IBM vLLM 배포형 채팅 요청 스키마 분기와 문서 반영

IBM/CP4D 인증을 사용하는 vLLM 등록 모델에서 배포형 /ml/v1/deployments/.../text/chat 계열 엔드포인트를 감지하도록 정리했다.

일반 OpenAI 호환 body 대신 messages+parameters 형태의 IBM deployment chat body를 사용하고 /v1/chat/completions를 강제로 붙이지 않도록 수정했다.

IBM 배포형 응답은 results.generated_text, output_text, choices.message.content를 함께 파싱하도록 보강했고 도구 호출 경로는 안전하게 일반 응답 폴백을 유도하도록 정리했다.

README와 DEVELOPMENT 문서를 2026-04-06 18:02 (KST) 기준으로 갱신했고 dotnet build 검증에서 경고 0 / 오류 0을 확인했다.

README.md

@@ -1311,3 +1311,7 @@ MIT License
 - 업데이트: 2026-04-06 17:52 (KST)
   - 런처 표시 체감 속도를 유지하기 위해 [LauncherWindow](/E:/AX%20Copilot%20-%20Codex/src/AxCopilot/Views/LauncherWindow.xaml.cs) 사전 생성은 다시 복원했다. 대신 무거운 후보를 색인으로 더 좁히기 위해, [App.xaml.cs](/E:/AX%20Copilot%20-%20Codex/src/AxCopilot/App.xaml.cs)의 인덱스 워밍업 진입점을 런처 표시 시점이 아니라 실제 검색 시점으로 옮겼다.
   - [LauncherViewModel.cs](/E:/AX%20Copilot%20-%20Codex/src/AxCopilot/ViewModels/LauncherViewModel.cs) 의 `SearchAsync(...)` 시작 시에만 `EnsureIndexWarmupStarted()`를 호출하도록 바꿔, 사용자가 런처를 단순 호출만 할 때는 전체 인덱스 스캔과 파일 감시가 돌지 않게 정리했다.
+- 업데이트: 2026-04-06 18:02 (KST)
+  - IBM 연결형 vLLM에서 `model_id` 또는 `mode`를 body에 넣지 말라는 응답이 오던 문제를 수정했다. [LlmService.cs](/E:/AX%20Copilot%20-%20Codex/src/AxCopilot/Services/LlmService.cs)에 IBM/CP4D 인증 + `/ml/v1/deployments/.../text/chat` 계열 엔드포인트를 감지하는 분기를 추가하고, 이 경우 일반 OpenAI 호환 body 대신 `messages + parameters` 형태의 IBM deployment chat body를 사용하도록 바꿨다.
+  - 같은 파일에서 IBM deployment chat 경로는 `/v1/chat/completions`를 더 이상 강제로 붙이지 않고, 스트리밍 여부에 따라 `/text/chat` 또는 `/text/chat_stream` URL을 사용하도록 정리했다. 응답 파싱도 `results[].generated_text`, `output_text`, `choices[].message.content`를 함께 지원하게 확장했다.
+  - [LlmService.ToolUse.cs](/E:/AX%20Copilot%20-%20Codex/src/AxCopilot/Services/LlmService.ToolUse.cs) 에서는 IBM deployment chat API가 감지되면 OpenAI function-calling body를 그대로 보내지 않고 `ToolCallNotSupportedException`으로 일반 응답 경로 폴백을 유도하도록 안전장치를 추가했다.

docs/DEVELOPMENT.md

@@ -4994,3 +4994,6 @@ ow + toggle ?쒓컖 ?몄뼱濡??ㅼ떆 ?뺣젹?덈떎.
 - Document update: 2026-04-06 17:43 (KST) - Removed the eager tray-menu warmup path (`PrepareForDisplay()`) from startup. This avoids doing popup layout/measure work for the tray menu until the user actually opens it, reducing idle desktop overhead further.
 - Document update: 2026-04-06 17:52 (KST) - Restored eager `LauncherWindow` construction so launcher open latency stays low, but kept the tray-menu warmup removed. This keeps the launcher responsive while continuing to trim startup work elsewhere.
 - Document update: 2026-04-06 17:52 (KST) - Moved index warmup from launcher-show time to actual search time. `LauncherViewModel.SearchAsync(...)` now triggers `App.EnsureIndexWarmupStarted()` only when the user performs a real search, so opening the launcher by itself no longer starts a full file scan and watcher bootstrap.
+- Document update: 2026-04-06 18:02 (KST) - Added IBM deployment-chat detection in `LlmService.cs` for vLLM registered models using `ibm_iam` or `cp4d_*` auth with `/ml/v1/deployments/...`-style endpoints. These requests now use IBM deployment chat URLs (`/text/chat` or `/text/chat_stream`) instead of appending `/v1/chat/completions`.
+- Document update: 2026-04-06 18:02 (KST) - Added an IBM deployment request body builder in `LlmService.cs` that omits OpenAI-style `model` and `stream` fields and sends `messages + parameters` instead. This directly addresses IBM responses complaining that `model_id` or `mode` must not be specified in the request body.
+- Document update: 2026-04-06 18:02 (KST) - Hardened vLLM response handling for IBM deployment endpoints by accepting `results[].generated_text`, `output_text`, and `choices[].message.content`, and by short-circuiting tool-use requests in `LlmService.ToolUse.cs` with a `ToolCallNotSupportedException` so IBM deployment chat connections do not receive an incompatible OpenAI function-calling payload.

src/AxCopilot/Services/LlmService.ToolUse.cs

@@ -431,13 +431,20 @@ public partial class LlmService
         List<ChatMessage> messages, IReadOnlyCollection<IAgentTool> tools, CancellationToken ct)
     {
         var activeService = ResolveService();
-        var body = BuildOpenAiToolBody(messages, tools);
 
         // 등록 모델의 커스텀 엔드포인트 우선 사용 (ResolveServerInfo)
         var (resolvedEp, _, allowInsecureTls) = ResolveServerInfo();
         var endpoint = string.IsNullOrEmpty(resolvedEp)
             ? ResolveEndpointForService(activeService)
             : resolvedEp;
+        var registered = GetActiveRegisteredModel();
+        if (UsesIbmDeploymentChatApi(activeService, registered, endpoint))
+        {
+            throw new ToolCallNotSupportedException(
+                "IBM 배포형 vLLM 연결은 OpenAI 도구 호출 형식과 다를 수 있어 일반 대화 경로로 폴백합니다.");
+        }
+
+        var body = BuildOpenAiToolBody(messages, tools);
 
         var url = activeService.ToLowerInvariant() == "ollama"
             ? endpoint.TrimEnd('/') + "/api/chat"

src/AxCopilot/Services/LlmService.cs

@@ -320,6 +320,103 @@ public partial class LlmService : IDisposable
              m.Alias == modelName));
     }
 
+    private Models.RegisteredModel? GetActiveRegisteredModel()
+    {
+        var llm = _settings.Settings.Llm;
+        return FindRegisteredModel(llm, ResolveService(), ResolveModel());
+    }
+
+    private static bool UsesIbmDeploymentChatApi(string service, Models.RegisteredModel? registered, string? endpoint)
+    {
+        if (!string.Equals(NormalizeServiceName(service), "vllm", StringComparison.OrdinalIgnoreCase))
+            return false;
+        if (registered == null)
+            return false;
+
+        var authType = (registered.AuthType ?? "").Trim().ToLowerInvariant();
+        if (authType is not ("ibm_iam" or "cp4d" or "cp4d_password" or "cp4d_api_key"))
+            return false;
+
+        var normalizedEndpoint = (endpoint ?? "").Trim().ToLowerInvariant();
+        return normalizedEndpoint.Contains("/ml/") ||
+               normalizedEndpoint.Contains("/deployments/") ||
+               normalizedEndpoint.Contains("/text/chat");
+    }
+
+    private string BuildIbmDeploymentChatUrl(string endpoint, bool stream)
+    {
+        var trimmed = (endpoint ?? "").Trim();
+        if (string.IsNullOrWhiteSpace(trimmed))
+            throw new InvalidOperationException("IBM 배포형 vLLM 엔드포인트가 비어 있습니다.");
+
+        var normalized = trimmed.ToLowerInvariant();
+        if (normalized.Contains("/text/chat_stream"))
+            return stream ? trimmed : trimmed.Replace("/text/chat_stream", "/text/chat", StringComparison.OrdinalIgnoreCase);
+        if (normalized.Contains("/text/chat"))
+            return stream ? trimmed.Replace("/text/chat", "/text/chat_stream", StringComparison.OrdinalIgnoreCase) : trimmed;
+        if (normalized.Contains("/deployments/"))
+            return trimmed.TrimEnd('/') + (stream ? "/text/chat_stream" : "/text/chat");
+
+        return trimmed;
+    }
+
+    private object BuildIbmDeploymentBody(List<ChatMessage> messages)
+    {
+        var msgs = new List<object>();
+        if (!string.IsNullOrWhiteSpace(_systemPrompt))
+            msgs.Add(new { role = "system", content = _systemPrompt });
+
+        foreach (var m in messages)
+        {
+            if (m.Role == "system")
+                continue;
+
+            msgs.Add(new
+            {
+                role = m.Role == "assistant" ? "assistant" : "user",
+                content = m.Content
+            });
+        }
+
+        return new
+        {
+            messages = msgs,
+            parameters = new
+            {
+                temperature = ResolveTemperature(),
+                max_new_tokens = ResolveOpenAiCompatibleMaxTokens()
+            }
+        };
+    }
+
+    private static string ExtractIbmDeploymentText(JsonElement root)
+    {
+        if (root.TryGetProperty("choices", out var choices) && choices.ValueKind == JsonValueKind.Array && choices.GetArrayLength() > 0)
+        {
+            var message = choices[0].TryGetProperty("message", out var choiceMessage) ? choiceMessage : default;
+            if (message.ValueKind == JsonValueKind.Object &&
+                message.TryGetProperty("content", out var content))
+                return content.GetString() ?? "";
+        }
+
+        if (root.TryGetProperty("results", out var results) && results.ValueKind == JsonValueKind.Array && results.GetArrayLength() > 0)
+        {
+            var first = results[0];
+            if (first.TryGetProperty("generated_text", out var generatedText))
+                return generatedText.GetString() ?? "";
+            if (first.TryGetProperty("output_text", out var outputText))
+                return outputText.GetString() ?? "";
+        }
+
+        if (root.TryGetProperty("generated_text", out var generated))
+            return generated.GetString() ?? "";
+
+        if (root.TryGetProperty("message", out var messageValue) && messageValue.ValueKind == JsonValueKind.String)
+            return messageValue.GetString() ?? "";
+
+        return "";
+    }
+
     /// <summary>
     /// 현재 활성 모델의 인증 헤더 값을 반환합니다.
     /// IBM IAM / CP4D 인증인 경우 토큰을 자동 발급/캐싱하여 반환합니다.
@@ -606,8 +703,14 @@ public partial class LlmService : IDisposable
         var llm  = _settings.Settings.Llm;
         var (endpoint, _, allowInsecureTls) = ResolveServerInfo();
         var ep = string.IsNullOrEmpty(endpoint) ? llm.Endpoint : endpoint;
-        var body = BuildOpenAiBody(messages, stream: false);
+        var registered = GetActiveRegisteredModel();
-        var url  = ep.TrimEnd('/') + "/v1/chat/completions";
+        var usesIbmDeploymentApi = UsesIbmDeploymentChatApi("vllm", registered, ep);
+        var body = usesIbmDeploymentApi
+            ? BuildIbmDeploymentBody(messages)
+            : BuildOpenAiBody(messages, stream: false);
+        var url = usesIbmDeploymentApi
+            ? BuildIbmDeploymentChatUrl(ep, stream: false)
+            : ep.TrimEnd('/') + "/v1/chat/completions";
         var json = JsonSerializer.Serialize(body);
 
         using var req = new HttpRequestMessage(HttpMethod.Post, url)
@@ -621,6 +724,12 @@ public partial class LlmService : IDisposable
         return SafeParseJson(respBody, root =>
         {
             TryParseOpenAiUsage(root);
+            if (usesIbmDeploymentApi)
+            {
+                var parsed = ExtractIbmDeploymentText(root);
+                return string.IsNullOrWhiteSpace(parsed) ? "(빈 응답)" : parsed;
+            }
+
             var choices = root.GetProperty("choices");
             if (choices.GetArrayLength() == 0) return "(빈 응답)";
             return choices[0].GetProperty("message").GetProperty("content").GetString() ?? "";
@@ -634,8 +743,14 @@ public partial class LlmService : IDisposable
         var llm  = _settings.Settings.Llm;
         var (endpoint, _, allowInsecureTls) = ResolveServerInfo();
         var ep = string.IsNullOrEmpty(endpoint) ? llm.Endpoint : endpoint;
-        var body = BuildOpenAiBody(messages, stream: true);
+        var registered = GetActiveRegisteredModel();
-        var url  = ep.TrimEnd('/') + "/v1/chat/completions";
+        var usesIbmDeploymentApi = UsesIbmDeploymentChatApi("vllm", registered, ep);
+        var body = usesIbmDeploymentApi
+            ? BuildIbmDeploymentBody(messages)
+            : BuildOpenAiBody(messages, stream: true);
+        var url = usesIbmDeploymentApi
+            ? BuildIbmDeploymentChatUrl(ep, stream: true)
+            : ep.TrimEnd('/') + "/v1/chat/completions";
 
         using var req  = new HttpRequestMessage(HttpMethod.Post, url) { Content = JsonContent(body) };
         await ApplyAuthHeaderAsync(req, ct);
@@ -657,12 +772,43 @@ public partial class LlmService : IDisposable
             {
                 using var doc = JsonDocument.Parse(data);
                 TryParseOpenAiUsage(doc.RootElement);
-                var choices = doc.RootElement.GetProperty("choices");
+                if (usesIbmDeploymentApi)
-                if (choices.GetArrayLength() > 0)
                 {
-                    var delta = choices[0].GetProperty("delta");
+                    if (doc.RootElement.TryGetProperty("status", out var status) &&
-                    if (delta.TryGetProperty("content", out var c))
+                        string.Equals(status.GetString(), "error", StringComparison.OrdinalIgnoreCase))
-                        text = c.GetString();
+                    {
+                        var detail = doc.RootElement.TryGetProperty("message", out var message)
+                            ? message.GetString()
+                            : "IBM vLLM 스트리밍 오류";
+                        throw new InvalidOperationException(detail);
+                    }
+
+                    if (doc.RootElement.TryGetProperty("results", out var results) &&
+                        results.ValueKind == JsonValueKind.Array &&
+                        results.GetArrayLength() > 0)
+                    {
+                        var first = results[0];
+                        if (first.TryGetProperty("generated_text", out var generatedText))
+                            text = generatedText.GetString();
+                        else if (first.TryGetProperty("output_text", out var outputText))
+                            text = outputText.GetString();
+                    }
+                    else if (doc.RootElement.TryGetProperty("choices", out var ibmChoices) && ibmChoices.GetArrayLength() > 0)
+                    {
+                        var delta = ibmChoices[0].GetProperty("delta");
+                        if (delta.TryGetProperty("content", out var c))
+                            text = c.GetString();
+                    }
+                }
+                else
+                {
+                    var choices = doc.RootElement.GetProperty("choices");
+                    if (choices.GetArrayLength() > 0)
+                    {
+                        var delta = choices[0].GetProperty("delta");
+                        if (delta.TryGetProperty("content", out var c))
+                            text = c.GetString();
+                    }
                 }
             }
             catch (JsonException ex)