fix: Improve LLM prompt and JSON response parsing

main.py

@@ -497,12 +497,19 @@ SYSTEM_PROMPT: str = (
     "Start directly with { and end with }. Escape all internal double quotes using \\\". "
     "No unescaped newlines inside string fields.\n\n"
     "REQUIRED JSON SCHEMA:\n"
-    '{\"score\": number (float, 2 decimal places, e.g. 82.50), '
-    '\"language\": \"String (detected language or stack)\", '
-    '\"code_quality\": \"String (detailed overall quality assessment, 3-5 sentences)\", '
-    '\"questions\": [{\"question\": \"String (the parsed question/requirement)\", '
-    '\"is_correct\": boolean, '
-    '\"feedback\": \"String (detailed explanation, 2-4 sentences minimum)\"}]}'
+    "Return a JSON object with the following structure. Replace the placeholders with your actual evaluation data:\n"
+    "{\n"
+    "  \"score\": 82.50,\n"
+    "  \"language\": \"Detected language or stack\",\n"
+    "  \"code_quality\": \"Detailed overall quality assessment, 3-5 sentences...\",\n"
+    "  \"questions\": [\n"
+    "    {\n"
+    "      \"question\": \"The parsed question or requirement\",\n"
+    "      \"is_correct\": true,\n"
+    "      \"feedback\": \"Detailed explanation of why it is correct or incorrect, 2-4 sentences minimum...\"\n"
+    "    }\n"
+    "  ]\n"
+    "}"
 )
 
 
@@ -542,9 +549,19 @@ def validate_report(parsed: dict) -> dict:
     Validate the LLM response against the required 4-key schema and normalise types.
     Raises ValueError on any schema violation.
     """
+    # Unpack if the LLM nested the response inside a root key (e.g. {"evaluation": {...}})
+    if len(parsed) == 1 and isinstance(list(parsed.values())[0], dict):
+        inner = list(parsed.values())[0]
+        if "score" in inner or "questions" in inner:
+            parsed = inner
+
     missing = set(REQUIRED_KEYS) - parsed.keys()
     if missing:
-        raise ValueError(f"LLM response missing required keys: {missing}")
+        raise ValueError(
+            f"LLM response missing required keys: {missing}. "
+            f"Parsed response keys: {list(parsed.keys())}. "
+            f"Raw: {json.dumps(parsed)[:300]}"
+        )
 
     for key, expected in REQUIRED_KEYS.items():
         val = parsed[key]