omnikey-cli 1.0.19 → 1.0.21

package/backend-dist/agent/agentPrompts.js

@@ -26,14 +26,16 @@ ${hasTaskInstructions
 - If a request needs BOTH machine data AND web search: emit a \`<shell_script>\` first → wait for \`TERMINAL OUTPUT:\` → then call the web tool with concrete values. Never use placeholders like "my IP" in a web query.
 
 **Incoming message tags:**
-- \`TERMINAL OUTPUT:\` — stdout/stderr from a prior script. Use it to continue reasoning or emit a follow-up.
+- \`TERMINAL OUTPUT:\` — stdout/stderr from a prior script. Analyze it immediately and respond with EITHER a follow-up \`<shell_script>\` (if more data is needed) OR a \`<final_answer>\` (if you have enough to conclude). You MUST pick one — never respond with plain text.
 - \`COMMAND ERROR:\` — script failed. Diagnose and emit a corrected \`<shell_script>\` or explain in \`<final_answer>\`.
 - No prefix — direct user message; treat as the primary request.
 
 **Response format — every response must be exactly one of:**
-1. \`<shell_script>...</shell_script>\` — to run commands.
+1. \`<shell_script>...</shell_script>\` — to run commands and gather more data.
 2. A \`web_search\` or \`web_fetch\` tool call — to fetch web context (use native tool calling, not XML tags).
-3. \`<final_answer>...</final_answer>\` — when done.
+3. \`<final_answer>...</final_answer>\` — your conclusion once you have enough information.
+
+**Critical rule:** After receiving \`TERMINAL OUTPUT:\` you MUST immediately produce either \`<shell_script>\` or \`<final_answer>\`. Never output raw text, markdown, or any other format. If the terminal output contains enough information to answer the user's request, output \`<final_answer>\` right away.
 
 No plain text, reasoning, or other tags outside these blocks. Never wrap in additional XML/JSON.
 

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,
@@ -254,7 +254,12 @@ async function authenticateFromAuthHeader(authHeader, log) {
     }
 }
 function createUserContent(content, hasStoredPrompt) {
-    return hasStoredPrompt ? content.replace(/@omniAgent/g, '').trim() : content;
+    return hasStoredPrompt
+        ? content
+            .toLowerCase()
+            .replace(/@omniagent/g, '')
+            .trim()
+        : content;
 }
 async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
     const { sessionState: session, hasStoredPrompt } = await getOrCreateSession(sessionId, subscription, clientMessage.platform, log);
@@ -290,9 +295,14 @@ 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) {
+        // Terminal output and command errors are always user-role messages — they
+        // represent environment feedback that the agent must reason about next.
+        // Pushing them as 'assistant' would create two consecutive assistant turns
+        // which breaks most LLM APIs and prevents the model from processing the output.
         session.history.push({
             role: 'user',
             content: isAssistance
@@ -306,7 +316,7 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
     const tools = isFinalTurn ? undefined : buildAvailableTools();
     const recordUsage = async (result) => {
         const usage = result.usage;
-        if (!usage || !subscription.id)
+        if (!usage || !subscription.id || config_1.config.isSelfHosted)
             return;
         try {
             await subscriptionUsage_1.SubscriptionUsage.create({
@@ -359,12 +369,58 @@ 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;
             }
         }
@@ -410,6 +466,31 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
             });
             sessionMessages.delete(sessionId);
         }
+        else if (content) {
+            // Fallback: the LLM returned content without any recognized tag and it
+            // is not the final turn (e.g. plain-text conclusion after terminal
+            // output). Treat it as a final answer so the client is never left
+            // hanging.
+            log.info('Agent returned untagged content on a non-final turn; treating as final answer', {
+                sessionId,
+                subscriptionId: subscription.id,
+                turn: session.turns,
+            });
+            session.history.push({ role: 'assistant', content });
+            send({
+                session_id: sessionId,
+                sender: 'agent',
+                content: `<final_answer>\n${content}\n</final_answer>`,
+            });
+            sessionMessages.delete(sessionId);
+        }
+        else {
+            log.warn('Agent returned empty content with no recognized tags; sending error', {
+                sessionId,
+            });
+            sendFinalAnswer(send, sessionId, 'The agent returned an empty response. Please try again.', true);
+            sessionMessages.delete(sessionId);
+        }
     }
     catch (err) {
         log.error('Agent LLM call failed', { error: err });

package/backend-dist/featureRoutes.js

@@ -118,7 +118,7 @@ async function enhanceText(logger, text, cmd, subscription) {
         const { rawResponse, usage, model } = result;
         // Record token usage for this subscription and model, if usage
         // data is available and we know which subscription made the call.
-        if (usage && subscription.id) {
+        if (usage && subscription.id && !config_1.config.isSelfHosted) {
             try {
                 await subscriptionUsage_1.SubscriptionUsage.create({
                     subscriptionId: subscription.id,
@@ -197,7 +197,7 @@ async function streamEnhanceResponse(res, text, cmd) {
             return;
         }
         const { usage, model } = result;
-        if (usage && subscription.id) {
+        if (usage && subscription.id && !config_1.config.isSelfHosted) {
             try {
                 await subscriptionUsage_1.SubscriptionUsage.create({
                     subscriptionId: subscription.id,

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": "1.0.19",
+  "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",