Merge pull request #380 from maheshgudi/beautify_assertions

Prettify assertion error output

18 files changed, 297 insertions, 159 deletions

diff --git a/yaksh/compare_stdio.py b/yaksh/error_messages.py
index c4076de..7ea8618 100644
--- a/yaksh/compare_stdio.py
+++ b/yaksh/error_messages.py
@@ -3,7 +3,24 @@ try:
 except ImportError:
     from itertools import izip_longest as zip_longest
 
-    
+def prettify_exceptions(exception, message, traceback=None, testcase=None):
+    err = {"type": "assertion",
+           "exception": exception,
+           "traceback": traceback,
+           "message": message
+           }
+    if exception == 'RuntimeError' or exception == 'RecursionError':
+        err["traceback"] = None
+
+    if exception == 'AssertionError':
+        value = ("Expected answer from the"  
+                + " test case did not match the output")
+        err["message"] = value
+        err["traceback"] = None
+        if testcase:
+            err["test_case"] = testcase
+    return err
+
 def _get_incorrect_user_lines(exp_lines, user_lines):
     err_line_numbers = []
     for line_no, (expected_line, user_line) in \
@@ -16,17 +33,19 @@ def _get_incorrect_user_lines(exp_lines, user_lines):
 def compare_outputs(expected_output, user_output, given_input=None):
     given_lines = user_output.splitlines()
     exp_lines = expected_output.splitlines()
-    msg = {"given_input":given_input,
+    msg = {"type": "stdio",
+           "given_input": given_input,
            "expected_output": exp_lines,
-           "user_output":given_lines
-            }
+           "user_output": given_lines
+           }
     ng = len(given_lines)
     ne = len(exp_lines)
     err_line_numbers = _get_incorrect_user_lines(exp_lines, given_lines)
     msg["error_line_numbers"] = err_line_numbers
     if ng != ne:
         msg["error_msg"] = ("Incorrect Answer: "
-                            + "We had expected {} number of lines. ".format(ne)
+                            + "We had expected {} number of lines. "\
+                                .format(ne)
                             + "We got {} number of lines.".format(ng)
                             )
         return False, msg
diff --git a/yaksh/evaluator_tests/test_bash_evaluation.py b/yaksh/evaluator_tests/test_bash_evaluation.py
index 2faa7bf..5542710 100644
--- a/yaksh/evaluator_tests/test_bash_evaluation.py
+++ b/yaksh/evaluator_tests/test_bash_evaluation.py
@@ -104,7 +104,9 @@ class BashAssertionEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assert_correct_output(self.timeout_msg, result.get("error"))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
@@ -533,7 +535,9 @@ class BashHookEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get('success'))
-        self.assert_correct_output(self.timeout_msg, result.get('error'))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
diff --git a/yaksh/evaluator_tests/test_c_cpp_evaluation.py b/yaksh/evaluator_tests/test_c_cpp_evaluation.py
index 0898b3f..162d90c 100644
--- a/yaksh/evaluator_tests/test_c_cpp_evaluation.py
+++ b/yaksh/evaluator_tests/test_c_cpp_evaluation.py
@@ -151,7 +151,9 @@ class CAssertionEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assert_correct_output(self.timeout_msg, result.get("error"))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
@@ -406,7 +408,9 @@ class CppStdIOEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assert_correct_output(self.timeout_msg, result.get("error"))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
@@ -616,7 +620,9 @@ class CppStdIOEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assert_correct_output(self.timeout_msg, result.get("error"))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
 
     def test_cpp_only_stdout(self):
         # Given
@@ -976,7 +982,9 @@ class CppHookEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get('success'))
-        self.assert_correct_output(self.timeout_msg, result.get('error'))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
diff --git a/yaksh/evaluator_tests/test_java_evaluation.py b/yaksh/evaluator_tests/test_java_evaluation.py
index 5ddf8cd..35b64d0 100644
--- a/yaksh/evaluator_tests/test_java_evaluation.py
+++ b/yaksh/evaluator_tests/test_java_evaluation.py
@@ -160,7 +160,9 @@ class JavaAssertionEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assert_correct_output(self.timeout_msg, result.get("error"))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
@@ -405,7 +407,9 @@ class JavaStdIOEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assert_correct_output(self.timeout_msg, result.get("error"))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
@@ -845,7 +849,9 @@ class JavaHookEvaluationTestCases(EvaluatorBaseTest):
         # Then
 
         self.assertFalse(result.get('success'))
-        self.assert_correct_output(self.timeout_msg, result.get('error'))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
diff --git a/yaksh/evaluator_tests/test_python_evaluation.py b/yaksh/evaluator_tests/test_python_evaluation.py
index a2faf77..71d7732 100644
--- a/yaksh/evaluator_tests/test_python_evaluation.py
+++ b/yaksh/evaluator_tests/test_python_evaluation.py
@@ -24,9 +24,15 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
             f.write('2'.encode('ascii'))
         tmp_in_dir_path = tempfile.mkdtemp()
         self.in_dir = tmp_in_dir_path
-        self.test_case_data = [{"test_case_type": "standardtestcase", "test_case": 'assert(add(1,2)==3)', 'weight': 0.0},
-                               {"test_case_type": "standardtestcase", "test_case": 'assert(add(-1,2)==1)', 'weight': 0.0},
-                               {"test_case_type": "standardtestcase", "test_case":  'assert(add(-1,-2)==-3)', 'weight': 0.0},
+        self.test_case_data = [{"test_case_type": "standardtestcase",
+                                "test_case": 'assert(add(1,2)==3)',
+                                 'weight': 0.0},
+                               {"test_case_type": "standardtestcase",
+                                "test_case": 'assert(add(-1,2)==1)',
+                                 'weight': 0.0},
+                               {"test_case_type": "standardtestcase",
+                                "test_case":  'assert(add(-1,-2)==-3)',
+                                 'weight': 0.0},
                                ]
         self.timeout_msg = ("Code took more than {0} seconds to run. "
                             "You probably have an infinite loop in"
@@ -76,23 +82,29 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get('success'))
-        self.assert_correct_output('AssertionError  in:\n assert(add(1,2)==3)',
-                        result.get('error')
-                      )
-        self.assert_correct_output('AssertionError  in:\n assert(add(-1,2)==1)',
-                        result.get('error')
-                      )
-        self.assert_correct_output('AssertionError  in:\n assert(add(-1,-2)==-3)',
-                        result.get('error')
-                      )
+        given_test_case_list = [tc["test_case"] for tc in self.test_case_data]
+        for error in result.get("error"):
+            self.assertEqual(error['exception'], 'AssertionError')
+            self.assertEqual(error['message'],
+                 "Expected answer from the test case did not match the output"
+                 )
+        error_testcase_list = [tc['test_case'] for tc in result.get('error')]
+        self.assertEqual(error_testcase_list, given_test_case_list)
+
 
     def test_partial_incorrect_answer(self):
         # Given
         user_answer = "def add(a,b):\n\treturn abs(a) + abs(b)"
-        test_case_data = [{"test_case_type": "standardtestcase", "test_case": 'assert(add(-1,2)==1)', 'weight': 1.0},
-                            {"test_case_type": "standardtestcase", "test_case":  'assert(add(-1,-2)==-3)', 'weight': 1.0},
-                            {"test_case_type": "standardtestcase", "test_case": 'assert(add(1,2)==3)', 'weight': 2.0}
-                               ]
+        test_case_data = [{"test_case_type": "standardtestcase",
+                           "test_case": 'assert(add(-1,2)==1)',
+                           'weight': 1.0},
+                          {"test_case_type": "standardtestcase",
+                           "test_case":  'assert(add(-1,-2)==-3)',
+                            'weight': 1.0},
+                          {"test_case_type": "standardtestcase",
+                           "test_case": 'assert(add(1,2)==3)',
+                           'weight': 2.0}
+                          ]
         kwargs = {
                   'metadata': {
                     'user_answer': user_answer,
@@ -110,13 +122,15 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # Then
         self.assertFalse(result.get('success'))
         self.assertEqual(result.get('weight'), 2.0)
-        self.assert_correct_output('AssertionError  in:\n assert(add(-1,2)==1)',
-                        result.get('error')
-                      )
-        self.assert_correct_output('AssertionError  in:\n assert(add(-1,-2)==-3)',
-                        result.get('error')
-                      )
-
+        given_test_case_list = [tc["test_case"] for tc in self.test_case_data]
+        given_test_case_list.remove('assert(add(1,2)==3)')
+        for error in result.get("error"):
+            self.assertEqual(error['exception'], 'AssertionError')
+            self.assertEqual(error['message'],
+                 "Expected answer from the test case did not match the output"
+                 )
+        error_testcase_list = [tc['test_case'] for tc in result.get('error')]
+        self.assertEqual(error_testcase_list, given_test_case_list)
     def test_infinite_loop(self):
         # Given
         user_answer = "def add(a, b):\n\twhile True:\n\t\tpass"
@@ -136,7 +150,9 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get('success'))
-        self.assert_correct_output(self.timeout_msg, result.get('error'))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
 
     def test_syntax_error(self):
         # Given
@@ -165,14 +181,12 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        error_as_str = ''.join(result.get("error"))
-        err = error_as_str.splitlines()
+        err = result.get("error")[0]['traceback']
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assertEqual(5, len(err))
         for msg in syntax_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+            self.assert_correct_output(msg, err)
 
     def test_indent_error(self):
         # Given
@@ -200,13 +214,15 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        err = result.get("error")[0].splitlines()
+        err = result.get("error")[0]["traceback"].splitlines()
 
         # Then
         self.assertFalse(result.get("success"))
         self.assertEqual(5, len(err))
         for msg in indent_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+            self.assert_correct_output(msg, 
+                                      result.get("error")[0]['traceback']
+                                      )
 
     def test_name_error(self):
         # Given
@@ -231,14 +247,9 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        error_as_str = ''.join(result.get("error"))
-        err = error_as_str.splitlines()
-
-        # Then
-        self.assertFalse(result.get("success"))
-        self.assertEqual(25, len(err))
+        err = result.get("error")[0]["traceback"]
         for msg in name_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+            self.assertIn(msg, err)
 
     def test_recursion_error(self):
         # Given
@@ -246,10 +257,7 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         def add(a, b):
             return add(3, 3)
         """)
-        recursion_error_msg = ["Traceback", 
-                               "maximum recursion depth exceeded"
-                               ]
-
+        recursion_error_msg = "maximum recursion depth exceeded"
         kwargs = {
                   'metadata': {
                     'user_answer': user_answer,
@@ -263,13 +271,11 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        error_as_str = ''.join(result.get("error"))
-        err = error_as_str.splitlines()
+        err = result.get("error")[0]['message']
 
         # Then
         self.assertFalse(result.get("success"))
-        for msg in recursion_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+        self.assert_correct_output(recursion_error_msg, err)
 
     def test_type_error(self):
         # Given
@@ -296,14 +302,12 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        error_as_str = ''.join(result.get("error"))
-        err = error_as_str.splitlines()
+        err = result.get("error")[0]['traceback']
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assertEqual(25, len(err))
         for msg in type_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+            self.assert_correct_output(msg, err)
 
     def test_value_error(self):
         # Given
@@ -332,18 +336,19 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        error_as_str = ''.join(result.get("error"))
-        err = error_as_str.splitlines()
+        err = result.get("error")[0]['traceback']
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assertEqual(28, len(err))
         for msg in value_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+            self.assert_correct_output(msg, err)
 
     def test_file_based_assert(self):
         # Given
-        self.test_case_data = [{"test_case_type": "standardtestcase", "test_case": "assert(ans()=='2')", "weight": 0.0}]
+        self.test_case_data = [{"test_case_type": "standardtestcase",
+                                "test_case": "assert(ans()=='2')",
+                                "weight": 0.0}
+                               ]
         self.file_paths = [(self.tmp_file, False)]
         user_answer = dedent("""
             def ans():
@@ -369,20 +374,17 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         self.assertTrue(result.get('success'))
 
     def test_single_testcase_error(self):
-        # Given
         """ Tests the user answer with just an incorrect test case """
 
+        # Given
         user_answer = "def palindrome(a):\n\treturn a == a[::-1]"
         test_case_data = [{"test_case_type": "standardtestcase",
-                             "test_case": 's="abbb"\nasert palindrome(s)==False',
-                             "weight": 0.0
-                            }
+                           "test_case": 's="abbb"\nasert palindrome(s)==False',
+                           "weight": 0.0
+                           }
                           ]
         syntax_error_msg = ["Traceback",
                             "call",
-                            "File",
-                            "line",
-                            "<string>",
                             "SyntaxError",
                             "invalid syntax"
                             ]
@@ -399,14 +401,12 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        error_as_str = ''.join(result.get("error"))
-        err = error_as_str.splitlines()
+        err =  result.get("error")[0]['traceback']
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assertEqual(13, len(err))
         for msg in syntax_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+            self.assert_correct_output(msg, err)
 
 
     def test_multiple_testcase_error(self):
@@ -415,13 +415,11 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # Given
         user_answer = "def palindrome(a):\n\treturn a == a[::-1]"
         test_case_data = [{"test_case_type": "standardtestcase",
-                             "test_case": 'assert(palindrome("abba")==True)',
-                             "weight": 0.0
-                            },
+                           "test_case": 'assert(palindrome("abba")==True)',
+                           "weight": 0.0},
                           {"test_case_type": "standardtestcase",
-                             "test_case": 's="abbb"\nassert palindrome(S)==False',
-                             "weight": 0.0
-                            }
+                           "test_case": 's="abbb"\nassert palindrome(S)==False',
+                           "weight": 0.0}
                           ]
         name_error_msg = ["Traceback",
                           "call",
@@ -441,14 +439,12 @@ class PythonAssertionEvaluationTestCases(EvaluatorBaseTest):
         # When
         grader = Grader(self.in_dir)
         result = grader.evaluate(kwargs)
-        error_as_str = ''.join(result.get("error"))
-        err = error_as_str.splitlines()
+        err = result.get("error")[0]['traceback']
 
         # Then
         self.assertFalse(result.get("success"))
-        self.assertEqual(11, len(err))
         for msg in name_error_msg:
-            self.assert_correct_output(msg, result.get("error"))
+            self.assertIn(msg, err)
 
     def test_unicode_literal_bug(self):
         # Given
@@ -674,7 +670,9 @@ class PythonStdIOEvaluationTestCases(EvaluatorBaseTest):
         result = grader.evaluate(kwargs)
 
         # Then
-        self.assert_correct_output(timeout_msg, result.get('error'))
+        self.assert_correct_output(timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         self.assertFalse(result.get('success'))
 
 
@@ -915,7 +913,9 @@ class PythonHookEvaluationTestCases(EvaluatorBaseTest):
 
         # Then
         self.assertFalse(result.get('success'))
-        self.assert_correct_output(self.timeout_msg, result.get('error'))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
 
     def test_assignment_upload(self):
         # Given
diff --git a/yaksh/evaluator_tests/test_python_stdio_evaluator.py b/yaksh/evaluator_tests/test_python_stdio_evaluator.py
index 8877544..9b8d702 100644
--- a/yaksh/evaluator_tests/test_python_stdio_evaluator.py
+++ b/yaksh/evaluator_tests/test_python_stdio_evaluator.py
@@ -1,4 +1,4 @@
-from yaksh.compare_stdio import compare_outputs
+from yaksh.error_messages import compare_outputs
 
 def test_compare_outputs():
     exp = "5\n5\n"
diff --git a/yaksh/evaluator_tests/test_scilab_evaluation.py b/yaksh/evaluator_tests/test_scilab_evaluation.py
index c3a1c83..f7a9925 100644
--- a/yaksh/evaluator_tests/test_scilab_evaluation.py
+++ b/yaksh/evaluator_tests/test_scilab_evaluation.py
@@ -137,7 +137,9 @@ class ScilabEvaluationTestCases(EvaluatorBaseTest):
         result = grader.evaluate(kwargs)
 
         self.assertFalse(result.get("success"))
-        self.assert_correct_output(self.timeout_msg, result.get("error"))
+        self.assert_correct_output(self.timeout_msg,
+                                   result.get("error")[0]["message"]
+                                   )
         parent_proc = Process(os.getpid()).children()
         if parent_proc:
             children_procs = Process(parent_proc[0].pid)
diff --git a/yaksh/grader.py b/yaksh/grader.py
index a9a3738..38cce8d 100644
--- a/yaksh/grader.py
+++ b/yaksh/grader.py
@@ -21,7 +21,7 @@ except ImportError:
 # Local imports
 from .settings import SERVER_TIMEOUT
 from .language_registry import create_evaluator_instance
-
+from .error_messages import prettify_exceptions
 
 MY_DIR = abspath(dirname(__file__))
 registry = None
@@ -141,7 +141,8 @@ class Grader(object):
             for idx, test_case_instance in enumerate(test_case_instances):
                 test_case_success = False
                 test_case_instance.compile_code()
-                test_case_success, err, mark_fraction = test_case_instance.check_code()
+                eval_result = test_case_instance.check_code()
+                test_case_success, err, mark_fraction = eval_result
                 if test_case_success:
                     weight += mark_fraction * test_case_instance.weight
                 else:
@@ -154,16 +155,20 @@ class Grader(object):
                 test_case_instance.teardown()
 
         except TimeoutException:
-            error.append(self.timeout_msg)
-        except OSError:
-            msg = traceback.format_exc(limit=0)
-            error.append("Error: {0}".format(msg))
+            error.append(prettify_exceptions("TimeoutException",
+                                             self.timeout_msg
+                                             )
+                         )
         except Exception:
             exc_type, exc_value, exc_tb = sys.exc_info()
             tb_list = traceback.format_exception(exc_type, exc_value, exc_tb)
             if len(tb_list) > 2:
                 del tb_list[1:3]
-            error.append("Error: {0}".format("".join(tb_list)))
+            error.append(prettify_exceptions(exc_type.__name__,
+                                         str(exc_value),
+                                         "".join(tb_list),
+                                         )
+                         )
         finally:
             # Set back any original signal handler.
             set_original_signal_handler(prev_handler)
diff --git a/yaksh/hook_evaluator.py b/yaksh/hook_evaluator.py
index f5364d6..41ef6e4 100644
--- a/yaksh/hook_evaluator.py
+++ b/yaksh/hook_evaluator.py
@@ -2,13 +2,13 @@
 import sys
 import traceback
 import os
-import signal
 import psutil
 
 # Local imports
 from .file_utils import copy_files, delete_files
 from .base_evaluator import BaseEvaluator
 from .grader import TimeoutException
+from .error_messages import prettify_exceptions
 
 
 class HookEvaluator(BaseEvaluator):
@@ -60,19 +60,32 @@ class HookEvaluator(BaseEvaluator):
         success = False
         mark_fraction = 0.0
         try:
-            tb = None
             _tests = compile(self.hook_code, '<string>', mode='exec')
             hook_scope = {}
             exec(_tests, hook_scope)
             check = hook_scope["check_answer"]
-            success, err, mark_fraction = check(self.user_answer)
+            try:
+                success, err, mark_fraction = check(self.user_answer)
+            except Exception:
+                raise
+
         except TimeoutException:
             processes = psutil.Process(os.getpid()).children(recursive=True)
             for process in processes:
                 process.kill()
             raise
         except Exception:
-            msg = traceback.format_exc(limit=0)
-            err = "Error in Hook code: {0}".format(msg)
-        del tb
+            exc_type, exc_value, exc_tb = sys.exc_info()
+            tb_list = traceback.format_exception(exc_type,
+                                                 exc_value,
+                                                 exc_tb
+                                                 )
+            if len(tb_list) > 2:
+                del tb_list[1:3]
+            err = prettify_exceptions(exc_type.__name__,
+                                      str(exc_value),
+                                      "Error in Hook Code:\n"
+                                        + "".join(tb_list)
+                                      )
+
         return success, err, mark_fraction
diff --git a/yaksh/python_assertion_evaluator.py b/yaksh/python_assertion_evaluator.py
index c8f2dd0..440f422 100644
--- a/yaksh/python_assertion_evaluator.py
+++ b/yaksh/python_assertion_evaluator.py
@@ -10,6 +10,7 @@ import importlib
 from .file_utils import copy_files, delete_files
 from .base_evaluator import BaseEvaluator
 from .grader import TimeoutException
+from .error_messages import prettify_exceptions
 
 
 class PythonAssertionEvaluator(BaseEvaluator):
@@ -68,39 +69,22 @@ class PythonAssertionEvaluator(BaseEvaluator):
         success = False
         mark_fraction = 0.0
         try:
-            tb = None
             _tests = compile(self.test_case, '<string>', mode='exec')
             exec(_tests, self.exec_scope)
         except TimeoutException:
             raise
         except Exception:
-            type, value, tb = sys.exc_info()
-            info = traceback.extract_tb(tb)
-            fname, lineno, func, text = info[-1]
-            text = str(self.test_case)
-
-            # Get truncated traceback
-            err_tb_lines = traceback.format_exc().splitlines()
-            stripped_tb_lines = []
-            for line in err_tb_lines:
-                line = re.sub(r'File\s+".*?",\s+line',
-                     'File <file>, line',
-                     line
-                    )
-                stripped_tb_lines.append(line)
-            stripped_tb = '\n'.join(stripped_tb_lines[-10::])
-
-            err = "Expected Test Case:\n{0}\n" \
-                "Error Traceback - {1} {2} in:\n {3}\n{4}".format(
-                    self.test_case,
-                    type.__name__,
-                    str(value),
-                    text,
-                    stripped_tb
-                    )
+            exc_type, exc_value, exc_tb = sys.exc_info()
+            tb_list = traceback.format_exception(exc_type, exc_value, exc_tb)
+            if len(tb_list) > 2:
+                del tb_list[1:3]
+            err = prettify_exceptions(exc_type.__name__,
+                                      str(exc_value),
+                                      "".join(tb_list),
+                                      self.test_case
+                                      )
         else:
             success = True
             err = None
             mark_fraction = 1.0 if self.partial_grading else 0.0
-        del tb
         return success, err, mark_fraction
diff --git a/yaksh/python_stdio_evaluator.py b/yaksh/python_stdio_evaluator.py
index 2b443a7..b08103a 100644
--- a/yaksh/python_stdio_evaluator.py
+++ b/yaksh/python_stdio_evaluator.py
@@ -9,7 +9,7 @@ except ImportError:
 # Local imports
 from .file_utils import copy_files, delete_files
 from .base_evaluator import BaseEvaluator
-from .compare_stdio import compare_outputs
+from .error_messages import compare_outputs
 
 
 @contextmanager
diff --git a/yaksh/static/yaksh/css/exam.css b/yaksh/static/yaksh/css/exam.css
index fff904e..ec48a14 100644
--- a/yaksh/static/yaksh/css/exam.css
+++ b/yaksh/static/yaksh/css/exam.css
@@ -2,6 +2,6 @@ table td, table th { border: black solid 1px !important;
 			word-wrap: break-word !important;
 			white-space: pre-wrap !important;
 			}
-output{
+#stdio, #assertion {
 	table-layout: fixed
 }
\ No newline at end of file
diff --git a/yaksh/stdio_evaluator.py b/yaksh/stdio_evaluator.py
index 5e4ce18..55adb5c 100644
--- a/yaksh/stdio_evaluator.py
+++ b/yaksh/stdio_evaluator.py
@@ -5,7 +5,7 @@ import signal
 # Local imports
 from .base_evaluator import BaseEvaluator
 from .grader import TimeoutException
-from .compare_stdio import compare_outputs
+from .error_messages import compare_outputs
 
 
 class StdIOEvaluator(BaseEvaluator):
diff --git a/yaksh/templates/exam.html b/yaksh/templates/exam.html
index 9596c1c..a1f0df4 100644
--- a/yaksh/templates/exam.html
+++ b/yaksh/templates/exam.html
@@ -80,27 +80,50 @@
                 {% block main %}
                 {% endblock %}
             </div>
+            <br/>
             {% if question.type == 'code' or question.type == 'upload' %}
                 {% if error_message %}
                     <div class="row" id="error_panel">
       {% for error in error_message %}
               <div class="panel panel-danger">
-      <div class="panel-heading">Testcase No. {{ forloop.counter }}</div>
+      <div class="panel-heading">Error No. {{ forloop.counter }}</div>
       <div class="panel-body">
       <div class="well well-sm">
-      {% if not error.expected_output %}
+      {% if not error.type %}
       <pre><code> {{error|safe}} </code></pre>
-      {% else %}
-        {% if error.given_input %}
-       <table class="table table-bordered">
-        <col width="30%">
-          <tr class = "active">
-          <td> For given Input value(s):</td>
-          <td>{{error.given_input}}</td>
+      {% elif error.type == 'assertion' %}
+            {% if error.test_case %}
+      <strong> We tried your code with the following test case:</strong><br/></br>
+              <pre><code><strong style="color:#d9534f">{{error.test_case}}</strong></code></pre>
+            {% endif %}
+          <p> <b>The following error took place: </b></p>
+        <table class="table table-bordered" width="100%" id='assertion'>
+          <col width="30%">
+            <tr class = "active">
+            <td><b>Exception Name: </b></td>
+            <td><span style="color: #d9534f">{{error.exception}}</span></td>
+          </tr>
+          <tr>
+            <td><b>Exception Message: </b></td><td>{{error.message}}</td>
           </tr>
-       </table>
+          <tr>
+            {% if error.traceback %}
+            <td><b>Full Traceback: </b></td>
+            <td><pre>{{error.traceback}}</pre></td>
+            {% endif %}
+          </tr>
+        </table>
+      {% elif error.type == 'stdio' %}
+        {% if error.given_input %}
+         <table class="table table-bordered">
+          <col width="30%">
+            <tr class = "active">
+            <td> For given Input value(s):</td>
+            <td>{{error.given_input}}</td>
+            </tr>
+         </table>
        {% endif %}
-        <table class="table table-bordered" width="100%" id="output">
+        <table class="table table-bordered" width="100%" id="stdio">
           <col width="10%">
           <col width="40%">
           <col width="40%">
diff --git a/yaksh/templates/yaksh/grade_user.html b/yaksh/templates/yaksh/grade_user.html
index 37bc788..3339177 100644
--- a/yaksh/templates/yaksh/grade_user.html
+++ b/yaksh/templates/yaksh/grade_user.html
@@ -218,9 +218,8 @@ Status : <b style="color: green;"> Passed </b><br/>
             {% endif %}
             {% with ans.error_list as err %}
             {% for error in err %}
-            {% if not error.expected_output %}
-      <pre><code> {{error|safe}} </code></pre>
-      {% else %}
+
+      {% if error.type == 'stdio' %}
         <div class = "well well-sm">
         {% if error.given_input %}
        <table class="table table-bordered">
@@ -262,6 +261,32 @@ Status : <b style="color: green;"> Passed </b><br/>
            </tr>
           </table>
         </div>
+        {% elif error.type == 'assertion' %}
+            {% if error.test_case %}
+      <strong> We tried you code with the following test case:</strong><br/></br>
+              <pre><code><strong style="color:#d9534f">{{error.test_case}}</strong></code></pre>
+            {% endif %}
+          <p> <b>The following error took place: </b></p>
+          <div class="well well-sm">
+        <table class="table table-bordered" width="100%">
+          <col width="30%">
+            <tr class = "active">
+            <td><b>Exception Name: </b></td>
+            <td><span style="color: #d9534f">{{error.exception}}</span></td>
+          </tr>
+          <tr>
+            <td><b>Exception Message: </b></td><td>{{error.message}}</td>
+          </tr>
+          <tr>
+            {% if error.traceback %}
+            <td><b>Full Traceback: </b></td>
+            <td><pre>{{error.traceback}}</pre></td>
+            {% endif %}
+          </tr>
+        </table>
+      </div> <!-- Closes well -->
+        {% else %}
+            <pre><code> {{error|safe}} </code></pre>
         {% endif %}
         {% endfor %}
         {% endwith %}
diff --git a/yaksh/templates/yaksh/user_data.html b/yaksh/templates/yaksh/user_data.html
index 6dfaac3..a0219dd 100644
--- a/yaksh/templates/yaksh/user_data.html
+++ b/yaksh/templates/yaksh/user_data.html
@@ -136,12 +136,10 @@ User IP address: {{ paper.user_ip }}
             <div class="panel-heading">Correct answer
         {% else %}
             <div class="panel panel-danger">
-            <div class="panel-heading">Error
+            <div class="panel-heading">Error<br/>
         {% with answer.error_list as err %}
             {% for error in err %}
-            {% if not error.expected_output %}
-      <pre><code> {{error|safe}} </code></pre>
-      {% else %}
+            {% if error.type == 'stdio' %}
         <div class = "well well-sm">
         {% if error.given_input %}
        <table class="table table-bordered">
@@ -183,6 +181,32 @@ User IP address: {{ paper.user_ip }}
            </tr>
           </table>
         </div>
+        {% elif error.type == 'assertion' %}
+            {% if error.test_case %}
+      <strong> We tried you code with the following test case:</strong><br/></br>
+              <pre><code><strong style="color:#d9534f">{{error.test_case}}</strong></code></pre>
+            {% endif %}
+          <p> <b>The following error took place: </b></p>
+          <div class="well well-sm">
+        <table class="table table-bordered" width="100%">
+          <col width="30%">
+            <tr class = "active">
+            <td><b>Exception Name: </b></td>
+            <td><span style="color: #d9534f">{{error.exception}}</span></td>
+          </tr>
+          <tr>
+            <td><b>Exception Message: </b></td><td>{{error.message}}</td>
+          </tr>
+          <tr>
+            {% if error.traceback %}
+            <td><b>Full Traceback: </b></td>
+            <td><pre>{{error.traceback}}</pre></td>
+            {% endif %}
+          </tr>
+        </table>
+      </div> <!-- Closes well -->
+        {% else %}
+          <pre><code> {{error|safe}} </code></pre>
         {% endif %}
         {% endfor %}
         {% endwith %}
diff --git a/yaksh/templates/yaksh/view_answerpaper.html b/yaksh/templates/yaksh/view_answerpaper.html
index 79987b1..fa16a08 100644
--- a/yaksh/templates/yaksh/view_answerpaper.html
+++ b/yaksh/templates/yaksh/view_answerpaper.html
@@ -131,9 +131,8 @@
     
         {% with answer.error_list as err %}
             {% for error in err %}
-            {% if not error.expected_output %}
-      <pre><code> {{error|safe}} </code></pre>
-      {% else %}
+            
+      {% if error.type == 'stdio' %}
         <div class = "well well-sm">
         {% if error.given_input %}
        <table class="table table-bordered">
@@ -175,6 +174,32 @@
            </tr>
           </table>
         </div>
+        {% elif error.type == 'assertion' %}
+            {% if error.test_case %}
+      <strong> We tried you code with the following test case:</strong><br/></br>
+              <pre><code><strong style="color:#d9534f">{{error.test_case}}</strong></code></pre>
+            {% endif %}
+          <p> <b>The following error took place: </b></p>
+          <div class="well well-sm">
+        <table class="table table-bordered" width="100%">
+          <col width="30%">
+            <tr class = "active">
+            <td><b>Exception Name: </b></td>
+            <td><span style="color: #d9534f">{{error.exception}}</span></td>
+          </tr>
+          <tr>
+            <td><b>Exception Message: </b></td><td>{{error.message}}</td>
+          </tr>
+          <tr>
+            {% if error.traceback %}
+            <td><b>Full Traceback: </b></td>
+            <td><pre>{{error.traceback}}</pre></td>
+            {% endif %}
+          </tr>
+        </table>
+      </div> <!-- Closes well -->
+        {% else %}
+          <pre><code> {{error|safe}} </code></pre>
         {% endif %}
         {% endfor %}
         {% endwith %}
diff --git a/yaksh/tests/test_code_server.py b/yaksh/tests/test_code_server.py
index 5f80f2d..1309624 100644
--- a/yaksh/tests/test_code_server.py
+++ b/yaksh/tests/test_code_server.py
@@ -61,7 +61,7 @@ class TestCodeServer(unittest.TestCase):
         # Then
         data = json.loads(result.get('result'))
         self.assertFalse(data['success'])
-        self.assertTrue('infinite loop' in data['error'][0])
+        self.assertTrue('infinite loop' in data['error'][0]['message'])
 
     def test_correct_answer(self):
         # Given
@@ -104,7 +104,7 @@ class TestCodeServer(unittest.TestCase):
         # Then
         data = json.loads(result.get('result'))
         self.assertFalse(data['success'])
-        self.assertTrue('AssertionError' in data['error'][0])
+        self.assertTrue('AssertionError' in data['error'][0]['exception'])
 
     def test_multiple_simultaneous_hits(self):
         # Given
@@ -143,7 +143,7 @@ class TestCodeServer(unittest.TestCase):
         for i in range(N):
             data = results.get()
             self.assertFalse(data['success'])
-            self.assertTrue('infinite loop' in data['error'][0])
+            self.assertTrue('infinite loop' in data['error'][0]['message'])
 
     def test_server_pool_status(self):
         # Given