                 run = []
             if assistant:
                 run.append(item)
-            elif item.get("type") == "function_call_output":
+            elif item.get("type") in (
+                "function_call_output",
+                "custom_tool_call_output",
+            ):
                 output = item.get("output")
                 content = (
                     parse_content(output)

{verifiers-0.2.2.dev31 → verifiers-0.2.2.dev33}/verifiers/v1/gepa/adapter.py

@@ -87,7 +87,7 @@ class GEPAAdapter:
 
     def make_reflective_dataset(
         self,
-        candidate: Candidate,  # noqa: ARG002 - required by GEPA's adapter protocol
+        candidate: Candidate,  # Required by GEPA's adapter protocol.
         eval_batch: EvaluationBatch[Episode, Episode],
         components_to_update: list[str],
     ) -> Mapping[str, Sequence[Mapping[str, Any]]]:

{verifiers-0.2.2.dev31 → verifiers-0.2.2.dev33}/verifiers/v1/utils/git.py

@@ -97,7 +97,7 @@ async def capture_patch(
             )
             return
         raw = await runtime.read(capped)
-    except Exception as exc:  # noqa: BLE001 - capture must never fail the rollout.
+    except Exception as exc:  # Capture must never fail the rollout.
         trace.info["patch_error"] = f"{type(exc).__name__}: {exc}"
         return
     finally:
@@ -105,7 +105,7 @@ async def capture_patch(
         # on shared-filesystem runtimes; removal is best-effort by design.
         try:
             await runtime.run(["rm", "-f", full, capped], env or {})
-        except Exception:  # noqa: BLE001,S110 - cleanup must never fail the rollout.
+        except Exception:  # Cleanup must never fail the rollout.
             pass
     if len(raw) > PATCH_CAP_BYTES:
         raw = raw[:PATCH_CAP_BYTES]