Hermeslike-Mixagent-MoA

DSH normally routes every LLM call to a single model, leaving the collaborative upside of multi-model generation completely untapped. hermeslike-moa grafts the Hermes Mixture-of-Agents pipeline directly into DSH's LlmAdapter layer: each request is first judged by a router, then fanned out to N parallel flash reference advisors who each produce an independent perspective, and finally synthesized by a streaming aggregator (pro or flash). You stay inside your existing DSH workflow and gain a noticeably higher-quality generation path without spinning up a separate Hermes instance. Two direct-connect options, deepseek-v4-pro and deepseek-v4-flash, let latency- or cost-sensitive calls bypass the MoA overhead entirely, and any unknown model id falls back to flash so the expensive path is never triggered silently.

Under the hood, the plugin faithfully mirrors Hermes' moa_loop.py reference semantics: it drops the primary system prompt, inlines tool_calls as rendered markers, trims tool results with a head+tail 4000-character budget, emits zero tool-role messages for strict-provider compatibility, and appends a synthetic user instruction to preserve the user-turn invariant. A dual-channel design lets the OpenCode Go subscription quota carry the entire pipeline preferentially, automatically failing over to the official DeepSeek channel after three consecutive failures. A three-gate V4P budget guard (forward-looking pool multiple, rolling-window Pro call cap, minimum interval) acts as a hard brake on the expensive model, with optional automatic downgrade to Flash when any gate trips. Centralized secret redaction, token-budget context trimming, per-slot reasoning-effort control, and two fanout modes (user_turn / per_iteration) are all tunable live from the DSH web settings page without a restart.

This plugin is built for developers who already rely heavily on DeepSeek inside DSH and want multi-model collaboration without leaving their toolchain, users who hold an OpenCode Go subscription and want to squeeze maximum value from it, and power users who treat V4 Pro spend as a strictly managed budget rather than an open tab.