verifiers 0.2.2.dev96__tar.gz → 0.2.2.dev97__tar.gz

{verifiers-0.2.2.dev96 → verifiers-0.2.2.dev97}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev96
+Version: 0.2.2.dev97
 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.dev96 → verifiers-0.2.2.dev97}/tests/v1/conftest.py

@@ -164,7 +164,10 @@ def _eval_config(
         seat_cfg.setdefault("max_output_tokens", max_tokens)
         seat_cfg.setdefault("timeout", {"rollout": rollout_timeout, "scoring": 60})
         # Flake resilience: retries are per-agent now (flat RetryConfig).
-        seat_cfg.setdefault("retries", {"max_retries": 2, "include": ["ProviderError"]})
+        seat_cfg.setdefault(
+            "retries",
+            {"max_retries": 2, "include": ["ProviderError", "HarnessError"]},
+        )
     return EvalConfig(
         env={
             "taskset": taskset_cfg,

{verifiers-0.2.2.dev96 → verifiers-0.2.2.dev97}/tests/v1/fixtures/echo_tool_v1.py

@@ -11,6 +11,7 @@ so it would also serve taskset-scoped (`Taskset.toolsets`).
 """
 
 import verifiers.v1 as vf
+from verifiers.v1.types import content_text
 
 PHRASE = "hello world"
 ECHO_TOKEN = "ok-7f3"  # the tool stamps this; only a real tool call can surface it
@@ -36,11 +37,9 @@ class EchoToolTask(vf.Task[vf.TaskData, vf.State, EchoToolTaskConfig]):
 
     @vf.reward(weight=1.0)
     async def echoed(self, trace: vf.Trace) -> float:
-        # The stamped token surfaces in an ASSISTANT message only if the model called
-        # the MCP tool and relayed its result — wherever in the exchange that happened
-        # (in a conversation the last turn may be a closing pleasantry).
-        replies = ((m.content or "").lower() for m in trace.assistant_messages)
-        return float(any(PHRASE in r and ECHO_TOKEN in r for r in replies))
+        # A stamped TOOL result proves the tool really ran with the phrase.
+        results = (content_text(m.content).lower() for m in trace.tool_messages)
+        return float(any(PHRASE in r and ECHO_TOKEN in r for r in results))
 
 
 class EchoToolConfig(vf.TasksetConfig):

{verifiers-0.2.2.dev96 → verifiers-0.2.2.dev97}/verifiers/v1/dialects/chat.py

@@ -141,7 +141,11 @@ def _content_to_wire(content):
 
 def message_to_wire(message: Message) -> dict:
     if message.role == "assistant":
-        wire: dict = {"role": "assistant", "content": message.content}
+        # Strict providers reject `content: null` without tool calls.
+        content = message.content
+        if content is None and not message.tool_calls:
+            content = ""
+        wire: dict = {"role": "assistant", "content": content}
         if message.provider_state:
             wire["reasoning_details"] = message.provider_state
         elif message.reasoning_content is not None:

{verifiers-0.2.2.dev96 → verifiers-0.2.2.dev97}/verifiers/v1/envs/user_sim/env.py

@@ -21,13 +21,14 @@ from pydantic import Field
 
 import verifiers.v1 as vf
 
-PERSONA = """You are role-playing a USER talking to an AI assistant. This is your situation and goal:
+PERSONA = """You are role-playing a USER talking to an AI assistant. This is your situation — what you want the ASSISTANT to do for you:
 
 {scenario}
 
 Rules:
 - Open the conversation with your request, in your own words.
 - Stay in character: short, natural user messages; never act as the assistant.
+- The assistant does the work, not you: never perform the task yourself — any tool call, answer, or output format it asks for must come from the assistant; ask for it instead.
 - Reveal details only when asked, as a real user would.
 - When the assistant has fully met your goal — or you are convinced it cannot — reply with exactly {done} and nothing else."""
 

{verifiers-0.2.2.dev96 → verifiers-0.2.2.dev97}/verifiers/v1/harnesses/openclaw/harness.py

@@ -117,10 +117,8 @@ class OpenClawHarness(Harness[OpenClawHarnessConfig]):
         data: TaskData,
     ) -> ProgramResult:
         system_prompt, prompt = self.resolve_prompt(data)
-        reasoning = ctx.sampling.reasoning_effort not in (
-            None,
-            "none",
-        ) or ctx.model.rsplit("/", 1)[-1].startswith(("gpt-5", "o1", "o3", "o4"))
+        # Opt-in via sampling: OpenClaw redacts persisted encrypted reasoning, so resumed sessions 400 replaying it.
+        reasoning = ctx.sampling.reasoning_effort not in (None, "none")
         directory = OPENCLAW_DIR.format(version=self.config.version)
         state_dir = f".vf-openclaw/{trace.id}"
         config_path = f"{state_dir}/openclaw.json"