How AutoDev Bridge Uses LLMs to Accelerate Legacy System Migration

Background

In 2023 the team built legacy‑system features under the name AutoDev. By 2025, after several iterations of the autonomous programming agent AutoDev Sketch, the concept of AutoDev Bridge was introduced to apply AI agents to legacy system migration.

Why Large Models Perform Better

Reasonable path planning – large models act as knowledge bases that can evaluate migration costs.

Architecture blueprint generation – by analyzing directory structures, dependencies, and APIs, the AI can draft an initial architecture diagram.

Business knowledge extraction – using AST analysis and call‑chain information, the AI uncovers existing business logic for rewrite.

Cross‑language translation – LLMs can translate code into target languages within seconds to minutes.

Migration safety net – automated tests verify migration correctness, providing precise regression testing (front‑end coverage remains limited).

AutoDev Bridge Architecture

The preliminary solution consists of several core components:

LLM‑generated migration plan based on an “Explore‑Perceive‑Respond” workflow.

C4‑based current architecture analysis driven by AI tool calls.

Business‑logic analysis using AST and call‑chain information.

Generation of migration test cases.

AI‑assisted code translation.

Tight integration with IDEs supplies accurate context, reducing hallucinations.

Explore‑Perceive‑Respond Framework

The process mirrors the Cynefin model for complex problems:

Explore : invoke tools to collect basic system information such as directory layout and dependencies.

Perceive : use exploration results to produce an initial architecture blueprint and migration plan.

Respond : validate the plan, generate migration test cases, and produce migration code.

In practice, a Chinese‑language model (V3) handles exploration, a planning model (R1) designs the plan, and V3 executes the response.

Architecture‑View Tooling

A set of view‑oriented tools helps the AI understand system structure:

componentView : lists all UI components (e.g., React, Vue).

containerView : lists all modules in the project.

webApiView : lists all Web APIs.

stylingView : lists all CSS/SCSS classes.

dir : retrieves the directory hierarchy at a given level.

history : retrieves commit history for a file.

knowledge : analyzes API call chains (default depth = 2, from Controller to Repository).

Example AI‑generated architecture‑view analysis:

Note: DeepSeek currently struggles with the C4 model and requires further optimization.

Business Knowledge Extraction and Understanding

The workflow extracts business logic by first using webApiView to list APIs, then knowledge to retrieve the call chain. Example output: /knowledge:GET#/api/blog/* With the Controller‑to‑Repository call chain, the AI can accurately infer the API’s business logic. A representative call‑tree diagram:

In real scenarios the AI also combines search tools to gather additional context.

Conclusion

The project is open‑source and can be followed on GitHub: https://github.com/unit-mesh/auto-dev