cs.SEMar 1, 2026
SaveSWE-Adept: An LLM-Based Agentic Framework for Deep Codebase Analysis and Structured Issue Resolution
Organizations: Electrical and Computer Engineering, Purdue University
Abstract
Large language models (LLMs) exhibit strong performance on self-contained programming tasks. However, they still struggle with repository-level software engineering (SWE), which demands (1) deep codebase navigation with effective context management for accurate localization, and (2) systematic approaches for iterative, test-driven code modification to resolve issues. To address these challenges, we propose SWE-Adept, an LLM-based two-agent framework where a localization agent identifies issue-relevant code locations and a resolution agent implements the corresponding fixes. For issue localization, we introduce agent-directed depth-first search that selectively traverses code dependencies. This minimizes issue-irrelevant content in the agent's context window and improves localization accuracy. For issue resolution, we employ adaptive planning and structured problem solving. We equip the agent with specialized tools for progress tracking and Git-based version control. These tools interface with a shared working memory that stores code-state checkpoints indexed by execution steps, facilitating precise checkpoint retrieval. This design enables reliable agent-driven version-control operations for systematic issue resolution, including branching to explore alternative solutions and reverting failed edits. Experiments on SWE-Bench Lite and SWE-Bench Pro demonstrate that SWE-Adept consistently outperforms prior approaches in both issue localization and resolution, improving the end-to-end resolve rate by up to 4.3%.
Figures & tables
Figure 1: Overview of Issue Localization framework. (a) Codebase is indexed and represented as code-structure tree in the structured database. Based on this representation, (b) Issue Localization Agent performs search and pinpointing: (b-1) agent-directed depth-first traversal for selective, dependency-aware exploration, with search tools (Table 9 ) returning lightweight structural information; (b-2) post-search two-stage filtering (code-preview and location heuristics followed by content-based analysis) for candidate re-ranking and final issue-relevant locations.
Figure 2: Overview of Issue Resolution framework. (a) Issue Resolution Agent takes identified code locations as input and performs structured issue resolution. The agent is equipped with two CLI (command-line interface)-based tool families (§ 3.3 ), hypothesis_plan (Table 10 ) and hypothesis_git (Table 11 ), for planning, progress tracking, and version control. (b) Backend working memory stores structured metadata for hypotheses, to-dos and code-state checkpoints. Both tool families interface with this memory to manage checkpoints for version-control operations, including (a-1) branching to explore alternative solutions (hypotheses) and (a-2) reverting failed edits. (c) The agent merges the selected hypothesis branch after comparing all implemented hypotheses and submits the final patch.
| Framework | Model | SWE-Bench Lite | SWE-Bench Pro | ||||
| File Acc@3 | Func Acc@5 | # Tokens | File Acc@3 | Func Acc@5 | # Tokens | ||
| Embedding | CodeSage-Large | 71.2% | 40.1% | N/A | 54.0% | 26.5% | N/A |
| Embedding | CodeRankEmbed | 76.6% | 50.0% | N/A | 61.0% | 33.0% | N/A |
| SWE-agent | GPT-5.2 | 84.3% | 60.0% | 78k | 71.0% | 37.8% | 125k |
| Claude-4.5 | 90.5% | 78.6% | 298k | 78.0% | 55.0% | 340k | |
| RepoGraph | GPT-5.2 | 86.8% | 60.6% | 150k | 71.5% | 40.0% | 272k |
Table 1: Issue localization performance of different frameworks on SWE-Bench Lite and SWE-Bench Pro with GPT-5.2 and Claude-Sonnet-4.5 (Claude-4.5). Accuracy is reported at file and function levels (§ 4.1 ). Best results are in bold and second-best is underlined. # Tokens denotes the total number of input and output tokens per instance.
| Framework | Model | SWE-Bench Lite | SWE-Bench Pro | ||
| Resolve Rate | # Tokens | Resolve Rate | # Tokens | ||
| SWE-agent | GPT-5.2 | 52.5% | 260k | 30.3% | 338k |
| Claude-4.5 | 66.0% | 2640k | 39.5% | 3320k | |
| RepoGraph | GPT-5.2 | 50.3% | 272k | 30.0% | 431k |
| Claude-4.5 | 66.2% | 2968k | 38.0% | 3815k | |
| OpenHands | GPT-5.2 | 51.5% | 275k | 29.5% | 365k |
Table 2: End-to-end issue resolve rate of different frameworks on SWE-Bench Lite and SWE-Bench Pro with GPT-5.2 and Claude-Sonnet-4.5 (Claude-4.5). Best results are in bold and second-best is underlined. # Tokens denotes the total number of input and output tokens per instance, including issue localization and issue resolution.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Venn diagram (left) of resolved-instance overlap between SWE-Adept and SWE-agent; and error breakdowns (right) for instances uniquely failed by each method. Reported results are on SWE-Bench Lite with Claude-Sonnet-4.5.
| Framework | Model | Python | JavaScript | Go | TypeScript | ||||
| Localize | Resolve | Localize | Resolve | Localize | Resolve | Localize | Resolve | ||
| SWE-agent | GPT-5.2 | 43.3% | 38.3% | 40.0% | 29.2% | 30.8% | 24.2% | 35.0% | 27.5% |
| Claude-4.5 | 61.7% | 51.7% | 54.2% | 37.5% | 50.8% | 30.8% | 50.0% | 35.0% | |
| OrcaLoca | GPT-5.2 | 50.0% | 42.5% | 45.8% | 34.2% | 34.2% | 25.0% | 40.0% | 30.0% |
| Claude-4.5 | 64.2% | 53.3% | 56.7% | 40.0% | 53.3% | 31.7% | 52.5% | 37.5% | |
| SWE-Adept | GPT-5.2 | 55.0% | 48.3% | 47.5% | 37.5% | 36.7% | 26.7% | 42.5% | 40.0% |
Table 3: Issue localization and resolution performance by programming language on SWE-Bench Pro subset with GPT-5.2 and Claude-Sonnet-4.5 (Claude-4.5). The subset contains 60 Python, 60 JavaScript, 60 Go, and 20 TypeScript instances. Localize denotes function-level localization accuracy (Func Acc@5), and Resolve denotes end-to-end issue resolve rate.
Figure 4: Search behavior of Issue Localization Agent and localization accuracy by maximum search depth. (a) Search-tool invocation distribution of Issue Localization Agent. (b) Instance distribution (bars) and function-level localization accuracy (lines) by maximum search depth on SWE-Bench Lite with Claude-Sonnet-4.5.
Figure 5: Problem-solving behavior of Issue Resolution Agent and resolve rate by number of explored hypotheses. (a) Prevalence of systematic problem-solving behaviors: multi-hypothesis branching, dynamic to-do expansion, and checkpoint-based reversion (behaviors are non-mutually exclusive and may overlap). (b) Instance distribution (bars) and resolve rate (lines) by number of explored hypotheses on SWE-Bench Lite with Claude-Sonnet-4.5.
| Framework | SWE-Bench Lite | SWE-Bench Pro |
| SWE-agent | 66.0% | 39.5% |
| SWE-Adept | 71.3% | 45.0% |
| - w/o designed issue localization | 69.0% | 43.3% |
| - w/o designed issue resolution | 68.7% | 41.5% |
Table 4: Ablation study evaluating the contribution of each agent in SWE-Adept. Results report end-to-end resolve rate on SWE-Bench Lite and SWE-Bench Pro with Claude-Sonnet-4.5. “w/o designed” replaces the corresponding SWE-Adept module with the SWE-agent module.
| SWE-Bench Lite | SWE-Bench Pro | |||
| Func Acc@5 | # Tokens | Func Acc@5 | # Tokens | |
| • GPT-5.2 | ||||
| Search returns full code content | 69.3% | 218k | 45.0% | 316k |
| Search returns code preview + specialized filtering | 70.8% | 197k | 46.0% | 252k |
| • Claude-Sonnet-4.5 | ||||
| Search returns full code content | 87.2% | 393k | 62.0% | 502k |
Table 5: Ablation study evaluating context management for issue localization . “Search returns full code content” provides complete source code during traversal and lets the agent directly select candidate locations from the full-code context. “Search returns code preview + specialized filtering” represents our proposed design: it provides compact code previews during traversal, first filters preview-level candidate locations, and then loads full source code only for shortlisted locations during final re-ranking.
| Framework | SWE-Bench Lite | SWE-Bench Pro |
| SWE-agent | 66.0% | 39.5% |
| SWE-agent + raw Git commands (prompted) | 67.0% | 37.0% |
| SWE-Adept | 71.3% | 45.0% |
| SWE-Adept + raw Git commands (no hypothesis_git , no memory) | 69.0% | 41.5% |
Table 6: Ablation study evaluating raw Git command usage for issue resolution . Results report end-to-end resolve rate on SWE-Bench Lite and SWE-Bench Pro with Claude-Sonnet-4.5.
| Framework | SWE-Bench Lite | SWE-Bench Pro | ||
| Localize | Resolve | Localize | Resolve | |
| mini-SWE-Agent | 75.9% | 64.2% | 50.5% | 36.8% |
| Claude Code | 85.4% | 70.2% | 59.5% | 42.5% |
| SWE-Adept | 87.8% | 71.3% | 63.3% | 45.0% |
Table 7: Comparison with Claude Code and mini-SWE-Agent using Claude-Sonnet-4.5. Localize denotes function-level localization accuracy (Func Acc@5), and Resolve denotes end-to-end issue resolve rate.
| Framework | Model | SWE-Bench Lite | SWE-Bench Pro | ||
| Localize | Resolve | Localize | Resolve | ||
| SWE-agent | GLM-4.7-Flash (30B) | 55.5% | 43.3% | 35.0% | 26.5% |
| SWE-agent | Qwen3.5-27B | 67.3% | 56.8% | 43.5% | 33.8% |
| SWE-Adept | GLM-4.7-Flash (30B) | 63.3% | 48.2% | 42.0% | 30.3% |
| SWE-Adept | Qwen3.5-27B | 75.4% | 62.7% | 51.0% | 38.3% |
Table 8: Issue localization and end-to-end issue resolution results on SWE-Bench Lite and SWE-Bench Pro using open-source models. Localize denotes function-level localization accuracy (Func Acc@5), and Resolve denotes end-to-end issue resolve rate.
| Tool | Description | Input Parameters | Output (Agent Observation) |
| find_file | Search file by exact file-name or glob pattern. Enumerate files within codebase (optionally restricted to a specified directory). | file_name † dir_path ‡ | Matched file paths; for each file, a skeleton listing the signatures of classes and functions defined in that file. |
| find_code_def | Search class/function definition via exact matching, with fallback to regex matching and fuzzy matching ranked by a weighted sum of character-level similarity metrics (n-gram, Jaro–Winkler distance, and longest common subsequence). Enumerate indexed code units (optionally restricted to a file). | definition_name † file_path ‡ | Ranked retrieved code definitions with file path, line span, child-unit identifiers, and a concise code preview (definition signature + child-unit invocation context). |
| find_code_content | Search variable name (robust to camel-case/snake-case variants) or an exact code snippet . Enumerate code lines (optionally restricted to a file or a line span). | content † file_path ‡ start_line ‡ end_line ‡ | Retrieved matches, each associated to a containing code unit (function/class/code chunk), with the unit name, file path, line span, child-unit identifiers, and a concise code preview (unit signature + matched lines). |
| find_child_unit | Search class/function definition via exact matching given its name and file path (specified by a child-unit identifier). Enumerate indexed code units restricted to a file. | definition_name † file_path † | Exact-match class/function definition with file path, line span, child-unit identifiers, and a concise code preview (definition signature + child-unit invocation context). |
| finish_search | Signal completion of the search and trigger subsequent filtering/ranking over the collected candidates. | None | None |
Table 9: List of tools for Issue Localization Agent . Superscripts of input parameters denote argument requirement: required † , optional ‡ . Gray text in the Description column provides auxiliary information on how each search tool performs enumeration.
Figure 6: System prompt and stepwise instruction template for Issue Localization Agent .
| Tool Family | Command (CLI-Based) | Description | Output (Agent Observation) |
| hypothesis_plan | update_hypothesis – hypotheses_markdown | • Create, progress-track, and update hypotheses (i.e., alternative solutions) sorted by agent-estimated confidence, and annotate with status tags: [ ] (pending) [-] (in-progress) [v] (successful) [!] (failed) Support dynamic expansion for adaptive planning. • Store each hypothesis’s content and status in working memory. | Hypothesis overview with current status and ordering. |
| update_todo – current_hypothesis – todos_markdown | • Create, progress-track and update a to-do list for the current hypothesis, and annotate with status tags: [ ] (pending) [-] (in-progress) [v] (successful) [!] (failed) Support dynamic expansion for adaptive planning. • Store each to-do’s content and status in working memory. | Hypothesis-corresponded to-do list with current status. | |
| log_insight – insight | • Generate insights for the current hypothesis based on execution feedback, which inform reflection-driven actions (e.g., revert) and final cross-hypothesis comparison. • Store the insights in working memory. | Insights attached to the current hypothesis. |
Table 10: Planning tools for Issue Resolution Agent . The hypothesis_plan tool family interfaces with backend working memory to support adaptive planning, progress tracking, and insight logging. Tool invocations follow the command-line interface (CLI) format: each command uses – parameter <value> syntax, e.g., hypothesis_plan log_insight – insight <insight_content> .
| Tool Family | Command (CLI-Based) | Description | Output (Agent Observation) |
| hypothesis_git | init_base (internally executes) git config git rev-parse HEAD git add -A git commit -m git rev-parse HEAD | • Obtain the Git hashes of (i) the original code state and (ii) a shared common working base git_hash_base (including issue reproduction code) for subsequent hypothesis branches. • Store the Git hashes in working memory. | Confirmation of saved original code state and common working base. |
| start_hypothesis – branch_name (internally executes) git stash push -m git checkout git checkout -b | • Checkout a new hypothesis branch from the common working base. • Retrieve the Git hash of the common working base from working memory. • Store the branch name for the corresponding hypothesis in working memory. | Confirmation of branch creation and checkout; workspace moves to the hypothesis branch. | |
| commit_todo – todo_content – commit_message (internally executes) git add -A git commit -m git rev-parse HEAD | • Commit code changes for one completed to-do step and obtain its checkpoint metadata (Git hash and commit message). • Store the Git hash and commit message for the corresponding to-do in working memory. | Confirmation of to-do commit. | |
| revert_to – source_hypothesis – source_todo – new_branch_name (internally executes) git stash push -m git checkout git checkout -b | • Checkout the specified to-do checkpoint and create a new branch from that state, then switch the workspace to the new branch and continue exploration under the new hypothesis branch. • Retrieve the Git hash of the specified to-do from working memory. • Store the new hypothesis’s branch name in working memory. | Confirmation of revert and new-branch creation. | |
| compare_hypotheses (internally executes) git diff --shortstat git diff --numstat | Retrieve records from working memory and compare implemented hypotheses. | Hypothesis comparison report - branch name - hypothesis content and status - to-dos content and status - commit messages - insights - code diff statistics (against the original state) | |
| merge_solution – branch_name (internally executes) git checkout --detach git merge | • Retrieve the selected hypothesis branch and the Git hash of the original code state from working memory. • Check out the original code state, then apply the selected branch’s changes to produce a clean, patch-ready state for submission. | Confirmation that changes are applied onto the original code state. |
Table 11: Version-control tools for Issue Resolution Agent . The hypothesis_git tool family interfaces with backend working memory to manage code-state information. Tool invocations follow the command-line interface (CLI) format: each command uses – parameter <value> syntax, e.g., hypothesis_git merge_solution – branch_name <name> . Gray text in the Command column lists the raw Git commands executed internally (not exposed to the agent).
Figure 7: System prompt and stepwise instruction template for Issue Resolution Agent .