vLLM 모델 해석 및 max_tokens 상한 보정

vLLM 연결 시 등록 모델 alias와 실제 모델 ID가 섞여 payload로 전달되던 경로를 보정해 RegisteredModel에서 실제 모델명을 우선 찾아 요청에 사용하도록 수정했다.

OpenAI-compatible 일반 대화와 도구 호출 모두 vLLM 서버 허용 범위를 넘지 않도록 max_tokens를 자동 보정하도록 통일했다.

검증: dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify\\ -p:IntermediateOutputPath=obj\\verify\\ (경고 0, 오류 0)

src/AxCopilot/Services/LlmService.ToolUse.cs

@@ -225,7 +225,7 @@ public partial class LlmService
             return new
             {
                 model = activeModel,
-                max_tokens = Math.Max(llm.MaxContextTokens, 4096),
+                max_tokens = ResolveOpenAiCompatibleMaxTokens(),
                 temperature = llm.Temperature,
                 system = systemPrompt,
                 messages = msgs,
@@ -237,7 +237,7 @@ public partial class LlmService
         return new
         {
             model = activeModel,
-            max_tokens = Math.Max(llm.MaxContextTokens, 4096),
+            max_tokens = ResolveOpenAiCompatibleMaxTokens(),
             temperature = llm.Temperature,
             messages = msgs,
             tools = toolDefs,
@@ -661,7 +661,7 @@ public partial class LlmService
             ["tools"] = toolDefs,
             ["stream"] = false,
             ["temperature"] = ResolveTemperature(),
-            ["max_tokens"] = llm.MaxContextTokens,
+            ["max_tokens"] = ResolveOpenAiCompatibleMaxTokens(),
         };
         var effort = ResolveReasoningEffort();
         if (!string.IsNullOrWhiteSpace(effort))

src/AxCopilot/Services/LlmService.cs

@@ -249,10 +249,32 @@ public partial class LlmService : IDisposable
         var llm = _settings.Settings.Llm;
         var service = NormalizeServiceName(llm.Service);
         if (service is "ollama" or "vllm" && !string.IsNullOrEmpty(llm.Model))
+        {
+            var registered = FindRegisteredModel(llm, service, llm.Model);
+            if (registered != null)
+            {
+                var registeredModelName = CryptoService.DecryptIfEnabled(registered.EncryptedModelName, llm.EncryptionEnabled);
+                if (!string.IsNullOrWhiteSpace(registeredModelName))
+                    return registeredModelName;
+            }
+
             return CryptoService.DecryptIfEnabled(llm.Model, llm.EncryptionEnabled);
+        }
         return llm.Model;
     }
 
+    private int ResolveOpenAiCompatibleMaxTokens()
+    {
+        var llm = _settings.Settings.Llm;
+        var requested = Math.Clamp(llm.MaxContextTokens, 1, 1_000_000);
+        var service = NormalizeServiceName(llm.Service);
+
+        if (service == "vllm")
+            return Math.Min(requested, 8192);
+
+        return requested;
+    }
+
     /// <summary>
     /// 현재 활성 모델에 매칭되는 RegisteredModel을 찾아 엔드포인트/API키를 반환합니다.
     /// RegisteredModel에 전용 서버 정보가 있으면 그것을 사용하고, 없으면 기본 설정을 사용합니다.
@@ -648,7 +670,7 @@ public partial class LlmService : IDisposable
             ["messages"] = msgs,
             ["stream"] = stream,
             ["temperature"] = ResolveTemperature(),
-            ["max_tokens"] = llm.MaxContextTokens
+            ["max_tokens"] = ResolveOpenAiCompatibleMaxTokens()
         };
         var effort = ResolveReasoningEffort();
         if (!string.IsNullOrWhiteSpace(effort))