Graph-Guided Repository Environment Construction

Jianying Pan; John Zhang; Hongyu Zhang

cs.SESep 27, 2026

Graph-Guided Repository Environment Construction

Authors: Jianying Pan, John Zhang, Hongyu Zhang

Organizations: Chongqing University · University of New South Wales

Abstract

Coding agents now increasingly rely on execution to validate their solutions, making the construction of reliable execution environments a critical enabling capability. However, repository environment construction is challenging because execution requirements are fragmented across repository artifacts and may only become apparent during execution. Existing agent-based approaches address this problem through iterative interaction, but information about the current construction state, including discovered requirements, satisfied and unresolved prerequisites, and their dependencies, can remain distributed across the interaction history. We present Graph2Env, an agent-based approach centered on DepGraph, a typed dependency graph that explicitly represents the environment requirements needed for repository execution, their dependency relations, and their states. Graph2Env uses DepGraph to guide environment construction and continuously refines it with execution feedback, while persisting successful repairs into a replayable construction procedure. The resulting artifacts are then applied in a fresh environment to verify that the constructed environment can be reproduced. We evaluate Graph2Env on a benchmark of 200 Python repositories drawn from RATBench and EnvBench, against a static dependency-inference baseline (pipreqs), three specialized environment-construction systems (Repo2Run, RAT, and SetupX), and two general-purpose coding agents (SWE-agent and Claude Code). Graph2Env achieves an 81.0% Environment Build Success Rate (EBSR) and a 59.3% Environment Setup Success Rate (ESSR), outperforming the strongest baseline by 9.5 and 9.0 percentage points, respectively.

Figures & tables

Figure 1: Motivating comparison between trajectory-based environment construction and Graph2Env on the wecode-ai/Wegent repository † † footnotemark: . Repo2Run loses track of earlier manifest evidence and installs the wrong docx package, whereas Graph2Env preserves the import-to-package mapping in DepGraph, compiles it into the build script, and succeeds under clean replay.

Figure 2: Overview of Graph2Env. Repository information is analyzed through dependency resolution and lightweight probing to construct DepGraph, which represents execution requirements, their relations, and satisfaction states. DepGraph is compiled into a replayable build script, and both provide structured context for execution and repair, while execution feedback refines the graph and updates the script. Once the designated target succeeds, Graph2Env replays the build script and restores runtime configuration in a fresh environment; replay success terminates construction, whereas replay failure returns additional feedback for further repair.

Type	Method	EBSR	ESSR	Tokens (K)	Steps
Static Dependency	pipreqs	10.0%	11.6%	N/A	N/A
Environment Building Agents	Repo2Run	42.5%	34.7%	1128.6	37.5
	RAT	44.0%	40.4%	2066.2	62.0
	SetupX	52.5%	45.9%	913.6	85.0
Coding agent	SWE-agent	50.0%	35.6%	1894.7	71.7
Coding agent	Claude Code	71.5%	50.3%	1849.5	36.5

Table 1: Environment construction success and agent token usage across baselines over the 200 repository benchmark. N/A marks a method that invokes no language model.

		Graph2Env		Claude Code		Repo2Run
Tier	$n$	EBSR	ESSR	EBSR	ESSR	EBSR	ESSR
Easy	62	93.5	79.8	87.1	70.1	66.1	53.8
Medium	86	82.6	57.5	75.6	48.3	40.7	30.5
Hard	43	69.8	41.9	51.2	33.1	18.6	20.8

Table 2: EBSR and ESSR (%) by construction difficulty tier.

n

is the number of repositories.

Configuration	EBSR	ESSR
SWE-agent	50.0%	35.6%
+ graph context	62.0% $\uparrow$ 12.0	42.4% $\uparrow$ 6.8

Table 3: SWE-agent with and without DepGraph as context, on all 200 repositories.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Action	Description
Inspection	inspect_graph	Returns the complete, filtered, or selected DepGraph view.
Inspection	inspect_setup	Returns the complete build script or a stable block for a specified node.
Inspection	inspect_observation	Returns a bounded exact slice of a prior raw execution observation.
Execution	run_shell	Runs an Agent-provided shell command in the selected construction container.
Execution	run_block	Runs the setup commands associated with one DepGraph node.
Execution	run_setup	Runs the current complete build script.

Appendix

Table 5: Agent-visible actions in the Graph2Env interface.

Figure 4: An environment construction example for rq/rq . (A) Initial requirement discovery and DepGraph-driven compilation. (B) Execution-guided repair updates the graph, build script, and runtime configuration. (C) Clean replay reconstructs the environment and successfully collects 690 tests.

Figure 5: Scale of the 200 benchmark repositories (RATBench- and EnvBench-sampled, pooled). Left: code size ( code_bytes , GitHub’s per-language byte count; median 1.28 MB, IQR 589 KB–3.62 MB). Right: non-blank lines of source code at the pinned commit, excluding vendored and minified files (median 27k, IQR 13k–88k). The two measures agree closely (Spearman

\rho=0.89

Stars	% of repos
$<$ 100	22%
100–1k	31%
1k–10k	36%
$>$ 10k	11%

Appendix

Table 6: GitHub stars.

	Data/ML	Dev tools	Scientific	Web	Infra/DB	Security	Other	Total
Repos	46	38	35	32	30	4	15	200

Appendix

Table 7: Application domains of the 200 benchmark repositories. Each repository is assigned one primary domain from its PyPI Topic classifiers where they give a unique majority, and otherwise by an LLM from its description, GitHub topics, and README.

Type	Method	# Built	Image (MB)
Static Dependency	pipreqs	111	1616.5
Environment Building Agents	Repo2Run	147	4145.5
	RAT	159	1577.0
	SetupX	126	1595.5
Coding agent	SWE-agent	111	1309.4
Coding agent	Claude Code	171	983.6

Appendix

Table 8: Built Dockerfiles and image size across baselines over the 200 repositories. # Built counts Dockerfiles that built successfully; Image (MB) is the mean over built images.

	Data/ML	Dev tools	Scientific	Web	Infra/DB	Security	Other
Method	( $n$ =46)	( $n$ =38)	( $n$ =35)	( $n$ =32)	( $n$ =30)	( $n$ =4)	( $n$ =15)
Repo2Run	23.9 / 22.0	60.5 / 58.5	45.7 / 39.2	43.8 / 33.1	43.3 / 26.6	50.0 / 24.9	40.0 / 24.8
Claude Code	60.9 / 40.0	73.7 / 71.5	85.7 / 56.5	65.6 / 40.6	66.7 / 41.9	75.0 / 32.4	86.7 / 55.4
Graph2Env	69.6 / 46.7	89.5 / 82.0	88.6 / 62.7	96.9 / 58.8	70.0 / 53.4	75.0 / 27.6	66.7 / 54.1

Appendix

Table 9: EBSR / ESSR (%) by application domain.

n

is the number of repositories in each domain.

	Small	Medium	Large
Method	( $<$ 500 KB, $n$ =35)	(500 KB–5 MB, $n$ =125)	( $>$ 5 MB, $n$ =40)
Repo2Run	74.3 / 58.1	41.6 / 35.1	17.5 / 12.8
Claude Code	91.4 / 71.0	72.8 / 51.7	50.0 / 27.6
Graph2Env	100.0 / 76.7	83.2 / 63.0	57.5 / 32.6

Appendix

Table 10: EBSR / ESSR (%) by repository code size.

Figure 6: Distribution of Graph2Env’s executed actions by category, for successful (EBSR = 1,

n

=162) and failed (

n

=27) runs. The remaining 11 failed runs (10 at the time limit, 1 crashed) were terminated before their action trace was written and are excluded. Each bar is the share of that group’s actions. Actions issued inside an action_batch are counted individually; the batch itself is not. Categories follow Table 5 .

Repository name	Commit sha	# Tests
adamchainz/django-mysql	d6f050	739
aiidateam/aiida-core	9cff5f	3,741
alliander-opensource/weather-provider-api	70f5d3	129
anthropics/anthropic-sdk-python	d2f654	4,217
anthropics/claude-quickstarts	882638	280
apiflask/apiflask	d71d64	376

Appendix

Table 11: Benchmark repositories with their pinned commit and gold test count (Section 4.1 ).

Explore similar work

Zero2Repo: Can Coding Agents Build Repositories from Scratch?
Sep 29, 2026Pei Yang, Tianyu Shi, Yuhang Yao +23Coding Agents Software
Gradient Data · McGill University · Carnegie Mellon University +6
SetupX: Can LLM Agents Learn from Past Failures in Functionality-Correct Code Repository Setup?
May 25, 2026Zihang Zhou, Ziqian Ren, Yukai Wu +7Repository-Level Code Understanding Large Language Model Agents
Shanghai Jiao Tong University · Beijing University of Posts and Telecommunications · Independent Researcher +2
RAT: RunAnyThing via Fully Automated Environment Configuration
Apr 25, 2026Renhong Huang, Dongdong Hua, Yifei Sun +4Repository-Level Code Understanding Coding Agents
Zhejiang University · Ant Group

Sep 29, 2026, cs.SEJ/K move · Enter open · S save

Zero2Repo: Can Coding Agents Build Repositories from Scratch?

Pei Yang, Tianyu Shi, Yuhang Yao, Wanyi Chen+22

Gradient Data · McGill University · Carnegie Mellon University+6

Coding Agents Software

Open paper AskPDF arXiv

Coding agents are increasingly asked to build software rather than patch it, yet benchmarks for from-scratch repository construction are mostly limited to a single language and depend on manually curated tasks. We introduce Zero2Repo, a benchmark in which an agent receives a product requirements document, an interface contract, and an empty workspace, and must deliver a complete repository in the project's native ecosystem. Tasks are produced by a language-agnostic authoring pipeline that converts real, version-pinned open-source projects into behavioral specifications, reproducible environments, and hidden acceptance tests. Each task is validated by execution: a reference implementation derived from the upstream project must pass, and adversarial validation must show that the tests reject incorrect implementations. Evaluation runs production coding agents in isolated containers, withholds the acceptance tests until an explicit submission, and assigns a binary reward only when every test passes, with no LLM judge. The pipeline and harness make no language-specific assumptions and apply to mainstream programming ecosystems; the current release contains Python, TypeScript, Go, and C++ tasks. Even on 11 tasks drawn from repositories that frontier models have very likely seen during training, the strongest agent solves only 10, and every failing submission passes 90-99% of the hidden tests; for the two strongest agents, 67-100% of failed tests trace to a single omission or a low-frequency rule stated in the specification rather than to a missing subsystem, so each failure is a concrete target for improvement.

Abstract

Explore similar work

Zero2Repo: Can Coding Agents Build Repositories from Scratch?

SetupX: Can LLM Agents Learn from Past Failures in Functionality-Correct Code Repository Setup?

RAT: RunAnyThing via Fully Automated Environment Configuration