Code extraction error for the previous substeps for specific questions 13, 62 and 76

[amaand-cerebras]

The entire code is not being used for the previous code which causes a cascading issue in the next steps.

if self.previous_llm_code[prev_step] is None:
                    if (
                        (prob_id == "13" and prev_step == 5) or 
                        (prob_id == "62" and prev_step == 0) or 
                        (prob_id == "76" and prev_step == 2)
                    ):
                        prev_file_path = Path(
                            "../data",
                            f"{prob_id}.{prev_step+1}.txt"
                        )
                    else:
                        prev_file_path = Path(
                            self.output_dir,
                            self._get_background_dir(),
                            f"{prob_id}.{prev_step + 1}.py"
                        )
                    if prev_file_path.is_file():
                        prev_file_content = prev_file_path.read_text(encoding='utf-8')
                        func_name = extract_function_name(
                            prob_data["sub_steps"][prev_step]["function_header"]
                        )
                        function_code = get_function_from_code(
                            prev_file_content, func_name
                        )
                        self.previous_llm_code[prev_step] = function_code
                    else:
                        raise Exception(f'Generating problem {prob_id} step {num_steps} ahead of step {prev_step + 1}.')

A simple fix to this is to not do any extraction from the .txt files and use the entire code as it is:

if self.previous_llm_code[prev_step] is None:
                    if (
                        (prob_id == "13" and prev_step == 5) or 
                        (prob_id == "62" and prev_step == 0) or 
                        (prob_id == "76" and prev_step == 2)
                    ):
                        prev_file_path = Path(
                            "../data",
                            f"{prob_id}.{prev_step+1}.txt"
                        )
                    else:
                        prev_file_path = Path(
                            self.output_dir,
                            self._get_background_dir(),
                            f"{prob_id}.{prev_step + 1}.py"
                        )

                    if prev_file_path.is_file():
                        prev_file_content = prev_file_path.read_text(encoding='utf-8')
                        
                        if (
                            (prob_id == "13" and prev_step == 5) or 
                            (prob_id == "62" and prev_step == 0) or 
                            (prob_id == "76" and prev_step == 2)
                        ):
                            # Use the entire .txt content for these special cases
                            self.previous_llm_code[prev_step] = prev_file_content
                        else:
                            # Extract the function code for all other cases
                            func_name = extract_function_name(
                                prob_data["sub_steps"][prev_step]["function_header"]
                            )
                            function_code = get_function_from_code(
                                prev_file_content, func_name
                            )
                            self.previous_llm_code[prev_step] = function_code
                    else:
                        raise Exception(f'Generating problem {prob_id} step {num_steps} ahead of step {prev_step + 1}.')

[prann-ab]

#61 should cover the 13 and 62 halves of this.

Root cause is in extract_function_name: it matches def ...( before class ...(, and a class-based function_header includes its own first method, so class Maxwell: / class EnlargedBlock: resolve to __init__. get_function_from_code then returns just that method and the class is dropped from the code accumulated into every later sub-step — which is why the dependent steps fail regardless of what the model generates.

Two smaller issues in the same function, both fixed there: the class pattern requires a base-class paren so class Foo: never matches, and the patterns are unanchored so they can match docstring prose (some headers document a parameter as env: class Block, which yields env as the "function name").

11 of 341 sub-step headers change, all genuine class headers; 330 are identical.

76.3 is a plain function and already extracted correctly (987/1020 chars, unchanged by the patch) — if you are still seeing failures on that one it is likely a separate cause, and I would be glad to look if you have a repro.