verifiers 0.3.2.dev97__tar.gz → 0.3.2.dev99__tar.gz

{verifiers-0.3.2.dev97 → verifiers-0.3.2.dev99}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev97
+Version: 0.3.2.dev99
 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.dev97 → verifiers-0.3.2.dev99}/tests/v1/test_e2e.py

@@ -352,6 +352,10 @@ async def test_acp_resume_with_tool(run_v1, harness, harness_runtime, tmp_path):
     # populates trace.tools; the ACP transcript is the source of truth for use.
     assert "tool" in segments[1]["roles"]
     assert segments[1]["tool_outputs"]
+    if harness.id == "pi":
+        recall = next(tool for tool in trace.tools if tool.name == "resume_recall")
+        assert recall.parameters["properties"]["codeword"]["type"] == "string"
+        assert "codeword" in recall.parameters["required"]
     if harness.id == "rlm":
         assert "turns_since_last_compaction" in trace.metrics
         assert all(call.acp is not None for call in trace.calls)

{verifiers-0.3.2.dev97 → verifiers-0.3.2.dev99}/tests/v1/test_trace.py

@@ -33,6 +33,41 @@ class MyState(vf.State):
     score: int = 0
 
 
+@pytest.mark.parametrize(
+    "reason,condition",
+    [
+        ("max_total_tokens", "max_total_tokens"),
+        ("max_total_turns", "max_turns"),
+        ("token_budget", "max_output_tokens"),
+        ("compaction_failed", "compaction_failed"),
+        ("done", "agent_completed"),
+        (None, "agent_completed"),
+    ],
+)
+def test_rlm_stop_reason_survives_completion_and_serialization(reason, condition):
+    trace = vf.Trace(
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="q")),
+    )
+    harness = RLMHarness(RLMHarnessConfig(id="rlm"))
+    metadata = {
+        RLM_SESSION_METADATA_KEY: {
+            "session_id": trace.id,
+            "metrics": {"turns": 1},
+            "last_stop_reason": reason,
+        }
+    }
+    harness.acp_turn_result(
+        trace, vf.ACPTurn(reply="partial", response_metadata=metadata)
+    )
+    trace.stop("agent_completed")
+    harness.acp_close_result(trace, metadata)
+
+    restored = vf.WireTrace.model_validate_json(trace.model_dump_json())
+    assert restored.stop_condition == condition
+    assert restored.is_truncated is (condition != "agent_completed")
+
+
 class FailingSegmentRollout:
     ok = Rollout.ok
     closed = Rollout.closed

{verifiers-0.3.2.dev97 → verifiers-0.3.2.dev99}/verifiers/v1/harnesses/pi/harness.py

@@ -137,7 +137,7 @@ class PiHarness(ACPHarness[PiHarnessConfig]):
             extension_path = f"{agent_dir}/mcp.js"
             mcp = {
                 "mcpServers": {
-                    name: {"url": url, "lifecycle": "eager"}
+                    name: {"url": url, "lifecycle": "eager", "directTools": True}
                     for name, url in mcp_urls.items()
                 }
             }

{verifiers-0.3.2.dev97 → verifiers-0.3.2.dev99}/verifiers/v1/harnesses/rlm/harness.py

@@ -41,6 +41,15 @@ class _SessionSnapshot(BaseModel):
 
     session_id: str = Field(pattern=r"^[A-Za-z0-9._:-]{1,128}$")
     metrics: dict[str, int | float]
+    last_stop_reason: str | None = None
+
+
+_TRUNCATION_STOPS = {
+    "max_total_tokens": "max_total_tokens",
+    "max_total_turns": "max_turns",
+    "token_budget": "max_output_tokens",
+    "compaction_failed": "compaction_failed",
+}
 
 
 class CompactionConfig(BaseConfig):
@@ -251,6 +260,8 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
         if snapshot.session_id != trace.id:
             raise ValueError("RLM session snapshot does not match the rollout")
         trace.record_metrics(snapshot.metrics)
+        if condition := _TRUNCATION_STOPS.get(snapshot.last_stop_reason or ""):
+            trace.stop(condition)
 
     def acp_turn_result(self, trace: Trace, result: ACPTurn) -> None:
         self._consume_snapshot(trace, result.response_metadata)

{verifiers-0.3.2.dev97 → verifiers-0.3.2.dev99}/verifiers/v1/trace.py

@@ -621,12 +621,13 @@ class Trace(BaseModel, Generic[DataT, StateT, AgentConfigT]):
 
     @property
     def is_truncated(self) -> bool:
-        """True for framework limits or a length-finished final response."""
+        """True for framework limits, failed compaction, or a length-finished response."""
         if self.stop_condition in (
             "max_turns",
             "max_input_tokens",
             "max_output_tokens",
             "max_total_tokens",
+            "compaction_failed",
         ):
             return True
         last = next((c for c in reversed(self.calls) if c.error is None), None)