dsh-my-go

Running multiple sub-agents inside a single DeepSeek Harness session quickly turns the main conversation into a bloated, hard-to-audit mess: it has to remember who is doing what, re-prompt failed work, and keep quality gates — all while burning context. dsh-my-go lifts that orchestration into a broker-level state machine. Sisyphus, the main session, becomes a pure dispatcher and quality reviewer; seven specialized sub-agents (quick execution, search, documentation, multimodal recognition, code authoring, planning, and architecture debugging) each persist as independent, continuable sessions. No two run at the same time, no two talk directly — everything flows through Sisyphus, which keeps the observable chain tight.

Dispatching happens at step granularity. Prometheus decomposes the goal into an ordered list of steps; Sisyphus then picks the lowest-token specialist for each one, escalating only when reasoning or design is required. A conclusion that misses the bar is rejected on the spot, and the offending sub-agent keeps its context to retry. Every specialist can be bound to its own provider, model, and reasoning tier, so cheap tasks ride lightweight models while heavy design work gets the strongest available. Toggle peak routing and the plugin automatically switches between two provider/model pairs based on Beijing peak hours — no manual config flipping needed.

If you already develop with DSH and want the think-then-act, review-then-retry, reject-then-continue loop codified without writing your own scheduling prompts, dsh-my-go ships a ready-made preset. Pick the MyGO!!!!! mode, describe the goal in plain language, and let the broker handle the rest. It does not alter DSH's core interaction model; it simply adds a stateful, auditable, and configurable orchestration layer on top.