Merge pull request #635 from Chris0Jeky/feature/616-chat-json-truncation

Chris0Jeky · web-flow · commit 34d632a93663 · 2026-03-31T15:52:22.000+01:00
Fix chat response truncation and raw JSON display
diff --git a/backend/src/Taskdeck.Application/Services/GeminiLlmProvider.cs b/backend/src/Taskdeck.Application/Services/GeminiLlmProvider.cs
@@ -81,12 +81,42 @@ public async Task<LlmCompletionResult> CompleteAsync(ChatCompletionRequest reque
                 return BuildFallbackResult(lastUserMessage, "Live provider request failed.", GetConfiguredModelOrDefault());
             }
 
-            if (!TryParseResponse(body, out var content, out var tokensUsed))
+            if (!TryParseResponse(body, out var content, out var tokensUsed, out var finishReason))
             {
                 _logger.LogWarning("Gemini completion response could not be parsed.");
                 return BuildFallbackResult(lastUserMessage, "Live provider response parsing failed.", GetConfiguredModelOrDefault());
             }
 
+            // Detect truncation: Gemini returns finishReason "MAX_TOKENS" when the
+            // response was cut off by the maxOutputTokens limit.
+            if (string.Equals(finishReason, "MAX_TOKENS", StringComparison.OrdinalIgnoreCase))
+            {
+                _logger.LogWarning("Gemini response was truncated (finishReason=MAX_TOKENS).");
+                return new LlmCompletionResult(
+                    content,
+                    tokensUsed,
+                    IsActionable: false,
+                    Provider: "Gemini",
+                    Model: GetConfiguredModelOrDefault(),
+                    IsDegraded: true,
+                    DegradedReason: "Response was truncated");
+            }
+
+            // When JSON mode was requested and the response starts with '{' but
+            // does not parse as valid JSON, the output was likely truncated.
+            if (useInstructionExtraction && LooksLikeTruncatedJson(content))
+            {
+                _logger.LogWarning("Gemini JSON-mode response is not valid JSON; treating as truncated.");
+                return new LlmCompletionResult(
+                    content,
+                    tokensUsed,
+                    IsActionable: false,
+                    Provider: "Gemini",
+                    Model: GetConfiguredModelOrDefault(),
+                    IsDegraded: true,
+                    DegradedReason: "Response was truncated");
+            }
+
             // Try to parse structured instruction extraction from the LLM response
             if (LlmInstructionExtractionPrompt.TryParseStructuredResponse(
                     content,
@@ -225,10 +255,11 @@ private static object MapMessage(ChatCompletionMessage message)
         };
     }
 
-    private static bool TryParseResponse(string responseBody, out string content, out int tokensUsed)
+    private static bool TryParseResponse(string responseBody, out string content, out int tokensUsed, out string? finishReason)
     {
         content = string.Empty;
         tokensUsed = 0;
+        finishReason = null;
 
         if (string.IsNullOrWhiteSpace(responseBody))
         {
@@ -248,6 +279,13 @@ private static bool TryParseResponse(string responseBody, out string content, ou
             }
 
             var firstCandidate = candidates[0];
+
+            if (firstCandidate.TryGetProperty("finishReason", out var finishReasonElement) &&
+                finishReasonElement.ValueKind == JsonValueKind.String)
+            {
+                finishReason = finishReasonElement.GetString();
+            }
+
             if (!firstCandidate.TryGetProperty("content", out var candidateContent) ||
                 !candidateContent.TryGetProperty("parts", out var parts) ||
                 parts.ValueKind != JsonValueKind.Array ||
@@ -324,6 +362,30 @@ private static LlmCompletionResult BuildFallbackResult(string userMessage, strin
             Instructions: instructions);
     }
 
+    /// <summary>
+    /// Returns true when <paramref name="text"/> starts with '{' but does not
+    /// parse as valid JSON — a strong signal the response was cut off mid-output.
+    /// </summary>
+    internal static bool LooksLikeTruncatedJson(string text)
+    {
+        if (string.IsNullOrWhiteSpace(text))
+            return false;
+
+        var trimmed = text.TrimStart();
+        if (!trimmed.StartsWith('{'))
+            return false;
+
+        try
+        {
+            using var doc = JsonDocument.Parse(trimmed);
+            return false;
+        }
+        catch (JsonException)
+        {
+            return true;
+        }
+    }
+
     private static int EstimateTokens(string text)
     {
         if (string.IsNullOrWhiteSpace(text))
diff --git a/backend/src/Taskdeck.Application/Services/ILlmProvider.cs b/backend/src/Taskdeck.Application/Services/ILlmProvider.cs
@@ -10,7 +10,7 @@ public interface ILlmProvider
 
 public record ChatCompletionRequest(
     List<ChatCompletionMessage> Messages,
-    int MaxTokens = 1024,
+    int MaxTokens = 2048,
     double Temperature = 0.7,
     LlmRequestAttribution? Attribution = null,
     string? SystemPrompt = null,
diff --git a/backend/src/Taskdeck.Application/Services/OpenAiLlmProvider.cs b/backend/src/Taskdeck.Application/Services/OpenAiLlmProvider.cs
@@ -52,12 +52,43 @@ public async Task<LlmCompletionResult> CompleteAsync(ChatCompletionRequest reque
                 return BuildFallbackResult(lastUserMessage, "Live provider request failed.", GetConfiguredModelOrDefault());
             }
 
-            if (!TryParseResponse(body, out var content, out var tokensUsed))
+            if (!TryParseResponse(body, out var content, out var tokensUsed, out var finishReason))
             {
                 _logger.LogWarning("OpenAI completion response could not be parsed.");
                 return BuildFallbackResult(lastUserMessage, "Live provider response parsing failed.", GetConfiguredModelOrDefault());
             }
 
+            // Detect truncation: OpenAI returns finish_reason "length" when the
+            // response was cut off by the max_tokens limit.
+            if (string.Equals(finishReason, "length", StringComparison.OrdinalIgnoreCase))
+            {
+                _logger.LogWarning("OpenAI response was truncated (finish_reason=length).");
+                return new LlmCompletionResult(
+                    content,
+                    tokensUsed,
+                    IsActionable: false,
+                    Provider: "OpenAI",
+                    Model: GetConfiguredModelOrDefault(),
+                    IsDegraded: true,
+                    DegradedReason: "Response was truncated");
+            }
+
+            // When JSON mode was requested and the response starts with '{' but
+            // does not parse as valid JSON, the output was likely truncated.
+            var useInstructionExtraction = request.SystemPrompt is null;
+            if (useInstructionExtraction && LooksLikeTruncatedJson(content))
+            {
+                _logger.LogWarning("OpenAI JSON-mode response is not valid JSON; treating as truncated.");
+                return new LlmCompletionResult(
+                    content,
+                    tokensUsed,
+                    IsActionable: false,
+                    Provider: "OpenAI",
+                    Model: GetConfiguredModelOrDefault(),
+                    IsDegraded: true,
+                    DegradedReason: "Response was truncated");
+            }
+
             // Try to parse structured instruction extraction from the LLM response
             if (LlmInstructionExtractionPrompt.TryParseStructuredResponse(
                     content,
@@ -227,10 +258,11 @@ private object BuildRequestPayload(ChatCompletionRequest request)
         return payload;
     }
 
-    private static bool TryParseResponse(string responseBody, out string content, out int tokensUsed)
+    private static bool TryParseResponse(string responseBody, out string content, out int tokensUsed, out string? finishReason)
     {
         content = string.Empty;
         tokensUsed = 0;
+        finishReason = null;
 
         if (string.IsNullOrWhiteSpace(responseBody))
         {
@@ -259,6 +291,12 @@ private static bool TryParseResponse(string responseBody, out string content, ou
                 return false;
             }
 
+            if (first.TryGetProperty("finish_reason", out var finishReasonElement) &&
+                finishReasonElement.ValueKind == JsonValueKind.String)
+            {
+                finishReason = finishReasonElement.GetString();
+            }
+
             if (root.TryGetProperty("usage", out var usage) &&
                 usage.TryGetProperty("total_tokens", out var totalTokens) &&
                 totalTokens.TryGetInt32(out var parsedTokens))
@@ -308,6 +346,30 @@ private static LlmCompletionResult BuildFallbackResult(string userMessage, strin
             Instructions: instructions);
     }
 
+    /// <summary>
+    /// Returns true when <paramref name="text"/> starts with '{' but does not
+    /// parse as valid JSON — a strong signal the response was cut off mid-output.
+    /// </summary>
+    internal static bool LooksLikeTruncatedJson(string text)
+    {
+        if (string.IsNullOrWhiteSpace(text))
+            return false;
+
+        var trimmed = text.TrimStart();
+        if (!trimmed.StartsWith('{'))
+            return false;
+
+        try
+        {
+            using var doc = JsonDocument.Parse(trimmed);
+            return false;
+        }
+        catch (JsonException)
+        {
+            return true;
+        }
+    }
+
     private static int EstimateTokens(string text)
     {
         if (string.IsNullOrWhiteSpace(text))
diff --git a/backend/tests/Taskdeck.Application.Tests/Services/GeminiLlmProviderTests.cs b/backend/tests/Taskdeck.Application.Tests/Services/GeminiLlmProviderTests.cs
@@ -432,6 +432,55 @@ await provider.CompleteAsync(new ChatCompletionRequest(
         hasSystemInstruction.Should().BeFalse("empty system prompt should not produce system_instruction field");
     }
 
+    [Fact]
+    public async Task CompleteAsync_ShouldReturnDegraded_WhenFinishReasonIsMaxTokens()
+    {
+        var settings = BuildSettings();
+        var handler = new StubHttpMessageHandler(_ =>
+        {
+            return new HttpResponseMessage(HttpStatusCode.OK)
+            {
+                Content = new StringContent(
+                    """
+                    {
+                      "candidates": [
+                        {
+                          "content": {
+                            "parts": [{ "text": "partial response" }]
+                          },
+                          "finishReason": "MAX_TOKENS"
+                        }
+                      ],
+                      "usageMetadata": { "totalTokenCount": 50 }
+                    }
+                    """,
+                    Encoding.UTF8,
+                    "application/json")
+            };
+        });
+
+        var provider = new GeminiLlmProvider(new HttpClient(handler), settings, NullLogger<GeminiLlmProvider>.Instance);
+        var result = await provider.CompleteAsync(new ChatCompletionRequest(
+            [new ChatCompletionMessage("User", "tell me something")],
+            SystemPrompt: string.Empty));
+
+        result.IsDegraded.Should().BeTrue();
+        result.DegradedReason.Should().Be("Response was truncated");
+        result.Content.Should().Be("partial response");
+        result.IsActionable.Should().BeFalse();
+    }
+
+    [Theory]
+    [InlineData("{\"reply\":\"incomplete", true)]
+    [InlineData("{}", false)]
+    [InlineData("plain text response", false)]
+    [InlineData("", false)]
+    [InlineData("  { broken json", true)]
+    public void LooksLikeTruncatedJson_ShouldDetectPartialJson(string input, bool expected)
+    {
+        GeminiLlmProvider.LooksLikeTruncatedJson(input).Should().Be(expected);
+    }
+
     private static LlmProviderSettings BuildSettings()
     {
         return new LlmProviderSettings
diff --git a/backend/tests/Taskdeck.Application.Tests/Services/OpenAiLlmProviderTests.cs b/backend/tests/Taskdeck.Application.Tests/Services/OpenAiLlmProviderTests.cs
@@ -269,6 +269,85 @@ public async Task CompleteAsync_ShouldRedactSensitiveDetails_WhenUnexpectedExcep
         message.Should().Contain($"Authorization: Bearer {SensitiveDataRedactor.RedactedValue}");
     }
 
+    [Fact]
+    public async Task CompleteAsync_ShouldReturnDegraded_WhenFinishReasonIsLength()
+    {
+        var settings = BuildSettings();
+        var handler = new StubHttpMessageHandler(_ =>
+        {
+            return new HttpResponseMessage(HttpStatusCode.OK)
+            {
+                Content = new StringContent(
+                    """
+                    {
+                      "choices": [
+                        {
+                          "message": { "content": "partial response" },
+                          "finish_reason": "length"
+                        }
+                      ],
+                      "usage": { "total_tokens": 50 }
+                    }
+                    """,
+                    Encoding.UTF8,
+                    "application/json")
+            };
+        });
+
+        var provider = new OpenAiLlmProvider(new HttpClient(handler), settings, NullLogger<OpenAiLlmProvider>.Instance);
+        var result = await provider.CompleteAsync(new ChatCompletionRequest(
+            [new ChatCompletionMessage("User", "tell me something")],
+            SystemPrompt: string.Empty));
+
+        result.IsDegraded.Should().BeTrue();
+        result.DegradedReason.Should().Be("Response was truncated");
+        result.Content.Should().Be("partial response");
+        result.IsActionable.Should().BeFalse();
+    }
+
+    [Fact]
+    public async Task CompleteAsync_ShouldReturnDegraded_WhenJsonModeResponseIsInvalidJson()
+    {
+        var settings = BuildSettings();
+        // Build a valid OpenAI response whose content value is truncated JSON.
+        // The inner content must be JSON-escaped so the outer envelope parses.
+        var truncatedContent = "{\\\"reply\\\":\\\"this is cut off";
+        var responseBody = $@"{{
+  ""choices"": [{{
+    ""message"": {{ ""content"": ""{truncatedContent}"" }},
+    ""finish_reason"": ""stop""
+  }}],
+  ""usage"": {{ ""total_tokens"": 50 }}
+}}";
+        var handler = new StubHttpMessageHandler(_ =>
+        {
+            return new HttpResponseMessage(HttpStatusCode.OK)
+            {
+                Content = new StringContent(responseBody, Encoding.UTF8, "application/json")
+            };
+        });
+
+        var provider = new OpenAiLlmProvider(new HttpClient(handler), settings, NullLogger<OpenAiLlmProvider>.Instance);
+
+        // SystemPrompt defaults to null -> JSON mode is requested
+        var result = await provider.CompleteAsync(new ChatCompletionRequest(
+            [new ChatCompletionMessage("User", "tell me something")]));
+
+        result.IsDegraded.Should().BeTrue();
+        result.DegradedReason.Should().Be("Response was truncated");
+    }
+
+    [Theory]
+    [InlineData("{\"reply\":\"incomplete", true)]
+    [InlineData("{}", false)]
+    [InlineData("plain text response", false)]
+    [InlineData("", false)]
+    [InlineData("  { broken json", true)]
+    public void LooksLikeTruncatedJson_ShouldDetectPartialJson(string input, bool expected)
+    {
+        OpenAiLlmProvider.LooksLikeTruncatedJson(input).Should().Be(expected);
+    }
+
     private static LlmProviderSettings BuildSettings()
     {
         return new LlmProviderSettings
diff --git a/frontend/taskdeck-web/src/views/AutomationChatView.vue b/frontend/taskdeck-web/src/views/AutomationChatView.vue
