DevIns (Development Instruction) is an AI agent language built into AutoDev (unit-mesh/auto-dev), introduced in AutoDev DevIns —— 开源 AI 智能体语言，构建 AI 驱动的自动编程. Its author describes it as an interaction language between natural language and instruction text: natural language describes a software development task, and instruction text is used to interact with agents and with the IDE. He calls it interactive, compilable and executable — when a DevIns script is run, the DevIns compiler generates instruction text from the instructions it contains and sends it to the agent, whose result is then applied to the editor or IDE.
The language grew out of the observation, in AutoDev, that every interaction with the model went through instruction text. DevIns aims to let users describe development tasks more quickly and to process what an AI agent returns automatically, rather than leaving the user to apply it by hand.
It was first named DevIn, from “AutoDev Input Language”, and renamed to DevIns shortly before release
because a similarly named Devin AI project had just published a demo video; at the time of the announcement
its default file extension was still .devin because of JetBrains’ plugin review process.
Details
A DevIns file mixes a natural-language request with instructions that pull in context. In the post’s
example, “explain code” followed by a /file: instruction naming a Java source file is compiled by AutoDev,
together with the surrounding context, into an instruction that reads that file’s contents.
Besides reading files, code changes and custom variables, DevIns has instructions that act: /write writes
code to a path (optionally a line range), /run runs the corresponding tests, /patch applies a patch from
the AI’s output, and /commit commits code. A /write:README.md instruction followed by a code block, for
example, writes that block into README.md. The IDE provides completion and hints for the instructions.
In AutoDev plugin version 1.7.2, a user creates a hello.devins file, writes instructions and clicks run.
The authors’ stated next steps were to strengthen how the language interacts with agents, add richer
instructions, combine it with AutoDev’s custom agents, and build a cross-platform DevIns compiler.