yanxu-dsh-plugin

AI code generation can produce large diffs in minutes, but the delivery steps around those diffs often lack structure: an agent may skip repository readiness checks, conflate a Draft PR with a merged one, or expose generated changes before a human reviews them. yanxu-dsh-plugin closes that gap by adding a single, loadable skill called yanxu-delivery-workflow to DeepSeek Harness, turning each AI-assisted coding task into an auditable, stage-gated pipeline.

The skill walks the agent through a defined sequence: run doctor and sign a task contract before any code is generated; establish an explicit source-file allowlist and a single test command for implementation; keep all generated changes in an isolated archive until a human approves; treat Draft PR, CI success, merge, and deployment as four distinct states; and measure efficiency claims only with real, quality-passing tasks. Every shell command is executed through the existing bash or pwsh tools inside the DSH sandbox, so the plugin never bypasses permissions, stores credentials, auto-merges, or deploys anything. The plugin supplies the workflow instructions; the DSH sandboxed toolchain does the actual work.

This plugin is aimed at developers and teams who already rely on DeepSeek Harness for daily work and want AI-assisted coding to leave a traceable evidence trail before code is merged. It does not replace DeepSeek Harness nor the Yanxu Dev host environment, which you still install in the target project. It simply adds one disciplined workflow entry point to the DSH skills catalog.