lmproto: carry thinking options on the wire

src/plyngent/lmproto/anthropic/client.py

@@ -34,6 +34,8 @@ class AnthropicClient:
     decoder: msgspec.json.Decoder[AnthropicMessageResponse]
     stream_decoder: msgspec.json.Decoder[AnthropicStreamEvent]
     models_decoder: msgspec.json.Decoder[AnthropicModelsResponse]
+    # Explicit ``thinking`` budget for this client (0 = derive from the effort).
+    thinking_budget_tokens: int = 0
     _api_key: str
     _api_version: str
 
@@ -43,6 +45,7 @@ class AnthropicClient:
         self.decoder = msgspec.json.Decoder(AnthropicMessageResponse)
         self.stream_decoder = msgspec.json.Decoder(AnthropicStreamEvent)
         self.models_decoder = msgspec.json.Decoder(AnthropicModelsResponse)
+        self.thinking_budget_tokens = max(0, config.thinking_budget_tokens)
         self._api_key = config.api_key
         self._api_version = config.anthropic_version
 

src/plyngent/lmproto/anthropic/config.py

@@ -20,3 +20,6 @@ class AnthropicConfig:
     base_url: str = DEFAULT_ANTHROPIC_BASE_URL
     anthropic_version: str = DEFAULT_ANTHROPIC_VERSION
     timeout: HttpTimeout = (DEFAULT_HTTP_CONNECT_TIMEOUT, DEFAULT_HTTP_READ_TIMEOUT)
+    # Explicit ``thinking.budget_tokens`` for reasoning models (0 = derive it
+    # from the configured reasoning_effort; see config/reasoning.py).
+    thinking_budget_tokens: int = 0

src/plyngent/lmproto/anthropic/model.py

@@ -62,6 +62,18 @@ class AnthropicMetadata(Struct, omit_defaults=True):
     user_id: str | Unset = UNSET
 
 
+class AnthropicThinkingConfig(Struct, omit_defaults=True):
+    """``thinking`` request block (extended thinking).
+
+    ``budget_tokens`` is the reasoning budget the model may spend before it
+    answers; the API requires it to be at least 1024 and smaller than
+    ``max_tokens`` (see ``agent.messages_bridge``, which keeps both valid).
+    """
+
+    type: Literal["enabled", "disabled"] = "enabled"
+    budget_tokens: int | Unset = UNSET
+
+
 class AnthropicMessagesParam(Struct, omit_defaults=True):
     model: str
     max_tokens: int = 8192
@@ -75,6 +87,7 @@ class AnthropicMessagesParam(Struct, omit_defaults=True):
     temperature: float | Unset = UNSET
     top_p: float | Unset = UNSET
     top_k: int | Unset = UNSET
+    thinking: AnthropicThinkingConfig | Unset = UNSET
 
 
 class AnthropicUsage(Struct, omit_defaults=True):

src/plyngent/lmproto/deepseek/openai_compat/client.py

@@ -16,16 +16,17 @@ if TYPE_CHECKING:
     from .model import ChatCompletionsParam
 
 
-def _inject_thinking(data: bytes) -> bytes:
-    """Inject ``thinking: {type: "enabled"}`` into the encoded request body.
+def _inject_thinking(data: bytes, param: ChatCompletionsParam) -> bytes:
+    """Inject the ``thinking`` flag into the encoded request body.
 
     DeepSeek's reasoning model requires the ``thinking`` parameter to be
-    explicitly set (default is ``enabled``). The agent loop constructs the
-    base ``ChatCompletionsParam`` which lacks this field, so we add it
-    after msgspec encoding.
+    explicitly set (default is ``enabled``); ``reasoning_effort = "none"`` is the
+    way to ask for it off. The agent loop constructs the base
+    ``ChatCompletionsParam`` which lacks this field, so we add it after msgspec
+    encoding.
     """
     body = json.loads(data)
-    body["thinking"] = {"type": "enabled"}
+    body["thinking"] = {"type": "disabled" if param.reasoning_effort == "none" else "enabled"}
     return json.dumps(body, separators=(",", ":")).encode("utf-8")
 
 
@@ -49,7 +50,7 @@ class DeepseekOpenAIClient(BaseOpenAIClient):
         self, param: ChatCompletionsParam, *, stream: bool = False
     ) -> ChatCompletionResponse | AsyncIterator[ChatCompletionChunk]:
         param = coerce_chat_completions_param_any(msgspec.structs.replace(param, stream=stream))
-        data = _inject_thinking(self.encoder.encode(param))
+        data = _inject_thinking(self.encoder.encode(param), param)
         if stream:
             resp = await self.session.post(
                 "/chat/completions",