Merge nucleic/plucky-north-vole-sdna into dev

tests/test_label_swe_chat_prompts.py

@@ -20,12 +20,12 @@ class LabelSWEChatPromptsTests(unittest.TestCase):
             "repoID": "repo",
             "userID": "user",
             "sessionID": "session",
-            "sourceTurnIDs": ["one", "two", "three"],
+            "sourceTurnIDs": ["one", "two"],
             "prompt": "What is making this test fail?",
-            "teacherContext": ["It fails only on CI.", "Please diagnose it."],
+            "teacherResponse": "It fails only on CI.",
         }
         value["promptHash"] = prompt_hash(value["prompt"])
-        value["contextPromptHashes"] = [prompt_hash(item) for item in value["teacherContext"]]
+        value["teacherResponseHash"] = prompt_hash(value["teacherResponse"])
         line = base.SourceLine(1, json.dumps(value), "line-hash")
         return label_swe_chat_prompts.candidates([line])[0]
 
@@ -39,7 +39,7 @@ class LabelSWEChatPromptsTests(unittest.TestCase):
             }]
         }
 
-    def test_context_hashes_are_checked_and_context_dependent_labels_become_vague_eval(self):
+    def test_response_hash_is_checked_and_context_dependent_labels_become_vague_eval(self):
         candidate = self.candidate()
         decisions = label_swe_chat_prompts.validate_decisions(
             [candidate], self.decision(candidate, False)
@@ -50,8 +50,8 @@ class LabelSWEChatPromptsTests(unittest.TestCase):
     def test_candidate_rejects_context_hash_mismatch(self):
         candidate = self.candidate()
         value = json.loads(candidate.line.raw)
-        value["contextPromptHashes"][0] = "bad"
-        with self.assertRaisesRegex(ValueError, "contextPromptHashes"):
+        value["teacherResponseHash"] = "bad"
+        with self.assertRaisesRegex(ValueError, "teacherResponseHash"):
             label_swe_chat_prompts.candidates(
                 [base.SourceLine(1, json.dumps(value), "line-hash")]
             )
@@ -63,7 +63,6 @@ class LabelSWEChatPromptsTests(unittest.TestCase):
         )
         encoded = json.dumps(state)
         self.assertNotIn("It fails only on CI.", encoded)
-        self.assertNotIn("Please diagnose it.", encoded)
 
 
 if __name__ == "__main__":