fix: Add structured output fallback for non-Claude models in app spec generation

apps/server/src/providers/codex-provider.ts

@@ -98,9 +98,14 @@ const TEXT_ENCODING = 'utf-8';
  * This is the "no output" timeout - if the CLI doesn't produce any JSONL output
  * for this duration, the process is killed. For reasoning models with high
  * reasoning effort, this timeout is dynamically extended via calculateReasoningTimeout().
+ *
+ * For feature generation (which can generate 50+ features), we use a much longer
+ * base timeout (5 minutes) since Codex models are slower at generating large JSON responses.
+ *
  * @see calculateReasoningTimeout from @automaker/types
  */
 const CODEX_CLI_TIMEOUT_MS = DEFAULT_TIMEOUT_MS;
+const CODEX_FEATURE_GENERATION_BASE_TIMEOUT_MS = 300000; // 5 minutes for feature generation
 const CONTEXT_WINDOW_256K = 256000;
 const MAX_OUTPUT_32K = 32000;
 const MAX_OUTPUT_16K = 16000;
@@ -827,7 +832,14 @@ export class CodexProvider extends BaseProvider {
       // Higher reasoning effort (e.g., 'xhigh' for "xtra thinking" mode) requires more time
       // for the model to generate reasoning tokens before producing output.
       // This fixes GitHub issue #530 where features would get stuck with reasoning models.
-      const timeout = calculateReasoningTimeout(options.reasoningEffort, CODEX_CLI_TIMEOUT_MS);
+      //
+      // For feature generation with 'xhigh', use the extended 5-minute base timeout
+      // since generating 50+ features takes significantly longer than normal operations.
+      const baseTimeout =
+        options.reasoningEffort === 'xhigh'
+          ? CODEX_FEATURE_GENERATION_BASE_TIMEOUT_MS
+          : CODEX_CLI_TIMEOUT_MS;
+      const timeout = calculateReasoningTimeout(options.reasoningEffort, baseTimeout);
 
       const stream = spawnJSONLProcess({
         command: commandPath,

apps/server/src/routes/app-spec/generate-features-from-spec.ts

@@ -8,10 +8,11 @@
 import * as secureFs from '../../lib/secure-fs.js';
 import type { EventEmitter } from '../../lib/events.js';
 import { createLogger } from '@automaker/utils';
-import { DEFAULT_PHASE_MODELS } from '@automaker/types';
+import { DEFAULT_PHASE_MODELS, supportsStructuredOutput, isCodexModel } from '@automaker/types';
 import { resolvePhaseModel } from '@automaker/model-resolver';
 import { streamingQuery } from '../../providers/simple-query-service.js';
 import { parseAndCreateFeatures } from './parse-and-create-features.js';
+import { extractJsonWithArray } from '../../lib/json-extractor.js';
 import { getAppSpecPath } from '@automaker/platform';
 import type { SettingsService } from '../../services/settings-service.js';
 import {
@@ -25,6 +26,64 @@ const logger = createLogger('SpecRegeneration');
 
 const DEFAULT_MAX_FEATURES = 50;
 
+/**
+ * Timeout for Codex models when generating features (5 minutes).
+ * Codex models are slower and need more time to generate 50+ features.
+ */
+const CODEX_FEATURE_GENERATION_TIMEOUT_MS = 300000; // 5 minutes
+
+/**
+ * Type for extracted features JSON response
+ */
+interface FeaturesExtractionResult {
+  features: Array<{
+    id: string;
+    category?: string;
+    title: string;
+    description: string;
+    priority?: number;
+    complexity?: 'simple' | 'moderate' | 'complex';
+    dependencies?: string[];
+  }>;
+}
+
+/**
+ * JSON schema for features output format (Claude/Codex structured output)
+ */
+const featuresOutputSchema = {
+  type: 'object',
+  properties: {
+    features: {
+      type: 'array',
+      items: {
+        type: 'object',
+        properties: {
+          id: { type: 'string', description: 'Unique feature identifier (kebab-case)' },
+          category: { type: 'string', description: 'Feature category' },
+          title: { type: 'string', description: 'Short, descriptive title' },
+          description: { type: 'string', description: 'Detailed feature description' },
+          priority: {
+            type: 'number',
+            description: 'Priority level: 1 (highest) to 5 (lowest)',
+          },
+          complexity: {
+            type: 'string',
+            enum: ['simple', 'moderate', 'complex'],
+            description: 'Implementation complexity',
+          },
+          dependencies: {
+            type: 'array',
+            items: { type: 'string' },
+            description: 'IDs of features this depends on',
+          },
+        },
+        required: ['id', 'title', 'description'],
+      },
+    },
+  },
+  required: ['features'],
+} as const;
+
 export async function generateFeaturesFromSpec(
   projectPath: string,
   events: EventEmitter,
@@ -136,23 +195,80 @@ Generate ${featureCount} NEW features that build on each other logically. Rememb
         provider: undefined,
         credentials: undefined,
       };
-  const { model, thinkingLevel } = resolvePhaseModel(phaseModelEntry);
+  const { model, thinkingLevel, reasoningEffort } = resolvePhaseModel(phaseModelEntry);
 
   logger.info('Using model:', model, provider ? `via provider: ${provider.name}` : 'direct API');
 
+  // Codex models need extended timeout for generating many features.
+  // Use 'xhigh' reasoning effort to get 5-minute timeout (300s base * 1.0x = 300s).
+  // The Codex provider has a special 5-minute base timeout for feature generation.
+  const isCodex = isCodexModel(model);
+  const effectiveReasoningEffort = isCodex ? 'xhigh' : reasoningEffort;
+
+  if (isCodex) {
+    logger.info('Codex model detected - using extended timeout (5 minutes for feature generation)');
+  }
+  if (effectiveReasoningEffort) {
+    logger.info('Reasoning effort:', effectiveReasoningEffort);
+  }
+
+  // Determine if we should use structured output based on model type
+  const useStructuredOutput = supportsStructuredOutput(model);
+  logger.info(
+    `Structured output mode: ${useStructuredOutput ? 'enabled (Claude/Codex)' : 'disabled (using JSON instructions)'}`
+  );
+
+  // Build the final prompt - for non-Claude/Codex models, include explicit JSON instructions
+  let finalPrompt = prompt;
+  if (!useStructuredOutput) {
+    finalPrompt = `${prompt}
+
+CRITICAL INSTRUCTIONS:
+1. DO NOT write any files. Return the JSON in your response only.
+2. After analyzing the spec, respond with ONLY a JSON object - no explanations, no markdown, just raw JSON.
+3. The JSON must have this exact structure:
+{
+  "features": [
+    {
+      "id": "unique-feature-id",
+      "category": "Category Name",
+      "title": "Short Feature Title",
+      "description": "Detailed description of the feature",
+      "priority": 1,
+      "complexity": "simple|moderate|complex",
+      "dependencies": ["other-feature-id"]
+    }
+  ]
+}
+
+4. Feature IDs must be unique, lowercase, kebab-case (e.g., "user-authentication", "data-export")
+5. Priority ranges from 1 (highest) to 5 (lowest)
+6. Complexity must be one of: "simple", "moderate", "complex"
+7. Dependencies is an array of feature IDs that must be completed first (can be empty)
+
+Your entire response should be valid JSON starting with { and ending with }. No text before or after.`;
+  }
+
   // Use streamingQuery with event callbacks
   const result = await streamingQuery({
-    prompt,
+    prompt: finalPrompt,
     model,
     cwd: projectPath,
     maxTurns: 250,
     allowedTools: ['Read', 'Glob', 'Grep'],
     abortController,
     thinkingLevel,
+    reasoningEffort: effectiveReasoningEffort, // Extended timeout for Codex models
     readOnly: true, // Feature generation only reads code, doesn't write
     settingSources: autoLoadClaudeMd ? ['user', 'project', 'local'] : undefined,
     claudeCompatibleProvider: provider, // Pass provider for alternative endpoint configuration
     credentials, // Pass credentials for resolving 'credentials' apiKeySource
+    outputFormat: useStructuredOutput
+      ? {
+          type: 'json_schema',
+          schema: featuresOutputSchema,
+        }
+      : undefined,
     onText: (text) => {
       logger.debug(`Feature text block received (${text.length} chars)`);
       events.emit('spec-regeneration:event', {
@@ -163,15 +279,51 @@ Generate ${featureCount} NEW features that build on each other logically. Rememb
     },
   });
 
-  const responseText = result.text;
+  // Get response content - prefer structured output if available
+  let contentForParsing: string;
+
+  if (result.structured_output) {
+    // Use structured output from Claude/Codex models
+    logger.info('✅ Received structured output from model');
+    contentForParsing = JSON.stringify(result.structured_output);
+    logger.debug('Structured output:', contentForParsing);
+  } else {
+    // Use text response (for non-Claude/Codex models or fallback)
+    // Pre-extract JSON to handle conversational text that may surround the JSON response
+    // This follows the same pattern used in generate-spec.ts and validate-issue.ts
+    const rawText = result.text;
+    logger.info(`Feature stream complete.`);
+    logger.info(`Feature response length: ${rawText.length} chars`);
+    logger.info('========== FULL RESPONSE TEXT ==========');
+    logger.info(rawText);
+    logger.info('========== END RESPONSE TEXT ==========');
 
-  logger.info(`Feature stream complete.`);
-  logger.info(`Feature response length: ${responseText.length} chars`);
-  logger.info('========== FULL RESPONSE TEXT ==========');
-  logger.info(responseText);
-  logger.info('========== END RESPONSE TEXT ==========');
+    // Pre-extract JSON from response - handles conversational text around the JSON
+    const extracted = extractJsonWithArray<FeaturesExtractionResult>(rawText, 'features', {
+      logger,
+    });
+    if (extracted) {
+      contentForParsing = JSON.stringify(extracted);
+      logger.info('✅ Pre-extracted JSON from text response');
+    } else {
+      // If pre-extraction fails, we know the next step will also fail.
+      // Throw an error here to avoid redundant parsing and make the failure point clearer.
+      logger.error(
+        '❌ Could not extract features JSON from model response. Full response text was:\n' +
+          rawText
+      );
+      const errorMessage =
+        'Failed to parse features from model response: No valid JSON with a "features" array found.';
+      events.emit('spec-regeneration:event', {
+        type: 'spec_regeneration_error',
+        error: errorMessage,
+        projectPath: projectPath,
+      });
+      throw new Error(errorMessage);
+    }
+  }
 
-  await parseAndCreateFeatures(projectPath, responseText, events);
+  await parseAndCreateFeatures(projectPath, contentForParsing, events);
 
   logger.debug('========== generateFeaturesFromSpec() completed ==========');
 }

apps/server/src/routes/app-spec/generate-spec.ts

@@ -9,7 +9,7 @@ import * as secureFs from '../../lib/secure-fs.js';
 import type { EventEmitter } from '../../lib/events.js';
 import { specOutputSchema, specToXml, type SpecOutput } from '../../lib/app-spec-format.js';
 import { createLogger } from '@automaker/utils';
-import { DEFAULT_PHASE_MODELS, isCursorModel } from '@automaker/types';
+import { DEFAULT_PHASE_MODELS, supportsStructuredOutput } from '@automaker/types';
 import { resolvePhaseModel } from '@automaker/model-resolver';
 import { extractJson } from '../../lib/json-extractor.js';
 import { streamingQuery } from '../../providers/simple-query-service.js';
@@ -120,10 +120,13 @@ ${prompts.appSpec.structuredSpecInstructions}`;
   let responseText = '';
   let structuredOutput: SpecOutput | null = null;
 
-  // Determine if we should use structured output (Claude supports it, Cursor doesn't)
-  const useStructuredOutput = !isCursorModel(model);
+  // Determine if we should use structured output based on model type
+  const useStructuredOutput = supportsStructuredOutput(model);
+  logger.info(
+    `Structured output mode: ${useStructuredOutput ? 'enabled (Claude/Codex)' : 'disabled (using JSON instructions)'}`
+  );
 
-  // Build the final prompt - for Cursor, include JSON schema instructions
+  // Build the final prompt - for non-Claude/Codex models, include JSON schema instructions
   let finalPrompt = prompt;
   if (!useStructuredOutput) {
     finalPrompt = `${prompt}