Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(llm): route configured reasoning effort through Responses API

cgycorey8a66de1190f4
diff --git a/README.md b/README.md
index 3aa58dfde..42a28780b 100644
--- a/README.md
+++ b/README.md
@@ -832,6 +832,8 @@ Copy `agent/.env.example` to `agent/.env` and uncomment the provider block you w
| `<PROVIDER>_API_KEY` | Yes* | API key (`OPENROUTER_API_KEY`, `DEEPSEEK_API_KEY`, etc.) |
| `<PROVIDER>_BASE_URL` | Yes | API endpoint URL |
| `LANGCHAIN_MODEL_NAME` | Yes | Model name (e.g. `deepseek-v4-pro`) |
+| `LANGCHAIN_REASONING_EFFORT` | No | Reasoning effort (`none`, `low`, `medium`, `high`, or `max`) |
+| `LANGCHAIN_USE_RESPONSES_API` | No | Responses transport override: unset auto-selects it when reasoning effort is set, `true` forces it, and `false` keeps legacy Chat Completions |
| `TUSHARE_TOKEN` | No | Tushare Pro token for A-share data (falls back to AKShare) |
| `TIMEOUT_SECONDS` | No | LLM call timeout, default 120s |
| `API_AUTH_KEY` | Recommended for network deployments | Bearer token required when the API is reachable from non-local clients |
diff --git a/agent/.env.example b/agent/.env.example
index e844e6e3c..b09b6e3c1 100644
--- a/agent/.env.example
+++ b/agent/.env.example
@@ -138,14 +138,15 @@ MAX_RETRIES=2
# explicit direct transports still honor SSL_CERT_FILE/SSL_CERT_DIR.
# VIBE_TRADING_DISABLE_HTTP_PROXY=0
-# Reasoning effort: none / low / medium / high / max. Two delivery paths:
-# * Relays requiring opt-in (OpenRouter, Requesty) get extra_body.reasoning.effort.
-# * Direct OpenAI gets a top-level reasoning_effort — gpt-5.6-* models require an
-# explicit value (including "none") before they accept function tools.
-# Leave unset for every other provider: Moonshot / DeepSeek official APIs return
-# reasoning_content by default, and other OpenAI-compatible providers may reject
-# the field outright.
+# Reasoning effort: none / low / medium / high / max.
+# ChatOpenAI-compatible providers use /v1/responses automatically when an effort
+# is set and send reasoning.effort. Native Moonshot / DeepSeek adapters keep
+# their provider-specific reasoning behavior.
# LANGCHAIN_REASONING_EFFORT=medium
+# When an effort is configured, the ChatOpenAI-compatible adapter automatically
+# uses /v1/responses and sends reasoning.effort. Set this to false only for a
+# legacy endpoint that supports Chat Completions but not Responses.
+# LANGCHAIN_USE_RESPONSES_API=false
# ============================================================================
# Data Sources (DataConfig)
diff --git a/agent/src/config/env_schema.py b/agent/src/config/env_schema.py
index d371e32fd..37062294f 100644
--- a/agent/src/config/env_schema.py
+++ b/agent/src/config/env_schema.py
@@ -133,6 +133,9 @@ class LLMConfig(_EnvBase):
timeout_seconds: int = Field(alias="TIMEOUT_SECONDS", default=120)
max_retries: int = Field(alias="MAX_RETRIES", default=2)
langchain_reasoning_effort: str = Field(alias="LANGCHAIN_REASONING_EFFORT", default="")
+ langchain_use_responses_api: EnvBool | None = Field(
+ alias="LANGCHAIN_USE_RESPONSES_API", default=None
+ )
vibe_trading_deepseek_adapter: str = Field(alias="VIBE_TRADING_DEEPSEEK_ADAPTER", default="auto")
moonshot_user_agent: str = Field(alias="MOONSHOT_USER_AGENT", default="")
openai_codex_base_url: str = Field(
diff --git a/agent/src/providers/llm.py b/agent/src/providers/llm.py
index 425cb4bb3..d62ba1e32 100644
--- a/agent/src/providers/llm.py
+++ b/agent/src/providers/llm.py
@@ -1253,9 +1253,14 @@ def build_llm(*, model_name: Optional[str] = None, callbacks: Any = None) -> Any
):
logger.info("Forcing temperature=1.0 for %s (provider requirement)", name)
temperature = 1.0
- # Optional reasoning activation for relays requiring opt-in (e.g. OpenRouter).
- # Moonshot/DeepSeek official APIs emit reasoning by default and ignore this field.
effort = get_env_config().llm.langchain_reasoning_effort.strip().lower()
+ # Moonshot/DeepSeek official APIs emit reasoning by default and ignore this field.
+ configured_responses_api = get_env_config().llm.langchain_use_responses_api
+ use_responses_api = (
+ configured_responses_api
+ if configured_responses_api is not None
+ else bool(effort)
+ )
creds = get_llm_credentials(provider, name)
api_key = creds["api_key"]
_validate_authorization_credential(
@@ -1270,9 +1275,16 @@ def build_llm(*, model_name: Optional[str] = None, callbacks: Any = None) -> Any
"timeout": get_env_config().llm.timeout_seconds,
"max_retries": get_env_config().llm.max_retries,
"callbacks": callbacks,
+ "output_version": "responses/v1" if use_responses_api else None,
+ "use_responses_api": use_responses_api,
+ "reasoning": (
+ {"effort": effort}
+ if use_responses_api and effort
+ else None
+ ),
"extra_body": (
{"reasoning": {"effort": effort}}
- if effort and caps.openrouter_reasoning_body
+ if effort and not use_responses_api and caps.openrouter_reasoning_body
else None
),
# Direct OpenAI takes the effort as a top-level request field instead
@@ -1281,7 +1293,11 @@ def build_llm(*, model_name: Optional[str] = None, callbacks: Any = None) -> Any
# payload without the field.
"reasoning_effort": (
effort
- if effort and _supports_top_level_reasoning_effort(provider, caps.name)
+ if (
+ effort
+ and not use_responses_api
+ and _supports_top_level_reasoning_effort(provider, caps.name)
+ )
else None
),
"vibe_provider": provider,
diff --git a/agent/tests/test_env_schema.py b/agent/tests/test_env_schema.py
index 562b4bbbc..0e1f20604 100644
--- a/agent/tests/test_env_schema.py
+++ b/agent/tests/test_env_schema.py
@@ -78,10 +78,22 @@ class TestEnvConfigDefaults:
assert c.llm.vibe_trading_disable_http_proxy is False
assert c.llm.max_retries == 2
assert c.llm.langchain_reasoning_effort == ""
+ assert c.llm.langchain_use_responses_api is None
assert c.llm.vibe_trading_deepseek_adapter == "auto"
assert c.llm.moonshot_user_agent == ""
assert c.llm.openai_codex_base_url == "https://chatgpt.com/backend-api/codex/responses"
+ def test_llm_responses_api_reads_boolean_env(self, monkeypatch: pytest.MonkeyPatch) -> None:
+ monkeypatch.setenv("LANGCHAIN_USE_RESPONSES_API", "true")
+ reset_env_config()
+
+ assert EnvConfig().llm.langchain_use_responses_api is True
+
+ monkeypatch.setenv("LANGCHAIN_USE_RESPONSES_API", "false")
+ reset_env_config()
+
+ assert EnvConfig().llm.langchain_use_responses_api is False
+
def test_data_defaults(self) -> None:
c = EnvConfig()
assert c.data.tushare_token == ""
diff --git a/agent/tests/test_llm.py b/agent/tests/test_llm.py
index 3a0581564..4640d16ec 100644
--- a/agent/tests/test_llm.py
+++ b/agent/tests/test_llm.py
@@ -647,7 +647,8 @@ class TestKimiTemperature:
class TestReasoningEffortPassthrough:
"""LANGCHAIN_REASONING_EFFORT is forwarded as extra_body.reasoning.effort
to the underlying OpenAI-compatible client. Used for OpenRouter-style
- relays that require opt-in to enable thinking."""
+ relays that require opt-in to enable thinking when Chat Completions is
+ selected explicitly."""
def _capture(self, env: dict[str, str]) -> dict:
import src.providers.llm as llm_mod
@@ -683,6 +684,7 @@ class TestReasoningEffortPassthrough:
"OPENROUTER_BASE_URL": "https://openrouter.ai/api/v1",
"LANGCHAIN_MODEL_NAME": "moonshotai/kimi-k2-thinking",
"LANGCHAIN_REASONING_EFFORT": "medium",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
assert captured["extra_body"] == {"reasoning": {"effort": "medium"}}
@@ -695,6 +697,7 @@ class TestReasoningEffortPassthrough:
"OPENROUTER_BASE_URL": "https://openrouter.ai/api/v1",
"LANGCHAIN_MODEL_NAME": "moonshotai/kimi-k2-thinking",
"LANGCHAIN_REASONING_EFFORT": "HIGH",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
assert captured["extra_body"]["reasoning"]["effort"] == "high"
diff --git a/agent/tests/test_llm_reasoning_effort.py b/agent/tests/test_llm_reasoning_effort.py
index 4f7cbeba9..4057e73d6 100644
--- a/agent/tests/test_llm_reasoning_effort.py
+++ b/agent/tests/test_llm_reasoning_effort.py
@@ -11,9 +11,9 @@ The setting has two mutually exclusive delivery paths:
in /v1/chat/completions. To use function tools, use /v1/responses or set
reasoning_effort to 'none'.
-Every other OpenAI-compatible provider must receive neither, because a stale
-global effort value would otherwise ride along on payloads that can reject the
-unknown field (DeepSeek is the documented case in ``test_provider_diagnostics``).
+Any ChatOpenAI-compatible provider/model with an effort and no explicit
+transport setting uses the Responses API; explicit ``false`` retains the
+legacy Chat Completions path for endpoints that do not expose Responses.
"""
from __future__ import annotations
@@ -121,6 +121,7 @@ class TestDirectOpenAI:
"OPENAI_API_KEY": "sk-test",
"LANGCHAIN_MODEL_NAME": "gpt-5.6-sol",
"LANGCHAIN_REASONING_EFFORT": "none",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -134,6 +135,7 @@ class TestDirectOpenAI:
"OPENAI_API_KEY": "sk-test",
"LANGCHAIN_MODEL_NAME": "gpt-5.6-sol",
"LANGCHAIN_REASONING_EFFORT": "high",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -165,6 +167,7 @@ class TestUnsupportedProviders:
"LANGCHAIN_MODEL_NAME": "deepseek-v4-pro",
"LANGCHAIN_REASONING_EFFORT": "high",
"VIBE_TRADING_DEEPSEEK_ADAPTER": "openai-compatible",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -180,6 +183,7 @@ class TestUnsupportedProviders:
"GEMINI_BASE_URL": "https://generativelanguage.googleapis.com/v1beta/openai",
"LANGCHAIN_MODEL_NAME": "gemini-3.5-flash",
"LANGCHAIN_REASONING_EFFORT": "high",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -195,6 +199,7 @@ class TestUnsupportedProviders:
"LANGCHAIN_MODEL_NAME": "deepseek-v4-pro",
"LANGCHAIN_REASONING_EFFORT": "high",
"VIBE_TRADING_DEEPSEEK_ADAPTER": "openai-compatible",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -208,6 +213,7 @@ class TestUnsupportedProviders:
"OPENAI_API_KEY": "sk-test",
"LANGCHAIN_MODEL_NAME": "house-model-1",
"LANGCHAIN_REASONING_EFFORT": "high",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -225,6 +231,7 @@ class TestRelayOptIn:
"OPENROUTER_BASE_URL": "https://openrouter.ai/api/v1",
"LANGCHAIN_MODEL_NAME": "deepseek/deepseek-v4-pro",
"LANGCHAIN_REASONING_EFFORT": "high",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -239,6 +246,7 @@ class TestRelayOptIn:
"REQUESTY_BASE_URL": "https://router.requesty.ai/v1",
"LANGCHAIN_MODEL_NAME": "openai/gpt-4o-mini",
"LANGCHAIN_REASONING_EFFORT": "medium",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
}
)
@@ -268,6 +276,73 @@ class TestRequestPayload:
assert "reasoning_effort" not in self._payload(None)
+class TestResponsesAPI:
+ def test_reasoning_effort_auto_selects_responses_for_any_chat_model(self) -> None:
+ kwargs = _capture_kwargs(
+ {
+ "LANGCHAIN_PROVIDER": "some-openai-compatible-gateway",
+ "OPENAI_API_KEY": "sk-test",
+ "OPENAI_BASE_URL": "https://gateway.example/v1",
+ "LANGCHAIN_MODEL_NAME": "arbitrary-reasoning-model",
+ "LANGCHAIN_REASONING_EFFORT": "high",
+ }
+ )
+
+ assert kwargs["use_responses_api"] is True
+ assert kwargs["output_version"] == "responses/v1"
+ assert kwargs["reasoning"] == {"effort": "high"}
+ assert kwargs["reasoning_effort"] is None
+
+ def test_any_provider_and_model_can_opt_into_responses_reasoning(self) -> None:
+ kwargs = _capture_kwargs(
+ {
+ "LANGCHAIN_PROVIDER": "openai",
+ "OPENAI_API_KEY": "sk-test",
+ "OPENAI_BASE_URL": "https://gateway.example/v1",
+ "LANGCHAIN_MODEL_NAME": "arbitrary-reasoning-model",
+ "LANGCHAIN_REASONING_EFFORT": "high",
+ "LANGCHAIN_USE_RESPONSES_API": "true",
+ }
+ )
+
+ assert kwargs["use_responses_api"] is True
+ assert kwargs["output_version"] == "responses/v1"
+ assert kwargs["reasoning"] == {"effort": "high"}
+ assert kwargs["reasoning_effort"] is None
+
+ def test_explicit_chat_mode_remains_available_for_legacy_endpoints(self) -> None:
+ kwargs = _capture_kwargs(
+ {
+ "LANGCHAIN_PROVIDER": "some-openai-compatible-gateway",
+ "OPENAI_API_KEY": "sk-test",
+ "OPENAI_BASE_URL": "https://gateway.example/v1",
+ "LANGCHAIN_MODEL_NAME": "arbitrary-reasoning-model",
+ "LANGCHAIN_REASONING_EFFORT": "high",
+ "LANGCHAIN_USE_RESPONSES_API": "false",
+ }
+ )
+
+ assert kwargs["use_responses_api"] is False
+ assert kwargs["reasoning"] is None
+
+ def test_responses_payload_does_not_assume_chat_messages(self) -> None:
+ if llm_mod.ChatOpenAIWithReasoning is None:
+ pytest.skip("langchain-openai is not installed")
+ from langchain_core.messages import HumanMessage
+
+ instance = llm_mod.ChatOpenAIWithReasoning(
+ model="arbitrary-reasoning-model",
+ api_key="sk-test",
+ use_responses_api=True,
+ output_version="responses/v1",
+ reasoning={"effort": "high"},
+ )
+
+ payload = instance._get_request_payload([HumanMessage(content="hi")])
+
+ assert payload["reasoning"] == {"effort": "high"}
+
+
class TestSettingsAllowlist:
"""The settings API admits the explicit 'none' value and says so."""