polydev-ai 1.5.2 → 1.6.0

package/lib/cliManager.js

@@ -42,6 +42,9 @@ class CLIManager {
         id: 'claude_code',
         name: 'Claude Code',
         command: process.env.CLAUDE_CODE_PATH || 'claude',
+        // Model configuration: use env var or default to 'haiku' for speed
+        // Options: haiku (fastest), sonnet (balanced), opus (most capable)
+        model: process.env.POLYDEV_CLAUDE_MODEL || 'haiku',
         subcommands: {
           chat: [],
           version: ['--version'],
@@ -55,6 +58,9 @@ class CLIManager {
         id: 'codex_cli',
         name: 'Codex CLI',
         command: process.env.CODEX_CLI_PATH || 'codex',
+        // Model configuration: use env var or default to 'gpt-4o-mini' for speed
+        // Options: gpt-4o-mini (fastest), gpt-4o, o3-mini, o1, etc.
+        model: process.env.POLYDEV_CODEX_MODEL || 'gpt-4o-mini',
         subcommands: {
           chat: ['chat'],
           version: ['--version'],
@@ -72,6 +78,8 @@ class CLIManager {
         id: 'gemini_cli',
         name: 'Gemini CLI',
         command: process.env.GEMINI_CLI_PATH || 'gemini',
+        // Model configuration for Gemini
+        model: process.env.POLYDEV_GEMINI_MODEL || 'gemini-2.0-flash',
         subcommands: {
           chat: ['chat'],
           version: ['--version'],
@@ -437,13 +445,15 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
       if (providerId === 'codex_cli') {
         const execArgs = promptVariants.find(args => args.includes('exec')) || promptVariants[0];
         try {
-          const content = await this.executeCodexExec(provider.command, execArgs, prompt, timeoutMs);
+          // Pass model configuration to Codex
+          const content = await this.executeCodexExec(provider.command, execArgs, prompt, timeoutMs, provider.model);
           return {
             success: true,
             content,
             tokens_used: this.estimateTokens(prompt + content),
             latency_ms: Date.now() - startTime,
             provider: providerId,
+            model: provider.model,
             mode: 'args',
             timestamp: new Date()
           };
@@ -453,6 +463,7 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
             error: `CLI execution failed: ${error instanceof Error ? error.message : String(error)}`,
             latency_ms: Date.now() - startTime,
             provider: providerId,
+            model: provider.model,
             mode,
             timestamp: new Date()
           };
@@ -462,7 +473,9 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
       let lastErrorMessage = null;
 
       for (const promptArgs of promptVariants) {
-        const args = Array.isArray(promptArgs) ? [...promptArgs, prompt] : [prompt];
+        // Add model parameter for Claude Code
+        const modelArgs = provider.model ? ['--model', provider.model] : [];
+        const args = Array.isArray(promptArgs) ? [...promptArgs, ...modelArgs, prompt] : [...modelArgs, prompt];
         try {
           const result = await this.executeCliCommand(
             provider.command,
@@ -480,6 +493,7 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
               tokens_used: this.estimateTokens(prompt + content),
               latency_ms: Date.now() - startTime,
               provider: providerId,
+              model: provider.model,
               mode: 'args',
               timestamp: new Date()
             };
@@ -609,7 +623,7 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
     });
   }
 
-  async executeCodexExec(executable, commandArgs, prompt, timeoutMs) {
+  async executeCodexExec(executable, commandArgs, prompt, timeoutMs, model) {
     if (!executable) {
       throw new Error('Missing Codex executable');
     }
@@ -619,8 +633,12 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
     }
 
     const workingDir = process.cwd();
+    
+    // Build args with model configuration if specified
+    const modelArgs = model ? ['-c', `model="${model}"`] : [];
     const args = [
       ...commandArgs,
+      ...modelArgs,
       '--sandbox',
       'workspace-write',
       '--skip-git-repo-check',
@@ -675,27 +693,73 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
       const parseCodexOutput = (output) => {
         if (!output || !output.trim()) return null;
         
+        // First, try to extract the response between "codex" marker and "tokens used"
+        // This is the most reliable pattern for Codex CLI output
+        const codexMarkerMatch = output.match(/\bcodex\s*\n([\s\S]*?)(?:\n\s*tokens used|\n\s*$)/i);
+        if (codexMarkerMatch && codexMarkerMatch[1]) {
+          const extracted = codexMarkerMatch[1].trim();
+          if (extracted.length > 0 && !extracted.startsWith('ERROR')) {
+            return extracted;
+          }
+        }
+        
+        // Fallback: Try to find bullet point responses
+        const bulletMatches = output.match(/•\s*(.+)/g);
+        if (bulletMatches && bulletMatches.length > 0) {
+          const bulletContent = bulletMatches
+            .map(m => m.replace(/^•\s*/, '').trim())
+            .filter(s => s.length > 0)
+            .join('\n');
+          if (bulletContent.length > 0) {
+            return bulletContent;
+          }
+        }
+        
+        // Last resort: Filter out known noise patterns line by line
         const lines = output.split('\n');
         const contentLines = [];
         
-        // Filter out MCP protocol noise and collect actual content
+        // Patterns to skip (Codex-specific noise)
+        const noisePatterns = [
+          /^\s*$/,                                          // Empty lines
+          /^\d{4}-\d{2}-\d{2}T[\d:]+.*?(ERROR|WARN|INFO)/i, // Timestamp logs
+          /^OpenAI Codex v[\d.]+/i,                         // Version banner
+          /^-{4,}$/,                                        // Separator lines
+          /^workdir:/i,                                     // Header fields
+          /^model:/i,
+          /^provider:/i,
+          /^approval:/i,
+          /^sandbox:/i,
+          /^reasoning effort:/i,
+          /^reasoning summaries:/i,
+          /^session id:/i,
+          /^user$/i,                                        // "user" marker
+          /^thinking$/i,                                    // "thinking" marker
+          /^codex$/i,                                       // "codex" marker
+          /^tokens used$/i,                                 // Token count header
+          /^[\d,]+$/,                                       // Just numbers (token counts)
+          /^ERROR:\s*MCP/i,                                 // MCP errors
+          /MCP client for .* failed/i,                      // MCP client failures
+          /handshake.*failed/i,                             // Handshake errors
+          /connection closed/i,                             // Connection errors
+          /\[MCP\]/i,                                       // MCP tags
+          /^MCP\s/i,                                        // MCP prefix
+          /rmcp::transport/i,                               // Rust MCP transport errors
+          /serde error/i,                                   // Serde parsing errors
+          /^\*\*.*\*\*$/,                                   // Bold status messages like **Awaiting next steps**
+        ];
+        
         for (const line of lines) {
           const trimmedLine = line.trim();
-          // Skip empty lines and MCP/protocol noise
-          if (!trimmedLine) continue;
-          if (trimmedLine.startsWith('MCP ')) continue;
-          if (trimmedLine.includes('handshake') && trimmedLine.includes('error')) continue;
-          if (trimmedLine.startsWith('[MCP]')) continue;
-          if (trimmedLine.startsWith('Error:') && trimmedLine.includes('MCP')) continue;
           
-          // Extract bullet content if present, otherwise keep the line
-          const bulletMatch = trimmedLine.match(/^•\s*(.+)/);
-          if (bulletMatch) {
-            contentLines.push(bulletMatch[1].trim());
-          } else if (!trimmedLine.startsWith('•')) {
-            // Keep non-bullet content that isn't just whitespace
-            contentLines.push(trimmedLine);
-          }
+          // Check if line matches any noise pattern
+          const isNoise = noisePatterns.some(pattern => pattern.test(trimmedLine));
+          if (isNoise) continue;
+          
+          // Skip if line looks like the echoed user prompt (heuristic: contains "?" and is long)
+          if (trimmedLine.includes('?') && trimmedLine.length > 20) continue;
+          
+          contentLines.push(trimmedLine);
         }
         
         const result = contentLines.join('\n').trim();
@@ -730,26 +794,49 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
 
         const trimmedStdout = stdout.trim();
         const trimmedStderr = stderr.trim();
+        
+        // Combine stdout and stderr for parsing - Codex may output to either
+        const combinedOutput = trimmedStdout + '\n' + trimmedStderr;
+
+        // Always try to parse stdout first, regardless of exit code
+        // MCP handshake failures cause non-zero exit but don't prevent valid responses
+        const parsedStdout = parseCodexOutput(trimmedStdout);
+        if (parsedStdout) {
+          resolve(parsedStdout);
+          return;
+        }
+        
+        // Try parsing combined output (some responses may appear in stderr)
+        const parsedCombined = parseCodexOutput(combinedOutput);
+        if (parsedCombined) {
+          resolve(parsedCombined);
+          return;
+        }
 
-        if (code === 0 && trimmedStdout) {
-          // Use the robust parser for consistent output handling
-          const parsed = parseCodexOutput(trimmedStdout);
-          if (parsed) {
-            resolve(parsed);
-            return;
-          }
-          // Fallback to raw output if parsing yields nothing
-          resolve(trimmedStdout);
-        } else if (code === 0) {
-          // Code 0 but no stdout - check if stderr has useful info
-          const parsedStderr = parseCodexOutput(trimmedStderr);
-          if (parsedStderr) {
-            resolve(parsedStderr);
-            return;
+        // Only now consider it a failure
+        if (code === 0) {
+          // Successful exit but no parseable output
+          if (trimmedStdout) {
+            resolve(trimmedStdout); // Return raw output as fallback
+          } else {
+            reject(new Error('Codex completed but produced no output'));
           }
-          reject(new Error('Codex completed but produced no output'));
         } else {
-          reject(new Error(trimmedStderr || trimmedStdout || `Codex exited with code ${code}`));
+          // Non-zero exit and no parseable response - this is a real error
+          // Filter error message to remove MCP noise
+          const errorLines = (trimmedStderr || trimmedStdout).split('\n')
+            .filter(line => {
+              const l = line.trim().toLowerCase();
+              return l.length > 0 && 
+                     !l.includes('mcp') && 
+                     !l.includes('handshake') &&
+                     !l.includes('rmcp::') &&
+                     !l.includes('serde error');
+            })
+            .slice(0, 3); // Only first 3 relevant error lines
+          
+          const cleanError = errorLines.join('; ') || `Codex exited with code ${code}`;
+          reject(new Error(cleanError));
         }
       });
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polydev-ai",
-  "version": "1.5.2",
+  "version": "1.6.0",
   "description": "Agentic workflow assistant with CLI integration - get diverse perspectives from multiple LLMs when stuck or need enhanced reasoning",
   "keywords": [
     "mcp",