Update training.py

training.py

@@ -175,13 +175,29 @@ def supervised_warmup(model, tokenizer, n_examples=500, epochs=2):
         ]
         last_output = random.choice(last_outputs)
         # Use same prompt structure as build_prompt
-        prompt = f"""You are a code review agent and name your proposed fix function fix.
+        prompt = f"""You are an AI code review agent. Your goal is to convince a simulated human developer to accept your proposed fix and name your proposed fix function fix.
+
+The developer has a **{author_personality}** personality and will only accept if you provide solid evidence:
+- Tests pass (high pass ratio)
+- Lint is clean (zero errors)
+- Documentation or references are provided
+- Your reasoning is clear, uses words like "because" or "therefore", and is detailed (over 30 words if needed)
+
+Workflow:
+1. Use `inspect` to understand the code.
+2. Use `run_tests` and `run_linter` to gather evidence.
+3. Propose a fix (`fix`) and explain why it works (`comment` or `question`).
+4. If the developer pushes back, read their response carefully and address their specific concern.
+5. Once convinced, use `done` to finish.
 
 Code:
-{code}
+{obs.code_snippet}
 
-Last Output:
-{last_output}
+Author says:
+{author_msg if author_msg else "(no response yet – start with inspection)"}
+
+Last tool output:
+{tool_output if tool_output else "(none)"}
 
 Available actions:
 run_tests, run_linter, inspect, fix, comment, question, done