Consume the typed RendererConfig surface

pyproject.toml

@@ -74,7 +74,7 @@ dev = [
     "aiohttp>=3.9.0",
     "python-dotenv>=1.0.0",
     "nltk",
-    "renderers>=0.1.8.dev4",
+    "renderers>=0.1.8.dev28",
 ]
 policy = [
     "semgrep>=1.150.0",
@@ -97,7 +97,7 @@ openenv = [
     "openenv-core>=0.3.0",
 ]
 renderers = [
-    "renderers>=0.1.8.dev4",
+    "renderers>=0.1.8.dev28",
 ]
 rl = [
     "torch>=2.8.0,<2.9.0",

tests/test_renderer_client.py

@@ -5,6 +5,7 @@
 
 import verifiers as vf
 from renderers import RendererPool
+from renderers import config_from_name
 from renderers.base import ParsedResponse, RenderedTokens, create_renderer
 from verifiers.clients.renderer_client import (
     RendererClient,
@@ -24,13 +25,16 @@
 )
 
 
-def test_renderer_client_honors_configured_renderer_name():
+def test_renderer_client_honors_configured_renderer_config():
+    from renderers import Qwen3VLRendererConfig
+
     RendererClient._shared_pools.clear()
 
+    cfg = Qwen3VLRendererConfig()
     client = object.__new__(RendererClient)
     client._renderer = None
     client._pool_size = 1
-    client._config = vf.ClientConfig(client_type="renderer", renderer="qwen3_vl")
+    client._config = vf.ClientConfig(client_type="renderer", renderer_config=cfg)
 
     sentinel_pool = RendererPool.__new__(RendererPool)
     with patch(
@@ -42,24 +46,23 @@ def test_renderer_client_honors_configured_renderer_name():
     assert pool is sentinel_pool
     create_pool_mock.assert_called_once_with(
         "Qwen/Qwen3-VL-4B-Instruct",
-        renderer="qwen3_vl",
+        cfg,
         size=1,
-        tool_parser=None,
-        reasoning_parser=None,
-        preserve_all_thinking=False,
-        preserve_thinking_between_tool_calls=False,
     )
 
 
 def test_renderer_client_uses_renderer_model_name_override():
+    from renderers import Qwen3VLRendererConfig
+
     RendererClient._shared_pools.clear()
 
+    cfg = Qwen3VLRendererConfig()
     client = object.__new__(RendererClient)
     client._renderer = None
     client._pool_size = 1
     client._config = vf.ClientConfig(
         client_type="renderer",
-        renderer="qwen3_vl",
+        renderer_config=cfg,
         renderer_model_name="Qwen/Qwen3-VL-4B-Instruct",
     )
 
@@ -73,12 +76,8 @@ def test_renderer_client_uses_renderer_model_name_override():
     assert pool is sentinel_pool
     create_pool_mock.assert_called_once_with(
         "Qwen/Qwen3-VL-4B-Instruct",
-        renderer="qwen3_vl",
+        cfg,
         size=1,
-        tool_parser=None,
-        reasoning_parser=None,
-        preserve_all_thinking=False,
-        preserve_thinking_between_tool_calls=False,
     )
 
 
@@ -524,7 +523,7 @@ def _load_tokenizer_and_renderer(model_name: str, renderer_name: str):
     from transformers import AutoTokenizer
 
     tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
-    renderer = create_renderer(tokenizer, renderer=renderer_name)
+    renderer = create_renderer(tokenizer, config_from_name(renderer_name))
     return tokenizer, renderer
 
 

tests/test_renderer_e2e.py

@@ -17,14 +17,15 @@
 are exercised. Tokenizers come from the local HF cache; no network.
 """
 
+import json
 import logging
 from typing import Any
 
 import pytest
 
 import verifiers as vf
 from datasets import Dataset
-from renderers import create_renderer
+from renderers import config_from_name, create_renderer
 from verifiers.clients.renderer_client import RendererClient, _to_renderer_message
 from verifiers.types import Messages, State
 
@@ -83,7 +84,7 @@ def _load(model_name: str, renderer_name: str):
         from transformers import AutoTokenizer
 
         tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
-        renderer = create_renderer(tokenizer, renderer=renderer_name)
+        renderer = create_renderer(tokenizer, config_from_name(renderer_name))
         _renderer_cache[key] = (tokenizer, renderer)
     return _renderer_cache[key]
 
@@ -106,6 +107,13 @@ def tokenizer_and_renderer(model_family):
 # ── Scripted vLLM stand-in ───────────────────────────────────────────
 
 
+class _ScriptedResponse:
+    """httpx.Response stand-in: ``parse_generate_response`` reads ``.content`` as bytes."""
+
+    def __init__(self, payload: dict[str, Any]):
+        self.content = json.dumps(payload).encode()
+
+
 class ScriptedVLLM:
     """Fake ``AsyncOpenAI``-compatible client serving canned
     /inference/v1/generate responses (vllm 0.20 wire shape).
@@ -124,22 +132,24 @@ async def post(self, path: str, *, cast_to=None, body=None, options=None):
         assert self._completions, "ScriptedVLLM ran out of canned completions"
         completion_ids = self._completions.pop(0)
 
-        return {
-            "request_id": f"resp-{len(self.requests)}",
-            "choices": [
-                {
-                    "index": 0,
-                    "token_ids": list(completion_ids),
-                    "logprobs": {
-                        "content": [
-                            {"token": f"token_id:{tid}", "logprob": -0.1}
-                            for tid in completion_ids
-                        ]
-                    },
-                    "finish_reason": "stop",
-                }
-            ],
-        }
+        return _ScriptedResponse(
+            {
+                "request_id": f"resp-{len(self.requests)}",
+                "choices": [
+                    {
+                        "index": 0,
+                        "token_ids": list(completion_ids),
+                        "logprobs": {
+                            "content": [
+                                {"token": f"token_id:{tid}", "logprob": -0.1}
+                                for tid in completion_ids
+                            ]
+                        },
+                        "finish_reason": "stop",
+                    }
+                ],
+            }
+        )
 
     async def close(self):
         pass