AutoDev DevIns —— 开源 AI 智能体语言，构建 AI 驱动的自动编程

This post, written in Chinese by Phodal Huang, announces DevIns, a new AI agent language built into AutoDev (unit-mesh/auto-dev). It follows a previous AutoDev release that had added custom agents, letting users build their own agents to assist with software development tasks. DevIns is presented as a way for users to describe development tasks more quickly while also automatically processing what an AI agent returns — for example, a /write:README.md instruction followed by a code block is translated by AutoDev into writing that content to README.md.

The post explains the motivation, what the language is, how to use it in the IDE, why it was renamed, and what the authors plan next.

Key Points

• The author reports that, as more and more agents were built in AutoDev, all interaction with the model turned out to happen through instruction text: the user interacts with an agent through instructions, and the agent returns content and operates on the editor or IDE. A custom prompt such as “explain the selected code: $selection” is given as an example in which “explain” can be read as an instruction.
• From this the authors asked whether users could interact with agents in natural language, with the model replying in instruction text that operates the editor or IDE.
• DevIns is defined in the post as an interaction language between natural language and instruction text: natural language describes the development task, and instruction text is used to interact with agents and the IDE. The post calls it interactive, compilable and executable.
• Beyond reading file contents, code changes and custom variables, the post lists agent instructions: /write to operate on code at a path, /run to run the corresponding tests, /patch to apply a patch from the AI’s output, and /commit to commit code.
• The authors say they drew on their IDE-development experience to give DevIns intelligent completion and hints, so that users need not worry about the complexity of the instructions.
• To try it, the post says to install version 1.7.2 of the AutoDev plugin, create a hello.devins file, write DevIns instructions and click run.
• The language was first called DevIn, from “AutoDev Input Language”. When it was close to release, a similarly named Devin AI project published a demo video, so the authors renamed it DevIns (Development Instruction). Because of JetBrains’ review process, the default file extension was still .devin at the time of writing.
• Planned next steps are strengthening how DevIns interacts with agents (the post wonders whether something like Jupyter Notebook fits), building more agents with AutoDev’s custom-agent capability, designing richer DevIns instructions, and building a cross-platform DevIns compiler.

Context

The post is the author’s own announcement of a feature of the open-source tool he works on, and its claims describe that feature as designed at the time rather than any measured result.