engine/llm.rs (2309 lines) mixes transport/streaming, prompts, scan_files, and a JSON repair toolbox. parse_review_response and parse_scan_response implement diverging strict→repair→partial paths. review_diff retries once with a stricter prompt; review_diff_stream does not. The stream path prints to stdout from inside llm.rs. #562 patched a symptom at the command level.
Note. A file split alone adds no depth; the gain is a single 'structured findings from a model response' step owning repair + retry, with a fake transport adapter for tests.
From the 2026-10 architecture review (deepening candidates, develop @ 147dd75).
engine/llm.rs(2309 lines) mixes transport/streaming, prompts,scan_files, and a JSON repair toolbox.parse_review_responseandparse_scan_responseimplement diverging strict→repair→partial paths.review_diffretries once with a stricter prompt;review_diff_streamdoes not. The stream path prints to stdout from inside llm.rs. #562 patched a symptom at the command level.Note. A file split alone adds no depth; the gain is a single 'structured findings from a model response' step owning repair + retry, with a fake transport adapter for tests.
From the 2026-10 architecture review (deepening candidates, develop @ 147dd75).