verifiers 0.3.2.dev57__tar.gz → 0.3.2.dev59__tar.gz

{verifiers-0.3.2.dev57 → verifiers-0.3.2.dev59}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev57
+Version: 0.3.2.dev59
 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.dev57 → verifiers-0.3.2.dev59}/verifiers/v1/dialects/chat.py

@@ -15,6 +15,7 @@ from typing import Any
 from urllib.parse import urlsplit
 
 from openai.types.chat import ChatCompletion
+from openai.types.chat.chat_completion import Choice
 
 from verifiers.v1.configs.runtime import NetworkPolicyConfig
 from verifiers.v1.dialects.base import (
@@ -44,12 +45,18 @@ from verifiers.v1.types import (
 )
 
 
+class ModdedChoice(Choice):
+    # Logprobs are relayed in Response.raw; trace conversion never reads them.
+    logprobs: Any = None
+
+
 class ModdedChatCompletion(ChatCompletion):
     """The OpenAI SDK closes `service_tier` to a fixed `Literal`, but providers return tiers
     outside it (e.g. Prime's `provisioned`), which makes `model_validate` reject an otherwise
     valid completion. Widen the field to a plain string — we don't consume it — so parsing stays
     lenient about the label instead of dropping it."""
 
+    choices: list[ModdedChoice]
     service_tier: str | None = None
 
 

{verifiers-0.3.2.dev57 → verifiers-0.3.2.dev59}/verifiers/v1/dialects/responses.py

@@ -315,7 +315,10 @@ def fold_assistant(items: list[dict] | None) -> AssistantMessage:
 def response_from_wire(response: OpenAIResponse) -> Response:
     """An OpenAI Responses object -> a vf `Response` (its `output` items folded into one
     assistant message)."""
-    data = response.model_dump()
+    # Copy the output snapshot without traversing echoed tools or request settings.
+    data = response.model_dump(
+        include={"id", "created_at", "model", "status", "error", "output"}
+    )
     status = data.get("status")
     if status not in (None, "completed", "incomplete"):
         error = data.get("error") or {}

{verifiers-0.3.2.dev57 → verifiers-0.3.2.dev59}/verifiers/v1/interception/server.py

@@ -490,13 +490,13 @@ class InterceptionServer(Interception):
             body = from_json(raw)
         except ValueError:
             body = json.loads(raw)
-        req_hash = await _request_digest(raw)
+        body = dialect.apply_overrides(body, session.ctx.model, session.ctx.sampling)
+        streaming = dialect.streaming(body)
+        req_hash = await _request_digest(raw) if not streaming else b""
         # Keep `read()` for aiohttp's size guard, then release its cache and our local
         # alias after parsing so the wire body does not survive model inference.
         request._read_bytes = None
         del raw
-        body = dialect.apply_overrides(body, session.ctx.model, session.ctx.sampling)
-        streaming = dialect.streaming(body)
         try:
             acp, upstream_headers = extract_acp_info(request.headers)
         except ValueError as error:
@@ -638,7 +638,9 @@ class InterceptionServer(Interception):
 
         try:
             body, policy_paths = self.mediate_capabilities(session, dialect, body)
-            model_request = dialect.parse_request(body)
+            # Restricted mediation can mutate the body without reporting policy paths.
+            if request_rewrites or session.network_policy.network_restricted:
+                model_request = dialect.parse_request(body)
             turn = graph.prepare_turn(session.trace, model_request.messages)
         except ValueError as error:
             return web.json_response(dialect.error_body(str(error)), status=400)

{verifiers-0.3.2.dev57 → verifiers-0.3.2.dev59}/verifiers/v1/session.py

@@ -158,22 +158,24 @@ class RolloutSession:
         """Run typed request interceptors and stops over one canonical request."""
         if not self.request_interceptors and (not run_stops or not self.request_stops):
             return request, [], None
-        turn = graph.prepare_turn(self.trace, request.messages)
-        prepared_users = self.prepared_users.copy()
         prepared: set[int] = set()
         candidates: set[int] = set()
-        for position in range(turn.tail_start, len(request.messages)):
-            message = request.messages[position]
-            if isinstance(message, UserMessage):
-                candidates.add(position)
-                key = graph.message_hash(message)
-                if prepared_users[key]:
-                    prepared_users[key] -= 1
-                    prepared.add(position)
-            elif isinstance(message, ToolMessage):
-                candidates.add(position)
-                if self.prepared_tool_results.get(message.tool_call_id) == message:
-                    prepared.add(position)
+        if self.request_interceptors:
+            turn = graph.prepare_turn(self.trace, request.messages)
+            prepared_users = self.prepared_users.copy()
+            for position in range(turn.tail_start, len(request.messages)):
+                message = request.messages[position]
+                if isinstance(message, UserMessage):
+                    candidates.add(position)
+                    if prepared_users:
+                        key = graph.message_hash(message)
+                        if prepared_users[key]:
+                            prepared_users[key] -= 1
+                            prepared.add(position)
+                elif isinstance(message, ToolMessage):
+                    candidates.add(position)
+                    if self.prepared_tool_results.get(message.tool_call_id) == message:
+                        prepared.add(position)
         already_intercepted = candidates and candidates == prepared
 
         current = request
@@ -240,7 +242,7 @@ class RolloutSession:
     def consume_prepared(self, messages: Messages) -> None:
         """Forget pre-harness rewrites only after their model request commits."""
         for message in messages:
-            if isinstance(message, UserMessage):
+            if isinstance(message, UserMessage) and self.prepared_users:
                 key = graph.message_hash(message)
                 if self.prepared_users[key]:
                     self.prepared_users[key] -= 1