From fd541a8174f9d8ef81313604db793dc4839678e9 Mon Sep 17 00:00:00 2001
From: Milad Khoshdel <miladkhoshdel@gmail.com>
Date: Sat, 8 Aug 2026 16:06:06 +0330
Subject: [PATCH] refactor: extract reusable LLM client from chat CLI

- move OpenAI client and streaming logic into the LLM module
- add structured response and usage data classes
- support streamed output through an optional callback
- simplify the chat CLI to use the reusable LLM interface
---
 llm.py  | 92 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
 main.py | 79 +++++--------------------------------------------
 2 files changed, 98 insertions(+), 73 deletions(-)

diff --git a/llm.py b/llm.py
index 85f23ab..49d6899 100644
--- a/llm.py
+++ b/llm.py
@@ -1,5 +1,95 @@
 from dataclasses import dataclass
 
+from openai import OpenAI
+
+
+class LLM:
+    def __init__(self, settings, client=None):
+        self.settings = settings
+
+        if client is None:
+            client = OpenAI(
+                base_url=settings.base_url,
+                api_key=settings.api_key,
+            )
+
+        self.client = client
+
+    def create_stream(self, messages):
+        return self.client.chat.completions.create(
+            model=self.settings.model_name,
+            messages=messages,
+            max_tokens=self.settings.max_tokens,
+            temperature=self.settings.temperature,
+            top_p=self.settings.top_p,
+            stream=True,
+            stream_options={
+                "include_usage": self.settings.show_usage,
+            },
+            extra_body={
+                "reasoning": {
+                    "effort": "low",
+                    "exclude": True,
+                }
+            },
+        )
+
+    @staticmethod
+    def build_usage(usage):
+        if usage is None:
+            return None
+
+        completion_details = getattr(
+            usage,
+            "completion_tokens_details",
+            None,
+        )
+
+        return LLMUsage(
+            prompt_tokens=usage.prompt_tokens,
+            completion_tokens=usage.completion_tokens,
+            total_tokens=usage.total_tokens,
+            reasoning_tokens=getattr(
+                completion_details,
+                "reasoning_tokens",
+                None,
+            ),
+            cost=getattr(usage, "cost", None),
+        )
+
+    def build_response(self, answer, finish_reason, usage):
+        return LLMResponse(
+            content="".join(answer),
+            finish_reason=finish_reason,
+            usage=self.build_usage(usage),
+        )
+
+    def complete(self, messages, on_text=None):
+        usage = None
+        finish_reason = None
+        answer = []
+
+        with self.create_stream(messages) as stream:
+            for chunk in stream:
+                if chunk.usage is not None:
+                    usage = chunk.usage
+
+                if not chunk.choices:
+                    continue
+
+                choice = chunk.choices[0]
+                content = choice.delta.content
+
+                if content:
+                    answer.append(content)
+                    if on_text is not None:
+                        on_text(content)
+
+                if choice.finish_reason is not None:
+                    finish_reason = choice.finish_reason
+
+        return self.build_response(answer, finish_reason, usage)
+
 
 @dataclass(frozen=True)
 class LLMUsage:
@@ -17,4 +107,4 @@ class LLMResponse:
     usage: LLMUsage | None = None
 
 
-__all__ = ["LLMResponse", "LLMUsage"]
+__all__ = ["LLM", "LLMResponse", "LLMUsage"]
diff --git a/main.py b/main.py
index 64cd2dd..fe65043 100644
--- a/main.py
+++ b/main.py
@@ -1,6 +1,6 @@
-from openai import OpenAI, OpenAIError
+from openai import OpenAIError
 
-from llm import LLMResponse, LLMUsage
+from llm import LLM
 from settings import load_settings
 from utils import print_usage, read_input
 
@@ -25,10 +25,7 @@ def main():
         print(f"Configuration error: {error}")
         return 1
 
-    client = OpenAI(
-        base_url=settings.base_url,
-        api_key=settings.api_key,
-    )
+    model = LLM(settings)
 
     while True:
         system_rule = read_input("System rule: ")
@@ -65,78 +62,16 @@ def main():
             }
         )
 
-        usage = None
-        finish_reason = None
-        answer = []
-
         try:
-            with client.chat.completions.create(
-                model=settings.model_name,
-                messages=messages,
-                max_tokens=settings.max_tokens,
-                temperature=settings.temperature,
-                top_p=settings.top_p,
-                stream=True,
-                stream_options={"include_usage": settings.show_usage},
-                extra_body={
-                    "reasoning": {
-                        "effort": "low",
-                        "exclude": True,
-                    }
-                },
-            ) as stream:
-
-                for chunk in stream:
-
-                    if chunk.usage is not None:
-                        usage = chunk.usage
-
-                    if not chunk.choices:
-                        continue
-
-                    choice = chunk.choices[0]
-                    content = choice.delta.content
-
-                    if content:
-                        answer.append(content)
-                        print(content, end="", flush=True)
-
-                    if choice.finish_reason is not None:
-                        finish_reason = choice.finish_reason
-
+            response = model.complete(
+                messages,
+                on_text=lambda text: print(text, end="", flush=True),
+            )
         except OpenAIError as error:
             messages.pop()
             print(f"\nRequest failed: {error}")
             continue
 
-        response_usage = None
-
-        if usage is not None:
-
-            completion_details = getattr(
-                usage,
-                "completion_tokens_details",
-                None,
-            )
-
-            response_usage = LLMUsage(
-                prompt_tokens=usage.prompt_tokens,
-                completion_tokens=usage.completion_tokens,
-                total_tokens=usage.total_tokens,
-                reasoning_tokens=getattr(
-                    completion_details,
-                    "reasoning_tokens",
-                    None,
-                ),
-                cost=getattr(usage, "cost", None),
-            )
-
-        response = LLMResponse(
-            content="".join(answer),
-            finish_reason=finish_reason,
-            usage=response_usage,
-        )
-
         if response.finish_reason == "length":
             print("Warning: the answer may be incomplete.")