RepoNav: From Snippet Retrieval to File-Centered Repository Navigation for Code Agents
Hongzheng Chai, Jiakun Li, Hongyue Yu, Yuan Yuan
cs.SE · cs.CL
Quick verdict
Claims a file-centered scaffold over snippets narrows the file-to-function gap; ablated evidence, but single-run stats and QA subset choice merit caution.
Summary
The paper diagnoses a 'file-to-function gap' in repository-level code localization: with a fixed mini-SWE-agent scaffold, 44–58% of function-level misses occur after the agent has already reached the correct file (Figure 1, three backbones). It proposes RepoNav, a post-retrieval interface that deterministically aggregates dense chunk hits into file-level candidates (hybrid peak/mean scoring, α=0.5, m=3) and serializes the top five files into a three-block scaffold — [ANCHORS], [GLIMPSE], [CANDIDATE_TARGETS] — plus an on-demand list_symbols browsing tool, leaving the retriever unchanged. On LocBench (560 Python instances, seven models), average Function Acc@5 rises from 0.441 (Bash) and 0.484 (Snippet) to 0.509 with the tool-matched Snippet+ListSym baseline and 0.557 with RepoNav; the file-to-function gap narrows from 26.9 to 21.7 points. A GPT-OSS-120B paired analysis reports +6.28pp Function Acc@5 over Snippet (p=0.017). Controlled ablations (File-Only vs Inline vs Tree Scaffold) support actionability over information volume, and SWE-QA-Bench shows +1.76 average total-score gain over Snippet. Limitations: Python-only (ast-based extraction), roughly doubled interaction tokens versus Snippet (62.4k vs 31.7k), single greedy run per configuration, significance tested only for GPT-OSS-120B, and SWE-QA scored on per-model intersection subsets.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A large share (44–58%) of function-level localization misses occur after the agent has already reached the correct file, identifying within-file navigation as the central bottleneck (Figure 1).
- RepoNav's file-centered scaffold improves function-level localization beyond a tool-matched baseline (average Function Acc@5 0.557 vs 0.509 for Snippet+ListSym), narrowing the file-to-function gap from 26.9 to 21.7 points.
- The gains arise from structured, action-oriented evidence organization rather than from exposing more structural information or adding tool access (Table 2: File-Only 32.61% → Inline 46.96% → Tree 52.13% Function Acc@5).
- The benefit transfers to repository-level question answering, with RepoNav averaging +1.76 total-score points over Snippet Search on SWE-QA-Bench across four models.
Strengths
Tool-matched causal design
The Snippet+ListSym baseline holds list_symbols access constant, so the decisive S+L→RepoNav increment (+4.8pp average Function Acc@5, Table 1/Figure 4a) isolates file-centered organization from mere tool availability.
Presentation-only RQ2 ablation directly tests the mechanism
Table 2's File-Only (32.61%) → Inline Scaffold (46.96%) → Tree Scaffold (52.13%) ladder under matched retrieval/tools/workflow, with Tree using 16% fewer tokens than Inline while invoking list_symbols +19.65pp more, supports actionability over information volume.
Consistency and corroboration across analyses
Honest cost accounting and precise parameterization
Retrieval-only reference separates retriever from interface
Weaknesses
Narrow statistical coverage
SWE-QA protocol under-specified and selection-prone
2x interaction cost vs Snippet with no accuracy-per-cost analysis
Crude lexical anchoring, Python-only
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RepoNav targets function-level localization in repository-scale code agents. The paper advances four claims: (1) a large share of function-level failures occur after the correct file is found — 44–58% of misses are Correct File, Wrong Function across GPT-OSS-120B, Qwen3-Next-80B-A3B, and Gemini-3-Flash under Bash and Snippet interfaces (Figure 1); (2) a post-retrieval interface aggregating chunk hits into file candidates and serializing them as a three-block scaffold ([ANCHORS], [GLIMPSE], [CANDIDATE_TARGETS]) plus a list_symbols browsing tool improves function-level localization; (3) gains come from structured organization, not extra structure volume or tool access; (4) the benefit transfers to repository QA (SWE-QA-Bench). Scope is explicitly Python-only; the Limitations section candidly excludes other languages, patch generation, and long-horizon maintenance.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the file-to-function gap in code localization?
It is the difference between an agent finding the right file and pinpointing the right function inside it. This paper measures it on LocBench as File Acc@5 minus Function Acc@5: about 29.2 points under a bash-only interface and 26.9 under snippet retrieval across seven models, because 44–58% of function-level misses happen after the correct file has already been reached.
What is RepoNav and how does it work?
RepoNav is a post-retrieval interface for LLM code agents. It keeps the underlying dense retriever unchanged but aggregates retrieved chunks into file-level candidates, serializes the top files as a plain-text scaffold with three blocks ([ANCHORS], [GLIMPSE], [CANDIDATE_TARGETS]), and adds a list_symbols tool for on-demand browsing of a file's structure (imports, classes, functions, line ranges) without reading full bodies.
How much does RepoNav improve localization?
On LocBench (560 Python instances) across seven models, average Function Acc@5 rises from 0.484 with flat snippet retrieval to 0.557 with RepoNav (0.509 with the tool-matched Snippet+ListSym baseline), and the file-to-function gap narrows from 26.9 to 21.7 percentage points. A paired test with GPT-OSS-120B reports +6.28pp Function Acc@5 over snippet search (p=0.017).
What are RepoNav's main limitations?
It currently parses only Python (via the ast module), roughly doubles interaction tokens versus snippet search (62.4k vs 31.7k on average), was evaluated with a single greedy run per configuration, has significance testing only for one model, and its question-answering gains are measured on per-model intersection subsets of SWE-QA-Bench. Other languages and patch-generation tasks are left to future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are embedded as loose text fragments in the extraction (e.g., Figure 1 percentages, Figure 4/5/6 annotations), so figure-level readings rely on those fragments rather than rendered axes.
- Table 8's GLM-4.7 block contains a mid-number line-break artifact ('0.7044 |.7595'), leaving a few cells in that row visually ambiguous.
- Agent prompts, the exact mini-SWE-agent configuration, and the SWE-QA judging protocol are not present in the provided text; assessment of those details rests on the paper's code-release claim.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.