dsh-moa

A DeepSeek Harness plugin that adds Mixture of Agents (MoA) as an on-demand tool: run_moa.

Instead of using one model, MoA sends the same prompt to several proposer models in parallel, then has a stronger aggregator model synthesize the best final answer from all their outputs. It is not on the normal request path — the agent calls the tool only when multi-model synthesis is worth the extra tokens/latency.

Install

dsh plugin --profile web add github:morphlinglan/dsh-moa

Then restart dsh web if it is already running.

Usage

Configure the proposer pool and aggregator in your profile's cordis.patch.yml:

- insert:
    - id: dsh-moa
      name: dsh-moa
      config:
        toolName: run_moa
        # Replace with your own providers/models.
        proposers:
          - provider: proposer-provider-a
            model: proposer-model-a
          - provider: proposer-provider-b
            model: proposer-model-b
        aggregator:
          provider: aggregator-provider
          model: aggregator-model
        minProposers: 2
        proposerMaxTokens: 1500
        aggregatorMaxTokens: 2500
        fallbackLongest: true
        maxRetries: 2

The providers/models above are placeholders only. Replace them with providers and models you actually have access to.

Then ask the agent to use run_moa for complex analysis, synthesis, translation, or review tasks.

Config

Field	Type	Default	Description
toolName	string	"run_moa"	Tool name registered in DSH.
proposers	{ provider, model }[]	[]	Models that independently answer the prompt in parallel.
aggregator	{ provider, model }	required	Stronger model that synthesizes the final answer.
minProposers	number	2	Minimum successful proposers required to run aggregation.
proposerMaxTokens	number	1500	Output cap for each proposer.
aggregatorMaxTokens	number	2500	Output cap for the aggregator.
fallbackLongest	boolean	true	If the aggregator fails, fall back to the longest proposer output.
temperature	number	—	Optional sampling temperature passed to every call.
reasoningEffort	string	—	Optional adapter-owned reasoning effort id.
maxRetries	number	0	Retries per proposer/aggregator on transient errors, with exponential backoff.
iterations	number	1	Iterative MoA rounds (1 = single layer + aggregator).

License

MIT