GooDAnDReaDY/dsh-moa
Mixture-of-Agents multi-model collaboration and synthesis engine for DeepSeek Harness.
dsh-moa brings the Mixture of Agents architecture to DeepSeek Harness through a /moa slash command: for broad or underspecified prompts, advisor models first formulate a 2–4 question clarifying questionnaire, then multiple proposer models evaluate the prompt concurrently, each writing files into its own isolated .moa/candidate-N/ sandbox. A flagship judge model benchmarks all proposals, selects the winner via a machine marker (WINNER_CANDIDATE_INDEX: N), and promotes the winning files into the project root while pruning the sandboxes. It ships 10 built-in presets (default, code-review, fast-audit, deep-architect, bug-hunter, refactor-cleanup, frontend-ui, security-audit, math-logic, creative-brainstorm) and candidate personas (minimalist, robustness, performance, tester, general). Optional extras include a pre-promotion dsh-time-machine git checkpoint (fail-closed if absent), an automated test-execution gate, a multi-judge panel with majority/highest-score/unanimous voting, composite hybrid synthesis, budget guardrails (trim/abort), a live cost counter, disk-level candidate diffing, a win-rate leaderboard, and a local Ollama/MiniPC fallback.
Install
dsh plugin --profile web add @goodandready/dsh-moaCopied verbatim from the README's '📦 Installation' section, which says to install into your DeepSeek Harness web profile, then restart the DeepSeek Harness instance and refresh the browser. Presets and model pipelines are configured in settings.yaml or through Settings → Plugins → Mixture of Agents.
Compatibility
Node 20+; DeepSeek Harness web profile (installed via dsh plugin --profile web add). It registers a /moa composer slash command, a settings card, REST endpoints under /dsh-moa/*, and /api/dsh-moa/update. No DSH version range is stated. The optional Live Canvas preview step is skipped silently unless @goodandready/dsh-live-canvas is installed in the same profile.
Details
- Repo: GooDAnDReaDY/dsh-moa
- Category: Agent Capabilities
- Stars: 5
- Version: GitHub repo GooDAnDReaDY/dsh-moa; npm package @goodandready/dsh-moa; MIT © GooDAnDReaDY; Node 20+ (README badge). The README highlights v0.2.13 for its 10 built-in presets but does not state a single current version.
FAQ
- How does the judge pick a winner and where do the files land?
- Each proposer writes files into its own .moa/candidate-N/ directory; the flagship judge compares implementations and emits WINNER_CANDIDATE_INDEX: N, then the winner's files are promoted to the workspace root and the temporary candidate directories are pruned. You can also promote a specific candidate with /moa promote <runId> <candidateIndex>.
- What protects the workspace before winner files overwrite it?
- Before promoting, dsh-moa invokes the local dsh-time-machine service to create a shadow Git checkpoint (moa-pre-promotion: candidate-N). The README says promotion is blocked (fail-closed) if dsh-time-machine is absent or unreachable, protecting uncheckpointed files, and that the check can only be overridden with force: true.
- Can I cap the cost of a run?
- Yes: budget_guard_enabled and max_budget_usd account for candidate generation, multi-judge voting and downstream synthesis overhead. Mode trim reduces the candidate pool to fit the budget, while abort cancels execution before tokens are consumed; in abort payloads all panel and candidate token metrics are preserved.
Alternatives
MkaliezZ/dsh-agentfuse-plugin · LeslieWylie/agent-loop-workflow · SnowAmberX/dsh-role-router