In DeepSeek Harness (DSH) development, asking multiple models the same question often leads to an anchoring effect, in which later models are influenced by earlier answers. dsh-ai-council is a role-driven enterprise AI deliberation plugin. It creates a temporary corporate decision board, with each seat having an independent system prompt and responsibilities. Models are dynamically assigned to these roles, draft independently in the first round, and then rebut unresolved objections in subsequent rounds until evidence-based consensus is reached or the chair makes a final ruling.

Core Design

Roles first, models second. Seats such as Principal Software Architect, Security Architect, and Commercial Director are persistent role definitions. Models may be replaced between councils or mid-run, but the seats remain unchanged.

Independent first round. Members do not see each other’s opinions in the first round, reducing anchoring effects and premature convergence.

Bounded deliberation. Subsequent rounds expose previously structured positions and unresolved focus items from the chair. Members must respond to objections from peers, but are explicitly instructed not to agree merely to create consensus.

Consensus is not a majority vote. The host validates the chair’s consensus claim against the configured weighted approval threshold and authoritative role blockers. If security, data, or release roles report domain-critical issues, consensus may be blocked.

Failures replace models, not roles. A failed model route is retried through the remaining candidates for that role. The replacement model receives the same role system prompt and the current council state.

Built-in Roles and Templates

The plugin includes 20 enterprise role presets by default, such as:

  • Chair / Technical Director
  • CTO / Technology Strategy
  • Principal Software Architect
  • Staff Implementation Engineer
  • Tech Stack Specialist
  • Security Architect / CISO
  • SRE / DevOps Lead
  • Performance Engineer
  • Database / Data Architect
  • QA / Test Architect
  • UI/UX Lead
  • Product Manager
  • Commercial / Business Director
  • Finance / Cost Controller
  • Customer Support Lead
  • Compliance / Privacy Officer
  • API / Integration Architect
  • Release Engineering Lead
  • Adversarial Reviewer / Red Team
  • Future Maintainer

All built-in roles can be edited, disabled, or deleted, and built-in roles can be restored without deleting custom roles. Each role stores the name, description, member/chair type, role-specific system prompt, expertise tags, council weight, blocking permission, veto categories, model routing hints, and optional provider/model pinning.

Built-in templates include:

  • Software Architecture Review
  • Production Readiness Review
  • Product / Feature Decision
  • Database / Migration Review
  • Security & Privacy Review
  • Commercial / Pricing Board
  • API / Integration Review
  • Full Corporate Board

Templates can also be managed with create, update, delete, and query operations.

Visualized Control Room and Integration

The plugin provides a visualized control room to display the real-time decision pipeline, active councils, the current phase/round, role-to-model configuration, provider/model routing, member positions and confidence levels, authoritative blockers, chair activity, consensus progress, and the persistent event timeline.

In conversation, the deliberation process is visible. ai_council tool calls are rendered as live council cards rather than generic loaders. Background commands use the same cards.

Commands include:

  • /council [proposal]: start deliberation
  • /council-result [council-id]: view result
  • /council-history: view history

Usage Examples

Command-line Usage

/council [proposal]

Native Tool Usage

Call the ai_council tool:

ai_council(proposal="...", template="auto", role_ids=[...])

Return Structure Examples

Member return structure:

{
  "position": "approve_with_changes",
  "confidence": 0.87,
  "summary": "...",
  "blocking_objections": [],
  "important_objections": ["..."],
  "recommendations": ["..."],
  "evidence": ["..."],
  "responses_to_peers": ["..."]
}

Chair return structure:

{
  "status": "consensus",
  "consensus_reached": true,
  "consensus_score": 0.91,
  "decision": "...",
  "rationale": "...",
  "unresolved_blocking_issues": [],
  "required_changes": ["..."],
  "dissent": ["..."],
  "next_round_focus": []
}

Installation and Enablement

The plugin relies on the DSH environment to run. The specific DSH plugin installation command was not found. Please use the following links to obtain the plugin source code or directory:

  • GitHub: https://github.com/AGSQ11/dsh-ai-council
  • Plugin directory: https://www.skillhub.cn/plugins/AGSQ11/dsh-ai-council

Applicable Scenarios and Notes

The plugin is suitable for scenarios that require structured multi-party decision-making, such as architecture reviews, security reviews, and product decisions.

Notes:

  1. Consensus mechanism: Consensus is not a simple majority vote; it must pass the weighted approval threshold and authoritative role blockers.
  2. Model failure: If model routing fails during deliberation, the model is replaced rather than the role, and the role definition remains unchanged.
  3. Permission dependency: The plugin must be correctly configured in the DSH environment before use.

The plugin is released under the MIT License.

dsh-ai-council: DeepSeek Harness Enterprise AI Deliberation Plugin

In DeepSeek Harness (DSH) development, asking multiple models the same question often leads to an anchoring effect, in which later models are influenced by earlier answers. dsh-ai-council is a role-driven enterprise AI deliberation plugin. It creates a temporary corporate decision board, with each seat having an independent system prompt and responsibilities. Models are dynamically assigned to these roles, draft independently in the first round, and then rebut unresolved objections in subsequent rounds until evidence-based consensus is reached or the chair makes a final ruling.

Core Design

Roles first, models second. Seats such as Principal Software Architect, Security Architect, and Commercial Director are persistent role definitions. Models may be replaced between councils or mid-run, but the seats remain unchanged.

Independent first round. Members do not see each other’s opinions in the first round, reducing anchoring effects and premature convergence.

Bounded deliberation. Subsequent rounds expose previously structured positions and unresolved focus items from the chair. Members must respond to objections from peers, but are explicitly instructed not to agree merely to create consensus.

Consensus is not a majority vote. The host validates the chair’s consensus claim against the configured weighted approval threshold and authoritative role blockers. If security, data, or release roles report domain-critical issues, consensus may be blocked.

Failures replace models, not roles. A failed model route is retried through the remaining candidates for that role. The replacement model receives the same role system prompt and the current council state.

Built-in Roles and Templates

The plugin includes 20 enterprise role presets by default, such as:

  • Chair / Technical Director
  • CTO / Technology Strategy
  • Principal Software Architect
  • Staff Implementation Engineer
  • Tech Stack Specialist
  • Security Architect / CISO
  • SRE / DevOps Lead
  • Performance Engineer
  • Database / Data Architect
  • QA / Test Architect
  • UI/UX Lead
  • Product Manager
  • Commercial / Business Director
  • Finance / Cost Controller
  • Customer Support Lead
  • Compliance / Privacy Officer
  • API / Integration Architect
  • Release Engineering Lead
  • Adversarial Reviewer / Red Team
  • Future Maintainer

All built-in roles can be edited, disabled, or deleted, and built-in roles can be restored without deleting custom roles. Each role stores the name, description, member/chair type, role-specific system prompt, expertise tags, council weight, blocking permission, veto categories, model routing hints, and optional provider/model pinning.

Built-in templates include:

  • Software Architecture Review
  • Production Readiness Review
  • Product / Feature Decision
  • Database / Migration Review
  • Security & Privacy Review
  • Commercial / Pricing Board
  • API / Integration Review
  • Full Corporate Board

Templates can also be managed with create, update, delete, and query operations.

Visualized Control Room and Integration

The plugin provides a visualized control room to display the real-time decision pipeline, active councils, the current phase/round, role-to-model configuration, provider/model routing, member positions and confidence levels, authoritative blockers, chair activity, consensus progress, and the persistent event timeline.

In conversation, the deliberation process is visible. ai_council tool calls are rendered as live council cards rather than generic loaders. Background commands use the same cards.

Commands include:

  • /council [proposal]: start deliberation
  • /council-result [council-id]: view result
  • /council-history: view history

Usage Examples

Command-line Usage

/council [proposal]

Native Tool Usage

Call the ai_council tool:

ai_council(proposal="...", template="auto", role_ids=[...])

Return Structure Examples

Member return structure:

{
  "position": "approve_with_changes",
  "confidence": 0.87,
  "summary": "...",
  "blocking_objections": [],
  "important_objections": ["..."],
  "recommendations": ["..."],
  "evidence": ["..."],
  "responses_to_peers": ["..."]
}

Chair return structure:

{
  "status": "consensus",
  "consensus_reached": true,
  "consensus_score": 0.91,
  "decision": "...",
  "rationale": "...",
  "unresolved_blocking_issues": [],
  "required_changes": ["..."],
  "dissent": ["..."],
  "next_round_focus": []
}

Installation and Enablement

The plugin relies on the DSH environment to run. The specific DSH plugin installation command was not found. Please use the following links to obtain the plugin source code or directory:

  • GitHub: https://github.com/AGSQ11/dsh-ai-council
  • Plugin directory: https://www.skillhub.cn/plugins/AGSQ11/dsh-ai-council

Applicable Scenarios and Notes

The plugin is suitable for scenarios that require structured multi-party decision-making, such as architecture reviews, security reviews, and product decisions.

Notes:

  1. Consensus mechanism: Consensus is not a simple majority vote; it must pass the weighted approval threshold and authoritative role blockers.
  2. Model failure: If model routing fails during deliberation, the model is replaced rather than the role, and the role definition remains unchanged.
  3. Permission dependency: The plugin must be correctly configured in the DSH environment before use.

The plugin is released under the MIT License.