verifiers 0.3.2.dev85__tar.gz → 0.3.2.dev87__tar.gz

{verifiers-0.3.2.dev85 → verifiers-0.3.2.dev87}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev85
+Version: 0.3.2.dev87
 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.dev85 → verifiers-0.3.2.dev87}/verifiers/v1/dialects/chat.py

@@ -549,14 +549,19 @@ class ChatDialect(Dialect[ChatCompletion]):
 
     def stream_events(self, raw: dict) -> list[bytes]:
         choice = (raw.get("choices") or [{}])[0]
+        message = dict(choice.get("message") or {"role": "assistant", "content": ""})
+        if message.get("tool_calls"):
+            message["tool_calls"] = [
+                {**call, "index": index}
+                for index, call in enumerate(message["tool_calls"])
+            ]
         chunk = {
             **{key: value for key, value in raw.items() if key != "choices"},
             "object": "chat.completion.chunk",
             "choices": [
                 {
                     "index": 0,
-                    "delta": choice.get("message")
-                    or {"role": "assistant", "content": ""},
+                    "delta": message,
                     "finish_reason": choice.get("finish_reason"),
                 }
             ],

{verifiers-0.3.2.dev85 → verifiers-0.3.2.dev87}/verifiers/v1/harnesses/utils/core.py

@@ -9,7 +9,8 @@ from pathlib import Path
 from typing import TYPE_CHECKING
 
 import httpx
-from openai import APIStatusError, AsyncOpenAI
+from openai import APIStatusError, AsyncOpenAI, omit
+from openai.lib.streaming.chat import AsyncChatCompletionStream
 
 if TYPE_CHECKING:
     # The harness bundles this module into the generated script before execution.
@@ -191,7 +192,29 @@ async def chat(
     kwargs = {"model": model, "messages": messages, "tools": tools or None}
     if tools and tool_choice is not None:
         kwargs["tool_choice"] = tool_choice
-    return await client.chat.completions.create(**kwargs)
+    raw_stream = await client.chat.completions.create(
+        **kwargs, stream=True, stream_options={"include_usage": True}
+    )
+    # Accumulate native deltas without auto-parsing tool arguments or treating
+    # finish_reason="length" as an exception: compaction owns that decision.
+    async with AsyncChatCompletionStream(
+        raw_stream=raw_stream, response_format=omit, input_tools=[]
+    ) as response:
+        completion = None
+        async for event in response:
+            if event.type == "chunk":
+                completion = event.snapshot
+        if (
+            completion is None
+            or not completion.choices
+            or any(choice.finish_reason is None for choice in completion.choices)
+        ):
+            raise RuntimeError("model stream ended before a completion finished")
+        for choice in completion.choices:
+            # Some providers repeat the role in each delta. The SDK concatenates
+            # these strings, but the role is metadata, not incremental content.
+            choice.message.role = "assistant"
+        return completion
 
 
 async def run_tool_hook(

{verifiers-0.3.2.dev85 → verifiers-0.3.2.dev87}/verifiers/v1/interception/server.py

@@ -41,7 +41,11 @@ from pydantic_core import PydanticSerializationError, from_json, to_json
 from verifiers.v1 import graph
 from verifiers.v1.clients import Client, resolve_client
 from verifiers.v1.clients.base import join_url
-from verifiers.v1.configs.client import BaseClientConfig, resolve_api_key
+from verifiers.v1.configs.client import (
+    BaseClientConfig,
+    TrainClientConfig,
+    resolve_api_key,
+)
 from verifiers.v1.dialects import DIALECTS, Dialect
 from verifiers.v1.dialects.base import (
     PROVIDER_CAPABILITY_POLICY_CODE,
@@ -127,15 +131,18 @@ def _capture_response(response: web.Response) -> ReplayResponse:
         data = bytes(body)
     else:
         raise TypeError("coalesced interception responses must have a byte body")
-    return ReplayResponse(status=response.status, body=data)
+    return ReplayResponse(
+        status=response.status,
+        body=data,
+        content_type=response.headers["Content-Type"],
+    )
 
 
 def _replay_response(response: ReplayResponse) -> web.Response:
     return web.Response(
         body=response.body,
         status=response.status,
-        content_type="application/json",
-        charset="utf-8",
+        headers={"Content-Type": response.content_type},
     )
 
 
@@ -492,7 +499,10 @@ class InterceptionServer(Interception):
             body = json.loads(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""
+        relay_streaming = streaming and not isinstance(
+            session.ctx.client, TrainClientConfig
+        )
+        req_hash = await _request_digest(raw) if not relay_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
@@ -507,7 +517,7 @@ class InterceptionServer(Interception):
             session.trace.id,
             streaming,
         )
-        # Graph atomicity under retries: one logical non-streaming call must commit at most
+        # Graph atomicity under retries: one logical buffered call must commit at most
         # one turn. An explicit key identifies that call directly; otherwise only the SDK's
         # retry marker activates body-digest replay, since an unmarked repeated body can be a
         # legitimate later turn.
@@ -517,14 +527,14 @@ class InterceptionServer(Interception):
         replay_key: str | None = None
         binding = (request.path, req_hash)
         if idempotency_key:
-            if streaming and acp is None:
+            if relay_streaming and acp is None:
                 return web.json_response(
                     dialect.error_body(
                         "Idempotency-Key is not supported for streaming requests"
                     ),
                     status=400,
                 )
-            if not streaming:
+            if not relay_streaming:
                 replay_key = f"explicit:{idempotency_key}"
             # This key identifies the harness-to-interception hop. The server owns its
             # replay semantics, and the body has since been rewritten with rollout model
@@ -534,7 +544,7 @@ class InterceptionServer(Interception):
                 for name, value in upstream_headers.items()
                 if name.lower() != IDEMPOTENCY_KEY_HEADER.lower()
             }
-        elif not streaming:
+        elif not relay_streaming:
             replay_key = f"retry:{request.path}:{req_hash.hex()}"
 
         if replay_key is not None:
@@ -648,7 +658,7 @@ class InterceptionServer(Interception):
             return self._fail(session, dialect, error)
 
         inspect_response = bool(session.response_interceptors or session.response_stops)
-        if streaming:
+        if relay_streaming:
             return await self._stream(
                 request,
                 session,
@@ -663,7 +673,15 @@ class InterceptionServer(Interception):
             )
 
         def serve(response: Response) -> web.Response:
-            served = _completion_response(response.raw)
+            if streaming:
+                # Training generates a complete response with token metadata.
+                # Commit it through the normal path, then frame it for SSE clients.
+                served = web.Response(
+                    body=b"".join(dialect.stream_events(response.raw or {})),
+                    content_type="text/event-stream",
+                )
+            else:
+                served = _completion_response(response.raw)
             if idempotent is not None:
                 idempotent.response = _capture_response(served)
                 idempotent.completed_at = time.monotonic()

{verifiers-0.3.2.dev85 → verifiers-0.3.2.dev87}/verifiers/v1/session.py

@@ -96,7 +96,7 @@ class RolloutLimits:
 
 @dataclass
 class IdempotentRequest:
-    """One non-streaming model request shared by its original call and retries."""
+    """One buffered model request shared by its original call and retries."""
 
     binding: tuple[str, bytes]
     response: "ReplayResponse | None" = None
@@ -110,6 +110,7 @@ class ReplayResponse:
 
     status: int
     body: bytes
+    content_type: str
 
 
 @dataclass