dsh-smart-profile

In the DeepSeek Harness ecosystem, nearly every capability is available as a plugin, which creates a practical problem: developers face a long catalogue of candidates and struggle to tell which ones actually fit their project. dsh-smart-profile flips the approach. Instead of starting from a generic plugin list, it begins with your repository, detects the language, framework, database, and CI/CD signals, and then trims the recommendation to the specific task you are working on.

The core of the plugin is an explainable, safety-first configuration pipeline. The detection stage returns evidence and confidence scores rather than bare labels. Candidate scoring weighs project fit, release recency, dependency size, lifecycle script risk, and DSH bundle declarations, and automatically refuses candidates below a threshold or flagged as review-required. Installation is preview-only by default; writing to a DSH profile requires both --apply and --approve. After each package is installed it is verified, and on failure the packages installed during the current run are rolled back in reverse order. The bundle also registers a set of read-only Harness tools so a model can query detection results and recommendations during a conversation, but it deliberately exposes no auto-install endpoint, ensuring a model call can never silently write host-level code.

dsh-smart-profile suits developers who already use DeepSeek Harness to build agent workflows and want plugin configuration driven by their actual project stack. It covers a broad range of languages and frameworks, from JavaScript, Python, Go, and Rust to React, Spring Boot, and FastAPI. With zero runtime npm dependencies, a read-only local dashboard bound to localhost, and a privacy-conscious scanner that skips .env files, private keys, and arbitrary source code, it is designed to feel safe in both solo and team environments.