dsh-init-command

Every project deserves an AGENTS.md that tells AI agents how to build, test, and follow the code style. Writing one by hand is tedious and easy to get wrong, especially when the repository is still a fresh skeleton. dsh-init-command adds an /init slash command to DeepSeek Harness that delegates the heavy lifting to an LLM.

The workflow runs in two phases. Phase one collects the first two levels of the directory tree and asks the model to classify the project type, language, framework, and build toolchain, returning structured JSON. Phase two embeds that classification together with the directory structure into the generation prompt and writes a complete AGENTS.md to disk, covering sections such as Project Overview, Build and Test Commands, Code Style Guidelines, Testing Instructions, Security Considerations, and AI Agent Guidelines. The classification phase defaults to thinking-off to save tokens, while the generation phase streams the model reasoning live into the session but writes the final text directly to the file, keeping the conversation clean. A mandatory rule inside the AI Agent Guidelines section instructs the agent to update AGENTS.md after every completed task so the document stays in sync as the project evolves.

Beyond core generation, the plugin handles common cold-start chores: git init with a main default branch, downloading a .gitignore matched to the detected project type, and creating an initial commit. An empty directory triggers a prompt for a project description instead of blind generation; an existing AGENTS.md is overwritten with the old content supplied as a rewrite reference. Zero runtime dependencies and no build step are required. It is best suited to developers who live in DSH as their coding environment and want AI agents to pick up project conventions quickly, whether bootstrapping a new repository or onboarding into a documentation-light codebase.