Contract-Coding: Towards Repo-Level Generation via Structured Symbolic Paradigm

Yi Lin; Lujin Zhao; Yijie Shi

cs.SEApr 10, 2026

Contract-Coding: Towards Repo-Level Generation via Structured Symbolic Paradigm

Authors: Yi Lin, Lujin Zhao, Yijie Shi

Organizations: State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, 100876, China

Abstract

The shift toward intent-driven software engineering (often termed "Vibe Coding") exposes a critical Context-Fidelity Trade-off: vague user intents overwhelm linear reasoning chains, leading to architectural collapse in complex repo-level generation. We propose Contract-Coding, a structured symbolic paradigm that bridges unstructured intent and executable code via Autonomous Symbolic Grounding. By projecting ambiguous intents into a formal Language Contract, our framework serves as a Single Source of Truth (SSOT) that enforces topological independence, effectively isolating inter-module implementation details, decreasing topological execution depth and unlocking Architectural Parallelism. Empirically, while state-of-the-art agents suffer from different hallucinations on the Greenfield-5 benchmark, Contract-Coding achieves 47% functional success while maintaining near-perfect structural integrity. Our work marks a critical step towards repository-scale autonomous engineering: transitioning from strict "specification-following" to robust, intent-driven architecture synthesis. Our code is available at https://github.com/imliinyi/Contract-Coding.

Figures & tables

Figure 1: Linear vs. Contract-Driven Parallelism. Unlike sequential workflows that suffer from context accumulation (Left), our Language Contract serves as a topological hypervisor, decoupling dependencies to enable parallel execution and global consistency via a hierarchical graph (Right).

Figure 2: Contract-driven execution. The Contract defines the global state matrix, which drives the Hierarchical Execution Graph (HEG) to schedule parallel agents.

Figure 3: Contract-Guided Self-Healing Case Study. When divergent assumptions (e.g., missing attributes) occur between parallel Backend/Algorithm agents, Contract Auditing detects schema violations during merging and triggers Normative Alignment to auto-correct semantic mismatches.

Legacy Multi-Agent Frameworks
Method	Gomoku			Plane Battle			City Sim			Snake++			Roguelike
Method	Overall	Files	Time	Overall	Files	Time	Overall	Files	Time	Overall	Files	Time	Overall	Files	Time
MetaGPT	70	5.2	126s	50	4.0	129s	30	3.0	185s	10	4.0	210s	10	10.9	261s
ChatDev	93	4.4	88s	90	5.1	79s	45	6.0	140s	25	7.2	165s	10	13.9	144s
FLOW	96	1.0	155s	76	1.0	76s	60	1.0	110s	30	1.0	130s	0	1.0	90s
OpenHands	100	3.3	173s	90	4.9	182s	63	7.3	296s	50	13.7	267s	30	14.1	311s

Table 1: Main Results on Repository-Level Generation. Comparison across five tasks of decreasing complexity. Overall : Average of Executability, Interactivity, and Rule Adherence. Files : Generated file count (closer to ground truth is better). Time : End-to-end latency.

Figure 4: Token Decoupling Analysis. As project complexity grows, the Language Contract size remains stable. Red markers show the compression efficiency.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

AI IDE	Version	VSCode Ver.	Backbone Model	Mode
Lingma	0.2.3	1.100.0	Default	Agent
CodeBuddy	4.1.1	1.100.0	GPT-5.1-Codex-Max	Craft
Trae	3.2.1	1.104.3	GPT-5 Medium	Solo Build
Gemini Studio	–	–	Gemini 3 Flash Preview	Build

Appendix

Table 2: Configuration of Commercial AI IDE Baselines. Detailed specifications of the IDE versions, underlying VSCode engines, and backbone models used in our comparative evaluation. All AI IDEs choose their corresponding Vibe coding mode.

Figure 5: Conflict Control via Differential Interval Analysis. Using $C_{base}$ as an immutable baseline, we decompose updates into Atomic Patches. Semantic conflicts are resolved via a Union-First Strategy, prioritizing information preservation to prevent silent data loss.

	Task 1: Gomoku (Logic)						Task 2: Plane Battle (Action)
Method	Exec	Inter	Rule	Succ.	Time(s)	Files	Exec	Inter	Rule	Succ.	Time(s)	Files	Parallel
Legacy Multi-Agent Frameworks
MetaGPT	70	70	70	70%	126.0	5.2	50	50	50	50%	128.5	4.0	$\times$
ChatDev	100	90	90	93%	87.9	4.4	90	90	90	90%	79.4	5.1	$\times$
FLOW	100	90	100	97%	155.0	1.0	100	80	50	77%	75.7	1.0	✓
OpenHands	100	100	100	100%	173.2	3.3	100	100	100	100%	182.0	4.9	$\times$

Appendix

Table 3: Detailed Metrics: Basic Tasks. Comparison on Logic-Oriented (Gomoku) and Action-Oriented (Plane Battle) tasks. Gemini Studio demonstrates high speed but produces bloated file structures compared to our optimized output.

	Task 3: City Sim (11 Files)						Task 4: Snake++ (16 Files)
Method	Exec	Inter	Rule	Succ.	Time(s)	Files	Exec	Inter	Rule	Succ.	Time(s)	Files	Parallel
Legacy Multi-Agent Frameworks
MetaGPT	50	20	20	20%	185	3.0	20	10	0	0%	210	4.0	$\times$
ChatDev	60	40	35	30%	140	6.0	30	30	15	10%	165	7.2	$\times$
FLOW	60	60	60	40%	110	1.0	30	30	30	0%	130	1.0	✓
OpenHands	100	40	50	63%	296.4	7.3	60	50	40	50%	266.8	13.7	$\times$

Appendix

Table 4: Detailed Metrics: Intermediate Tasks. Comparison on City Sim (Resource Management) and Snake++ (Event Driven). Note the sharp decline in Gemini Studio ’s rule adherence due to context drift, while Contract-Coding maintains high fidelity.

Method	Arch. Integrity (%)		Func. Fidelity (%)			Overall	Time	Qual.	Key Failure Mode
Method	Struc.	Dep.	Exec.	Inter.	Rules	Success	(s)	Files	Key Failure Mode
Legacy Multi-Agent Frameworks
MetaGPT	60	40	20	10	0	10%	261.0	10.9	Implementation Sparsity
ChatDev	0	10	20	10	0	10%	144.2	13.9	Flat Structure Fallacy
FLOW	–	–	0	0	0	0%	90.6	1.0	Aggregative Collapse
OpenHands	80	50	70	10	10	30%	310.7	14.1	Lost-in-the-Middle

Appendix

Table 5: Detailed Results on Roguelike. High-complexity tasks trigger distinct pathological behaviors in baselines.

Method	Overall Success (%)		Diagnostic
Method	Plane Battle	Gomoku	Consistency $(V(C))$	Existence $(E(C))$
Ours w/o Contract	65	85	-	82%(Low)
Ours (Full)	100	100	100% (High)	100% (High)

Appendix

Table 6: Component Ablation Study. Removing the Language Contract significantly impacts Overall Success and Task Existence.

Task	Backbone Model	Success Rate	Diff.	Outcome
Plane Battle	GPT-4o (Main)	100%	-	Parity
Plane Battle	Qwen-Plus (Variant)	100%	0%	Parity
Gomoku	GPT-4o (Main)	100%	-	Parity
Gomoku	Qwen-Plus (Variant)	100%	0%	Parity

Appendix

Table 7: Robustness Across Model Families. The framework achieves identical stability with Qwen-Plus, demonstrating model agnosticism.

Panel of Figure 6 — Figure 6: Visualizing the Language Contract. The Contract bridges the Product Requirements Documents (PRD) and Technical Manifold (API Specs). This structure allows the Contract-Guided Auditing mechanism to strictly validate implementation compliance via parsing, ensuring parallel context consistency.

Code Generation

Code generation studies methods for turning natural-language, visual, or formal specifications into executable programs and code edits.

See all 313 papers

Repo0: Design-Driven Zero-to-All Code Generation
Aug 20, 2026Silin Chen, Haoyi Teng, Xiaodong Gu +5Code Generation Agentic Code Generation
Shanghai Jiao Tong University · Chongqing University
CodeTeam: An LLM-Powered Multi-Agent Framework for Repository-Level Code Generation
Jun 20, 2026Yifei Wang, Ruiyin Li, Peng Liang +4Repository-Level Code Generation Multi-Agent Coding
School of Computer Science, Wuhan University, China · School of Computer Science, Nanjing University of Science and Technology, China · School of Computer Science, Central China Normal University, China +2
ReproAgent: Contract-Guided Paper-to-Code Reproduction
Aug 25, 2026Xue Hu, Zewei Pan, Zhongyuan Wang +3Scientific Code Generation Agentic Code Generation
Beihang University · Shanghai Jiao Tong University · Peking University +3

Aug 20, 2026J/K move · Enter open · S save

Repo0: Design-Driven Zero-to-All Code Generation

Silin Chen, Haoyi Teng, Xiaodong Gu, Yuling Shi+4

Shanghai Jiao Tong University · Chongqing University

Code Generation Agentic Code Generation Repository-Level Code Generation Software Architecture Design

Open paper AskPDF arXiv

Large language model agents have made substantial progress in code generation, yet most existing systems assume a predefined repository architecture. This assumption does not hold in zero-to-all code generation, where an agent must construct an entire software project directly from natural-language requirements while maintaining a modular repository architecture throughout development. We present Repo0, a continuous structural evolution framework for zero-to-all code generation. Repo0 maintains an explicit architectural state instantiated as a Dual-Directed-Acyclic-Graph (Dual-DAG), consisting of a requirement-level DAG, a component-level DAG, and their alignment relation. Starting from natural-language requirements, it iteratively evolves component boundaries through structural actions guided by modularity metrics until structural convergence, after which the converged architecture guides test-driven development code generation. We evaluate Repo0 on six real-world repositories from RepoCraft using GPT-5 mini and DeepSeek V3.2. Repo0 achieves the highest Functionality Coverage and Pass Rate across all settings. Compared with RPG, the strongest repository-planning baseline, Repo0 improves Functionality Coverage by up to 20.08 percentage points and Pass Rate by up to 29.74 percentage points. Ablation and structural-evolution analyses further demonstrate the importance of the Dual-DAG architectural state, modularity-guided structural evolution, and explicit structural convergence.

Abstract

Code Generation

Explore similar work

Repo0: Design-Driven Zero-to-All Code Generation

CodeTeam: An LLM-Powered Multi-Agent Framework for Repository-Level Code Generation

ReproAgent: Contract-Guided Paper-to-Code Reproduction