verifiers 0.1.15.dev375__tar.gz → 0.1.15.dev377__tar.gz

{verifiers-0.1.15.dev375 → verifiers-0.1.15.dev377}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.1.15.dev375
+Version: 0.1.15.dev377
 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-0.1.15.dev375 → verifiers-0.1.15.dev377}/tests/v1/fixtures/echo_v1.py

@@ -34,7 +34,17 @@ class EchoConfig(vf.TasksetConfig):
 class EchoTaskset(vf.Taskset[EchoTask, EchoConfig]):
     def load_tasks(self) -> list[EchoTask]:
         return [
-            EchoTask(idx=i, prompt=phrase, system_prompt=SYSTEM, answer=phrase)
+            # Keep coding-agent harnesses on the direct-response path instead of
+            # spending this single-turn smoke task on a tool call.
+            EchoTask(
+                idx=i,
+                prompt=(
+                    "Do not call tools or execute code. Reply immediately and include "
+                    f"this exact phrase in your final response: {phrase}"
+                ),
+                system_prompt=SYSTEM,
+                answer=phrase,
+            )
             for i, phrase in enumerate(self.config.phrases)
         ]
 

{verifiers-0.1.15.dev375 → verifiers-0.1.15.dev377}/verifiers/v1/serve/client.py

@@ -50,8 +50,9 @@ class EnvClient:
         self.socket.setsockopt(zmq.SNDHWM, 0)
         self.socket.setsockopt(zmq.RCVHWM, 0)
         self.socket.connect(address)
-        self._pending: dict[str, asyncio.Future] = {}
+        self._pending: dict[str, asyncio.Future[bytes]] = {}
         self._receiver: asyncio.Task | None = None
+        self._decode_slots = asyncio.BoundedSemaphore(1)
 
     def _ensure_receiver(self) -> None:
         if self._receiver is None:
@@ -65,7 +66,7 @@ class EnvClient:
                 break
             future = self._pending.pop(request_id_bytes.decode(), None)
             if future is not None and not future.done():
-                future.set_result(msgpack.unpackb(data, raw=False))
+                future.set_result(data)
 
     async def _request(
         self,
@@ -77,18 +78,38 @@ class EnvClient:
         `timeout` is only used for health polling — rollouts run untimed."""
         self._ensure_receiver()
         request_id = uuid.uuid4().hex
-        future: asyncio.Future = asyncio.get_running_loop().create_future()
+        future: asyncio.Future[bytes] = asyncio.get_running_loop().create_future()
         self._pending[request_id] = future
         payload = msgpack.packb(request.model_dump(mode="json"), use_bin_type=True)
         await self.socket.send_multipart(
             [request_id.encode(), request.method.encode(), payload]
         )
         try:
-            raw = await asyncio.wait_for(future, timeout)
+            data = await asyncio.wait_for(future, timeout)
         except (asyncio.TimeoutError, asyncio.CancelledError):
             self._pending.pop(request_id, None)
             raise
-        response = response_type.model_validate(raw)
+        if response_type in (HealthResponse, InfoResponse):
+            response = response_type.model_validate(msgpack.unpackb(data, raw=False))
+        else:
+            # Keep large trace replies compact on the loop and expand only one at a time.
+            await self._decode_slots.acquire()
+            decoding = asyncio.create_task(
+                asyncio.to_thread(
+                    lambda: response_type.model_validate(
+                        msgpack.unpackb(data, raw=False)
+                    )
+                )
+            )
+            # Hold the slot until the worker finishes so cancellation cannot overlap decodes.
+            decoding.add_done_callback(lambda _: self._decode_slots.release())
+            try:
+                response = await asyncio.shield(decoding)
+            except asyncio.CancelledError:
+                decoding.add_done_callback(
+                    lambda task: None if task.cancelled() else task.exception()
+                )
+                raise
         if not response.success:
             raise RuntimeError(response.error or "env server request failed")
         return response