verifiers 0.2.1.dev32__tar.gz → 0.2.1.dev34__tar.gz

{verifiers-0.2.1.dev32 → verifiers-0.2.1.dev34}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.1.dev32
+Version: 0.2.1.dev34
 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-0.2.1.dev32 → verifiers-0.2.1.dev34}/verifiers/v1/graph.py

@@ -36,6 +36,7 @@ from verifiers.v1.types import (
     Response,
     StrictBaseModel,
     TextContentPart,
+    Tool,
     ToolMessage,
     Usage,
 )
@@ -341,8 +342,10 @@ class PendingTurn:
             for span in tail_spans
         ]
 
-    def commit(self, response: Response) -> None:
+    def commit(self, response: Response, tools: list[Tool] | None = None) -> None:
         _commit_turn(self, response)
+        if tools:
+            self.trace.tools = tools
 
 
 def prepare_turn(trace: Trace, prompt: list[Message]) -> PendingTurn:

{verifiers-0.2.1.dev32 → verifiers-0.2.1.dev34}/verifiers/v1/interception/pool.py

@@ -62,14 +62,11 @@ class StaticInterceptionPool(Interception):
 
     @asynccontextmanager
     async def acquire(self, session: RolloutSession) -> AsyncIterator[Slot]:
-        # min + register have no await between them, so concurrent acquires can't all
-        # land on the same "least-loaded" server.
+        # server.acquire registers before its first yield, so concurrent acquires see the
+        # updated load before choosing their own least-loaded server.
         server = min(self.servers, key=lambda s: s.load)
-        secret = server.register(session)
-        try:
-            yield server.base_url, secret
-        finally:
-            server.unregister(secret)
+        async with server.acquire(session) as slot:
+            yield slot
 
 
 class ElasticInterceptionPoolConfig(BaseInterceptionConfig):

{verifiers-0.2.1.dev32 → verifiers-0.2.1.dev34}/verifiers/v1/interception/server.py

@@ -335,14 +335,8 @@ class InterceptionServer(Interception):
                 session.trace.id,
                 len(response.message.tool_calls or []),
             )
-            turn.commit(response)  # one node per new message;
+            turn.commit(response, tools)  # one node per new message;
             # branches fall out of walking the graph (see Trace.branches / verifiers.v1.graph)
-            if tools:
-                # Persist the advertised tools now that the model actually saw them (see
-                # `Trace.tools`). Only a tools-bearing turn updates it (truthy, so an empty
-                # parse or a trailing tool-less call never erases the schemas the recorded
-                # tool calls need).
-                session.trace.tools = tools
             # Hand back to the program when the model wants a tool (the program runs it) or
             # when there's no user simulator to keep the conversation going.
             if response.message.tool_calls or session.user is None:
@@ -489,9 +483,7 @@ class InterceptionServer(Interception):
         try:
             if parser_error is not None:
                 raise parser_error
-            turn.commit(parser.finish())
-            if tools:  # record only on commit, like the non-streaming path
-                session.trace.tools = tools
+            turn.commit(parser.finish(), tools)
             logger.debug("intercept stream turn: id=%s", session.trace.id)
         finally:
             # Release the withheld events only now — after the commit — then close.