dsh-moa

Single-model AI generation often falls short on ambiguous or complex engineering tasks: it makes premature assumptions, produces monolithic unvetted implementations, and suffers from single-perspective biases that a human reviewer only catches too late. dsh-moa brings the Mixture of Agents architecture natively to DeepSeek Harness, fanning a prompt out to multiple proposer models in parallel, then entrusting a frontier reasoning model to critically benchmark every candidate and promote the winning files straight into your project root.

The whole workflow is triggered with a single /moa command. When a prompt is underspecified, advisor models first compose a structured two-to-four-question clarification questionnaire so you can steer the architecture before any code is written. Once requirements are locked in, independent proposers generate files inside isolated .moa/candidate-N sandboxes on disk, eliminating cross-pollution. The judge selects the optimal implementation via a machine-readable marker, the winner is promoted to the workspace root, and temporary directories are pruned automatically. For web artifacts, dsh-moa plugs into Live Canvas to spawn a one-click browser preview. Incremental-edit mode detects an existing codebase and emits precise delta changes instead of destructive full-file rewrites, while a built-in cost tracker estimates token usage and run pricing without any external dependency.

It is built for engineers who regularly face open-ended coding briefs, perform multi-approach code reviews, or want a structured, repeatable workflow that surfaces several independent solutions and lets a reasoning model adjudicate quality before anything lands in the repository.