illuma-agents 1.0.63 → 1.0.64

package/dist/esm/graphs/Graph.mjs

@@ -7,7 +7,7 @@ import { SystemMessage, HumanMessage, AIMessageChunk } from '@langchain/core/mes
 import { GraphNodeKeys, ContentTypes, Providers, GraphEvents, MessageTypes, StepTypes, Constants } from '../common/enum.mjs';
 import { convertMessagesToContent, modifyDeltaProperties, formatAnthropicArtifactContent, formatArtifactPayload } from '../messages/core.mjs';
 import { createPruneMessages } from '../messages/prune.mjs';
-import { ensureThinkingBlockInMessages, stripThinkingBlocksFromMessages } from '../messages/format.mjs';
+import { ensureThinkingBlockInMessages } from '../messages/format.mjs';
 import { addCacheControl, addBedrockCacheControl } from '../messages/cache.mjs';
 import { formatContentStrings } from '../messages/content.mjs';
 import { extractToolDiscoveries } from '../messages/tools.mjs';
@@ -825,11 +825,11 @@ class StandardGraph extends Graph {
                         provider: agentContext.provider,
                         clientOptions: structuredClientOptions,
                     });
-                    // Strip thinking blocks from messages since structured output has thinking disabled
-                    const cleanMessages = stripThinkingBlocksFromMessages(finalMessages);
+                    // For no-tools case, we pass the original messages directly
+                    // There shouldn't be thinking blocks since this is the first invocation
                     const { structuredResponse, rawMessage } = await this.attemptStructuredInvoke({
                         currentModel: structuredModel,
-                        finalMessages: cleanMessages,
+                        finalMessages,
                         schema,
                         structuredOutputConfig: agentContext.structuredOutput,
                         provider: agentContext.provider,
@@ -1076,17 +1076,39 @@ If I seem to be missing something we discussed earlier, just give me a quick rem
                             provider: agentContext.provider,
                             clientOptions: structuredClientOptions,
                         });
-                        // Include the current result in the messages so the model knows what it just said
-                        // IMPORTANT: Strip thinking blocks from messages since structured output has thinking disabled
-                        // Bedrock will error if assistant messages contain thinking blocks when thinking is disabled
-                        let messagesWithResult = [...finalMessages];
+                        // Following LangGraph's Option 2 pattern:
+                        // Instead of passing the full conversation (which has thinking blocks),
+                        // just pass the model's final response text as a HumanMessage asking for structured output.
+                        // This avoids issues with thinking blocks and is more token-efficient.
+                        let responseContent = '';
                         if (resultMessage) {
-                            messagesWithResult.push(resultMessage);
+                            if (typeof resultMessage.content === 'string') {
+                                responseContent = resultMessage.content;
+                            }
+                            else if (Array.isArray(resultMessage.content)) {
+                                // Extract text parts only (skip thinking blocks, tool_use, etc.)
+                                responseContent = resultMessage.content
+                                    .filter((c) => typeof c === 'object' && c !== null &&
+                                    c.type === 'text' &&
+                                    typeof c.text === 'string')
+                                    .map(c => c.text)
+                                    .join('\n');
+                            }
                         }
-                        messagesWithResult = stripThinkingBlocksFromMessages(messagesWithResult);
+                        if (!responseContent) {
+                            console.log('[Graph] No response content to structure, skipping');
+                            throw new Error('No response content available for structured output');
+                        }
+                        console.log('[Graph] Structuring response content (first 200 chars):', responseContent.substring(0, 200));
+                        // Create a simple message asking to structure the response
+                        const structuredMessages = [
+                            new HumanMessage({
+                                content: `Based on the following response, extract and return the data in the required structured format:\n\n${responseContent}`,
+                            }),
+                        ];
                         const { structuredResponse, rawMessage } = await this.attemptStructuredInvoke({
                             currentModel: structuredModel,
-                            finalMessages: messagesWithResult,
+                            finalMessages: structuredMessages,
                             schema,
                             structuredOutputConfig: agentContext.structuredOutput,
                             provider: agentContext.provider,