fix(llm): structured output for DeepSeek V4 and reasoner

DeepSeek V4 and reasoner reject tool_choice but accept tools.
Route via a per-model capability table that suppresses tool_choice
for thinking-mode models.

#678 #689

tradingagents/llm_clients/capabilities.py

@@ -0,0 +1,95 @@
+"""Declarative per-model capability table for OpenAI-compatible providers.
+
+This is the single place that knows which model IDs reject which API
+parameters or require which structured-output method. The LLM client
+subclasses consult ``get_capabilities(model_name)`` instead of hardcoding
+model-name ``if`` ladders, so adding a new model (or a new provider quirk)
+means editing this table — not the client code.
+
+Pattern adapted from the per-model ``compat:`` flags DeepSeek themselves
+publish in their integration guides (e.g. the Oh My Pi config schema
+documents ``supportsToolChoice``, ``requiresReasoningContentForToolCalls``
+as declarative per-model fields).
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from typing import Literal
+
+
+StructuredMethod = Literal[
+    "function_calling",  # uses tools; respects supports_tool_choice
+    "json_mode",         # uses response_format={"type":"json_object"}
+    "json_schema",       # uses response_format={"type":"json_schema",...}
+    "none",              # no structured output available; caller falls back to free-text
+]
+
+
+@dataclass(frozen=True)
+class ModelCapabilities:
+    """What an OpenAI-compatible model accepts at the API level."""
+
+    supports_tool_choice: bool
+    supports_json_mode: bool
+    supports_json_schema: bool
+    preferred_structured_method: StructuredMethod
+    # DeepSeek thinking-mode models 400 if reasoning_content from prior
+    # assistant turns is not echoed back on the next request.
+    requires_reasoning_content_roundtrip: bool = False
+
+
+# DeepSeek's thinking models accept the ``tools`` array but reject the
+# ``tool_choice`` parameter (official Oh My Pi integration guide and the
+# 400 response in issue #678). Their official tool-calling examples
+# (api-docs.deepseek.com/guides/tool_calls) pass ``tools=[...]`` without
+# ``tool_choice`` — we mirror that pattern by setting supports_tool_choice
+# to False and letting the client suppress the kwarg.
+_DEEPSEEK_THINKING = ModelCapabilities(
+    supports_tool_choice=False,
+    supports_json_mode=True,
+    supports_json_schema=False,
+    preferred_structured_method="function_calling",
+    requires_reasoning_content_roundtrip=True,
+)
+
+_DEEPSEEK_CHAT = ModelCapabilities(
+    supports_tool_choice=True,
+    supports_json_mode=True,
+    supports_json_schema=False,
+    preferred_structured_method="function_calling",
+)
+
+_DEFAULT = ModelCapabilities(
+    supports_tool_choice=True,
+    supports_json_mode=True,
+    supports_json_schema=True,
+    preferred_structured_method="function_calling",
+)
+
+
+# Exact-ID matches take precedence over pattern matches.
+_BY_ID: dict[str, ModelCapabilities] = {
+    "deepseek-chat": _DEEPSEEK_CHAT,
+    "deepseek-reasoner": _DEEPSEEK_THINKING,
+    "deepseek-v4-flash": _DEEPSEEK_THINKING,
+    "deepseek-v4-pro": _DEEPSEEK_THINKING,
+}
+
+# Forward-compat patterns. A new ``deepseek-v5-*`` or ``deepseek-reasoner-*``
+# variant inherits the thinking-mode quirks automatically.
+_BY_PATTERN: list[tuple[re.Pattern[str], ModelCapabilities]] = [
+    (re.compile(r"^deepseek-v\d"), _DEEPSEEK_THINKING),
+    (re.compile(r"^deepseek-reasoner"), _DEEPSEEK_THINKING),
+]
+
+
+def get_capabilities(model_name: str) -> ModelCapabilities:
+    """Resolve capabilities by exact ID, then pattern, then default."""
+    if model_name in _BY_ID:
+        return _BY_ID[model_name]
+    for pattern, caps in _BY_PATTERN:
+        if pattern.match(model_name):
+            return caps
+    return _DEFAULT

tradingagents/llm_clients/openai_client.py

@@ -5,30 +5,45 @@ from langchain_core.messages import AIMessage
 from langchain_openai import ChatOpenAI
 
 from .base_client import BaseLLMClient, normalize_content
+from .capabilities import get_capabilities
 from .validators import validate_model
 
 
 class NormalizedChatOpenAI(ChatOpenAI):
-    """ChatOpenAI with normalized content output.
+    """ChatOpenAI with normalized content output and capability-aware binding.
 
     The Responses API returns content as a list of typed blocks
     (reasoning, text, etc.). ``invoke`` normalizes to string for
-    consistent downstream handling. ``with_structured_output`` defaults
-    to function-calling so the Responses-API parse path is avoided
-    (langchain-openai's parse path emits noisy
-    PydanticSerializationUnexpectedValue warnings per call without
-    affecting correctness).
+    consistent downstream handling.
 
-    Provider-specific quirks (e.g. DeepSeek's thinking mode) live in
-    purpose-built subclasses below so this base class stays small.
+    ``with_structured_output`` consults the per-model capability table
+    (``capabilities.get_capabilities``) to pick the method and to decide
+    whether ``tool_choice`` may be sent. Models that reject ``tool_choice``
+    (e.g. DeepSeek V4 and reasoner — per their official tool-calling
+    guide) still bind the schema as a tool, but no ``tool_choice``
+    parameter is sent.
+
+    Provider-specific quirks beyond structured-output (e.g. DeepSeek's
+    reasoning_content roundtrip) live in subclasses so this base class
+    stays small.
     """
 
     def invoke(self, input, config=None, **kwargs):
         return normalize_content(super().invoke(input, config, **kwargs))
 
     def with_structured_output(self, schema, *, method=None, **kwargs):
-        if method is None:
-            method = "function_calling"
+        caps = get_capabilities(self.model_name)
+        if caps.preferred_structured_method == "none":
+            raise NotImplementedError(
+                f"{self.model_name} has no structured-output method available; "
+                f"agent factories will fall back to free-text generation."
+            )
+        method = method or caps.preferred_structured_method
+        # When the model rejects tool_choice, suppress langchain's hardcoded
+        # value. The schema is still bound as a tool — exactly what
+        # DeepSeek's official tool-calling examples do.
+        if method == "function_calling" and not caps.supports_tool_choice:
+            kwargs.setdefault("tool_choice", None)
         return super().with_structured_output(schema, method=method, **kwargs)
 
 
@@ -52,18 +67,16 @@ def _input_to_messages(input_: Any) -> list:
 class DeepSeekChatOpenAI(NormalizedChatOpenAI):
     """DeepSeek-specific overrides on top of the OpenAI-compatible client.
 
-    Two quirks that don't apply to other OpenAI-compatible providers:
+    Thinking-mode round-trip is the only DeepSeek-specific behavior that
+    stays here. When DeepSeek's thinking models return a response with
+    ``reasoning_content``, that field must be echoed back as part of the
+    assistant message on the next turn or the API fails with HTTP 400.
+    ``_create_chat_result`` captures it on receive and
+    ``_get_request_payload`` re-attaches it on send.
 
-    1. **Thinking-mode round-trip.** When DeepSeek's thinking models return
-       a response with ``reasoning_content``, that field must be echoed
-       back as part of the assistant message on the next turn or the API
-       fails with HTTP 400. ``_create_chat_result`` captures the field on
-       receive and ``_get_request_payload`` re-attaches it on send.
-
-    2. **deepseek-reasoner has no tool_choice.** Structured output via
-       function-calling is unavailable, so we raise NotImplementedError
-       and let the agent factories fall back to free-text generation
-       (see ``tradingagents/agents/utils/structured.py``).
+    Tool-choice handling for V4 and reasoner — those models reject the
+    ``tool_choice`` parameter — is handled by the capability dispatch in
+    ``NormalizedChatOpenAI.with_structured_output``, not here.
     """
 
     def _get_request_payload(self, input_, *, stop=None, **kwargs):
@@ -94,15 +107,6 @@ class DeepSeekChatOpenAI(NormalizedChatOpenAI):
                 generation.message.additional_kwargs["reasoning_content"] = reasoning
         return chat_result
 
-    def with_structured_output(self, schema, *, method=None, **kwargs):
-        if self.model_name == "deepseek-reasoner":
-            raise NotImplementedError(
-                "deepseek-reasoner does not support tool_choice; structured "
-                "output is unavailable. Agent factories fall back to "
-                "free-text generation automatically."
-            )
-        return super().with_structured_output(schema, method=method, **kwargs)
-
 # Kwargs forwarded from user config to ChatOpenAI
 _PASSTHROUGH_KWARGS = (
     "timeout", "max_retries", "reasoning_effort",