verifiers 0.3.2.dev152__tar.gz → 0.3.2.dev153__tar.gz

{verifiers-0.3.2.dev152 → verifiers-0.3.2.dev153}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev152
+Version: 0.3.2.dev153
 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.dev152 → verifiers-0.3.2.dev153}/tests/v1/conftest.py

@@ -155,7 +155,10 @@ def _eval_config(
         runtime_cfg = dict(runtime)
         _configure_prime_runtimes(runtime_cfg)
         env_cfg.setdefault("agent", {})["runtime"] = runtime_cfg
-    retries = {"max_retries": 2, "include": ["ProviderError", "HarnessError"]}
+    retries = {
+        "max_retries": 2,
+        "include": ["ProviderError", "InterceptionError", "HarnessError"],
+    }
     env_cfg.setdefault("retries", retries)
     # Per-run caps live on the seats: resolve the env's declared roles and cap
     # each one (a test's own seat dict wins over the shared defaults).

{verifiers-0.3.2.dev152 → verifiers-0.3.2.dev153}/verifiers/v1/acp/__init__.py

@@ -13,7 +13,7 @@ from pydantic import BaseModel, ConfigDict, Field
 
 from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.harness import HarnessConfig
-from verifiers.v1.errors import HarnessError
+from verifiers.v1.errors import HarnessError, InterceptionError
 from verifiers.v1.harness import Harness, HarnessSession
 from verifiers.v1.runtimes import ProgramResult, Runtime, RuntimeProcess
 from verifiers.v1.semantic import (
@@ -308,6 +308,12 @@ class ACPHarnessSession(HarnessSession):
             detail = response.get("error") or "ACP session request failed"
             if stderr := self._stderr():
                 detail = f"{detail}\n\nACP process stderr:\n{stderr}"
+            error_data = response.get("error_data")
+            if (
+                isinstance(error_data, dict)
+                and error_data.get("kind") == "model_transport"
+            ):
+                raise InterceptionError(detail)
             raise RuntimeError(detail)
         harness = cast(ACPHarness, self.harness)
         harness._consume_protocol_metadata(self.trace, turn.response_metadata)

{verifiers-0.3.2.dev152 → verifiers-0.3.2.dev153}/verifiers/v1/acp/runner.py

@@ -224,9 +224,6 @@ async def prompt(
         if client.stop_reason != "cancelled":
             raise
         await connection.cancel(session_id=session_id)
-    except RequestError as error:
-        detail = error.data.get("details") if isinstance(error.data, dict) else None
-        raise RuntimeError(detail or str(error)) from error
     finally:
         client.prompt_task = None
     return client.turn_result()
@@ -377,6 +374,9 @@ async def serve_stream() -> None:
                     "ok": False,
                     "error": f"{type(error).__name__}: {error}",
                 }
+                if isinstance(error, RequestError) and isinstance(error.data, dict):
+                    response["error"] = error.data.get("details") or str(error)
+                    response["error_data"] = error.data
                 if operation == "prompt":
                     response["result"] = asdict(session.client.turn_result())
             write_packet(sys.stdout.buffer, response)

{verifiers-0.3.2.dev152 → verifiers-0.3.2.dev153}/verifiers/v1/errors.py

@@ -42,6 +42,11 @@ class ProviderError(RolloutError):
         self.status_code = status_code
 
 
+MODEL_TRANSPORT_ERROR_EXIT_CODE = 97
+"""Reserved harness exit code for an exhausted model transport failure.
+`Harness._check_result` maps it to `InterceptionError`."""
+
+
 class HarnessError(RolloutError):
     """The harness failed to install or launch, or its agent process exited unsuccessfully."""
 
@@ -65,7 +70,7 @@ class TaskError(RolloutError):
 
 
 class InterceptionError(RolloutError):
-    """The host interception server (model calls + `/state` + `/task` channels) couldn't be reached."""
+    """Communication with the host interception server failed."""
 
 
 class TunnelError(InterceptionError):

{verifiers-0.3.2.dev152 → verifiers-0.3.2.dev153}/verifiers/v1/harness.py

@@ -9,7 +9,13 @@ from typing import TYPE_CHECKING, ClassVar, Generic, TypeVar
 
 from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.harness import HarnessConfig, skill_destination
-from verifiers.v1.errors import HarnessError, SandboxError, boundary
+from verifiers.v1.errors import (
+    MODEL_TRANSPORT_ERROR_EXIT_CODE,
+    HarnessError,
+    InterceptionError,
+    SandboxError,
+    boundary,
+)
 from verifiers.v1.runtimes import ProgramResult, Runtime
 from verifiers.v1.task import TaskData
 from verifiers.v1.types import Messages
@@ -178,6 +184,10 @@ copy_failed=$(
                 f"runtime died under harness {self.config.id!r} "
                 f"(exit {result.exit_code}): {detail}"
             )
+        if result.exit_code == MODEL_TRANSPORT_ERROR_EXIT_CODE:
+            raise InterceptionError(
+                f"harness {self.config.id!r} model connection to interception failed: {detail}"
+            )
         raise HarnessError(
             f"harness {self.config.id!r} exited {result.exit_code}: {detail}"
         )

{verifiers-0.3.2.dev152 → verifiers-0.3.2.dev153}/verifiers/v1/harnesses/rlm/harness.py

@@ -69,7 +69,7 @@ class CompactionConfig(BaseConfig):
 
 class RLMHarnessConfig(HarnessConfig):
     version: str = Field(
-        default="ec0335baa434eb726146e330be391ee92bf3a0c8", min_length=1
+        default="38c44a06fdef484a19d8d6c4e5697100971f9ecd", min_length=1
     )
     """Git ref (branch, tag, or commit) of nano-rlm to install. Must know every
     field this harness puts on the wire, i.e. be at least the default ref."""

{verifiers-0.3.2.dev152 → verifiers-0.3.2.dev153}/verifiers/v1/harnesses/utils/core.py

@@ -5,6 +5,7 @@ import asyncio
 import json
 import logging
 import subprocess
+import traceback
 from contextlib import AsyncExitStack
 from pathlib import Path
 from typing import TYPE_CHECKING
@@ -39,6 +40,10 @@ if TYPE_CHECKING:
     )
     from verifiers.v1.harnesses.utils.mcp import call_mcp, connect_mcp  # noqa: TC004
 
+# Mirror of verifiers.v1.errors.MODEL_TRANSPORT_ERROR_EXIT_CODE;
+# this bundled script cannot import verifiers.
+MODEL_TRANSPORT_ERROR_EXIT_CODE = 97
+
 SERPER_URL = "https://google.serper.dev/search"
 
 BASH_TOOL = {
@@ -260,25 +265,32 @@ async def chat(
         kwargs["tools"] = tools
     if tools and tool_choice is not None:
         kwargs["tool_choice"] = tool_choice
-    async for attempt in AsyncRetrying(
-        retry=retry_if_exception_type((APIConnectionError, httpx.TransportError)),
-        stop=stop_after_attempt(client.max_retries + 1),
-        wait=wait_random_exponential(multiplier=0.5, max=8.0),
-        before_sleep=before_sleep_log(logging.getLogger(__name__), logging.WARNING),
-        reraise=True,
-    ):
-        # Reuse the interception server's body-digest replay guard on stream retries.
-        retry_count = attempt.retry_state.attempt_number - 1
-        headers = {"x-stainless-retry-count": str(retry_count)} if retry_count else omit
-        raw_stream = await client.chat.completions.create(
-            **kwargs,
-            stream=True,
-            stream_options={"include_usage": True},
-            extra_headers=headers,
-        )
-        # The SDK retries request setup; only stream consumption is retried here.
-        with attempt:
-            return await _read_chat_completion(raw_stream)
+    try:
+        async for attempt in AsyncRetrying(
+            retry=retry_if_exception_type((APIConnectionError, httpx.TransportError)),
+            stop=stop_after_attempt(client.max_retries + 1),
+            wait=wait_random_exponential(multiplier=0.5, max=8.0),
+            before_sleep=before_sleep_log(logging.getLogger(__name__), logging.WARNING),
+            reraise=True,
+        ):
+            # Reuse the interception server's body-digest replay guard on stream retries.
+            retry_count = attempt.retry_state.attempt_number - 1
+            headers = (
+                {"x-stainless-retry-count": str(retry_count)} if retry_count else omit
+            )
+            raw_stream = await client.chat.completions.create(
+                **kwargs,
+                stream=True,
+                stream_options={"include_usage": True},
+                extra_headers=headers,
+            )
+            # The SDK retries request setup; only stream consumption is retried here.
+            with attempt:
+                return await _read_chat_completion(raw_stream)
+    except (APIConnectionError, httpx.TransportError) as error:
+        # Preserve the original transport error in stderr for the host's diagnostic.
+        traceback.print_exc()
+        raise SystemExit(MODEL_TRANSPORT_ERROR_EXIT_CODE) from error
 
 
 async def _read_chat_completion(raw_stream):