From 4e55594a3528c81a6f144b54a48bda8a1f6ab5f2 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Sun, 15 Mar 2026 00:37:33 +0530
Subject: [PATCH 01/13] remove redundant pydantic calls (mostly
 normalize_messages)

---
 verifiers/envs/environment.py     |  4 +---
 verifiers/envs/multiturn_env.py   | 38 +++++++++++++------------------
 verifiers/utils/response_utils.py | 14 +++++-------
 3 files changed, 23 insertions(+), 33 deletions(-)

diff --git a/verifiers/envs/environment.py b/verifiers/envs/environment.py
index 848683418a..89ff1ea878 100644
--- a/verifiers/envs/environment.py
+++ b/verifiers/envs/environment.py
@@ -539,12 +539,10 @@ def resolve_optional_args(
             client, model, tool_defs, sampling_args
         )
 
-        normalized_prompt = normalize_messages(prompt, field_name="prompt")
-
         self._get_usage_tracker(state, create_if_missing=True)
 
         response = await client.get_response(
-            prompt=normalized_prompt,
+            prompt=prompt,
             model=model,
             tools=tool_defs,
             sampling_args=sampling_args,
diff --git a/verifiers/envs/multiturn_env.py b/verifiers/envs/multiturn_env.py
index 4e2c867c8b..e316ffe7cd 100644
--- a/verifiers/envs/multiturn_env.py
+++ b/verifiers/envs/multiturn_env.py
@@ -71,13 +71,9 @@ async def setup_state(self, state: State) -> State:
     async def get_prompt_messages(self, state: State) -> Messages:
         """Override for rollouts with non-linear message sequences."""
         if len(state["trajectory"]) == 0:
-            return normalize_messages(state["prompt"], field_name="state.prompt")
-        prev_turn_prompt = normalize_messages(
-            state["trajectory"][-1]["prompt"], field_name="trajectory.prompt"
-        )
-        prev_turn_completion = normalize_messages(
-            state["trajectory"][-1]["completion"], field_name="trajectory.completion"
-        )
+            return state["prompt"]
+        prev_turn_prompt = state["trajectory"][-1]["prompt"]
+        prev_turn_completion = state["trajectory"][-1]["completion"]
         messages = concat_messages([prev_turn_prompt, prev_turn_completion])
         env_response = await self.env_response(messages, state)
         env_response_messages = normalize_messages(
@@ -90,23 +86,21 @@ async def render_completion(self, state: State):
         if len(state["trajectory"]) == 0:
             state["completion"] = []
             return
-        last_prompt = normalize_messages(
-            state["trajectory"][-1]["prompt"], field_name="trajectory.prompt"
-        )
-        last_completion = normalize_messages(
-            state["trajectory"][-1]["completion"], field_name="trajectory.completion"
-        )
+        last_prompt = state["trajectory"][-1]["prompt"]
+        last_completion = state["trajectory"][-1]["completion"]
         full_conversation = concat_messages([last_prompt, last_completion])
         if state.get("final_env_response"):
-            full_conversation = concat_messages(
-                [
-                    full_conversation,
-                    normalize_messages(
-                        state["final_env_response"], field_name="final_env_response"
-                    ),
-                ]
-            )
-        prompt_messages = normalize_messages(state["prompt"], field_name="state.prompt")
+            final_resp = state["final_env_response"]
+            if isinstance(final_resp, str) or (
+                isinstance(final_resp, list)
+                and final_resp
+                and isinstance(final_resp[0], dict)
+            ):
+                final_resp = normalize_messages(
+                    final_resp, field_name="final_env_response"
+                )
+            full_conversation = concat_messages([full_conversation, final_resp])
+        prompt_messages = state["prompt"]
         state["completion"] = full_conversation[len(prompt_messages) :]
 
     async def add_trajectory_step(self, state: State, trajectory_step: TrajectoryStep):
diff --git a/verifiers/utils/response_utils.py b/verifiers/utils/response_utils.py
index b24f1c2f77..03de5fc2ea 100644
--- a/verifiers/utils/response_utils.py
+++ b/verifiers/utils/response_utils.py
@@ -9,14 +9,12 @@
 async def parse_response_message(response: Response) -> Messages:
     """Parse a vf.Response into a vf.Messages list (single vf.AssistantMessage)."""
     response_message = response.message
-    message_payload = {
-        "role": "assistant",
-        "content": response_message.content,
-        "reasoning_content": response_message.reasoning_content,
-        "thinking_blocks": response_message.thinking_blocks,
-        "tool_calls": response_message.tool_calls,
-    }
-    message = AssistantMessage.model_validate(message_payload)
+    message = AssistantMessage(
+        content=response_message.content,
+        reasoning_content=response_message.reasoning_content,
+        thinking_blocks=response_message.thinking_blocks,
+        tool_calls=response_message.tool_calls,
+    )
     return [message]
 
 

From 8ee58594efc644da0aa45ea928c043a8fcee3f1e Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Tue, 17 Mar 2026 10:31:36 +0000
Subject: [PATCH 02/13] keep normalizer in get_model_response

---
 verifiers/envs/environment.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/verifiers/envs/environment.py b/verifiers/envs/environment.py
index 89ff1ea878..848683418a 100644
--- a/verifiers/envs/environment.py
+++ b/verifiers/envs/environment.py
@@ -539,10 +539,12 @@ def resolve_optional_args(
             client, model, tool_defs, sampling_args
         )
 
+        normalized_prompt = normalize_messages(prompt, field_name="prompt")
+
         self._get_usage_tracker(state, create_if_missing=True)
 
         response = await client.get_response(
-            prompt=prompt,
+            prompt=normalized_prompt,
             model=model,
             tools=tool_defs,
             sampling_args=sampling_args,

From eb5a5ca70c339cf8ce7600ac223f313160d5627b Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Tue, 17 Mar 2026 10:32:08 +0000
Subject: [PATCH 03/13] Revert "keep normalizer in get_model_response"

This reverts commit 8b26a4250a581954a3edbb12268deb4b0ebeeed9.
---
 verifiers/envs/environment.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/verifiers/envs/environment.py b/verifiers/envs/environment.py
index 848683418a..89ff1ea878 100644
--- a/verifiers/envs/environment.py
+++ b/verifiers/envs/environment.py
@@ -539,12 +539,10 @@ def resolve_optional_args(
             client, model, tool_defs, sampling_args
         )
 
-        normalized_prompt = normalize_messages(prompt, field_name="prompt")
-
         self._get_usage_tracker(state, create_if_missing=True)
 
         response = await client.get_response(
-            prompt=normalized_prompt,
+            prompt=prompt,
             model=model,
             tools=tool_defs,
             sampling_args=sampling_args,

From 4bb8a63ed47f62836d19d4cc30465be2b2093a10 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Tue, 17 Mar 2026 10:33:29 +0000
Subject: [PATCH 04/13] fix ty

---
 verifiers/envs/environment.py                | 2 +-
 verifiers/envs/experimental/cli_agent_env.py | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/verifiers/envs/environment.py b/verifiers/envs/environment.py
index 89ff1ea878..4b91933161 100644
--- a/verifiers/envs/environment.py
+++ b/verifiers/envs/environment.py
@@ -494,7 +494,7 @@ def get_state_usage(self, state: State) -> TokenUsage | None:
     async def get_model_response(
         self,
         state: State,
-        prompt: Messages | str,
+        prompt: Messages,
         client: Client | None = None,
         model: str | None = None,
         tool_defs: list[Tool] | None = None,
diff --git a/verifiers/envs/experimental/cli_agent_env.py b/verifiers/envs/experimental/cli_agent_env.py
index 1db7005204..8f5d86fb0d 100644
--- a/verifiers/envs/experimental/cli_agent_env.py
+++ b/verifiers/envs/experimental/cli_agent_env.py
@@ -436,7 +436,7 @@ async def get_prompt_messages(self, state: State) -> Messages:
     async def get_model_response(
         self,
         state: State,
-        prompt: Messages | str,
+        prompt: Messages,
         client: Client | None = None,
         model: str | None = None,
         tool_defs: list[Tool] | None = None,

From 1877e7b5f81d4624122947a7b71fdd2fcd025392 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 11:54:43 +0000
Subject: [PATCH 05/13] add warning and auto-normalization

---
 verifiers/envs/environment.py | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

diff --git a/verifiers/envs/environment.py b/verifiers/envs/environment.py
index 4b91933161..4261f94c3f 100644
--- a/verifiers/envs/environment.py
+++ b/verifiers/envs/environment.py
@@ -494,7 +494,7 @@ def get_state_usage(self, state: State) -> TokenUsage | None:
     async def get_model_response(
         self,
         state: State,
-        prompt: Messages,
+        prompt: Messages | str,
         client: Client | None = None,
         model: str | None = None,
         tool_defs: list[Tool] | None = None,
@@ -541,6 +541,18 @@ def resolve_optional_args(
 
         self._get_usage_tracker(state, create_if_missing=True)
 
+        if not isinstance(prompt, list) or not all(
+            isinstance(m, vf.Message) for m in prompt
+        ):
+            self.logger.warning(
+                "get_model_response() received raw dicts/strings instead of "
+                "vf.Message objects. This triggers normalize_messages() on every "
+                "call, which can cause unnecessary Pydantic validation overhaead. "
+                "Return vf.Message types (e.g. vf.UserMessage, vf.AssistantMessage) "
+                "from get_prompt_messages() to avoid this overhead."
+            )
+            prompt = normalize_messages(prompt)
+
         response = await client.get_response(
             prompt=prompt,
             model=model,

From e219eeb75302668005d57c948a2cb69c54bdeac2 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 11:56:07 +0000
Subject: [PATCH 06/13] fix ty

---
 verifiers/envs/experimental/cli_agent_env.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/verifiers/envs/experimental/cli_agent_env.py b/verifiers/envs/experimental/cli_agent_env.py
index 8f5d86fb0d..1db7005204 100644
--- a/verifiers/envs/experimental/cli_agent_env.py
+++ b/verifiers/envs/experimental/cli_agent_env.py
@@ -436,7 +436,7 @@ async def get_prompt_messages(self, state: State) -> Messages:
     async def get_model_response(
         self,
         state: State,
-        prompt: Messages,
+        prompt: Messages | str,
         client: Client | None = None,
         model: str | None = None,
         tool_defs: list[Tool] | None = None,

From f7b61318073a31730af617986024dc1caded8086 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 12:13:44 +0000
Subject: [PATCH 07/13] move normalization into env loop to ensure trajectory
 list has correct types

---
 verifiers/envs/environment.py    | 14 +-------------
 verifiers/envs/multiturn_env.py  | 30 ++++++++++++++++++++++++------
 verifiers/utils/message_utils.py |  5 +++++
 3 files changed, 30 insertions(+), 19 deletions(-)

diff --git a/verifiers/envs/environment.py b/verifiers/envs/environment.py
index 4261f94c3f..4b91933161 100644
--- a/verifiers/envs/environment.py
+++ b/verifiers/envs/environment.py
@@ -494,7 +494,7 @@ def get_state_usage(self, state: State) -> TokenUsage | None:
     async def get_model_response(
         self,
         state: State,
-        prompt: Messages | str,
+        prompt: Messages,
         client: Client | None = None,
         model: str | None = None,
         tool_defs: list[Tool] | None = None,
@@ -541,18 +541,6 @@ def resolve_optional_args(
 
         self._get_usage_tracker(state, create_if_missing=True)
 
-        if not isinstance(prompt, list) or not all(
-            isinstance(m, vf.Message) for m in prompt
-        ):
-            self.logger.warning(
-                "get_model_response() received raw dicts/strings instead of "
-                "vf.Message objects. This triggers normalize_messages() on every "
-                "call, which can cause unnecessary Pydantic validation overhaead. "
-                "Return vf.Message types (e.g. vf.UserMessage, vf.AssistantMessage) "
-                "from get_prompt_messages() to avoid this overhead."
-            )
-            prompt = normalize_messages(prompt)
-
         response = await client.get_response(
             prompt=prompt,
             model=model,
diff --git a/verifiers/envs/multiturn_env.py b/verifiers/envs/multiturn_env.py
index e316ffe7cd..c79779bf95 100644
--- a/verifiers/envs/multiturn_env.py
+++ b/verifiers/envs/multiturn_env.py
@@ -13,7 +13,11 @@
     State,
     TrajectoryStep,
 )
-from verifiers.utils.message_utils import concat_messages, normalize_messages
+from verifiers.utils.message_utils import (
+    concat_messages,
+    normalize_messages,
+    requires_normalize_messages,
+)
 from verifiers.utils.response_utils import (
     parse_response_message,
     parse_response_tokens,
@@ -91,11 +95,14 @@ async def render_completion(self, state: State):
         full_conversation = concat_messages([last_prompt, last_completion])
         if state.get("final_env_response"):
             final_resp = state["final_env_response"]
-            if isinstance(final_resp, str) or (
-                isinstance(final_resp, list)
-                and final_resp
-                and isinstance(final_resp[0], dict)
-            ):
+            if requires_normalize_messages(final_resp):
+                self.logger.warning(
+                    "final_env_response returned raw dicts/strings instead of "
+                    "vf.Message objects. This triggers normalize_messages() on every "
+                    "turn, which can cause unnecessary Pydantic validation overhead. "
+                    "Return vf.Message types (e.g. vf.UserMessage, vf.AssistantMessage) "
+                    "from env_response() to avoid this overhead."
+                )
                 final_resp = normalize_messages(
                     final_resp, field_name="final_env_response"
                 )
@@ -150,6 +157,17 @@ async def rollout(
             while not await self.is_completed(state):
                 try:
                     prompt_messages = await self.get_prompt_messages(state)
+                    if requires_normalize_messages(prompt_messages):
+                        self.logger.warning(
+                            "get_prompt_messages() returned raw dicts/strings instead of "
+                            "vf.Message objects. This triggers normalize_messages() on every "
+                            "turn, which can cause unnecessary Pydantic validation overhead. "
+                            "Return vf.Message types (e.g. vf.UserMessage, vf.AssistantMessage) "
+                            "from get_prompt_messages() to avoid this overhead."
+                        )
+                        prompt_messages = normalize_messages(
+                            prompt_messages, field_name="prompt_messages"
+                        )
                     if state.get("final_env_response") is not None:
                         continue
                     response = await self.get_model_response(state, prompt_messages)
diff --git a/verifiers/utils/message_utils.py b/verifiers/utils/message_utils.py
index b3c55318bd..d0d86f6912 100644
--- a/verifiers/utils/message_utils.py
+++ b/verifiers/utils/message_utils.py
@@ -139,6 +139,11 @@ def normalize_messages(
     return normalized
 
 
+def requires_normalize_messages(value: list | str) -> bool:
+    """Check if a message list requires normalization into a list of Message objects."""
+    return not isinstance(value, list) or not all(isinstance(m, Message) for m in value)
+
+
 def concat_messages(messages_list: list[Messages]) -> Messages:
     """Concatenate multiple Messages lists into one."""
     result = []

From e86b558352cfe267f722fa62800d1756a15d21e2 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 12:31:46 +0000
Subject: [PATCH 08/13] maybe_normalize_messages and updated env_response
 return type

---
 verifiers/envs/multiturn_env.py  | 36 +++++++++-----------------------
 verifiers/utils/message_utils.py | 25 +++++++++++++++++++---
 2 files changed, 32 insertions(+), 29 deletions(-)

diff --git a/verifiers/envs/multiturn_env.py b/verifiers/envs/multiturn_env.py
index c79779bf95..f8a6c3d4a9 100644
--- a/verifiers/envs/multiturn_env.py
+++ b/verifiers/envs/multiturn_env.py
@@ -15,8 +15,7 @@
 )
 from verifiers.utils.message_utils import (
     concat_messages,
-    normalize_messages,
-    requires_normalize_messages,
+    maybe_normalize_messages,
 )
 from verifiers.utils.response_utils import (
     parse_response_message,
@@ -45,7 +44,7 @@ def __init__(self, max_turns: int = -1, **kwargs):
     @abstractmethod
     async def env_response(
         self, messages: Messages, state: State, **kwargs
-    ) -> Messages | str:
+    ) -> Messages:
         """
         Generate a response from the environment.
         """
@@ -80,7 +79,7 @@ async def get_prompt_messages(self, state: State) -> Messages:
         prev_turn_completion = state["trajectory"][-1]["completion"]
         messages = concat_messages([prev_turn_prompt, prev_turn_completion])
         env_response = await self.env_response(messages, state)
-        env_response_messages = normalize_messages(
+        env_response_messages = maybe_normalize_messages(
             env_response, field_name="env_response"
         )
         return concat_messages([messages, env_response_messages])
@@ -95,17 +94,9 @@ async def render_completion(self, state: State):
         full_conversation = concat_messages([last_prompt, last_completion])
         if state.get("final_env_response"):
             final_resp = state["final_env_response"]
-            if requires_normalize_messages(final_resp):
-                self.logger.warning(
-                    "final_env_response returned raw dicts/strings instead of "
-                    "vf.Message objects. This triggers normalize_messages() on every "
-                    "turn, which can cause unnecessary Pydantic validation overhead. "
-                    "Return vf.Message types (e.g. vf.UserMessage, vf.AssistantMessage) "
-                    "from env_response() to avoid this overhead."
-                )
-                final_resp = normalize_messages(
-                    final_resp, field_name="final_env_response"
-                )
+            final_resp = maybe_normalize_messages(
+                final_resp, field_name="final_env_response"
+            )
             full_conversation = concat_messages([full_conversation, final_resp])
         prompt_messages = state["prompt"]
         state["completion"] = full_conversation[len(prompt_messages) :]
@@ -157,17 +148,10 @@ async def rollout(
             while not await self.is_completed(state):
                 try:
                     prompt_messages = await self.get_prompt_messages(state)
-                    if requires_normalize_messages(prompt_messages):
-                        self.logger.warning(
-                            "get_prompt_messages() returned raw dicts/strings instead of "
-                            "vf.Message objects. This triggers normalize_messages() on every "
-                            "turn, which can cause unnecessary Pydantic validation overhead. "
-                            "Return vf.Message types (e.g. vf.UserMessage, vf.AssistantMessage) "
-                            "from get_prompt_messages() to avoid this overhead."
-                        )
-                        prompt_messages = normalize_messages(
-                            prompt_messages, field_name="prompt_messages"
-                        )
+                    prompt_messages = maybe_normalize_messages(
+                        prompt_messages,
+                        field_name="prompt_messages",
+                    )
                     if state.get("final_env_response") is not None:
                         continue
                     response = await self.get_model_response(state, prompt_messages)
diff --git a/verifiers/utils/message_utils.py b/verifiers/utils/message_utils.py
index d0d86f6912..e081b43588 100644
--- a/verifiers/utils/message_utils.py
+++ b/verifiers/utils/message_utils.py
@@ -1,4 +1,5 @@
 import json
+import logging
 import re
 from collections.abc import Mapping
 from typing import Any, cast
@@ -19,6 +20,8 @@
     UserMessage,
 )
 
+logger = logging.getLogger(__name__)
+
 
 def from_raw_content_part(part: dict[str, Any]) -> ContentPart:
     """Convert a raw content-part dict to a typed content part when possible."""
@@ -139,9 +142,25 @@ def normalize_messages(
     return normalized
 
 
-def requires_normalize_messages(value: list | str) -> bool:
-    """Check if a message list requires normalization into a list of Message objects."""
-    return not isinstance(value, list) or not all(isinstance(m, Message) for m in value)
+def maybe_normalize_messages(
+    value: Messages | str,
+    *,
+    field_name: str = "messages",
+) -> Messages:
+    """Normalize messages only if needed, logging a warning on first occurrence."""
+    requires_normalize = not isinstance(value, list) or not all(
+        isinstance(m, Message) for m in value
+    )
+    if not requires_normalize:
+        return cast(Messages, value)
+    logger.warning(
+        "%s returned raw dicts/strings instead of vf.Messages. This"
+        " repeatedly triggers normalize_messages(), causing unnecessary"
+        " Pydantic validation overhead. Return vf.Message types (e.g."
+        " vf.UserMessage, vf.AssistantMessage) to avoid this.",
+        field_name,
+    )
+    return normalize_messages(value, field_name=field_name)
 
 
 def concat_messages(messages_list: list[Messages]) -> Messages:

From f1dbfcb12b6886e4acbba2b6f84704fd23f9b8f5 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 12:35:28 +0000
Subject: [PATCH 09/13] fix envs + api

---
 environments/alphabet_sort/alphabet_sort.py         | 2 +-
 environments/doublecheck/doublecheck.py             | 2 +-
 environments/sentence_repeater/sentence_repeater.py | 7 +------
 verifiers/envs/experimental/cli_agent_env.py        | 2 +-
 verifiers/envs/experimental/gym_env.py              | 6 +++---
 5 files changed, 7 insertions(+), 12 deletions(-)

diff --git a/environments/alphabet_sort/alphabet_sort.py b/environments/alphabet_sort/alphabet_sort.py
index 75565121c6..4c5cc1cfef 100644
--- a/environments/alphabet_sort/alphabet_sort.py
+++ b/environments/alphabet_sort/alphabet_sort.py
@@ -224,7 +224,7 @@ async def env_response(
             assistant_count = len([m for m in messages if m["role"] == "assistant"])
             follow_ups = state["info"]["follow_ups"]
             follow_up_idx = assistant_count - 1
-            return [{"role": "user", "content": follow_ups[follow_up_idx]}]
+            return [vf.UserMessage(content=follow_ups[follow_up_idx])]
 
     def score_response(
         predicted: List[str], expected: List[str], apply_power: bool = True
diff --git a/environments/doublecheck/doublecheck.py b/environments/doublecheck/doublecheck.py
index a56b820aa1..bf7162ca7e 100644
--- a/environments/doublecheck/doublecheck.py
+++ b/environments/doublecheck/doublecheck.py
@@ -27,7 +27,7 @@ async def env_response(
         self, messages: Messages, state: State, **kwargs
     ) -> Messages:
         """Generate a response from the environment."""
-        return [{"role": "user", "content": "Are you sure?"}]
+        return [vf.UserMessage(content="Are you sure?")]
 
 
 def load_environment(
diff --git a/environments/sentence_repeater/sentence_repeater.py b/environments/sentence_repeater/sentence_repeater.py
index e189cedad9..e7ba1152a2 100644
--- a/environments/sentence_repeater/sentence_repeater.py
+++ b/environments/sentence_repeater/sentence_repeater.py
@@ -83,12 +83,7 @@ async def env_response(
         self, messages: Messages, state: State, **kwargs
     ) -> Messages:
         num_turn = len(state["trajectory"])
-        return [
-            {
-                "role": "user",
-                "content": state["info"]["questions"][num_turn],
-            }
-        ]
+        return [vf.UserMessage(content=state["info"]["questions"][num_turn])]
 
 
 def load_environment(**kwargs) -> vf.Environment:
diff --git a/verifiers/envs/experimental/cli_agent_env.py b/verifiers/envs/experimental/cli_agent_env.py
index 1db7005204..8f5d86fb0d 100644
--- a/verifiers/envs/experimental/cli_agent_env.py
+++ b/verifiers/envs/experimental/cli_agent_env.py
@@ -436,7 +436,7 @@ async def get_prompt_messages(self, state: State) -> Messages:
     async def get_model_response(
         self,
         state: State,
-        prompt: Messages | str,
+        prompt: Messages,
         client: Client | None = None,
         model: str | None = None,
         tool_defs: list[Tool] | None = None,
diff --git a/verifiers/envs/experimental/gym_env.py b/verifiers/envs/experimental/gym_env.py
index fe4781568a..0d0096319b 100644
--- a/verifiers/envs/experimental/gym_env.py
+++ b/verifiers/envs/experimental/gym_env.py
@@ -142,12 +142,12 @@ def obs_to_text(self, obs: Any) -> str:
             return self.obs_to_text_fn(obs)
         return str(obs)
 
-    def wrap_response(self, text: str) -> vf.Messages | str:
-        return cast(vf.Messages, [{"role": "user", "content": text}])
+    def wrap_response(self, text: str) -> vf.Messages:
+        return [vf.UserMessage(content=text)]
 
     async def env_response(
         self, messages: vf.Messages, state: State, **kwargs: Any
-    ) -> vf.Messages | str:
+    ) -> vf.Messages:
         if "gym_env" not in state:
             env = self.env_cls(**self.env_kwargs)
             seed = int(state["answer"])

From 559278afcde7049a6e03c51cf7b69980983a641b Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 12:40:01 +0000
Subject: [PATCH 10/13] style

---
 verifiers/envs/multiturn_env.py | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)

diff --git a/verifiers/envs/multiturn_env.py b/verifiers/envs/multiturn_env.py
index f8a6c3d4a9..d43f69d95c 100644
--- a/verifiers/envs/multiturn_env.py
+++ b/verifiers/envs/multiturn_env.py
@@ -79,10 +79,8 @@ async def get_prompt_messages(self, state: State) -> Messages:
         prev_turn_completion = state["trajectory"][-1]["completion"]
         messages = concat_messages([prev_turn_prompt, prev_turn_completion])
         env_response = await self.env_response(messages, state)
-        env_response_messages = maybe_normalize_messages(
-            env_response, field_name="env_response"
-        )
-        return concat_messages([messages, env_response_messages])
+        env_response = maybe_normalize_messages(env_response, field_name="env_response")
+        return concat_messages([messages, env_response])
 
     async def render_completion(self, state: State):
         """Override for rollouts with non-linear message sequences."""
@@ -149,8 +147,7 @@ async def rollout(
                 try:
                     prompt_messages = await self.get_prompt_messages(state)
                     prompt_messages = maybe_normalize_messages(
-                        prompt_messages,
-                        field_name="prompt_messages",
+                        prompt_messages, field_name="prompt_messages"
                     )
                     if state.get("final_env_response") is not None:
                         continue

From 162b0ac9e96ef35f56456446b05b5c7c42605c90 Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 13:06:07 +0000
Subject: [PATCH 11/13] log once utils

---
 verifiers/utils/logging_utils.py | 27 ++++++++++++++++++++++++++-
 verifiers/utils/message_utils.py |  5 ++++-
 2 files changed, 30 insertions(+), 2 deletions(-)

diff --git a/verifiers/utils/logging_utils.py b/verifiers/utils/logging_utils.py
index 6d0084f644..cdd43f780e 100644
--- a/verifiers/utils/logging_utils.py
+++ b/verifiers/utils/logging_utils.py
@@ -12,10 +12,33 @@
 from verifiers.errors import Error
 from verifiers.types import ErrorInfo, Messages
 from verifiers.utils.error_utils import ErrorChain
-from verifiers.utils.message_utils import format_messages
 
 LOGGER_NAME = "verifiers"
 
+_seen_once_keys: set[tuple[str, str]] = set()
+
+
+def log_once(
+    logger: logging.Logger,
+    level: int,
+    msg: str,
+    *args: object,
+    **kwargs: object,
+) -> None:
+    """Log a message only once per (logger name, message) pair for the process lifetime."""
+    key = (logger.name, msg)
+    if key in _seen_once_keys:
+        return
+    _seen_once_keys.add(key)
+    logger.log(level, msg, *args, **kwargs)
+
+
+def warning_once(
+    logger: logging.Logger, msg: str, *args: object, **kwargs: object
+) -> None:
+    """Shorthand for ``log_once(logger, logging.WARNING, ...)``."""
+    log_once(logger, logging.WARNING, msg, *args, **kwargs)
+
 
 class JsonFormatter(logging.Formatter):
     """JSON formatter for structured logging."""
@@ -144,6 +167,8 @@ def print_prompt_completions_sample(
     step: int,
     num_samples: int = 1,
 ) -> None:
+    from verifiers.utils.message_utils import format_messages
+
     def format_error(error: ErrorInfo | BaseException) -> Text:
         out = Text()
         if isinstance(error, BaseException):
diff --git a/verifiers/utils/message_utils.py b/verifiers/utils/message_utils.py
index e081b43588..7e82e35534 100644
--- a/verifiers/utils/message_utils.py
+++ b/verifiers/utils/message_utils.py
@@ -148,12 +148,15 @@ def maybe_normalize_messages(
     field_name: str = "messages",
 ) -> Messages:
     """Normalize messages only if needed, logging a warning on first occurrence."""
+    from verifiers.utils.logging_utils import warning_once
+
     requires_normalize = not isinstance(value, list) or not all(
         isinstance(m, Message) for m in value
     )
     if not requires_normalize:
         return cast(Messages, value)
-    logger.warning(
+    warning_once(
+        logger,
         "%s returned raw dicts/strings instead of vf.Messages. This"
         " repeatedly triggers normalize_messages(), causing unnecessary"
         " Pydantic validation overhead. Return vf.Message types (e.g."

From 74b0b78a34e4077bc066b98c2c1b62a4c3fa0d1d Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 13:39:00 +0000
Subject: [PATCH 12/13] address bugbot

---
 verifiers/utils/message_utils.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/verifiers/utils/message_utils.py b/verifiers/utils/message_utils.py
index 7e82e35534..cdb01a05b3 100644
--- a/verifiers/utils/message_utils.py
+++ b/verifiers/utils/message_utils.py
@@ -157,11 +157,10 @@ def maybe_normalize_messages(
         return cast(Messages, value)
     warning_once(
         logger,
-        "%s returned raw dicts/strings instead of vf.Messages. This"
+        f"{field_name} returned raw dicts/strings instead of vf.Messages. This"
         " repeatedly triggers normalize_messages(), causing unnecessary"
         " Pydantic validation overhead. Return vf.Message types (e.g."
         " vf.UserMessage, vf.AssistantMessage) to avoid this.",
-        field_name,
     )
     return normalize_messages(value, field_name=field_name)
 

From cd87b2c5ce14ea8b396844fea2220cf9eb78c6ad Mon Sep 17 00:00:00 2001
From: Mika Senghaas <mail@mikasenghaas.de>
Date: Wed, 18 Mar 2026 14:38:29 +0000
Subject: [PATCH 13/13] fix ty

---
 verifiers/utils/logging_utils.py | 16 ++++------------
 1 file changed, 4 insertions(+), 12 deletions(-)

diff --git a/verifiers/utils/logging_utils.py b/verifiers/utils/logging_utils.py
index cdd43f780e..830f5f3d4a 100644
--- a/verifiers/utils/logging_utils.py
+++ b/verifiers/utils/logging_utils.py
@@ -18,26 +18,18 @@
 _seen_once_keys: set[tuple[str, str]] = set()
 
 
-def log_once(
-    logger: logging.Logger,
-    level: int,
-    msg: str,
-    *args: object,
-    **kwargs: object,
-) -> None:
+def log_once(logger: logging.Logger, level: int, msg: str) -> None:
     """Log a message only once per (logger name, message) pair for the process lifetime."""
     key = (logger.name, msg)
     if key in _seen_once_keys:
         return
     _seen_once_keys.add(key)
-    logger.log(level, msg, *args, **kwargs)
+    logger.log(level, msg)
 
 
-def warning_once(
-    logger: logging.Logger, msg: str, *args: object, **kwargs: object
-) -> None:
+def warning_once(logger: logging.Logger, msg: str) -> None:
     """Shorthand for ``log_once(logger, logging.WARNING, ...)``."""
-    log_once(logger, logging.WARNING, msg, *args, **kwargs)
+    log_once(logger, logging.WARNING, msg)
 
 
 class JsonFormatter(logging.Formatter):