Hermeslike-Mixagent-MoA

Most DSH LLM calls go straight to one model, silently discarding the quality lift you get from letting several models reason in parallel and then merge their answers. hermeslike-moa bolts the Hermes Mixture-of-Agents pipeline into DSH's LlmAdapter at the transport layer: a router picks the strategy, N flash reference advisors generate independent perspectives in parallel, and a streaming aggregator (pro or flash) stitches them into a single coherent reply. The whole chain runs inside your existing DSH session, so you never leave your toolchain just to get multi-model reasoning. Two direct-connect shortcuts (deepseek-v4-pro, deepseek-v4-flash) skip the MoA hop when speed or cost matters, and any unrecognized model id defaults to flash, so the expensive pipeline is never invoked by accident.

Fidelity to Hermes is not just cosmetic. lib/reference.js re-implements the moa_loop.py advisor view verbatim: the primary system prompt is dropped, tool_calls are rendered as inline markers, tool results are cropped to a 4000-character head+tail budget and folded into the preceding assistant message, no tool-role messages are emitted (keeping strict providers happy), and a synthetic user instruction guarantees the conversation still ends on a user turn. Operationally the plugin runs a dual-channel stack that prefers the OpenCode Go subscription and fails the entire pipeline over to the official DeepSeek endpoint after three consecutive misses. A three-gate V4P guard (forward-looking budget-pool multiple, rolling-window Pro-call cap, minimum inter-call interval) hard-caps spend on the expensive model and can auto-downgrade to Flash. Secret redaction, token-budget context trimming, per-slot reasoning-effort dials, and two fanout cadences (user_turn / per_iteration) are all adjustable live from the DSH web settings page, no restart needed.

If you depend on DeepSeek inside DSH day-in, day-out and want multi-model collaboration that feels native, or you hold an OpenCode Go subscription and want every token counted twice before the expensive model even wakes up, this is the plugin that keeps you in your workflow and out of a second toolchain.