Fix: Pass user query to LLM context in Fast Lane and Deep Dive stages

prompts.py

@@ -988,16 +988,19 @@ def format_minimal_diagnosis(result) -> str:
 # =============================================================================
 
 FAST_LANE_PROMPT = """You are Agrow-AI. 
-TASK: Diagnose the crop issue based on the provided context.
+TASK: Answer the user's question and diagnose any crop issues based on the provided context.
 PRIORITY: SPEED & ACCURACY.
 
+USER QUESTION:
+{query}
+
 CONTEXT:
 {context}
 
 INSTRUCTIONS:
 1. [Hypothesis]: Briefly state what the primary signals (NDVI, NDRE, etc.) suggest.
 2. [Check]: Verify if supporting data (Moisture, Weather) aligns or contradicts.
-3. [Diagnosis]: State the final conclusion.
+3. [Diagnosis]: State the final conclusion that ANSWERS THE USER'S QUESTION.
 4. [Action]: One specific corrective action.
 
 OUTPUT JSON ONLY:
@@ -1012,11 +1015,14 @@ OUTPUT JSON ONLY:
 DEEP_DIVE_HYPOTHESIS_PROMPT = """You are Agrow-AI, conducting a DEEP DIVE diagnosis.
 STAGE A: HYPOTHESIS GENERATION
 
+USER QUESTION:
+{query}
+
 CONTEXT:
 {context}
 
 TASK:
-Identify top 3 possible causes for the observed issues. Do not conclude yet.
+Identify top 3 possible causes that could answer the user's question. Do not conclude yet.
 Think broadly (Nutrients, Pests, Water, Soil, Disease).
 
 OUTPUT JSON ONLY:
@@ -1032,6 +1038,9 @@ OUTPUT JSON ONLY:
 DEEP_DIVE_ADVERSARY_PROMPT = """You are Agrow-AI.
 STAGE B: ADVERSARIAL CHECK
 
+USER QUESTION:
+{query}
+
 HYPOTHESES:
 {hypotheses}
 
@@ -1041,6 +1050,7 @@ NEW EVIDENCE (Adversarial Data):
 TASK:
 Actively try to DISPROVE each hypothesis using the new evidence (SAR, Soil, Pests).
 If evidence contradicts a hypothesis, mark it as INVALID.
+Remember to focus on answering the user's question.
 
 OUTPUT JSON ONLY:
 {{
@@ -1056,6 +1066,9 @@ OUTPUT JSON ONLY:
 DEEP_DIVE_JUDGE_PROMPT = """You are Agrow-AI.
 STAGE C: FINAL VERDICT
 
+USER QUESTION:
+{query}
+
 WINNING HYPOTHESIS:
 {hypothesis}
 
@@ -1063,7 +1076,7 @@ CONSTRAINTS & HISTORY:
 {context}
 
 TASK:
-Provide the final diagnostic report and a detailed action plan.
+Provide the final diagnostic report and a detailed action plan that DIRECTLY ANSWERS the user's question.
 Consider farmer constraints (budget, machinery) and historical trends.
 
 OUTPUT JSON ONLY:

reasoning_engine.py

@@ -148,7 +148,8 @@ class ReasoningEngine:
         # Build ultra-compact context
         compact_ctx = self.aggregator.build_ultra_compact_context(context)
         
-        prompt = FAST_LANE_PROMPT.format(context=compact_ctx)
+        # Include user query in prompt
+        prompt = FAST_LANE_PROMPT.format(query=query, context=compact_ctx)
         full_prompt = f"{SYSTEM_PROMPT}\n\n{prompt}"
         
         response = self.llm(full_prompt)
@@ -181,21 +182,21 @@ class ReasoningEngine:
         """Execute 3-Stage Deep Dive."""
         logger.info("Executing DEEP DIVE (3-Call)...")
         
-        # 1. Hypothesis Generation
+        # 1. Hypothesis Generation - Include query
         ctx_hyp = self.aggregator.build_deep_dive_context(context, "hypothesis")
-        resp_hyp = self.llm(f"{SYSTEM_PROMPT}\n{DEEP_DIVE_HYPOTHESIS_PROMPT.format(context=ctx_hyp)}")
+        resp_hyp = self.llm(f"{SYSTEM_PROMPT}\n\nUSER QUERY: {query}\n\n{DEEP_DIVE_HYPOTHESIS_PROMPT.format(query=query, context=ctx_hyp)}")
         out_hyp = self._parse_json_safe(resp_hyp, {"hypotheses": []})
         
-        # 2. Adversarial Check
+        # 2. Adversarial Check - Include query context
         ctx_adv = self.aggregator.build_deep_dive_context(context, "adversary")
         hyp_str = json.dumps(out_hyp, indent=2)
-        resp_adv = self.llm(f"{SYSTEM_PROMPT}\n{DEEP_DIVE_ADVERSARY_PROMPT.format(hypotheses=hyp_str, context=ctx_adv)}")
+        resp_adv = self.llm(f"{SYSTEM_PROMPT}\n\nUSER QUERY: {query}\n\n{DEEP_DIVE_ADVERSARY_PROMPT.format(query=query, hypotheses=hyp_str, context=ctx_adv)}")
         out_adv = self._parse_json_safe(resp_adv, {"surviving_hypothesis": "Unknown"})
         
-        # 3. Final Verdict
+        # 3. Final Verdict - Include query
         ctx_judge = self.aggregator.build_deep_dive_context(context, "judge")
         winner = out_adv.get("surviving_hypothesis", "Unknown")
-        resp_judge = self.llm(f"{SYSTEM_PROMPT}\n{DEEP_DIVE_JUDGE_PROMPT.format(hypothesis=winner, context=ctx_judge)}")
+        resp_judge = self.llm(f"{SYSTEM_PROMPT}\n\nUSER QUERY: {query}\n\n{DEEP_DIVE_JUDGE_PROMPT.format(query=query, hypothesis=winner, context=ctx_judge)}")
         out_judge = self._parse_json_safe(resp_judge, {"final_diagnosis": winner, "action_plan": {}})
         
         # Map to ReasoningResult