verifiers 0.2.2.dev59__tar.gz → 0.2.2.dev60__tar.gz

{verifiers-0.2.2.dev59 → verifiers-0.2.2.dev60}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev59
+Version: 0.2.2.dev60
 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-0.2.2.dev59 → verifiers-0.2.2.dev60}/verifiers/v1/agent.py

@@ -25,7 +25,7 @@ from verifiers.v1.harness import Harness
 from verifiers.v1.interception import Interception, InterceptionServer
 from verifiers.v1.mcp import SharedToolServer
 from verifiers.v1.retries import backoff, trace_should_retry
-from verifiers.v1.rollout import RolloutRun, _as_messages
+from verifiers.v1.rollout import RolloutRun, RolloutTimeouts, _as_messages
 from verifiers.v1.runtimes import (
     NetworkPolicyConfig,
     Runtime,
@@ -533,23 +533,23 @@ class Agent:
             "harness": self.harness,
             "ctx": self.ctx,
             "runtime_config": runtime_config,
-            "setup_timeout": (
-                self.timeout.setup
-                if self.timeout.setup is not None
-                else task.data.timeout.setup
-            ),
-            "agent_timeout": cap_remote_agent_timeout(
-                agent_timeout, runtime_config, task
-            ),
-            "finalize_timeout": (
-                self.timeout.finalize
-                if self.timeout.finalize is not None
-                else task.data.timeout.finalize
-            ),
-            "scoring_timeout": (
-                self.timeout.scoring
-                if self.timeout.scoring is not None
-                else task.data.timeout.scoring
+            "timeouts": RolloutTimeouts(
+                setup=(
+                    self.timeout.setup
+                    if self.timeout.setup is not None
+                    else task.data.timeout.setup
+                ),
+                agent=cap_remote_agent_timeout(agent_timeout, runtime_config, task),
+                finalize=(
+                    self.timeout.finalize
+                    if self.timeout.finalize is not None
+                    else task.data.timeout.finalize
+                ),
+                scoring=(
+                    self.timeout.scoring
+                    if self.timeout.scoring is not None
+                    else task.data.timeout.scoring
+                ),
             ),
             "limits": self.limits,
             "shared_tools": shared_tools,

{verifiers-0.2.2.dev59 → verifiers-0.2.2.dev60}/verifiers/v1/rollout.py

@@ -23,6 +23,7 @@ import logging
 import time
 from collections.abc import AsyncIterator, Callable
 from contextlib import AsyncExitStack, asynccontextmanager
+from dataclasses import dataclass
 
 from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.agent import AgentConfig
@@ -57,6 +58,19 @@ from verifiers.v1.types import Messages
 logger = logging.getLogger(__name__)
 
 
+@dataclass(frozen=True)
+class RolloutTimeouts:
+    """Per-stage rollout timeouts in seconds (None = unlimited), each bounding one
+    lifecycle stage: `setup` covers task + harness setup in `open()`, `agent` is the
+    cumulative budget the run's segments draw down across `step()`s, `finalize` and
+    `scoring` bound their `close()` stages."""
+
+    setup: float | None = None
+    agent: float | None = None
+    finalize: float | None = None
+    scoring: float | None = None
+
+
 def _as_messages(raw: Messages) -> Messages:
     """A turn's messages may arrive typed or as wire dicts (env code naturally
     writes `{"role": "user", ...}`); the trace speaks typed, so normalize here."""
@@ -115,10 +129,7 @@ class RolloutRun:
         runtime_config: RuntimeConfig,
         wire_data: TaskData | None = None,
         has_user: bool = False,
-        setup_timeout: float | None = None,
-        agent_timeout: float | None = None,
-        finalize_timeout: float | None = None,
-        scoring_timeout: float | None = None,
+        timeouts: RolloutTimeouts | None = None,
         limits: RolloutLimits | None = None,
         shared_tools: dict[str, SharedToolServer] | None = None,
         interception: Interception | None = None,
@@ -130,11 +141,8 @@ class RolloutRun:
         self.ctx = ctx
         self.runtime_config = runtime_config
         self._has_user = has_user
-        self._setup_timeout = setup_timeout
-        self._agent_timeout = agent_timeout
-        self._agent_time_remaining = agent_timeout
-        self._finalize_timeout = finalize_timeout
-        self._scoring_timeout = scoring_timeout
+        self._timeouts = timeouts or RolloutTimeouts()
+        self._agent_time_remaining = self._timeouts.agent
         self._shared_tools = shared_tools or {}
         self._interception = interception
         self.runtime = runtime
@@ -164,7 +172,7 @@ class RolloutRun:
         self.deadline_at: float | None = None
         """The active harness segment's absolute deadline (event-loop clock), or
         None between segments / when unbounded. An interaction spends one cumulative
-        `agent_timeout` budget only while its own segments run, so time awaiting
+        `timeouts.agent` budget only while its own segments run, so time awaiting
         the caller (including another interleaved agent) cannot starve it."""
 
     @property
@@ -244,8 +252,8 @@ class RolloutRun:
             # Task setup and harness provisioning share one setup-stage deadline.
             setup_deadline = (
                 None
-                if self._setup_timeout is None
-                else asyncio.get_running_loop().time() + self._setup_timeout
+                if self._timeouts.setup is None
+                else asyncio.get_running_loop().time() + self._timeouts.setup
             )
             async with (
                 boundary(TaskError, "task setup"),
@@ -340,7 +348,7 @@ class RolloutRun:
                 self.fail(
                     HarnessError(
                         f"agent timeout: rollout exceeded its "
-                        f"{self._agent_timeout:g}s budget"
+                        f"{self._timeouts.agent:g}s budget"
                     )
                 )
             else:
@@ -406,7 +414,7 @@ class RolloutRun:
                         invoke(
                             self.task.finalize, {"trace": trace, "runtime": runtime}
                         ),
-                        self._finalize_timeout,
+                        self._timeouts.finalize,
                     )
                 now = time.time()
                 trace.timing.finalize.end = now
@@ -418,7 +426,7 @@ class RolloutRun:
                             self.task.score(trace, runtime),
                             self.harness.score(trace, runtime),
                         ),
-                        self._scoring_timeout,
+                        self._timeouts.scoring,
                     )
                 trace.timing.scoring.end = time.time()
         except Exception as e:  # noqa: BLE001 - finalize boundary records every rollout failure