omnikey-cli 1.0.20 → 1.0.21

package/backend-dist/agent/agentServer.js

@@ -112,7 +112,7 @@ async function runToolLoop(initialResult, session, sessionId, send, log, tools,
         session.history.push(result.assistantMessage);
         session.history.push({
             role: 'user',
-            content: 'You have reached the maximum number of tool calls. Based on all the information gathered so far, provide a single, final, concise answer. Do not call any more tools.',
+            content: 'You have reached the maximum number of tool calls. Do NOT make any further tool calls or web searches. You MUST now provide a final answer directly. If you still need to gather information from the system, generate a `<shell_scripts>` block instead of making tool calls.',
         });
         result = await ai_client_1.aiClient.complete(aiModel, session.history, {
             tools: undefined,
@@ -123,9 +123,6 @@ async function runToolLoop(initialResult, session, sessionId, send, log, tools,
     log.info('Finished reasoning and tool calls: ', {
         reason: result.finish_reason,
     });
-    if (result.assistantMessage) {
-        session.history.push(result.assistantMessage);
-    }
     return result;
 }
 function buildAvailableTools() {
@@ -298,6 +295,7 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
         isError: isErrorFlag,
         rawContentLength: (clientMessage.content || '').length,
         userContentLength: userContent.length,
+        isRecursiveCall: clientMessage.is_web_call,
     });
     const isAssistance = isTerminalOutput || isErrorFlag;
     if (!clientMessage?.is_web_call) {
@@ -371,21 +369,60 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
                 subscriptionId: subscription.id,
                 turn: session.turns,
             });
-            result = await runToolLoop(result, session, sessionId, send, log, buildAvailableTools(), recordUsage);
-            content = result.content.trim();
-            if (!content) {
-                log.warn('Agent returned empty content after tool loop; sending generic error.');
-                sendFinalAnswer(send, sessionId, 'The agent returned an empty response. Please try again.', true);
-                sessionMessages.delete(sessionId);
+            const toolLoopResult = await runToolLoop(result, session, sessionId, send, log, buildAvailableTools(), recordUsage);
+            const toolLoopContent = toolLoopResult.content.trim();
+            const toolLoopHasShell = toolLoopContent.includes('<shell_script>');
+            const toolLoopHasFinal = toolLoopContent.includes('<final_answer>');
+            const webToolFailed = session.history.some((msg) => msg.role === 'tool' && typeof msg.content === 'string' && msg.content.startsWith('Error'));
+            if (toolLoopHasShell || (toolLoopHasFinal && !webToolFailed)) {
+                // The tool loop already produced a shell script — use it directly.
+                // This avoids a redundant AI call and handles the case where the model
+                // emits a <shell_script> immediately after its web tool calls.
+                log.info('Tool loop produced shell script; processing inline', { sessionId });
+                content = toolLoopContent;
+                result = toolLoopResult;
+                // Fall through to the <shell_script> handling below.
+            }
+            else {
+                // The tool loop returned either plain text or a <final_answer>.
+                // We always make one more AI turn here so the model has a chance to
+                // correct itself — specifically when web tools failed (404 / error) the
+                // model tends to wrap a "please run this manually" message in
+                // <final_answer>. The directive below tells it to use <shell_script> as
+                // a fallback instead of asking the user to run commands.
+                if (toolLoopResult.assistantMessage) {
+                    session.history.push(toolLoopResult.assistantMessage);
+                }
+                session.history.push({
+                    role: 'user',
+                    content: webToolFailed
+                        ? [
+                            'IMPORTANT: The web search tool failed and is unavailable. Do NOT attempt any further web calls or ask the user to run commands manually.',
+                            'You MUST retrieve any needed data by generating a <shell_script> that runs terminal commands (curl, grep, cat, etc.).',
+                            'The shell script output will be returned to you automatically.',
+                            '',
+                            'Respond with exactly one of:',
+                            '- <shell_script>...</shell_script> — to fetch or retrieve data via terminal commands',
+                            '- <final_answer>...</final_answer> — only if you already have enough information',
+                            'No plain text. No web tool calls. No other format.',
+                        ].join('\n')
+                        : [
+                            'Web research is complete. The results are in the conversation above.',
+                            '',
+                            'Now respond with exactly one of:',
+                            '- <shell_script>...</shell_script> — to run terminal commands (output will be returned to you automatically)',
+                            '- <final_answer>...</final_answer> — only if you genuinely have enough information',
+                            'No plain text. No other format.',
+                        ].join('\n'),
+                });
+                await runAgentTurn(sessionId, subscription, {
+                    sender: 'agent',
+                    session_id: sessionId,
+                    content: '',
+                    is_web_call: true,
+                }, send, logger_1.logger);
                 return;
             }
-            await runAgentTurn(sessionId, subscription, {
-                sender: 'agent',
-                session_id: sessionId,
-                content: '',
-                is_web_call: true,
-            }, send, logger_1.logger);
-            return;
         }
         // Ensure that a proper <final_answer> block is produced for the
         // desktop clients once we reach the final turn. If the model did

package/backend-dist/index.js

@@ -64,8 +64,8 @@ app.get('/macos/appcast', (req, res) => {
     const appcastUrl = `${baseUrl}/macos/appcast`;
     // These should match the values embedded into the macOS app
     // Info.plist in macOS/build_release_dmg.sh.
-    const bundleVersion = '16';
-    const shortVersion = '1.0.15';
+    const bundleVersion = '17';
+    const shortVersion = '1.0.16';
     const xml = `<?xml version="1.0" encoding="utf-8"?>
 <rss version="2.0"
      xmlns:sparkle="http://www.andymatuschak.org/xml-namespaces/sparkle"

package/package.json

@@ -4,7 +4,7 @@
     "access": "public",
     "registry": "https://registry.npmjs.org/"
   },
-  "version": "1.0.20",
+  "version": "1.0.21",
   "description": "CLI for onboarding users to Omnikey AI and configuring OPENAI_API_KEY. Use Yarn for install/build.",
   "engines": {
     "node": ">=14.0.0",