dsh-workflows

Chaining multiple Agent steps inside DeepSeek Harness often leaves you with a fragile, one-off conversation: the logic is not saved, not reusable, and impossible to babysit in the background. dsh-workflows solves that by letting you define a saved JavaScript workflow in a flat .workflow.json file, launch it with a single slash command into a supervised background run, and then watch phases, Agent spend, and artifacts through the bundled Web dashboard. You can pause, resume, stop, or save a run inline; the package-owned compatibility evaluator guarantees replay safety so committed journal calls never re-spend an Agent while uncommitted external effects can retry cleanly, and a process crash settles every active row into a fully inspectable Interrupted state.

Authoring is first-class: the built-in /create-workflow skill walks you through intent and fan-out, emits JavaScript with JSON-as-data metadata, validates one representative args-selected path against canned Agent results, and saves only after the smoke check passes. The v2 storage layout holds a process-wide native advisory lease so cooperating hosts never corrupt each other's run data. The plugin ships no Rhai parser, no external CLI binary, and no account dependency; it is compiled and smoke-tested exclusively against official DeepSeek Harness 0.1.2-rc.1, keeping the trust surface as small as possible.

Best suited for Harness developers who treat multi-Agent orchestration as a versionable engineering asset and want a Web control plane for live monitoring, background execution, and safe recovery without leaving their existing profile setup.