dsh-product-subagent-console

The weakest link in a multi-agent workflow is rarely a single task execution; it is the gap between the plan you approved and what actually ran. You sign off on a task graph yet cannot see each child task's real progress. When a run ends ambiguous, tracing the first deviation is a needle in a haystack. And retrying only a safe subset feels risky without full context. dsh-product-subagent-console closes that loop by bringing plan design, live observation, deviation tracing, and evidence-backed recovery into one DeepSeek Harness conversation.

The workbench offers four lenses. Plan turns a goal into editable roles, dependencies, budgets, and tool bindings, then runs a preflight check before you approve any revision. Live renders every compatible child session as a real parent-child tree with lifecycle states and timing on a single canvas. Deviation maps each approved task to its actual attempts, replays the authoritative event timeline, and highlights the first evidence-backed mismatch. Recovery previews what would be retried versus reused, exports a redacted Run Capsule, and surfaces reusable workflow candidates — all before anything re-executes.

It fits teams running parallel code reviews, documentation passes, or multi-repo work where ownership and dependencies must stay visible. It helps you diagnose a run that is stuck, incomplete, or unexpectedly branched. And if you keep verifying the same pipeline, you can distill it into a safe, reusable starting point without auto-running it.