Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software Repair

Mehdi Bahrami; Kosaku Kimura; Satoshi Munakata; Satoshi Nakashima; Yu Ishikawa; Kosuke Maeda; Nao Soma; Kenichi Kobayashi

cs.SEAug 16, 2026

Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software Repair

Authors: Mehdi Bahrami, Kosaku Kimura, Satoshi Munakata, Satoshi Nakashima, Yu Ishikawa, Kosuke Maeda, Nao Soma, Kenichi Kobayashi,

Organizations: Fujitsu Research of America Santa Clara, CA, USA · Fujitsu Research Kawasaki, Japan · Fujitsu Research of Europe Slough, UK · Fujitsu Research & Development Beijing, China

Abstract

Industrial software-engineering teams increasingly need LLM agents that turn bug reports into correct patches, yet benchmark-scale operation adds long horizons, tool-use discipline, context persistence, heterogeneous clusters, and evaluation reuse. We present Kozuchi Agent, a language-agnostic open-weight repair agent and CI-operated evaluation pipeline. Explicit phases, persistent state, deterministic tools, a model-independent action interface, and cross-agent test-time selection make runs auditable and repeatable. With locally hosted Qwen3.5-27B, no fine-tuning, and TTS@8, Kozuchi resolves 374/500 SWE-bench Verified instances on the official evaluator. Unchanged on Multi-SWE-bench Java, the same 27-billion-parameter agent resolves 41/128 instances (32.03%), ranking first among strict open-weight submissions and fourth of 42 overall; on Python it ranks 12th of 135 and first among open-weight systems. Per-phase behavior remains within +/-5 percentage points across languages. Remaining failures mainly reflect semantic correctness, Java-specific harness issues, and selection errors. Across both tracks, results compare favorably with open/local peers by parameter count. Analysis of candidate diversity, selector regret, and patch reliability shows that the remaining gap is primarily semantic correctness and selection rather than edit formatting or proprietary-model access. Operationally, reusable CI stages reduce operator touch-points from five to one across heterogeneous internal clusters.

Figures & tables

Figure 1. Kozuchi Agent is parameter efficient on both Python SWE-bench Verified and Multi-SWE-bench Java. Dual-axis scatter plot of agent success rate against total parameter count on a log scale. The left (blue) axis shows SWE-bench Verified Python results for Kozuchi Agent and open/local peers with the Python Pareto frontier as a dashed blue line; the right (red) axis shows Multi-SWE-Bench Verified Java results for open-weight peers with the per-size-class Java envelope as a dashed red line. Kozuchi Agent + Qwen3.5-27B is highlighted with star markers on both axes (74.8 percent Python, 32.03 percent Java) at 27 billion parameters.

Mechanism / claim	Evidence type	Where
Candidate count $K{=}8$	Ablated: pass@1 334–343 vs. oracle 408	Tab. 7
Cross-agent selector	Ablated: 376 vs. 362 (order) vs. 346 (self-tests)	Tab. 7
Selector weights $w_{B},w_{R}$	Sensitivity sweep: $\leq$ 5/500 outcome change	§ 7
Phase graph, handover state, formatter, tool sandbox, CI reuse	Operational signatures; not ablated	§ 6 , § 9
Engineer-minute saving	Derived order-of-magnitude estimate	Tab. 2
GPU-hours	Wall-clock-derived estimate	§ 8.5

Table 1. Evidence map: mechanisms and claims by type.

Aspect	Pre-Kozuchi agent	Kozuchi agent
Operator touch-points / cycle	5 (vLLM, launch, wait, grade, log)	1 (CI push)
Operator effort / cycle (est. min)	$\approx$ 75	$\approx$ 5
Backend integrations	separate integration effort	4 action formats $\times$ 17 model configs
Reusable pipeline stages	0 of 9	6 of 9
Cluster targets / abstraction	per-cluster scripts	5 ( ENV_NAME )
Score provenance	manual score log	495 traj. + selection JSON

Table 2. Per-evaluation-cycle workflow comparison.

Figure 2. Kozuchi Agent workflow from candidate generation to cross-agent selection. Two-lane system diagram. The upper lane shows the broad issue-to-patch workflow through candidate generation, archived artifacts, selection, and submitted patch. The lower lane expands one candidate trajectory with durable shared memory, turn-level action normalization, tools, and the cross-run test application that produces the selector signals B_ij and R_ij.

Quantity	Value	Quantity	Value
Expected (Verified)	500	$w_{B}$	0.3
Selected (covered)	495	$w_{R}$	0.7
Resolved (xcheck selector)	376	Tie epsilon	0.001
Score over expected (xcheck)	0.752	Tie-breaker	shortest_patch_raw
Resolved (order baseline)	362	Tie-break applied	304/495
Score over expected (order)	0.724	Mean unique patches	5.59/8

Table 3. Cross-agent selector summary.

Quantity	Interpretation	Value
Mean pass@1	Average over eight runs	$67.7\text{\,}\mathrm{\%}$
All-run solved	$r_{i}=8$ scaffold-easy instances	234 / 500
Zero-run solved	$r_{i}=0$ scaffold-hard instances	92 / 500
Marginal instances	$1\leq r_{i}\leq 7$ selector-sensitive	174 / 500
Oracle pass@8	Any run resolves	408 / 500 = $81.6\text{\,}\mathrm{\%}$
Cross-agent selector	Official-evaluator result	374 / 500 = $74.8\text{\,}\mathrm{\%}$

Table 4. Candidate-level TTS@8 decomposition.

Repository	Resolved/Total	Rate	Repository	Resolved/Total	Rate
astropy/astropy	13/22	59.1%	pydata/xarray	18/22	81.8%
django/django	177/231	76.6%	pylint-dev/pylint	3/10	30.0%
matplotlib/matplotlib	23/34	67.7%	pytest-dev/pytest	16/19	84.2%
mwaskom/seaborn	1/2	50.0%	scikit-learn/scikit-learn	27/32	84.4%
pallets/flask	1/1	100.0%	sphinx-doc/sphinx	30/44	68.2%
psf/requests	8/8	100.0%	sympy/sympy	57/75	76.0%

Table 5. SWE-bench Verified resolution by repository.

Figure 3. Per-repository resolution rate for Kozuchi agent (first column on the left) and the curated open-weight peers. Heatmap comparing per-repository resolution rates for Kozuchi agent and open-weight peer agents across the 12 repositories in SWE-bench Verified. Kozuchi agent leads or ties the strongest peer on several high- and medium-volume repositories including django, sympy, sphinx, and pytest, while pylint remains low for most systems.

Category	Count	Share
WRONG_FIX	115	$91.3\text{\,}\mathrm{\%}$
REGRESSION	6	$4.8\text{\,}\mathrm{\%}$
MISSING_ARTEFACT	5	$4.0\text{\,}\mathrm{\%}$
PATCH_DID_NOT_APPLY	0	$0.0\text{\,}\mathrm{\%}$
EMPTY_PATCH	0	$0.0\text{\,}\mathrm{\%}$
Total	126	$100.0\text{\,}\mathrm{\%}$

Table 6. Failure taxonomy for unresolved instances.

Variant	Resolved	Rate	Note
$K{=}1$ , no selector (mean)	338.6	67.7%	eight-run mean
$K{=}1$ , no selector (best)	343	68.6%	best leg
$K{=}8$ , order baseline	362	72.4%	no cross-agent ranking
$K{=}8$ , self-tests only	346	69.2%	estimated
$K{=}8$ , cross-agent selector	376	75.2%	selected
$K{=}8$ , oracle	408	81.6%	upper bound

Table 7. Artifact-based TTS@8 ablation. Scope: this table isolates candidate count (

K

) and selector design only; phase graph, formatter, persistent state, and tool sandbox are not ablated and are evidenced operationally in § 6 and § 9 .

Figure 4. Cross-language transfer summary. Two-panel comparison. The first panel shows Python at 374 of 500 resolved and Java at 41 of 128 resolved with Wilson confidence intervals. The second panel overlays the eight phase message shares for the Python and Java trajectories, showing similar shape across the phase graph.

Signal	Python	Java
Benchmark	SWE-bench Verified ( $n{=}500$ )	Multi-SWE-bench Java ( $n{=}128$ )
Resolved (Wilson 95% CI)	374/500 = 74.80% [70.82, 78.41]	41/128 = 32.03% [24.57, 40.54]
Leaderboard position	12/135 overall	4/42 overall
Open-weight rank	1st curated open-weight	1st strict open-weight
BH-FDR peer wins	16/17 open-weight peers	34/41 Java peers
Same-family open margin	+26 vs. Qwen3-Coder-480B	+38 vs. Qwen2.5-72B

Table 8. Python–Java cross-track evidence for the same 27-billion-parameter Kozuchi agent configuration.

Software Engineering Agents

Software Engineering Agents are AI systems that use code, tools, and feedback to perform software development tasks.

See all 347 papers

PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents
Jul 21, 2026Tianyue Jiang, Yanlin Wang, Xin He +7LLM Self-Correction Software Engineering Agents
Sun Yat-sen University, China · State Key Laboratory of Blockchain and Data Security, Zhejiang University, China · School of Cyber Science and Engineering, Huazhong University of Science and Technology, China +1
SWE-Doctor: Guiding Software Engineering Agents with Runtime Diagnosis from Multi-Faceted Bug Reproduction Tests
Jul 1, 2026Yaoqi Guo, Yang Liu, Jie M. Zhang +3Software Engineering Agents Automated Program Repair
Nanyang Technology University · King’s College London · Peking University +2
From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents
Apr 2, 2026Nikolai Ludwig, Wasi Uddin Ahmad, Somshubra Majumdar +1Software Engineering Agents LLM Agent Training
NVIDIA Santa Clara, CA 95051, USA

Jul 21, 2026J/K move · Enter open · S save

PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents

Tianyue Jiang, Yanlin Wang, Xin He, Daya Guo+6

Sun Yat-sen University, China · State Key Laboratory of Blockchain and Data Security, Zhejiang University, China · School of Cyber Science and Engineering, Huazhong University of Science and Technology, China+1

LLM Self-Correction Software Engineering Agents Automated Program Repair Fault Localization Multi-Agent Coding

Open paper AskPDF arXiv

While Large Language Models have greatly advanced automated issue resolution, existing agent-based methods exhibit a fundamental limitation in their insufficient exploration of repair strategies. This insufficiency manifests in two key aspects. First, the exploration of multiple potential edit locations is limited. Second, the exploration of repair attempts at each location is also insufficient. To address these challenges, we present PhoenixRepair, a multi-agent framework that systematically explores multiple candidate edit locations and performs iterative reflection and refinement on patch generation, thereby expanding the search space of repair strategies. Our framework begins with multi-location sampling, optionally augmented with graph-based localization information for difficult tasks, followed by iterative reflection and refinement to generate better patches, culminating in final-round generation guided by distilled insights from all historical attempts. Experiments on SWE-bench-Verified demonstrate that PhoenixRepair achieves the largest relative improvement of 7.8% over SWE-agent under DeepSeek-V3.1, and attains the highest resolved rate of 76.0% Pass@1 under MiniMax-M2.5. Meanwhile, it achieves higher fault localization accuracy than existing approaches. Our code is available at https://github.com/DeepSoftwareAnalytics/PhoenixRepair.

Abstract

Software Engineering Agents

Explore similar work

PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents

SWE-Doctor: Guiding Software Engineering Agents with Runtime Diagnosis from Multi-Faceted Bug Reproduction Tests

From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents