From de2c5db39855e7e619c8e2d6098ad30b599f7084 Mon Sep 17 00:00:00 2001
From: Patrick Erichsen <patrick.a.erichsen@gmail.com>
Date: Wed, 23 Sep 2026 22:17:13 -0700
Subject: [PATCH] fix(skill-scan): forward optional reasoning effort to the
 model API

---
 skill-scan/README.md                   |  1 +
 skill-scan/README_zh.md                |  1 +
 skill-scan/pytests/test_llm_request.py | 24 ++++++++++++++++++++++++
 skill-scan/skill_scan/utils/llm.py     |  6 ++++++
 4 files changed, 32 insertions(+)

diff --git a/skill-scan/README.md b/skill-scan/README.md
index 1a7cf86d0..03b9ddea4 100644
--- a/skill-scan/README.md
+++ b/skill-scan/README.md
@@ -144,6 +144,7 @@ Key environment variables:
 | `LLM_MODEL` / `OPENAI_MODEL` | Default model | `deepseek-v4-flash` |
 | `LLM_BASE_URL` / `OPENAI_BASE_URL` | Default base URL | `https://openrouter.ai/api/v1` |
 | `DEFAULT_MODEL_CONTEXT_WINDOW` | Main model context window | `128000` |
+| `REASONING_EFFORT` | Optional reasoning effort (for example, `high`); supported values depend on the model and provider. Unset or blank values omit the API parameter. | Unset |
 | `LOG_LEVEL` | Log level | `INFO` |
 
 ---
diff --git a/skill-scan/README_zh.md b/skill-scan/README_zh.md
index 0dac273f5..626159090 100644
--- a/skill-scan/README_zh.md
+++ b/skill-scan/README_zh.md
@@ -144,6 +144,7 @@ asyncio.run(run())
 | `LLM_MODEL` / `OPENAI_MODEL` | 默认模型 | `deepseek-v4-flash` |
 | `LLM_BASE_URL` / `OPENAI_BASE_URL` | 默认 Base URL | `https://openrouter.ai/api/v1` |
 | `DEFAULT_MODEL_CONTEXT_WINDOW` | 主模型上下文窗口 | `128000` |
+| `REASONING_EFFORT` | 可选推理强度（例如 `high`）；支持的值取决于模型和服务提供商。未设置或为空白时不发送该 API 参数。 | 未设置 |
 | `LOG_LEVEL` | 日志级别 | `INFO` |
 
 ---
diff --git a/skill-scan/pytests/test_llm_request.py b/skill-scan/pytests/test_llm_request.py
index 4c1ff6734..15f3999c8 100644
--- a/skill-scan/pytests/test_llm_request.py
+++ b/skill-scan/pytests/test_llm_request.py
@@ -6,6 +6,30 @@
 from skill_scan.utils.llm import LLM
 
 
+@pytest.mark.parametrize(
+    ("setting", "expected"),
+    [("high", "high"), (" medium ", "medium"), (None, None), ("", None), (" \t ", None)],
+)
+def test_chat_request_uses_optional_reasoning_effort(monkeypatch, setting, expected):
+    if setting is None:
+        monkeypatch.delenv("REASONING_EFFORT", raising=False)
+    else:
+        monkeypatch.setenv("REASONING_EFFORT", setting)
+    create = Mock(return_value=[])
+    llm = object.__new__(LLM)
+    llm.model = "test-reasoning-model"
+    llm.client = SimpleNamespace(
+        chat=SimpleNamespace(completions=SimpleNamespace(create=create))
+    )
+
+    llm.chat_stream([{"role": "user", "content": "test"}])
+
+    if expected is None:
+        assert "reasoning_effort" not in create.call_args.kwargs
+    else:
+        assert create.call_args.kwargs["reasoning_effort"] == expected
+
+
 def test_chat_request_omits_temperature():
     create = Mock(return_value=[])
     llm = object.__new__(LLM)
diff --git a/skill-scan/skill_scan/utils/llm.py b/skill-scan/skill_scan/utils/llm.py
index e4dc57f8e..e36bb57eb 100644
--- a/skill-scan/skill_scan/utils/llm.py
+++ b/skill-scan/skill_scan/utils/llm.py
@@ -22,6 +22,7 @@
 
 import openai
 
+from skill_scan.utils.config import get_env
 from skill_scan.utils.loging import logger
 
 
@@ -74,11 +75,16 @@ def chat(self, message: list[dict], p=False, ret_usage=False) -> str | tuple[str
 
     def chat_stream(self, message: list[dict]) -> tuple[str, dict]:
         started = time.monotonic()
+        request_options = {}
+        reasoning_effort = (get_env("REASONING_EFFORT") or "").strip()
+        if reasoning_effort:
+            request_options["reasoning_effort"] = reasoning_effort
         response = self.client.chat.completions.create(
             model=self.model,
             messages=message,
             stream=True,
             stream_options={"include_usage": True},
+            **request_options,
         )
 
         ret = ""