Introduction

In the plugin ecosystem of DeepSeek Harness (DSH), multi-model collaboration usually faces a barrier of complex configuration or the need for model training. The devmom/fusion-symphony plugin provides a multi-model fusion orchestration solution based on the LLM Provider registered with DSH itself, using the MoA (Mixture of Agents) paradigm to achieve zero-training model fusion.

The following sections describe how to install, configure, and use the plugin.

What It Is

fusion-symphony is a DSH plugin package positioned as a multi-model fusion orchestration engine. It does not rely on external gateways and directly invokes models using the ctx.llm Provider already registered in DSH (such as deepseek, pi-ai, etc.). By replacing training-based coordinators with pure prompt orchestration, it reduces the advantages of the OpenRouter Fusion, Sakana Fugu Ultra, and Hermes MoA paradigms to three configurable axes: topology, tool permissions, and verification loops.

Core Capabilities

The plugin mainly includes the following features:

  • MoA fusion based on registered Providers: Uses LLM Providers registered in DSH to build multi-model collaboration.
  • Zero-training orchestration: Roles such as Conductor, Worker, Verifier, and Aggregator are defined through prompts and do not require weight updates.
  • Conductor task decomposition and evidence ledger: The Conductor is responsible for task decomposition. Claims made by Agent roles carry evidence (proof), and the aggregator references the ledger to generate the report.
  • Deep verification and scope-narrowing reassignment: The Verifier checks coverage and depth, and performs targeted reassignment for subtasks with insufficient depth.
  • Zero-compression assembly: Report length has no single-call limit; the aggregator generates the framework, and each section draft is concatenated verbatim.
  • Degradation chain: Supports L2 degradation (single-model direct answer when all workers fail) and L3 degradation (marking output as unverified when verification/aggregation fails).
  • Custom cards and process context injection: Injects execution briefings through processBriefing, and provides presentCall and presentResult custom cards.
  • Web Search and cross-Provider fusion: Supports Agent roles using ctx.web.search, and supports designating different model roles across Providers.

Installation and Enablement

Installing the plugin requires using DSH’s bundle mechanism to anchor the plugin directory to a profile.

dsh plugin --profile default add file:D:\work\DeepSeekHarness\Plugins\fusion-symphony

After execution, the plugin declares dsh.bundle.patch in package.json and, combined with cordis.patch.yml, inserts an entry into the Cordis composition. Configuration can be modified in the config field of cordis.patch.yml, or overridden through tool parameters.

Key configuration items include:

  • enabled: Whether to register the tool (default true).
  • provider / model: Default Provider and model ID.
  • conductorModel / aggregatorModel: Dedicated models for Conductor and aggregator (if empty, the default model is used).
  • workerModels: Worker model list, supports cross-Provider usage, and is key to fusion.
  • preset: Preset mode (auto, fusion, hermes, fugu).
  • maxRounds / maxWorkers: Maximum verification rounds and maximum concurrent workers.

Invocation

Invoke the fusion_symphony_orchestrate tool provided by the plugin. The following is an example of typical parameters:

{
  "query": "2026 年 X 行业市场规模官方最新数据是多少？",
  "preset": "auto",
  "return_mode": "synthesized",
  "worker_models": [
    { "provider": "openrouter", "model": "anthropic/claude-...", "role": "agent" },
    { "provider": "openrouter", "model": "openai/gpt-...", "role": "agent" },
    { "provider": "openrouter", "model": "google/gemini-...", "role": "advisor" }
  ]
}

Preset semantics:
* fusion: All agents, with tools, single-round synthesis.
* hermes: All advisors, no tools, single-round advisory, returns materials.
* fugu / auto: Mixed roles, tools assigned by role, multi-round verification + reassignment.

Return modes:
* synthesized: Returns the final answer after assembly.
* materials: Returns claims from each role for the host Agent to synthesize.

Notes

  • Real-device verification: Installation-state verification passed, but real-dialog-level verification requires restarting the Web GUI to actually consume the API quota of deepseek-official.
  • Billing and limits: maxTotalToolCalls and est_cost_usd are currently approximate values and are not included in billing; maxTokensPerWorker is actually enforced by a floor of at least 16,000.
  • Conductor state: useConductor is a reserved switch. True dynamic Plan has not yet been integrated; the current behavior is static multi-worker + verification/reassignment.
  • Model fusion requirement: With a single model, the degree of fusion decreases. To demonstrate fusion advantages, different models should be listed in workerModels (models across Providers or different models under the same Provider).