Why is the post-processing of generated fixes so time-consuming?

[lycfight]

I tested a single data point in the generate fix step，https://github.com/OpenAutoCoder/Agentless/blob/main/README_swebench.md#-repair:

for i in {0..3}; do python agentless/repair/repair.py --loc_file results/swe-bench-verified/edit_location_individual/loc_merged_${i}-${i}_outputs.jsonl \ --output_folder results/swe-bench-verified/repair_sample_$((i+1)) \ --loc_interval \ --top_n=3 \ --context_window=10 \ --max_samples 10 \ --cot \ --diff_format \ --gen_and_process \ --num_threads 1 \ --dataset=princeton-nlp/SWE-bench_Verified \ --target_id=django__django-10914 done

The generation took around ten seconds, but the post-processing took about three minutes. What is the post-processing doing that makes it so time-consuming?