Externalizing Requirement-to-Repair Artifacts as Observable Traces for LLM-Based Program Repair

Automated program repair (APR) often focuses on generating patches that pass test suites, but this "generate-and-validate" approach frequently treats the internal reasoning process as a black box. THEMIS addresses this by externalizing the requirement-to-repair process, creating a traceable record that connects issue requirements, semantic interpretations, code edits, and post-edit audit evidence. By making these transitions explicit, researchers can inspect how repair decisions evolve across stages.

THEMIS structures the repair process into four distinct modules that produce serialized, inspectable artifacts:

• Advanced Code Analyzer: Extracts semantic hypotheses from issue text and code, keeping them distinct from structural graph relations.
• Enhanced Graph Manager: Builds a requirement-code graph that maps issue requirements to specific code nodes, flagging violations and advisory constraints.
• Developer: Consumes distilled graph guidance to produce a bounded edit, while recording the rationale behind the change.
• Judge: Rebuilds the graph after the edit to generate post-edit audit records, including conflict history and repair briefs.

In an audit of 300 SWE-bench Lite cases, THEMIS demonstrated that its stage-aware projection provides broad support for inspection: 71.3% of cases retained a complete set of audited fields connecting the pre-edit, edit, and post-edit stages. Lexical analysis showed that target symbols identified by the system recur in 62.8% of patches, rising to 75.8% when related symbols are included. Furthermore, in a paired 100-case comparison, the relational workflow resolved 19 cases compared to 9 for a direct, non-relational baseline, suggesting that structuring the repair context significantly aids the model in producing successful patches.