dsh-council: A Multi-Model Deliberation Flow for DeepSeek Harness
a1exsun/dsh-council
Adds multi-model deliberation to DeepSeek Harness: models answer independently, review each other anonymously, and an arbiter delivers an inspectable final decision.
dsh-council is a DSH plugin that adds multi-model deliberation to DeepSeek Harness, splitting one question into independent answers, anonymous peer review, and a final arbitration. It targets the weakness of a lone model grading its own homework by having several models answer in parallel, then letting reviewers compare those anonymous answers, spot gaps, and rank them before an arbiter weighs everything into a recommendation. Answerers, reviewers, and the arbiter are picked per run, and the output keeps confidence notes, consensus, disagreements, blind spots, and average rankings, with original answers and reviews inspectable in child sessions while reusing your configured providers and credentials.
How to Install
dsh plugin --profile web add @a1exsun/dsh-council- Category
- Development & Operations
- Platform
- DSH Plugin
- Author
- a1exsun
- Distribution
- Plugin
dsh-council Key Features
dsh-council Repo Summary
What Does It Do?
DSH Council is a DSH plugin for DeepSeek Harness that turns a single question into a full deliberation cycle: independent answers, anonymous peer review, and one reasoned decision. It solves the problem of a lone model grading its own homework by having 2–8 models answer in parallel, 1–8 reviewers compare those anonymous answers and rank them, and a single arbiter weigh everything into a final recommendation. It reuses the providers and credentials you already configured in DSH, and every result stays inspectable.
Core Features
- Three-stage deliberation: answer, review, and decide run in sequence, with models chosen separately for each role and allowed to be reused across roles.
- Inspectable output: confidence notes, consensus, disagreements, blind spots, model identities, and average rankings, while original answers and reviews remain in DSH child sessions.
- Adjustable panel: up to 17 model participants per run, with answerers, reviewers, and the arbiter selected per run.
- Bilingual by default: English and Chinese follow DSH's language setting.
- Configurable limits: per-stage token caps and timeouts can be tuned in the DSH profile, and the defaults work without extra configuration.
How to Use This Plugin?
Enter /council in a DSH Web conversation to start a run. Pick your answerers, reviewers, and arbiter first, then ask your question — every participant starts with fresh context, so include everything it needs, for example comparing two task-leasing designs and asking how each recovers when a worker crashes after an external side effect but before acknowledging the job. Keep in mind that a run involves 4–17 model participants, each possibly making multiple requests and using Web tools, so cost rises accordingly; your question and intermediate answers go to the selected providers; and model agreement does not guarantee correctness, so review the evidence and confidence notes. Defaults need no changes, but you can set token limits and millisecond timeouts on the plugin's entry in your DSH profile.
How to Troubleshoot This Plugin?
Before running, make sure at least two models are configured and authenticated, otherwise no valid panel can be formed. If you changed the timeout settings, keep the overall run timeout at least as long as the child timeout, or the configuration will not behave as expected. The plugin relies on a Node.js and package-manager environment, so upgrade per the official instructions if versions do not match. If DSH Web is already running, restart the Web interface after enabling the plugin so the new capability loads.
This page is an independent rewrite of the plugin's official README — for authoritative documentation and the latest changes, refer to the source: a1exsun/dsh-council. The plugin is third-party code that runs on your machine once installed; inclusion does not imply endorsement — please review the source before installing.