omnikey-cli 1.0.15 → 1.0.16

package/backend-dist/agent/agentPrompts.js

@@ -42,11 +42,17 @@ When you generate shell scripts, make them clear, efficient, and focused on gath
 - If there is a conflict, follow: system rules first, then stored instructions, then ad-hoc guidance in the current input.
 
 **Web tools:**
-You have access to web tools you can call at any time during a turn:
-- \`web_fetch(url)\`: Fetches the text content of any publicly accessible URL. Use it to retrieve documentation, error references, API guides, release notes, or any other web resource that would help answer the user's question.
-- \`web_search(query)\`: Searches the web and returns a list of relevant results (title, URL, snippet). Use it when you need to discover the right URL before fetching, or when a quick summary of search results is sufficient.
+You have access to web tools, but you must use them sparingly and only when explicitly required:
+- \`web_fetch(url)\`: Only call this when the user has provided a specific URL in their current input or stored instructions and you need to retrieve its contents.
+- \`web_search(query)\`: Only call this when the user has explicitly asked you to search the web or look something up online.
 
-Use these tools proactively whenever the question involves current information, external documentation, or anything not already available in the conversation or machine output. You may call web tools multiple times in a single turn; call \`web_fetch\` on a promising URL from \`web_search\` results to get full details. Web tool results are injected back into the conversation automatically; continue reasoning and then emit your shell script or final answer as normal.
+Do NOT use web tools proactively. Do NOT call them to look up documentation, error references, or general information you could infer from the machine output or your own knowledge. Your primary workflow is to generate shell scripts, wait for the terminal output, and reason from that output. Only reach for web tools when there is a clear, explicit instruction or a URL provided by the user.
+
+**User message tags:**
+User messages may be prefixed with special tags that indicate their origin:
+- \`TERMINAL OUTPUT:\` — the content is stdout/stderr returned from a previously requested \`<shell_script>\`. Parse it as machine output and use it to continue your reasoning toward a \`<final_answer>\` or a follow-up \`<shell_script>\`.
+- \`COMMAND ERROR:\` — the shell script failed or the terminal returned a non-zero exit code. Treat the content as error output: diagnose the failure, then either emit a corrected \`<shell_script>\` or explain the issue in a \`<final_answer>\`.
+- No prefix — the content is a direct message from the user; treat it as the primary request or question to address.
 
 **Interaction rules:**
 - When you need to execute ANY shell command, respond with a single \`<shell_script>\` block that contains the FULL script to run.

package/backend-dist/agent/agentServer.js

@@ -255,8 +255,13 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
         let toolIterations = 0;
         while (result.finish_reason === 'tool_calls' && toolIterations < MAX_TOOL_ITERATIONS) {
             toolIterations++;
-            session.history.push(result.assistantMessage);
             const toolCalls = result.tool_calls ?? [];
+            // If the model claims tool_calls but sent none, treat it as a normal text
+            // response — pushing an assistant message with no following tool results
+            // would leave the history ending with an assistant turn, causing a 400.
+            if (!toolCalls.length)
+                break;
+            session.history.push(result.assistantMessage);
             log.info('Agent executing tool calls', {
                 sessionId,
                 turn: session.turns,
@@ -265,6 +270,18 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
             });
             const toolResults = await Promise.all(toolCalls.map(async (tc) => {
                 const args = tc.arguments;
+                // Notify the frontend that a web tool call is about to execute.
+                const webCallContent = tc.name === 'web_search'
+                    ? `Searching the web for: "${args.query ?? ''}"`
+                    : `Fetching URL: ${args.url ?? ''}`;
+                send({
+                    session_id: sessionId,
+                    sender: 'agent',
+                    content: webCallContent,
+                    is_terminal_output: false,
+                    is_error: false,
+                    is_web_call: true,
+                });
                 const toolResult = await (0, web_search_provider_1.executeTool)(tc.name, args, log);
                 log.info('Tool call completed', {
                     sessionId,
@@ -287,10 +304,14 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
             });
             await recordUsage(result);
         }
+        log.info('Finished reasoning and tool calls: ', {
+            reason: result.finish_reason,
+        });
+        const content = result.content.trim();
         // If the tool loop was exhausted while the model still wants more tool calls,
         // the last result has empty content. Force one final no-tools call so the model
         // must synthesize a text answer from everything gathered so far.
-        if (result.finish_reason === 'tool_calls') {
+        if (result.finish_reason === 'tool_calls' || !content) {
             log.warn('Tool iteration limit reached with pending tool calls; forcing final text response', {
                 sessionId,
                 turn: session.turns,
@@ -309,7 +330,6 @@ async function runAgentTurn(sessionId, subscription, clientMessage, send, log) {
             });
             await recordUsage(result);
         }
-        const content = result.content.trim();
         if (!content) {
             log.warn('Agent LLM returned empty content; sending generic error to client.');
             const errorMessage = 'The agent returned an empty response. Please try again.';

package/backend-dist/index.js

@@ -19,19 +19,31 @@ const app = (0, express_1.default)();
 const PORT = Number(config_1.config.port);
 app.use((0, cors_1.default)());
 app.use(express_1.default.json());
+// Landing page
+app.use(express_1.default.static(path_1.default.join(process.cwd(), 'public')));
 app.use('/api/subscription', (0, subscriptionRoutes_1.createSubscriptionRouter)(logger_1.logger));
 app.use('/api/feature', (0, featureRoutes_1.createFeatureRouter)());
 app.use('/api/instructions', (0, taskInstructionRoutes_1.taskInstructionRouter)());
-app.get('/macos/download', (req, res) => {
+app.get('/macos/download', (_req, res) => {
     const dmgPath = path_1.default.join(process.cwd(), 'macOS', 'OmniKeyAI.dmg');
-    res.download(dmgPath, 'OmniKeyAI.dmg', (err) => {
-        if (err) {
-            logger_1.logger.error('Failed to send OmniKeyAI.dmg for download.', { error: err });
-            if (!res.headersSent) {
-                res.status(500).send('Unable to download file.');
-            }
+    if (!fs_1.default.existsSync(dmgPath)) {
+        res.status(404).send('File not found.');
+        return;
+    }
+    res.set({
+        'Content-Type': 'application/octet-stream',
+        'Content-Disposition': 'attachment; filename="OmniKeyAI.dmg"',
+        'Content-Encoding': 'gzip',
+    });
+    const fileStream = fs_1.default.createReadStream(dmgPath);
+    const gzip = zlib_1.default.createGzip();
+    fileStream.on('error', (err) => {
+        logger_1.logger.error('Failed to send OmniKeyAI.dmg for download.', { error: err });
+        if (!res.headersSent) {
+            res.status(500).send('Unable to download file.');
         }
     });
+    fileStream.pipe(gzip).pipe(res);
 });
 // Sparkle appcast feed for macOS updates.
 // This feed uses the existing /macos/download endpoint as the
@@ -52,8 +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 = '13';
-    const shortVersion = '1.0.12';
+    const bundleVersion = '14';
+    const shortVersion = '1.0.13';
     const xml = `<?xml version="1.0" encoding="utf-8"?>
 <rss version="2.0"
      xmlns:sparkle="http://www.andymatuschak.org/xml-namespaces/sparkle"
@@ -81,7 +93,7 @@ app.get('/macos/appcast', (req, res) => {
 // ── Windows distribution endpoints ───────────────────────────────────────────
 // These should match the values in windows/OmniKey.Windows.csproj
 // <Version> and windows/build_release_zip.ps1 $APP_VERSION.
-const WIN_VERSION = '1.1';
+const WIN_VERSION = '1.2';
 const WIN_ZIP_FILENAME = 'OmniKeyAI-windows-win-x64.zip';
 const WIN_ZIP_PATH = path_1.default.join(process.cwd(), 'windows', WIN_ZIP_FILENAME);
 // Serves the pre-built ZIP produced by windows/build_release_zip.ps1.
@@ -128,6 +140,9 @@ app.get('/windows/update', (req, res) => {
 app.get('/health', (_req, res) => {
     res.json({ status: 'ok' });
 });
+app.get('*', (_req, res) => {
+    res.sendFile(path_1.default.join(process.cwd(), 'public', 'index.html'));
+});
 let server = null;
 async function start() {
     try {

package/package.json

@@ -4,7 +4,7 @@
     "access": "public",
     "registry": "https://registry.npmjs.org/"
   },
-  "version": "1.0.15",
+  "version": "1.0.16",
   "description": "CLI for onboarding users to Omnikey AI and configuring OPENAI_API_KEY. Use Yarn for install/build.",
   "engines": {
     "node": ">=14.0.0",