Fixed some minor bugs

FRRONTEEEND/components/ChatInterface.tsx

@@ -65,6 +65,12 @@ const cleanMarkdown = (content: string): string => {
   
   let cleaned = content;
   
+  // PHASE 0: Strip wrapping code fences that LLMs add around markdown
+  // This causes ReactMarkdown to render the ENTIRE response as a <code> block
+  // instead of parsing the markdown. Must be done FIRST.
+  cleaned = cleaned.replace(/^\s*```(?:markdown|md|text)?\s*\n/, '');
+  cleaned = cleaned.replace(/\n\s*```\s*$/, '');
+  
   // PHASE 1: Fix inline code that got split across lines
   // Pattern: `code` followed by newline(s) then comma
   cleaned = cleaned.replace(/`([^`\n]+)`\s*\n+\s*,/g, '`$1`, ');

server logs 3.txt

@@ -0,0 +1,270 @@
+===== Application Startup at 2026-02-19 20:07:11 =====
+
+INFO:src.api.app:👥 Multi-user session isolation initialized (fast mode)
+INFO:src.api.app:🔒 Workflow lock initialized for request queuing
+INFO:src.api.app:✅ Frontend assets mounted from /home/user/app/FRRONTEEEND/dist
+INFO:     Started server process [1]
+INFO:     Waiting for application startup.
+INFO:src.api.app:Initializing legacy global agent for health checks...
+🤖 Initialized with Mistral provider - Model: mistral-large-latest
+✅ Cache database initialized at /home/user/app/cache_db/cache.db
+🧠 Loading SBERT model: all-MiniLM-L6-v2...
+INFO:sentence_transformers.SentenceTransformer:Use pytorch device_name: cpu
+INFO:sentence_transformers.SentenceTransformer:Load pretrained SentenceTransformer: all-MiniLM-L6-v2
+✅ SBERT loaded on CPU
+📊 Token Budget: 120,000 tokens available (128,000 - 8,000 reserve)
+✅ Sessions database initialized at ./cache_db/sessions.db
+✅ Created new session: 5c7c1a12-66c5-44b5-9f22-08c8e758dbb0
+📁 Output directory: /tmp/data_science_agent/outputs
+INFO:src.api.app:✅ Health check agent initialized with provider: mistral
+INFO:src.api.app:👥 Per-session agents enabled - each user gets isolated instance
+INFO:src.api.app:🤖 Multi-agent architecture enabled with 5 specialists
+INFO:     Application startup complete.
+INFO:     Uvicorn running on http://0.0.0.0:7860 (Press CTRL+C to quit)
+INFO:     10.16.4.123:49303 - "GET / HTTP/1.1" 200 OK
+INFO:     10.16.34.155:4158 - "GET / HTTP/1.1" 200 OK
+INFO:     10.16.34.155:4158 - "GET /index.css HTTP/1.1" 200 OK
+INFO:     10.16.4.123:49303 - "GET /assets/index-C_AP0pY-.js HTTP/1.1" 200 OK
+INFO:     10.16.34.155:55847 - "GET / HTTP/1.1" 200 OK
+INFO:     10.16.34.155:55847 - "GET /index.css HTTP/1.1" 200 OK
+INFO:     10.16.4.123:28155 - "GET /assets/index-C_AP0pY-.js HTTP/1.1" 304 Not Modified
+INFO:     10.16.34.155:55847 - "GET /favicon.ico HTTP/1.1" 200 OK
+[SSE] ENDPOINT: Client connected for session_id=01c885ba-15c3-4d2b-9fc9-2c21b11554fc
+[SSE] Queue registered, total subscribers: 1
+INFO:     10.16.4.123:62662 - "GET /api/progress/stream/01c885ba-15c3-4d2b-9fc9-2c21b11554fc HTTP/1.1" 200 OK
+[SSE] SENDING connection event to client
+[SSE] No history to replay (fresh session)
+[SSE] Starting event stream loop for session 01c885ba-15c3-4d2b-9fc9-2c21b11554fc
+INFO:src.api.app:SSE stream cancelled for session 01c885ba-15c3-4d2b-9fc9-2c21b11554fc
+INFO:src.api.app:SSE stream closed for session 01c885ba-15c3-4d2b-9fc9-2c21b11554fc
+INFO:src.api.app:[ASYNC] Created new session: c96b385f...
+INFO:src.api.app:[ASYNC] File saved: housing.csv
+INFO:     10.16.34.155:11860 - "POST /run-async HTTP/1.1" 200 OK
+INFO:src.api.app:[BACKGROUND] Starting analysis for session c96b385f...
+[🧹] Clearing SSE history for c96b385f...
+INFO:src.api.app:[🆕] Creating lightweight session for c96b385f...
+INFO:src.api.app:✅ Session created for c96b385f (cache: 1/50) - <1s init
+[DEBUG] Session fallback triggered - has_ambiguous_ref=True, resolved={}
+[DEBUG] resolve_ambiguity returning: {}
+[DEBUG] Orchestrator received resolved_params: {}
+[DEBUG] Current file_path: '/tmp/data_science_agent/housing.csv', target_col: 'None'
+📝 User provided new file: /tmp/data_science_agent/housing.csv (ignoring session file: none)
+🔍 Extracting dataset schema locally (no LLM)...
+🧠 Semantic layer: Embedded 10 columns
+   Found 9 similar column pairs (potential duplicates)
+🧠 Semantic layer enriched 10 columns
+✅ Schema extracted: 20640 rows × 10 cols
+   File size: 1.14 MB
+
+🎯 Intent Classification:
+   Mode: EXPLORATORY
+   Confidence: 40%
+   Reasoning: No strong pattern match, defaulting to exploratory analysis
+   Sub-intent: default
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=intent_classified, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 1
+
+🧠 Routing to REASONING LOOP (exploratory mode)
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=reasoning_mode, msg=🧠 Reasoning Loop activated (exploratory mode)
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 2
+
+============================================================
+🧠 REASONING LOOP (EXPLORATORY mode)
+   Question: give me the most important feature of it 
+   Max iterations: 8
+============================================================
+
+🔬 Generating hypotheses from data profile...
+   📋 Final parameters: ['file_path']
+[SSE] ENDPOINT: Client connected for session_id=c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Queue registered, total subscribers: 1
+INFO:     10.16.34.155:11860 - "GET /api/progress/stream/c96b385f-e6ca-40ce-88ee-9fe7e109bd94 HTTP/1.1" 200 OK
+[SSE] SENDING connection event to client
+[SSE] Replaying 2 missed events for late-joining subscriber
+[SSE] Starting event stream loop for session c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Generated 5 hypotheses:
+   1. [0.9] The 'median_house_value' exhibits unexpected spatial clustering (e.g., sharp discontinuities at specific longitude/latitude boundaries) that may indicate data quality issues or unaccounted regional policies (e.g., zoning laws). This could distort predictive modeling if not addressed.
+   2. [0.9] 'Total_bedrooms' has a suspiciously high correlation with 'total_rooms' (likely >0.9), suggesting multicollinearity or potential data leakage (e.g., 'total_bedrooms' being derived from 'total_rooms'). This could inflate model performance metrics if not validated.
+   3. [0.8] The 'ocean_proximity' categorical variable may have imbalanced classes (e.g., 'ISLAND' is underrepresented), leading to poor generalization for rare categories. This could bias models toward majority classes like 'INLAND'.
+   4. [0.8] 'Median_income' shows a non-linear relationship with 'median_house_value' (e.g., diminishing returns at higher income levels), which linear models may fail to capture. This could explain underperformance in high-value predictions.
+   5. [0.8] Missing values in 'total_bedrooms' (if any) are not randomly distributed but instead cluster in older properties ('housing_median_age' > 50 years), indicating systematic data collection gaps for historic homes.
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=hypotheses_generated, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 3
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+
+── Iteration 1/8 ──
+🤔 REASON: Deciding next action...
+[SSE] GOT event from queue: hypotheses_generated
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: The user asks for the most important feature, which requires understanding feature relationships with the target. Analyzing correlations is the most direct and efficient first step to identify the strongest predictor.
+   Tool: analyze_correlations
+   Hypothesis: We expect to identify the feature with the highest absolute correlation to median_house_value, which will likely be the most important feature.
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=reasoning_step, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 4
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+⚡ ACT: Executing analyze_correlations...
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=tool_executing, msg=🔧 Executing: analyze_correlations
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 5
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+   📋 Final parameters: ['file_path', 'target']
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=tool_completed, msg=✓ Completed: analyze_correlations
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 6
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+   ✓ Tool completed successfully
+💾 Checkpoint saved: iteration 1, last tool: analyze_correlations
+📊 EVALUATE: Interpreting results...
+[SSE] GOT event from queue: reasoning_step
+[SSE] GOT event from queue: tool_executing
+[SSE] GOT event from queue: tool_completed
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Interpretation: The correlation analysis identifies which feature has the strongest statistical relationship with median_house_value, but the raw result alone doesn't reveal the specific feature or its practical significance without further inspection.
+   Answered: False (confidence: 30%)
+   Should stop: False
+   Next questions: ['Which specific feature has the highest correlation with median_house_value, and what is the correlation coefficient?', 'Are there non-linear relationships or interactions between features that might better explain median_house_value?', 'Does the most correlated feature have a causal relationship or is it confounded by other variables?']
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=finding_discovered, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 7
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+
+── Iteration 2/8 ──
+🤔 REASON: Deciding next action...
+[SSE] GOT event from queue: finding_discovered
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: 
+⚠️  Invalid tool: None, skipping iteration
+
+── Iteration 3/8 ──
+🤔 REASON: Deciding next action...
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: 
+⚠️  Invalid tool: None, skipping iteration
+
+── Iteration 4/8 ──
+🤔 REASON: Deciding next action...
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: 
+⚠️  Invalid tool: None, skipping iteration
+
+── Iteration 5/8 ──
+🤔 REASON: Deciding next action...
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: 
+⚠️  Invalid tool: None, skipping iteration
+
+── Iteration 6/8 ──
+🤔 REASON: Deciding next action...
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: 
+⚠️  Invalid tool: None, skipping iteration
+
+── Iteration 7/8 ──
+🤔 REASON: Deciding next action...
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: The user asked for the most important feature, and we have already run a correlation analysis. To directly identify the most important feature with statistical backing, we should use an automated feature selection tool that ranks features by their predictive power for the target variable 'median_house_value'.
+   Tool: auto_feature_selection
+   Hypothesis: We expect this tool to rank features by their importance for predicting 'median_house_value', revealing the single most important feature with quantifiable evidence.
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=reasoning_step, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 8
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+⚡ ACT: Executing auto_feature_selection...
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=tool_executing, msg=🔧 Executing: auto_feature_selection
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 9
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+   📋 Final parameters: ['file_path', 'target_col', 'task_type', 'method']
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=tool_completed, msg=✓ Completed: auto_feature_selection
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 10
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+   ✓ Tool completed successfully
+💾 Checkpoint saved: iteration 7, last tool: auto_feature_selection
+📊 EVALUATE: Interpreting results...
+[SSE] GOT event from queue: reasoning_step
+[SSE] GOT event from queue: tool_executing
+[SSE] GOT event from queue: tool_completed
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Interpretation: The feature selection result identifies the most statistically important feature for predicting 'median_house_value' based on its relative importance score, but the raw output does not explicitly reveal which feature it is or its practical significance. This directly addresses the user's question, but the lack of detailed output (e.g., feature name, importance score) limits immediate actionability.
+   Answered: False (confidence: 50%)
+   Should stop: False
+   Next questions: ['Which specific feature was identified as the most important, and what is its importance score?', "How does this feature's importance compare to its correlation with 'median_house_value' (from Step 1)?", "Are there potential confounders (e.g., multicollinearity, spatial clustering) that could distort this feature's apparent importance?"]
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=finding_discovered, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 11
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+
+── Iteration 8/8 ──
+🤔 REASON: Deciding next action...
+[SSE] GOT event from queue: finding_discovered
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Status: investigating
+   Reasoning: The previous steps identified feature importance scores but did not explicitly reveal the single most important feature or its practical significance. Generating a feature importance plot will visually confirm the top feature and provide clear, actionable evidence.
+   Tool: generate_feature_importance_plot
+   Hypothesis: We expect the plot to reveal the single most important feature for predicting 'median_house_value' with a clear visual ranking, addressing the user's question directly.
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=reasoning_step, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 12
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+⚡ ACT: Executing generate_feature_importance_plot...
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=tool_executing, msg=🔧 Executing: generate_feature_importance_plot
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 13
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+   ✓ Stripped hallucinated parameter 'file_path': /tmp/data_science_agent/housing.csv
+   ℹ️ Valid parameters for generate_feature_importance_plot: ['feature_importances', 'top_n', 'output_path']
+   📋 Final parameters: ['feature_importances', 'output_path', 'top_n']
+⚠️ <src.orchestrator.DataScienceCopilot object at 0x7f904c1efbf0> failed (attempt 1/2): 'str' object has no attribute 'keys'
+   Retrying in 1.0s...
+[SSE] GOT event from queue: reasoning_step
+[SSE] GOT event from queue: tool_executing
+   📋 Final parameters: ['feature_importances', 'output_path', 'top_n']
+❌ <src.orchestrator.DataScienceCopilot object at 0x7f904c1efbf0> failed after 2 attempts
+💡 Suggested fallback tools: []
+   ❌ Tool failed: 'str' object has no attribute 'keys'
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=tool_failed, msg=❌ FAILED: generate_feature_importance_plot
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 14
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+📊 EVALUATE: Interpreting results...
+[SSE] GOT event from queue: tool_failed
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+   Interpretation: The tool failure indicates a technical issue (likely a file path or data format problem) preventing direct visualization of the most important feature. However, prior `auto_feature_selection` results already identified the top feature for predicting 'median_house_value'—though its name/score weren't explicitly extracted. This suggests the answer exists but requires manual inspection of the raw output file.
+   Answered: False (confidence: 50%)
+   Should stop: False
+   Next questions: ['What is the exact name and importance score of the top-ranked feature in the `auto_feature_selection_results.csv` file?', 'Does the top feature align with domain expectations (e.g., income or location-based variables), or does it reveal unexpected patterns?', 'Are there data quality issues (e.g., multicollinearity, spatial clustering) that could distort the feature importance ranking?']
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=finding_discovered, msg=
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 15
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+
+============================================================
+📝 SYNTHESIZE: Building final answer from 3 findings...
+============================================================
+[SSE] GOT event from queue: finding_discovered
+INFO:httpx:HTTP Request: POST https://api.mistral.ai/v1/chat/completions "HTTP/1.1 200 OK"
+[DEBUG] Merging 0 reports into plots array
+[DEBUG] Final plots array length: 0
+
+✅ Reasoning loop completed in 74.7s
+   Iterations: 3
+   Tools used: analyze_correlations, auto_feature_selection, generate_feature_importance_plot
+   API calls: 13
+INFO:src.api.app:[BACKGROUND] Analysis completed for session c96b385f...
+[SSE] PROGRESS_MANAGER EMIT: session=c96b385f-e6ca-40ce-88ee-9fe7e109bd94, event_type=analysis_complete, msg=✅ Analysis completed successfully!
+[SSE] History stored, total events for c96b385f-e6ca-40ce-88ee-9fe7e109bd94: 16
+[SSE] Found 1 subscribers for c96b385f-e6ca-40ce-88ee-9fe7e109bd94
+[SSE] Successfully queued event to subscriber 1
+[SSE] GOT event from queue: analysis_complete
+INFO:src.api.app:SSE stream closed for session c96b385f-e6ca-40ce-88ee-9fe7e109bd94

src/reasoning/reasoner.py

@@ -319,6 +319,15 @@ class Reasoner:
         tool_name = data.get("tool_name")
         arguments = data.get("arguments", {})
         
+        # 🛡️ SAFETY: If LLM says "investigating" but provides no tool, treat as "done"
+        # This prevents wasting iterations on empty responses (seen in logs: 5 consecutive skips)
+        if status == "investigating" and not tool_name:
+            print(f"   ⚠️  Reasoner returned 'investigating' with no tool — forcing done")
+            return ReasoningOutput.done(
+                reasoning=data.get("reasoning", "No further tool selected. Synthesizing available findings."),
+                confidence=max(0.4, float(data.get("confidence", 0.4)))
+            )
+        
         # Ensure file_path is in arguments if tool needs it
         if tool_name and "file_path" not in arguments and tool_name not in [
             "execute_python_code", "get_smart_summary"

src/reasoning/synthesizer.py

@@ -40,6 +40,7 @@ RULES:
 - Be honest about confidence levels
 - Keep it under 500 words unless complex analysis warrants more
 - Use markdown formatting (headers, bullets, bold for emphasis)
+- Do NOT wrap your response in code fences (``` or ```markdown) — output raw markdown directly
 - ONLY report findings from SUCCESSFUL investigation steps
 - Do NOT invent numbers, statistics, or insights that are not present in the findings
 - If a step is marked [FAILED], ignore its results entirely — do not fabricate data from it
@@ -112,7 +113,7 @@ class Synthesizer:
             max_tokens=max_tokens
         )
         
-        return response.strip()
+        return self._strip_code_fences(response.strip())
 
     def synthesize_exploratory(
         self,
@@ -144,6 +145,7 @@ RULES:
 - Mention all generated visualizations with file paths
 - Suggest actionable next analysis steps
 - Keep it engaging but data-driven
+- Do NOT wrap your response in code fences (``` or ```markdown) — output raw markdown directly
 - ONLY report findings from SUCCESSFUL investigation steps
 - Do NOT invent numbers or statistics not present in the findings
 - If a step is marked [FAILED], ignore it entirely"""
@@ -169,7 +171,21 @@ Write the exploratory analysis report."""
             max_tokens=max_tokens
         )
         
-        return response.strip()
+        return self._strip_code_fences(response.strip())
+
+    def _strip_code_fences(self, text: str) -> str:
+        """
+        Remove wrapping code fences from LLM output.
+        
+        LLMs sometimes wrap markdown in ```markdown ... ``` which causes
+        ReactMarkdown to render the entire response as a code block
+        instead of parsing the markdown.
+        """
+        import re
+        # Strip leading ```markdown or ``` and trailing ```
+        stripped = re.sub(r'^\s*```(?:markdown|md|text)?\s*\n', '', text)
+        stripped = re.sub(r'\n\s*```\s*$', '', stripped)
+        return stripped.strip()
 
     def _format_artifacts(self, artifacts: Dict[str, Any], findings: FindingsAccumulator) -> str:
         """Format artifacts for the synthesis prompt."""