fix: use last step output instead of longest in MCP response extraction

extractResponseText() was picking the longest text across all workflow
steps, which could return routing/intent classification output instead
of the final generate-response output. Changed to pick the last step's
text, matching the natural execution order.

This fixes a bug where MCP clients would receive intent classifications
(e.g. "engineering-task, skills: api-gateway") instead of the actual
AI response when the routing step produced longer output than the
final response.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: buger

src/runners/mcp-server-runner.ts

@@ -576,41 +576,42 @@ export class McpServerRunner implements Runner {
 
     // The output history (reviewSummary.history) contains outputs keyed by step ID.
     // For assistant workflows the steps are e.g. chat.route-intent, chat.build-config,
-    // chat.generate-response. The final AI response is the LAST step that has a .text
-    // field with substantial content (not a short routing label).
+    // chat.generate-response. The final AI response is the LAST step that produced
+    // a text output — not the longest, since routing/classification steps can produce
+    // longer text than the actual response.
     const history = result?.reviewSummary?.history;
     if (history && typeof history === 'object') {
-      // Collect all candidate text outputs, keeping the last (deepest) one
-      let bestText = '';
+      // Pick the last step's text output (iteration order preserves insertion order)
+      let lastText = '';
       for (const [, outputs] of Object.entries(history)) {
         if (!Array.isArray(outputs)) continue;
         for (const item of outputs as any[]) {
           const text = item?.text ?? item?.output?.text;
-          if (typeof text === 'string' && text.length > bestText.length) {
-            bestText = text;
+          if (typeof text === 'string' && text.trim().length > 0) {
+            lastText = text;
           }
         }
       }
-      if (bestText) return bestText;
+      if (lastText) return lastText;
     }
 
-    // Grouped results from execution statistics
+    // Grouped results from execution statistics — pick the last text output
     const grouped = result?.executionStatistics?.groupedResults;
     if (grouped && typeof grouped === 'object') {
-      let bestText = '';
+      let lastText = '';
       for (const checkResults of Object.values(grouped)) {
         if (!Array.isArray(checkResults)) continue;
         for (const cr of checkResults as any[]) {
           const text =
             cr?.output?.text ??
             (typeof cr?.output === 'string' ? cr.output : null) ??
             (typeof cr?.content === 'string' && cr.content.trim() ? cr.content : null);
-          if (typeof text === 'string' && text.length > bestText.length) {
-            bestText = text;
+          if (typeof text === 'string' && text.trim().length > 0) {
+            lastText = text;
           }
         }
       }
-      if (bestText) return bestText;
+      if (lastText) return lastText;
     }
 
     // Direct properties on result

tests/unit/runners/mcp-server-runner.test.ts

@@ -213,7 +213,7 @@ describe('extractResponseText', () => {
     expect(extractResponseText(undefined)).toBe('No response from workflow.');
   });
 
-  it('picks the longest text from history (final AI response, not intent)', () => {
+  it('picks the last step text from history (final AI response, not routing)', () => {
     const result = {
       reviewSummary: {
         issues: [],
@@ -243,7 +243,7 @@ describe('extractResponseText', () => {
     expect(text).not.toBe('Short routing label');
   });
 
-  it('does not pick short intent/routing text over long response', () => {
+  it('does not pick short intent/routing text over last response', () => {
     const result = {
       reviewSummary: {
         issues: [],
@@ -261,6 +261,36 @@ describe('extractResponseText', () => {
     expect(text).not.toBe('chat');
   });
 
+  it('picks last step even when routing output is longer than final response', () => {
+    const result = {
+      reviewSummary: {
+        issues: [],
+        history: {
+          'chat.route-intent': [
+            {
+              text:
+                'Based on analysis of the user query, I have determined this is a request about ' +
+                'API gateway configuration. The user wants to understand how to set up rate limiting ' +
+                'with multiple policies across different API endpoints. Classifying as: engineering-task. ' +
+                'Relevant skills: api-gateway, rate-limiting, policy-management.',
+            },
+          ],
+          'chat.build-config': [{ mcp_servers: {}, text: 'config built' }],
+          'chat.generate-response': [
+            {
+              text: 'To set up rate limiting, use the Tyk Dashboard.',
+            },
+          ],
+        },
+      },
+    };
+
+    const text = extractResponseText(result);
+    // Should pick the last step (generate-response), NOT the longest (route-intent)
+    expect(text).toBe('To set up rate limiting, use the Tyk Dashboard.');
+    expect(text).not.toContain('Based on analysis');
+  });
+
   it('falls back to grouped results when history has no text', () => {
     const result = {
       reviewSummary: { issues: [], history: {} },
@@ -307,7 +337,7 @@ describe('extractResponseText', () => {
     expect(() => JSON.parse(text)).not.toThrow();
   });
 
-  it('handles multi-step workflow with multiple text outputs, picks longest', () => {
+  it('handles multi-step workflow with multiple text outputs, picks last', () => {
     const result = {
       reviewSummary: {
         issues: [],
@@ -322,6 +352,29 @@ describe('extractResponseText', () => {
     const text = extractResponseText(result);
     expect(text).toContain('comprehensive response');
   });
+
+  it('picks last grouped result, not longest', () => {
+    const result = {
+      reviewSummary: { issues: [], history: {} },
+      executionStatistics: {
+        groupedResults: {
+          routing: [
+            {
+              checkName: 'route',
+              output: {
+                text: 'Detailed routing analysis with lots of context about the user intent and classification',
+              },
+              issues: [],
+            },
+          ],
+          response: [{ checkName: 'respond', output: { text: 'Short final answer.' }, issues: [] }],
+        },
+      },
+    };
+
+    const text = extractResponseText(result);
+    expect(text).toBe('Short final answer.');
+  });
 });
 
 // ---------------------------------------------------------------------------