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",

tests/test_lmproto/test_deepseek_chat.py

@@ -0,0 +1,79 @@
+"""Tests for the DeepSeek chat-completions client's ``thinking`` flag."""
+
+from __future__ import annotations
+
+import json
+from typing import TYPE_CHECKING, Any, cast
+
+import pytest
+from msgspec import UNSET
+
+from plyngent.lmproto.deepseek import DeepseekOpenAIClient
+from plyngent.lmproto.openai_compatible.config import OpenAIConfig
+from plyngent.lmproto.openai_compatible.model import ChatCompletionsParam, UserChatMessage
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator
+
+    from plyngent.lmproto.openai_compatible.model import ReasoningEffort
+    from plyngent.typedef import Unset
+
+
+async def _posted_body(
+    monkeypatch: pytest.MonkeyPatch,
+    effort: ReasoningEffort | Unset = UNSET,
+) -> dict[str, Any]:
+    """Run one streamed request and return the JSON body the client posted."""
+    client = DeepseekOpenAIClient(OpenAIConfig(access_key_or_token="sk-test", base_url="https://api.deepseek.com"))
+    captured: dict[str, Any] = {}
+
+    class _Resp:
+        status_code = 200
+
+        async def iter_lines(self) -> AsyncIterator[bytes]:
+            yield b"data: [DONE]"
+
+        def close(self) -> None:
+            return None
+
+    async def fake_post(path: str, **kwargs: object) -> _Resp:
+        assert path == "/chat/completions"
+        captured["data"] = kwargs.get("data")
+        return _Resp()
+
+    monkeypatch.setattr(client.session, "post", fake_post)
+    param = ChatCompletionsParam(
+        model="deepseek-v4-pro",
+        messages=[UserChatMessage(content="hi")],
+        reasoning_effort=effort,
+    )
+    # Production dispatches the base param through the OpenAI-compatible surface
+    # (not the DeepSeek-specific annotation), so mirror that here.
+    cc: Any = client
+    stream = cast("AsyncIterator[Any]", await cc.chat_completions(param, stream=True))
+    async for _chunk in stream:
+        pass
+    data = captured["data"]
+    assert isinstance(data, bytes)
+    return json.loads(data)
+
+
+@pytest.mark.asyncio
+async def test_thinking_enabled_without_configured_effort(monkeypatch: pytest.MonkeyPatch) -> None:
+    body = await _posted_body(monkeypatch)
+    assert body["thinking"] == {"type": "enabled"}
+    assert "reasoning_effort" not in body
+
+
+@pytest.mark.asyncio
+async def test_thinking_enabled_with_effort(monkeypatch: pytest.MonkeyPatch) -> None:
+    body = await _posted_body(monkeypatch, "high")
+    assert body["thinking"] == {"type": "enabled"}
+    assert body["reasoning_effort"] == "high"
+
+
+@pytest.mark.asyncio
+async def test_thinking_disabled_for_none_effort(monkeypatch: pytest.MonkeyPatch) -> None:
+    body = await _posted_body(monkeypatch, "none")
+    assert body["thinking"] == {"type": "disabled"}
+    assert body["reasoning_effort"] == "none"