nothumanallowed 13.5.128 → 13.5.129

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "13.5.128",
+  "version": "13.5.129",
   "description": "NotHumanAllowed — 38 AI agents, 80 tools, Studio (visual agentic workflows). Email, calendar, browser automation, screen capture, canvas, cron/heartbeat, Alexandria E2E messaging, GitHub, Notion, Slack, voice chat, free AI (Liara), 28 languages. Zero-dependency CLI.",
   "type": "module",
   "bin": {

package/src/commands/ui.mjs

@@ -2111,9 +2111,33 @@ export async function cmdUI(args) {
               clean = clean.replace(/\[Screenshot[^\]]*\]/g, '').replace(/!\[.*?\]\(data:image[^)]+\)/g, '').slice(0, 3000);
               return `[${t.action} result]: ${clean.trim() || 'Done.'}`;
             }).join('\n\n');
-            const followUp = `The user asked: "${body.message}"\n\nI executed these tools and got REAL results:\n\n${toolContext}\n\nNow respond conversationally based ONLY on the REAL data above. Do NOT output any JSON blocks, base64, or image markdown — just natural text.`;
+            const followUp = `The user asked: "${body.message}"\n\nI executed these tools and got REAL results:\n\n${toolContext}\n\nNow respond to the user conversationally based ONLY on the REAL data above. If the user's request has multiple steps and the first step is done, execute the next step using a JSON tool block. Do NOT embed base64 data or image markdown — just natural text.`;
             try {
               fullResponse = await callLLM(config, enrichedSystemPrompt, followUp);
+
+              // Round 2: execute any tool calls emitted in the synthesis response
+              const { textParts: synthText2, actions: synthActions2 } = parseActions(fullResponse);
+              if (synthActions2.length > 0) {
+                const round2Results = [];
+                for (const { action: a2, params: p2 } of synthActions2) {
+                  try {
+                    const r2 = await executeTool(a2, p2, config);
+                    round2Results.push({ action: a2, result: typeof r2 === 'object' ? JSON.stringify(r2) : String(r2) });
+                  } catch (e2) {
+                    round2Results.push({ action: a2, result: `Error: ${e2.message}` });
+                  }
+                }
+                const round2Context = round2Results.map(t => `[${t.action} result]: ${t.result.slice(0, 2000)}`).join('\n\n');
+                try {
+                  const r2Summary = await callLLM(config, enrichedSystemPrompt, `The user asked: "${body.message}"\n\n${toolContext}\n\nRound 2 tool results:\n\n${round2Context}\n\nGive the user a final natural-language summary of everything. Do NOT output JSON blocks.`);
+                  fullResponse = synthText2.join('\n').replace(/```json[\s\S]*?```/g, '').trim() + (synthText2.join('').trim() ? '\n\n' : '') + r2Summary.trim();
+                } catch {
+                  fullResponse = synthText2.join('\n').replace(/```json[\s\S]*?```/g, '').trim() + '\n\n' + round2Results.map(t => `${t.action}: ${t.result}`).join('\n');
+                }
+              } else {
+                fullResponse = fullResponse.replace(/```json[\s\S]*?```/g, '').trim();
+              }
+
               // Prepend preserved markers so the UI can render canvas/screenshots
               if (preservedMarkers) fullResponse = preservedMarkers + fullResponse;
             } catch {
@@ -2615,17 +2639,50 @@ export async function cmdUI(args) {
               sendSSE('canvas', { markers: preservedMarkers });
             }
 
-            const followUp = `The user asked: "${msg}"\n\nI executed these tools and got REAL results:\n\n${toolContext}\n\nNow respond to the user conversationally based ONLY on the REAL data above. Present the results clearly. Do NOT output any JSON blocks, any base64 data, or any image markdown — just natural text. If a screenshot was taken, just mention "Screenshot captured" without embedding it.`;
+            const followUp = `The user asked: "${msg}"\n\nI executed these tools and got REAL results:\n\n${toolContext}\n\nNow respond to the user conversationally based ONLY on the REAL data above. Present the results clearly. If the user's request has multiple steps and the first step is done, execute the next step using a JSON tool block. Do NOT embed base64 data or image markdown — just natural text. If a screenshot was taken, just mention "Screenshot captured" without embedding it.`;
             sendSSE('tool_synthesis', {});
             try {
               finalResponse = await callLLMStream(config, enrichedPrompt, followUp, (chunk) => {
                 sendSSE('token', { content: chunk });
               });
               finalResponse = finalResponse
-                .replace(/```json[\s\S]*?```/g, '')
                 .replace(/!\[.*?\]\(data:image\/[^)]+\)/g, '')
                 .replace(/data:image\/[a-z]+;base64,[A-Za-z0-9+/=]{100,}/g, '[image]')
                 .trim();
+
+              // Round 2: execute any tool calls emitted in the synthesis response
+              const { textParts: synthText, actions: synthActions } = parseActions(finalResponse);
+              if (synthActions.length > 0) {
+                const round2Results = [];
+                for (const { action: a2, params: p2 } of synthActions) {
+                  sendSSE('tool', { action: a2, status: 'executing' });
+                  try {
+                    const r2 = await executeTool(a2, p2, config);
+                    const r2str = typeof r2 === 'object' ? JSON.stringify(r2) : String(r2);
+                    round2Results.push({ action: a2, result: r2str });
+                    sendSSE('tool', { action: a2, status: 'done', result: r2str.slice(0, 200) });
+                  } catch (e2) {
+                    round2Results.push({ action: a2, result: `Error: ${e2.message}` });
+                    sendSSE('tool', { action: a2, status: 'error', error: e2.message });
+                  }
+                }
+                const round2Context = round2Results.map(t => `[${t.action} result]: ${t.result.slice(0, 2000)}`).join('\n\n');
+                const round2Prompt = `${toolContext}\n\nRound 2 tool results:\n\n${round2Context}\n\nNow give the user a final natural-language summary of everything that was done. Do NOT output JSON blocks.`;
+                sendSSE('tool_synthesis', {});
+                try {
+                  finalResponse = synthText.join('\n').replace(/```json[\s\S]*?```/g, '').trim();
+                  const r2Summary = await callLLMStream(config, enrichedPrompt, `The user asked: "${msg}"\n\n${round2Prompt}`, (chunk) => {
+                    sendSSE('token', { content: chunk });
+                  });
+                  finalResponse = (finalResponse ? finalResponse + '\n\n' : '') + r2Summary.trim();
+                } catch {
+                  finalResponse = synthText.join('\n').replace(/```json[\s\S]*?```/g, '').trim() + '\n\n' + round2Results.map(t => `${t.action}: ${t.result}`).join('\n');
+                }
+              } else {
+                // No new tool calls — strip any leftover JSON blocks from display text
+                finalResponse = finalResponse.replace(/```json[\s\S]*?```/g, '').trim();
+              }
+
               // Prepend preserved markers for persistence
               if (preservedMarkers) finalResponse = preservedMarkers + finalResponse;
             } catch {

package/src/constants.mjs

@@ -5,7 +5,7 @@ import { fileURLToPath } from 'url';
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = path.dirname(__filename);
 
-export const VERSION = '13.5.128';
+export const VERSION = '13.5.129';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';