dsh-plugin-development

This plugin addresses a common pain point in DeepSeek Harness plugin development: the same workflow knowledge often gets duplicated across Codex, Claude Code, and DSH, while DSH plugins themselves can fall into quite different forms, such as live dynamic Cordis plugins, source-backed workspace packages, or external installable bundles. Without a shared process, it is easy to mix design, implementation, packaging, review, and runtime validation into an inconsistent workflow. dsh-plugin-development packages that full process into one portable Agent Skill.

Its core strength is distinguishing the plugin mode first and then applying the right checks. Dynamic runtime work focuses on process-local Host or Client behavior and live Cordis state. Workspace plugin work defers to the current DSH repository, tests, Loader composition, lifecycle proof, and snapshots when product-visible. Out-of-tree bundle work centers on packed files, profile-scoped installation, configuration dumping, packaged boot, and cleanup evidence. The workflow also covers Loader export failures, Service Definition/Provider/Consumer ownership, model-visible logging, Client Slots, CLI surfaces, configuration, lifecycle disposal, profile precedence, Git build risk, and audit-only authorization boundaries, without adding MCP servers, accounts, credentials, or remote-service dependencies.

It is suited to authors building or maintaining DeepSeek Harness plugins, as well as agent-workflow maintainers who want one canonical Skill to work across Codex, Claude Code, and DSH. During the DSH developer-preview phase, the Skill emphasizes evidence and boundaries: it helps verify design, packaging, and diagnosis where possible, and when runtime activation cannot be checked, it should stop and report that limitation rather than overclaim. An optional DSH bundle adapter can register the Skill in a profile and remove it reversibly, but the central product remains the shared portable Agent Skill itself.