feat: v1.3.0 — plan-first workflow, OpenRouter provider, enhanced prompt engine

Major changes:
- Plan-first workflow: AI generates structured plan before code, with
  plan review card (Modify Plan / Start Coding / Skip to Code)
- Post-coding UX: Preview + Request Modifications buttons after code gen
- OpenRouter integration: 4th AI provider with 20+ model support
- Enhanced prompt engine: 9 strategies, 11+ intent patterns, modular
- PLAN MODE system prompt block in all 4 services
- Fixed stale React closure in approveAndGenerate with isApproval flag
- Fixed canvas auto-opening during plan phase with wasIdle gate
- Updated README, CHANGELOG, .env.example, version bump to 1.3.0

lib/services/ollama-cloud.ts

@@ -164,27 +164,82 @@ export class OllamaCloudService {
     return this.availableModels.length > 0 ? this.availableModels : DEFAULT_MODELS;
   }
 
-  async enhancePrompt(prompt: string, model?: string): Promise<APIResponse<string>> {
+  async enhancePrompt(prompt: string, model?: string, options?: { toolCategory?: string; template?: string; diagnostics?: string }): Promise<APIResponse<string>> {
+    const toolCategory = options?.toolCategory || 'reasoning';
+    const template = options?.template || 'rtf';
+    const diagnostics = options?.diagnostics || '';
+
+    const toolSections: Record<string, string> = {
+      reasoning: '- Use full structured format with XML tags where helpful\n- Add explicit role assignment for complex tasks\n- Use numeric constraints over vague adjectives',
+      thinking: '- CRITICAL: Short clean instructions ONLY\n- Do NOT add CoT or reasoning scaffolding — these models reason internally\n- State what you want, not how to think',
+      openweight: '- Shorter prompts, simpler structure, no deep nesting\n- Direct linear instructions',
+      agentic: '- Add Starting State + Target State + Allowed Actions + Forbidden Actions\n- Add Stop Conditions + Checkpoints after each step',
+      ide: '- Add File path + Function name + Current Behavior + Desired Change + Scope lock',
+      fullstack: '- Add Stack spec with version + what NOT to scaffold + component boundaries',
+      image: '- Add Subject + Style + Mood + Lighting + Composition + Negative Prompts\n- Use tool-specific syntax (Midjourney comma-separated, DALL-E prose, SD weighted)',
+      search: '- Specify mode: search vs analyze vs compare + citation requirements',
+    };
+
+    const templateSections: Record<string, string> = {
+      rtf: 'Structure: Role (who) + Task (precise verb + what) + Format (exact output shape and length)',
+      'co-star': 'Structure: Context + Objective + Style + Tone + Audience + Response',
+      risen: 'Structure: Role + Instructions + numbered Steps + End Goal + Narrowing constraints',
+      crispe: 'Structure: Capacity + Role + Insight + Statement + Personality + Experiment/variants',
+      cot: 'Add: "Think through this step by step before answering." Only for standard reasoning models, NOT for o1/o3/R1.',
+      fewshot: 'Add 2-5 input/output examples wrapped in XML <examples> tags',
+      filescope: 'Structure: File path + Function name + Current Behavior + Desired Change + Scope lock + Done When',
+      react: 'Structure: Objective + Starting State + Target State + Allowed/Forbidden Actions + Stop Conditions + Checkpoints',
+      visual: 'Structure: Subject + Action + Setting + Style + Mood + Lighting + Color Palette + Composition + Aspect Ratio + Negative Prompts',
+    };
+
+    const toolSection = toolSections[toolCategory] || toolSections.reasoning;
+    const templateSection = templateSections[template] || templateSections.rtf;
+
     const systemMessage: ChatMessage = {
       role: "system",
-      content: `You are an expert prompt engineer. Your task is to enhance user prompts to make them more precise, actionable, and effective for AI coding agents.
+      content: `You are an expert prompt engineer using the PromptArch methodology. Enhance the user\'s prompt to be production-ready.
 
-Apply these principles:
-1. Add specific context about project and requirements
-2. Clarify constraints and preferences
-3. Define expected output format clearly
-4. Include edge cases and error handling requirements
-5. Specify testing and validation criteria
+STEP 1 — DIAGNOSE AND FIX these failure patterns:
+- Vague task verb -> replace with precise operation
+- Two tasks in one -> keep primary task, note the split
+- No success criteria -> add "Done when: [specific measurable condition]"
+- Missing output format -> add explicit format lock (structure, length, type)
+- No role assignment (complex tasks) -> add domain-specific expert identity
+- Vague aesthetic ("professional", "clean") -> concrete measurable specs
+- No scope boundary -> add explicit scope lock
+- Over-permissive language -> add constraints and boundaries
+- Emotional description -> extract specific technical fault
+- Implicit references -> restate fully
+- No grounding for factual tasks -> add certainty constraint
+- No CoT for logic tasks -> add step-by-step reasoning
 
-Return ONLY the enhanced prompt, no explanations or extra text.`,
+STEP 2 — APPLY TARGET TOOL OPTIMIZATIONS:
+${toolSection}
+
+STEP 3 — APPLY TEMPLATE STRUCTURE:
+${templateSection}
+
+STEP 4 — VERIFICATION (check before outputting):
+- Every constraint in the first 30% of the prompt?
+- MUST/NEVER over should/avoid?
+- Every sentence load-bearing with zero padding?
+- Format explicit with stated length?
+- Scope bounded?
+- Would this produce correct output on first try?
+
+STEP 5 — OUTPUT:
+Output ONLY the enhanced prompt. No explanations, no commentary, no markdown code fences.
+The prompt must be ready to paste directly into the target AI tool.${diagnostics ? '\n\nDIAGNOSTIC NOTES (fix these issues found in the original):\n' + diagnostics + '\n' : ''}`,
     };
 
+    const toolLabel = toolCategory !== 'reasoning' ? ` for ${toolCategory} AI tool` : '';
+
     const userMessage: ChatMessage = {
       role: "user",
-      content: `Enhance this prompt for an AI coding agent:\n\n${prompt}`,
+      content: `Enhance this prompt${toolLabel}:\n\n${prompt}`,
     };
 
-    return this.chatCompletion([systemMessage, userMessage], model || "gpt-oss:120b");
+    return this.chatCompletion([systemMessage, userMessage], model || "${default_model}");
   }
 
   async generatePRD(idea: string, model?: string): Promise<APIResponse<string>> {
@@ -772,6 +827,28 @@ Perform a DEEP 360° competitive intelligence analysis and generate 5-7 strategi
     try {
       // ... existing prompt logic ...
       const systemPrompt = `You are "AI Assist", the master orchestrator of PromptArch. Your goal is to provide intelligent support with a "Canvas" experience.
+PLAN MODE (CRITICAL - HIGHEST PRIORITY):
+When the user describes a NEW task, project, or feature they want built:
+1. DO NOT generate any code, [PREVIEW] tags, or implementation details.
+2. Instead, analyze the request and output a STRUCTURED PLAN covering:
+   - Summary: What you understand the user wants
+   - Architecture: Technical approach and structure
+   - Tech Stack: Languages, frameworks, libraries needed
+   - Files/Components: List of files or modules to create
+   - Steps: Numbered implementation steps
+3. Format the plan in clean Markdown with headers and bullet points.
+4. Keep plans concise but thorough. Focus on the WHAT and HOW, not the actual code.
+5. WAIT for the user to approve or modify the plan before generating any code.
+
+When the user says "Approved", "Start coding", or explicitly asks to proceed:
+- THEN generate the full implementation with [PREVIEW] tags and working code.
+- Follow the approved plan exactly.
+
+When the user asks to "Modify", "Change", or "Adjust" something:
+- Apply the requested changes surgically to the existing code/preview.
+- Output updated [PREVIEW] with the full modified code.
+
+
 
 AGENTS & CAPABILITIES:
 - content: Expert copywriter. Use [PREVIEW:content:markdown] for articles, posts, and long-form text.
@@ -833,7 +910,7 @@ CHANGE LOG (CRITICAL - MUST BE OUTSIDE PREVIEW):
 - Modified component Y
 - Fixed issue Z
 
-IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;
+IMPORTANT: IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;
 
       const messages: ChatMessage[] = [
         { role: "system", content: systemPrompt },