verifiers 0.1.15.dev394__tar.gz → 0.1.15.dev395__tar.gz

{verifiers-0.1.15.dev394 → verifiers-0.1.15.dev395}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.1.15.dev394
+Version: 0.1.15.dev395
 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.dev394 → verifiers-0.1.15.dev395}/verifiers/v1/dialects/base.py

@@ -151,6 +151,14 @@ class Dialect(ABC, Generic[ReqT, RespT]):
         """Whether the request asks for a streamed (SSE) response."""
         return bool(body.get("stream"))
 
+    def is_terminal_event(self, chunk: bytes) -> bool:
+        """Whether this complete SSE event ends the model's turn for the client. The
+        interception server withholds the terminal event (and anything after it) until the
+        turn is recorded, so a client that ends its turn on it can't race ahead to scoring
+        with the turn still uncommitted. Defaults to the `[DONE]` sentinel; a dialect whose
+        client ends on an earlier event (e.g. Responses' `response.completed`) overrides this."""
+        return is_sse_done_event(chunk)
+
     def error_body(self, message: str) -> dict:
         """An error payload in this format's error shape (OpenAI by default)."""
         return {"error": {"message": message, "type": "invalid_request_error"}}

{verifiers-0.1.15.dev394 → verifiers-0.1.15.dev395}/verifiers/v1/dialects/responses.py

@@ -47,6 +47,13 @@ from verifiers.v1.types import (
 )
 
 FINAL_EVENTS = ("response.completed", "response.incomplete", "response.failed")
+# Byte markers for the terminal event types above, in both compact and spaced JSON, so the
+# interception server can cheaply spot the turn-ending event without parsing each delta.
+_TERMINAL_MARKERS = tuple(
+    marker.encode()
+    for event in FINAL_EVENTS
+    for marker in (f'"type":"{event}"', f'"type": "{event}"')
+)
 # Sampling knobs the eval owns, in this format's shape (Responses uses `max_output_tokens`).
 _SAMPLING_KEYS = frozenset({"temperature", "top_p", "max_output_tokens", "max_tokens"})
 
@@ -263,6 +270,11 @@ class ResponsesDialect(Dialect[dict, OpenAIResponse]):
     upstream_path = "/responses"
     response_type = OpenAIResponse
 
+    def is_terminal_event(self, chunk: bytes) -> bool:
+        # A Responses client (e.g. codex) ends its turn on `response.completed`, before the
+        # trailing `[DONE]`, so the turn-ending event is the final event, not the sentinel.
+        return any(marker in chunk for marker in _TERMINAL_MARKERS)
+
     def parse_request(self, body: dict) -> tuple[Messages, list[Tool] | None]:
         prompt: Messages = []
         if instructions := body.get("instructions"):

{verifiers-0.1.15.dev394 → verifiers-0.1.15.dev395}/verifiers/v1/interception/server.py

@@ -472,6 +472,11 @@ class InterceptionServer:
         ready = asyncio.Event()
         producer = asyncio.create_task(_queue_chunks(reply.chunks, queue, ready))
         parser_error: Exception | None = None
+        # SSE events from the turn-ending one onward (the terminal event and any trailing
+        # `[DONE]`), withheld until the turn is committed: a client that ends its turn on the
+        # terminal event (e.g. codex on `response.completed`) would otherwise reach scoring
+        # with the turn still unrecorded.
+        deferred: list[bytes] = []
         try:
             await resp.prepare(request)
             while True:
@@ -487,11 +492,20 @@ class InterceptionServer:
                 if chunk is None:
                     await producer
                     break
+                if deferred or dialect.is_terminal_event(chunk):
+                    if parser_error is None:
+                        try:
+                            if on_done is not None and is_sse_done_event(chunk):
+                                on_done()
+                            feed_event(chunk)
+                        except Exception as e:
+                            parser_error = e
+                    # forwarded after the turn is committed, below
+                    deferred.append(chunk)
+                    continue
                 await resp.write(chunk)
                 if parser_error is None:
                     try:
-                        if on_done is not None and is_sse_done_event(chunk):
-                            on_done()
                         feed_event(chunk)
                     except Exception as e:
                         parser_error = e
@@ -511,7 +525,10 @@ class InterceptionServer:
             turn.commit(parser.finish())
             logger.debug("intercept stream turn: id=%s", session.trace.id)
         finally:
+            # Release the withheld events only now — after the commit — then close.
             with contextlib.suppress(ConnectionResetError):
+                for event in deferred:
+                    await resp.write(event)
                 await resp.write_eof()
         return resp