1

prisma/services/deepThink/manager.ts

@@ -1,7 +1,7 @@
 import { Type } from "@google/genai";
-import { ModelOption, AnalysisResult } from '../../types';
+import { ModelOption, AnalysisResult, ExpertResult, ReviewResult } from '../../types';
 import { cleanJsonString } from '../../utils';
-import { MANAGER_SYSTEM_PROMPT } from './prompts';
+import { MANAGER_SYSTEM_PROMPT, MANAGER_REVIEW_SYSTEM_PROMPT } from './prompts';
 
 export const executeManagerAnalysis = async (
   ai: any,
@@ -59,3 +59,66 @@ export const executeManagerAnalysis = async (
     return { thought_process: "Direct processing.", experts: [] };
   }
 };
+
+export const executeManagerReview = async (
+  ai: any,
+  model: ModelOption,
+  query: string,
+  currentExperts: ExpertResult[],
+  budget: number
+): Promise<ReviewResult> => {
+  const reviewSchema = {
+    type: Type.OBJECT,
+    properties: {
+      satisfied: { type: Type.BOOLEAN, description: "True if the experts have fully answered the query with high quality." },
+      critique: { type: Type.STRING, description: "If not satisfied, explain why and what is missing." },
+      next_round_strategy: { type: Type.STRING, description: "Plan for the next iteration." },
+      refined_experts: {
+        type: Type.ARRAY,
+        description: "The list of experts for the next round. Can be the same roles or new ones.",
+        items: {
+          type: Type.OBJECT,
+          properties: {
+            role: { type: Type.STRING },
+            description: { type: Type.STRING },
+            temperature: { type: Type.NUMBER },
+            prompt: { type: Type.STRING }
+          },
+          required: ["role", "description", "temperature", "prompt"]
+        }
+      }
+    },
+    required: ["satisfied", "critique"]
+  };
+
+  const expertOutputs = currentExperts.map(e => 
+    `--- [Round ${e.round}] Expert: ${e.role} ---\nOutput: ${e.content?.slice(0, 2000)}...`
+  ).join('\n\n');
+
+  const content = `User Query: "${query}"\n\nCurrent Expert Outputs:\n${expertOutputs}`;
+
+  const resp = await ai.models.generateContent({
+    model: model,
+    contents: content,
+    config: {
+      systemInstruction: MANAGER_REVIEW_SYSTEM_PROMPT,
+      responseMimeType: "application/json",
+      responseSchema: reviewSchema,
+      thinkingConfig: { 
+         includeThoughts: true, 
+         thinkingBudget: budget 
+      }
+    }
+  });
+
+  const rawText = resp.text || '{}';
+  const cleanText = cleanJsonString(rawText);
+
+  try {
+    return JSON.parse(cleanText) as ReviewResult;
+  } catch (e) {
+    console.error("Review JSON Parse Error:", e);
+    // Fallback: Assume satisfied if JSON fails to avoid infinite loops due to format errors
+    return { satisfied: true, critique: "JSON Error, proceeding to synthesis." };
+  }
+};

prisma/services/deepThink/prompts.ts

@@ -1,13 +1,35 @@
+
 import { ExpertResult } from '../../types';
 
-export const MANAGER_SYSTEM_PROMPT = `
-You are the "Dynamic Planning Engine". Your goal is to analyze a user query (considering the conversation context) and decompose it into a set of specialized expert personas (2 to 4) who can collaboratively solve specific aspects of the problem.
+export const MANAGER_SYSTEM_PROMPT = `You are the "Dynamic Planning Engine". Your goal is to analyze a user query (considering the conversation context) and decompose it into a set of specialized expert personas (2 to 4) who can collaboratively solve specific aspects of the problem.
 
-Your job is to create SUPPLEMENTARY experts to aid the Primary Responder.
-DO NOT create an expert that just repeats the user query. The Primary Responder is already doing that.
-Focus on specialized angles: specific coding patterns, historical context, devil's advocate, security analyst, etc.
+Your job is to create SUPPLEMENTARY experts
 
-For each expert, you must assign a specific 'temperature' (0.0 to 2.0).
+For each expert, you must assign a specific 'temperature' (0.0 to 2.0) based on the nature of their task:
+
+*   High temperature (1.0 - 2.0) 
+*   Low temperature (0.0 - 0.4) 
+*   Medium temperature (0.4 - 1.0)`;
+
+export const MANAGER_REVIEW_SYSTEM_PROMPT = `
+You are the "Quality Assurance & Orchestration Engine". 
+You have just received outputs from a team of AI experts.
+Your goal is to evaluate if these outputs are sufficient to fully answer the user's complex request with high quality.
+
+Criteria for "Not Satisfied":
+- Conflicting information between experts that isn't resolved.
+- Missing code implementation details or edge cases.
+- Shallow analysis that doesn't go deep enough.
+- Logic errors or hallucinations.
+
+If you are NOT satisfied:
+1. Provide a "critique" explaining exactly what is missing or wrong.
+2. Define a "next_round_strategy" (briefly) to fix it.
+3. Define the *refined_experts* for the next round. You can keep the same roles or create new ones. Their prompts MUST include the feedback/critique.
+
+If you ARE satisfied:
+1. Set satisfied to true.
+2. Leave refined_experts empty.
 `;
 
 export const getExpertSystemInstruction = (role: string, description: string, context: string) => {
@@ -23,11 +45,11 @@ ${recentHistory}
 
 Original User Query: "${query}"
 
-Here are the analyses from your expert panel:
-${expertResults.map(e => `--- Expert: ${e.role} (Temp: ${e.temperature}) ---\n${e.content || "(No output)"}\n`).join('\n')}
+Here are the analyses from your expert panel (potentially across multiple rounds of refinement):
+${expertResults.map(e => `--- [Round ${e.round || 1}] Expert: ${e.role} (Temp: ${e.temperature}) ---\n${e.content || "(No output)"}\n`).join('\n')}
 
 Your Task:
-1. Reflect on the experts' inputs. Identify conflicts and consensus.
+1. Reflect on the experts' inputs. Identify conflicts, consensus, and evolution of thought across rounds.
 2. Synthesize a final, comprehensive, and high-quality answer to the user's original query.
 3. Do not simply summarize; integrate the knowledge into a cohesive response.
 `;