fix(v1): de-flake the live E2E suite

tests/v1/conftest.py

@@ -164,7 +164,13 @@ 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"]})
+        # HarnessError covers live-model harness flakes (an agent ending a turn with
+        # no output, an agent-timeout stall); deterministic failures still fail every
+        # attempt.
+        seat_cfg.setdefault(
+            "retries",
+            {"max_retries": 2, "include": ["ProviderError", "HarnessError"]},
+        )
     return EvalConfig(
         env={
             "taskset": taskset_cfg,

tests/v1/fixtures/echo_tool_v1.py

@@ -11,6 +11,7 @@
 """
 
 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 +37,12 @@ def toolsets(cls, config: EchoToolTaskConfig) -> list[vf.Toolset]:
 
     @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 model called the MCP tool with the phrase —
+        # wherever in the exchange that happened. Scoring the assistant's relay instead
+        # would grade model obedience (a conversational reply may drop the stamp), not
+        # the tool loop this fixture exists to prove.
+        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/v1/dialects/chat.py

@@ -141,7 +141,13 @@ def _content_to_wire(content):
 
 def message_to_wire(message: Message) -> dict:
     if message.role == "assistant":
-        wire: dict = {"role": "assistant", "content": message.content}
+        # An empty completion records `content=None`; replayed as `"content": null`
+        # without tool calls, strict providers reject the whole request (422) — an
+        # empty string is the lossless wire form.
+        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/v1/envs/user_sim/env.py

@@ -21,13 +21,14 @@
 
 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/v1/harnesses/openclaw/harness.py

@@ -117,10 +117,13 @@ async def launch(
         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"))
+        # Reasoning replay is opt-in (`sampling.reasoning_effort`), never guessed from
+        # the model name. With `reasoning` on, OpenClaw asks for
+        # `reasoning.encrypted_content` and replays it from its persisted transcript on
+        # resume — where its secret redaction masks Prime's dotted `<token>.<metadata>`
+        # payloads (they fail OpenClaw's opaque-token allowlist), so every resumed
+        # session then dies on an upstream 400 replaying the mangled token.
+        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"