Repository‑level program repair demands not only correct patches but also inspectable records that explain how issue requirements are turned into code changes and post‑edit evidence. We introduce THEMIS, a stage‑aware repair workflow that externalizes the requirement‑to‑repair process through semantic interpretation, a runtime requirement‑code graph, graph‑derived developer guidance, retained repair rationale and patches, and post‑edit audit records.
THEMIS consists of: (1) a semantic interpretation layer that maps natural‑language requirements to code semantics; (2) a runtime requirement‑code graph that captures associations between requirement entities and code symbols; (3) graph‑derived developer guidance that supplies contextual prompts to the LLM repair model; (4) retained rationale and patches forming a traceable audit chain; (5) post‑edit audit records for cross‑stage correspondence checks.
A retrospective audit of 300 SWE‑bench Lite cases shows that a complete developer rationale is available for 288 cases (96%), and 214 cases (71.3%) retain a full set of audited fields linking the selected stages. Target symbols recur in 62.6% of rationales and 62.8% of patches, rising to 75.8% when related symbols are included.
In a paired comparison of 100 cases, the relational workflow resolves 19 cases versus 9 for the direct same‑input condition. Because the two conditions also differ in Analyzer output, graph‑derived distillation, and Judge records, we present this as preliminary workflow‑level evidence rather than a causal effect of the graph component.
Overall, THEMIS makes otherwise implicit requirement‑to‑repair transitions observable, enabling systematic examination of how repair decisions persist, align, and evolve across stages.
Review