fusion-symphony

Most frontier models carry their own blind spots and stylistic biases; stitching outputs together by hand, or training a dedicated coordinator, is costly and engineering-heavy. Fusion Symphony removes that burden entirely. With zero training and pure prompt orchestration, it turns the LLM Providers already registered in DeepSeek Harness into a collaborative advisor-and-researcher team that delivers a single, evidence-backed, multi-round-verified in-depth report. No external gateway, no weight updates, no hand-coded agent graph.

The engine is organized around three configurable axes: topology (all-agent, all-advisor, or mixed roles), tool permissions (web-search evidence on or off), and verification loop (single-round synthesis or multi-round verify with targeted redispatch). Workers draft chapters in parallel; every claim carries an evidence-ledger entry. The Verifier checks not only for contradictions and gaps but also for coverage and depth; when depth is insufficient, only the flagged sub-task is redispatched to keep costs in check. Report assembly uses a zero-compaction strategy: the aggregator generates the scaffolding (title, abstract, TOC, conclusion) in one lightweight call, while per-chapter drafts are concatenated verbatim into the final answer, so report length is no longer clamped by a single maxTokens limit. A graceful degradation chain falls back to a single-model direct answer if every worker fails, guaranteeing a usable result.

It is built for users who already have two or more Providers wired into DSH and need cross-perspective surveys, multi-source research briefs, or technical evaluation reports without the overhead of training a coordinator or manually wiring an agent graph. Configure workerModels, pick the fusion, hermes, or fugu preset, and call the tool in-session to get the fused output.