dsh-refine

Refining prompts, memories, skills, and subagents inside dsh often feels like chasing scattered terminal outputs, with no single place to inspect what changed, why it changed, or how to roll it back. dsh-refine fills that gap by porting the prime-agent /refine concept into the dsh ecosystem and wrapping it in a familiar UX layer: a /refine command for quick, instruction-driven refinements, and a settings panel where entries, a refinement timeline, one-click rollback, and auto-gate audit all live side by side.

The command path is deliberately non-blocking. When you type /refine , the engine and agent are validated synchronously, the acknowledgement returns immediately, and the input box is never frozen for a planner round-trip. Results land later in the panel timeline and /refine history. Rollback, by contrast, is fully synchronous: it applies only the stored inverse edits with no LLM call, so undoing the last refinement is a single click. Most importantly, the dsh-continual-harness engine is optional at runtime; without it, every operation degrades into actionable setup guidance rather than an error, and session logs written by the engine stay loadable through a one-time event-type registration with no data migration.

If you manage multiple entry kinds in dsh and want refinement to become an auditable, reversible workflow rather than a stream of ad-hoc edits, dsh-refine adds that lightweight, complete UX layer on top of the existing harness.