@probelabs/probe 0.6.0-rc246 → 0.6.0-rc248

package/build/agent/ProbeAgent.js

@@ -85,7 +85,7 @@ import {
   validateAndFixMermaidResponse,
   tryAutoWrapForSimpleSchema
 } from './schemaUtils.js';
-import { removeThinkingTags } from './xmlParsingUtils.js';
+import { removeThinkingTags, extractThinkingContent } from './xmlParsingUtils.js';
 import { predefinedPrompts } from './shared/prompts.js';
 import {
   MCPXmlBridge,
@@ -2904,10 +2904,11 @@ Follow these instructions carefully:
       // Track initial history length for storage
       const oldHistoryLength = this.history.length;
 
-      // Reset output buffer for this answer() call — but NOT during schema correction recursion
-      // When _schemaFormatted is true, this is a recursive call to fix JSON formatting,
-      // and we must preserve the output buffer so the parent call can append it
-      if (this._outputBuffer && !options?._schemaFormatted) {
+      // Reset output buffer for this answer() call — but NOT during recursive calls.
+      // _schemaFormatted: recursive call to fix JSON formatting
+      // _completionPromptProcessed: recursive call for completionPrompt follow-up
+      // Both must preserve the output buffer so the parent call can append it.
+      if (this._outputBuffer && !options?._schemaFormatted && !options?._completionPromptProcessed) {
         this._outputBuffer.items = [];
       }
 
@@ -3554,8 +3555,25 @@ Follow these instructions carefully:
                   continue; // Don't use broken response, continue the loop
                 }
 
-                finalResult = prevContent;
-                if (this.debug) console.log(`[DEBUG] Using previous response as completion: ${finalResult.substring(0, 100)}...`);
+                // Pre-strip thinking tags to avoid losing content at final cleanup stage
+                const strippedContent = removeThinkingTags(prevContent);
+                if (strippedContent.length > 50) {
+                  // Enough content outside thinking tags — use stripped version directly
+                  finalResult = strippedContent;
+                  if (this.debug) console.log(`[DEBUG] Using previous response (thinking-stripped) as completion: ${finalResult.substring(0, 100)}...`);
+                } else {
+                  // Content was mostly/entirely inside thinking tags.
+                  // Extract thinking content and use it as the actual answer.
+                  const thinkingContent = extractThinkingContent(prevContent);
+                  if (thinkingContent && thinkingContent.length > 50) {
+                    finalResult = thinkingContent;
+                    if (this.debug) console.log(`[DEBUG] Previous response was mostly in thinking tags — using thinking content as completion: ${finalResult.substring(0, 100)}...`);
+                  } else {
+                    // Neither stripped nor thinking content is substantive — use raw as fallback
+                    finalResult = prevContent;
+                    if (this.debug) console.log(`[DEBUG] Using previous response as completion (raw): ${finalResult.substring(0, 100)}...`);
+                  }
+                }
               } else {
                 finalResult = 'Error: No previous response found to use as completion.';
                 if (this.debug) console.log(`[DEBUG] No suitable previous response found for attempt_complete shorthand`);
@@ -4296,10 +4314,16 @@ After reviewing, provide your final answer using attempt_completion.`;
 
           // Make a follow-up call with the completion prompt
           // Pass _completionPromptProcessed to prevent infinite loops
+          // Save output buffer — the recursive answer() must not destroy DSL output() content
+          const savedOutputItems = this._outputBuffer ? [...this._outputBuffer.items] : [];
           const completionResult = await this.answer(completionPromptMessage, [], {
             ...options,
             _completionPromptProcessed: true
           });
+          // Restore output buffer so the parent call can append it to the final result
+          if (this._outputBuffer) {
+            this._outputBuffer.items = savedOutputItems;
+          }
 
           // Update finalResult with the result from the completion prompt
           finalResult = completionResult;

package/build/agent/index.js

@@ -29346,6 +29346,7 @@ ${RAW_OUTPUT_END}`;
     output += `
 
 [The above raw output (${rawContent.length} chars) will be passed directly to the final response. Do NOT repeat, summarize, or modify it.]`;
+    outputBuffer.items = [];
   }
   return output;
 }
@@ -83496,7 +83497,7 @@ You are working with a workspace. Available paths: ${workspaceDesc}
         }
         try {
           const oldHistoryLength = this.history.length;
-          if (this._outputBuffer && !options?._schemaFormatted) {
+          if (this._outputBuffer && !options?._schemaFormatted && !options?._completionPromptProcessed) {
             this._outputBuffer.items = [];
           }
           if (this.enableTasks) {
@@ -83941,8 +83942,20 @@ You are working with a workspace. Available paths: ${workspaceDesc}
                       completionAttempted = false;
                       continue;
                     }
-                    finalResult = prevContent;
-                    if (this.debug) console.log(`[DEBUG] Using previous response as completion: ${finalResult.substring(0, 100)}...`);
+                    const strippedContent = removeThinkingTags(prevContent);
+                    if (strippedContent.length > 50) {
+                      finalResult = strippedContent;
+                      if (this.debug) console.log(`[DEBUG] Using previous response (thinking-stripped) as completion: ${finalResult.substring(0, 100)}...`);
+                    } else {
+                      const thinkingContent = extractThinkingContent(prevContent);
+                      if (thinkingContent && thinkingContent.length > 50) {
+                        finalResult = thinkingContent;
+                        if (this.debug) console.log(`[DEBUG] Previous response was mostly in thinking tags \u2014 using thinking content as completion: ${finalResult.substring(0, 100)}...`);
+                      } else {
+                        finalResult = prevContent;
+                        if (this.debug) console.log(`[DEBUG] Using previous response as completion (raw): ${finalResult.substring(0, 100)}...`);
+                      }
+                    }
                   } else {
                     finalResult = "Error: No previous response found to use as completion.";
                     if (this.debug) console.log(`[DEBUG] No suitable previous response found for attempt_complete shorthand`);
@@ -84514,10 +84527,14 @@ ${finalResult}
 </result>
 
 After reviewing, provide your final answer using attempt_completion.`;
+              const savedOutputItems = this._outputBuffer ? [...this._outputBuffer.items] : [];
               const completionResult = await this.answer(completionPromptMessage, [], {
                 ...options,
                 _completionPromptProcessed: true
               });
+              if (this._outputBuffer) {
+                this._outputBuffer.items = savedOutputItems;
+              }
               finalResult = completionResult;
               if (this.debug) {
                 console.log(`[DEBUG] Completion prompt finished. New result length: ${finalResult?.length || 0}`);

package/build/tools/executePlan.js

@@ -525,6 +525,9 @@ function formatSuccess(result, description, attempt, outputBuffer) {
     const rawContent = outputBuffer.items.join('\n');
     output += `\n\n${RAW_OUTPUT_START}\n${rawContent}\n${RAW_OUTPUT_END}`;
     output += `\n\n[The above raw output (${rawContent.length} chars) will be passed directly to the final response. Do NOT repeat, summarize, or modify it.]`;
+    // Clear the buffer after reading to prevent re-wrapping on subsequent execute_plan calls
+    // Without this, extractRawOutputBlocks pushes content back to buffer, causing exponential duplication
+    outputBuffer.items = [];
   }
 
   return output;