Repository‑level issue localization aims to identify and rank the files and functions relevant to a reported problem. LLM agents typically work iteratively: they first select a set of potentially relevant locations, inspect the corresponding code, and then revise their judgments as new evidence emerges. Existing environments provide limited support for this loop, forcing agents to search for unresolved relation targets, reconstruct entity semantics from raw source, and revise candidates without evidential grounding. To overcome these gaps we introduce SemNav, a framework that seeds a broad candidate set via deterministic retrieval and lets an LLM agent continuously refine it, thus combining initial coverage with evidence‑driven revision. SemNav realizes this through three core components: a Semantic Navigation Graph that resolves program relations on demand via a language server, enabling direct cross‑file navigation; Issue‑conditioned Semantic Cards that offer concise, source‑grounded interpretations of each entity’s role and relevance to the issue; and a persistent Candidate Workspace that records each candidate together with its evidential basis, supporting grounded verification, revision, and ranking. Across SWE‑bench Lite and PLocBench, SemNav with Gemma 4B raises File Hit@10 from 68.33% to 82.67%. Ablation studies and trajectory analysis confirm the complementary contributions of all three components, while Semantic Cards cut working‑context load by 48.2% compared with full‑source reading. SemNav also tops all seven evidence‑quality metrics on SWE‑Explore and improves downstream issue resolution from 44.00% to 52.33%.
Review