improve capture of openai messages

braintrust-sdk/instrumentation/genai_1_18_0/build.gradle

@@ -5,7 +5,8 @@ muzzle {
   pass {
     group = 'com.google.genai'
     module = 'google-genai'
-    versions = '[1.18.0,)'
+    // TODO: autoinstrumentation muzzle fails in 1.65.0 because slf4j api is no longer a transitive dep
+    versions = '[1.18.0,1.65.0)'
   }
 }
 

braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/WrappedHttpClient.java

@@ -1,10 +1,8 @@
 package dev.braintrust.instrumentation.langchain.v1_8_0;
 
-import static dev.braintrust.json.BraintrustJsonMapper.toJson;
-
-import com.fasterxml.jackson.databind.JsonNode;
 import dev.braintrust.bootstrap.BraintrustBridge;
 import dev.braintrust.instrumentation.InstrumentationSemConv;
+import dev.braintrust.instrumentation.SseResponseAccumulator;
 import dev.braintrust.json.BraintrustJsonMapper;
 import dev.langchain4j.exception.HttpException;
 import dev.langchain4j.http.client.HttpClient;
@@ -126,9 +124,8 @@ static class WrappedServerSentEventListener implements ServerSentEventListener {
         private final String providerName;
         private final long startNanos = System.nanoTime();
         private final AtomicLong timeToFirstTokenNanos = new AtomicLong();
-        private final StringBuilder contentBuffer = new StringBuilder();
-        private String finishReason = null;
-        private JsonNode usageData = null;
+        private final SseResponseAccumulator accumulator =
+                new SseResponseAccumulator(BraintrustJsonMapper.get());
 
         WrappedServerSentEventListener(
                 ServerSentEventListener delegate, Span span, String providerName) {
@@ -182,52 +179,17 @@ public void onClose() {
 
         private void accumulateChunk(String data) {
             if (data == null || data.isEmpty() || "[DONE]".equals(data)) return;
-            try {
-                if (timeToFirstTokenNanos.get() == 0L) {
-                    timeToFirstTokenNanos.compareAndExchange(0L, System.nanoTime() - startNanos);
-                }
-                JsonNode chunk = BraintrustJsonMapper.get().readTree(data);
-                if (chunk.has("choices") && chunk.get("choices").size() > 0) {
-                    JsonNode choice = chunk.get("choices").get(0);
-                    if (choice.has("delta")) {
-                        JsonNode delta = choice.get("delta");
-                        if (delta.has("content")) {
-                            contentBuffer.append(delta.get("content").asText());
-                        }
-                    }
-                    if (choice.has("finish_reason") && !choice.get("finish_reason").isNull()) {
-                        finishReason = choice.get("finish_reason").asText();
-                    }
-                }
-                if (chunk.has("usage") && !chunk.get("usage").isNull()) {
-                    usageData = chunk.get("usage");
-                }
-            } catch (Exception e) {
-                log.debug("Failed to parse SSE chunk: {}", data, e);
+            if (timeToFirstTokenNanos.get() == 0L) {
+                timeToFirstTokenNanos.compareAndExchange(0L, System.nanoTime() - startNanos);
             }
+            accumulator.merge(data);
         }
 
         private void finalizeSpan() {
             try {
-                var root = BraintrustJsonMapper.get().createObjectNode();
-
-                var choicesArray = BraintrustJsonMapper.get().createArrayNode();
-                var choice = BraintrustJsonMapper.get().createObjectNode();
-                choice.put("index", 0);
-                if (finishReason != null) choice.put("finish_reason", finishReason);
-                var message = BraintrustJsonMapper.get().createObjectNode();
-                message.put("role", "assistant");
-                message.put("content", contentBuffer.toString());
-                choice.set("message", message);
-                choicesArray.add(choice);
-                root.set("choices", choicesArray);
-
-                if (usageData != null) {
-                    root.set("usage", usageData);
-                }
-
                 Long ttft = timeToFirstTokenNanos.get();
-                InstrumentationSemConv.tagLLMSpanResponse(span, providerName, toJson(root), ttft);
+                InstrumentationSemConv.tagLLMSpanResponse(
+                        span, providerName, accumulator.build(), ttft);
             } catch (Exception e) {
                 log.debug("Failed to finalize streaming span", e);
             }

braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/auto/LangchainInstrumentationModule.java

@@ -36,6 +36,7 @@ public List<String> getHelperClassNames() {
                 MANUAL_PACKAGE + "TracingProxy",
                 MANUAL_PACKAGE + "TracingToolExecutor",
                 MANUAL_PACKAGE + "OtelContextPassingExecutor",
+                "dev.braintrust.instrumentation.SseResponseAccumulator",
                 "dev.braintrust.instrumentation.InstrumentationSemConv",
                 "dev.braintrust.json.BraintrustJsonMapper");
     }

braintrust-sdk/instrumentation/langchain_1_8_0/src/test/java/dev/braintrust/instrumentation/langchain/v1_8_0/BraintrustLangchainTest.java

@@ -7,9 +7,12 @@
 import dev.braintrust.TestHarness;
 import dev.braintrust.instrumentation.Instrumenter;
 import dev.langchain4j.agent.tool.Tool;
+import dev.langchain4j.agent.tool.ToolSpecification;
 import dev.langchain4j.data.message.UserMessage;
 import dev.langchain4j.model.chat.ChatModel;
 import dev.langchain4j.model.chat.StreamingChatModel;
+import dev.langchain4j.model.chat.request.ChatRequest;
+import dev.langchain4j.model.chat.request.json.JsonObjectSchema;
 import dev.langchain4j.model.chat.response.ChatResponse;
 import dev.langchain4j.model.chat.response.StreamingChatResponseHandler;
 import dev.langchain4j.model.openai.OpenAiChatModel;
@@ -109,6 +112,9 @@ void testSyncChatCompletion() {
         assertNotNull(
                 output.get(0).get("message").get("content"),
                 "Output should contain assistant response content");
+
+        // The serialized span output should reflect the full response the client received.
+        assertSpanOutputReflects(response, span);
     }
 
     @Test
@@ -239,6 +245,129 @@ public void onError(Throwable error) {
                 choice.get("message").get("content"),
                 "Output should contain the complete streamed response");
         assertNotNull(choice.get("finish_reason"), "Output should have finish_reason");
+
+        // The reconstructed streaming span output should reflect the full response the client
+        // received — the instrumentation must feed every SSE event to the accumulator.
+        assertSpanOutputReflects(response, llmSpan);
+    }
+
+    @Test
+    @SneakyThrows
+    void testStreamingChatCompletionWithTools() {
+        // Auto-instrumentation intercepts OpenAiStreamingChatModel.Builder.build()
+        StreamingChatModel model =
+                OpenAiStreamingChatModel.builder()
+                        .apiKey(testHarness.openAiApiKey())
+                        .baseUrl(testHarness.openAiBaseUrl())
+                        .modelName("gpt-4o")
+                        .temperature(0.0)
+                        .build();
+
+        var weatherTool =
+                ToolSpecification.builder()
+                        .name("get_weather")
+                        .description("Get the current weather for a location")
+                        .parameters(
+                                JsonObjectSchema.builder()
+                                        .addStringProperty(
+                                                "location",
+                                                "The city and state, e.g. San" + " Francisco, CA")
+                                        .required("location")
+                                        .build())
+                        .build();
+
+        var chatRequest =
+                ChatRequest.builder()
+                        .messages(UserMessage.from("What is the weather in Paris, France?"))
+                        .toolSpecifications(weatherTool)
+                        .build();
+
+        var future = new CompletableFuture<ChatResponse>();
+        model.chat(
+                chatRequest,
+                new StreamingChatResponseHandler() {
+                    @Override
+                    public void onPartialResponse(String token) {}
+
+                    @Override
+                    public void onCompleteResponse(ChatResponse response) {
+                        future.complete(response);
+                    }
+
+                    @Override
+                    public void onError(Throwable error) {
+                        future.completeExceptionally(error);
+                    }
+                });
+        var response = future.get();
+
+        // The stream must carry tool-call deltas (merged by index) all the way to the span — the
+        // original bug dropped tool_calls entirely from streaming reconstruction.
+        assertTrue(
+                response.aiMessage().hasToolExecutionRequests(),
+                "Model should have requested a tool call");
+
+        var llmSpan =
+                testHarness.awaitExportedSpans(1).stream()
+                        .filter(s -> s.getName().equals("Chat Completion"))
+                        .findFirst()
+                        .orElseThrow(() -> new AssertionError("no 'Chat Completion' llm span"));
+
+        assertSpanOutputReflects(response, llmSpan);
+    }
+
+    /**
+     * Asserts that the llm span's serialized output ({@code braintrust.output_json}) reflects the
+     * full response the langchain client received — comparing the reconstructed assistant message
+     * against the client's parsed {@link ChatResponse} (content, thinking, and tool calls) rather
+     * than hand-asserting individual fields per test. langchain decodes the same stream
+     * independently of our accumulator, so agreement is a meaningful end-to-end check.
+     */
+    @SneakyThrows
+    private void assertSpanOutputReflects(ChatResponse clientResponse, SpanData llmSpan) {
+        String outputJson =
+                llmSpan.getAttributes().get(AttributeKey.stringKey("braintrust.output_json"));
+        assertNotNull(outputJson, "Span should have braintrust.output_json");
+        JsonNode message = JSON_MAPPER.readTree(outputJson).get(0).get("message");
+        assertNotNull(message, "Span output should contain a choice message");
+
+        var aiMessage = clientResponse.aiMessage();
+
+        if (aiMessage.text() != null) {
+            assertEquals(
+                    aiMessage.text(),
+                    message.path("content").asText(),
+                    "Span output content should match the client's assistant text");
+        }
+        if (aiMessage.thinking() != null) {
+            assertEquals(
+                    aiMessage.thinking(),
+                    message.path("reasoning_content").asText(),
+                    "Span output reasoning_content should match the client's thinking");
+        }
+        if (aiMessage.hasToolExecutionRequests()) {
+            JsonNode toolCalls = message.get("tool_calls");
+            assertNotNull(toolCalls, "Span output should contain tool_calls");
+            var requests = aiMessage.toolExecutionRequests();
+            assertEquals(
+                    requests.size(), toolCalls.size(), "tool_calls count should match the client");
+            for (int i = 0; i < requests.size(); i++) {
+                var request = requests.get(i);
+                JsonNode function = toolCalls.get(i).get("function");
+                assertEquals(
+                        request.name(), function.get("name").asText(), "tool name should match");
+                assertEquals(
+                        JSON_MAPPER.readTree(request.arguments()),
+                        JSON_MAPPER.readTree(function.get("arguments").asText()),
+                        "tool arguments should match");
+                if (request.id() != null) {
+                    assertEquals(
+                            request.id(),
+                            toolCalls.get(i).get("id").asText(),
+                            "tool id should match");
+                }
+            }
+        }
     }
 
     @Test