open-conclave

Multi-agent debate orchestrator plugin for OpenCode. Runs parallel sub-agents (Harper, Benjamin, Lucas), moderates debate rounds with a Captain, and applies early stopping when consensus is reached.

How it works

1. You select the Conclave agent tab in OpenCode and ask a question
2. Three sub-agents run in parallel — each from their own perspective:

• Harper — Research & Facts
• Benjamin — Logic, Math & Code
• Lucas — Creative & Alternative Perspectives

3. The Captain critiques all responses and scores consensus
4. Debate continues for up to 3 rounds; stops early when consensus ≥ 0.83
5. The Captain synthesizes a final answer from the full debate

Installation

Add open-conclave to the plugin array in ~/.config/opencode/opencode.json. OpenCode installs it automatically on next startup — no separate install command needed.

{
  "plugin": ["open-conclave"]
}

You can also pin a version or use a git URL:

{
  "plugin": [
    "[email protected]",
    "open-conclave@git+https://github.com/martinzokov/open-conclave.git"
  ]
}

Local / development

git clone https://github.com/martinzokov/open-conclave.git
cd open-conclave
bun install
bun run build
cp dist/index.js ~/.config/opencode/plugins/conclave.js

Then register by file path:

{
  "plugin": ["/Users/you/.config/opencode/plugins/conclave.js"]
}

Configuration

Conclave uses whichever model you have active in OpenCode as the default for all agents. You can override each agent individually in your OpenCode config:

{
  "plugins": ["open-conclave"],
  "agent": {
    "conclave-captain": {
      "model": "github-copilot/claude-sonnet-4.5"
    },
    "conclave-harper": {
      "model": "github-copilot/gpt-4.1"
    },
    "conclave-benjamin": {
      "model": "github-copilot/claude-sonnet-4.5"
    },
    "conclave-lucas": {
      "model": "github-copilot/gpt-4.1"
    }
  }
}

Any provider/model string that OpenCode supports works here (e.g. anthropic/claude-opus-4-6, openai/gpt-4o, github-copilot/claude-sonnet-4.5).

Default agent behaviour

Captain (conclave-captain)

Expert debate moderator and synthesizer. Handles three internal tasks:

• CRITIQUE — after each round, scores consensus (0.0–1.0), measures uncertainty improvement, and lists up to 2 open issues. Stops the debate early when consensus ≥ 0.83.
• SYNTHESIZE — produces the final answer from the full debate history. This is what the user sees.

The Captain's persona controls the language and tone of the final output.

Harper (conclave-harper)

Research & Facts — provides accurate, well-sourced factual analysis. Cites evidence where possible. Focuses on empirical data, historical context, verified claims, and source reliability.

Benjamin (conclave-benjamin)

Logic, Math & Code — provides rigorous logical analysis, mathematical reasoning, and technical evaluation. Focuses on formal correctness, algorithmic thinking, edge cases, and code quality.

Lucas (conclave-lucas)

Creative & Alternative Perspectives — challenges assumptions, offers creative solutions, and considers user experience. Focuses on unconventional approaches, human impact, design thinking, and unexplored angles.

Sub-agent response format

Each sub-agent always responds with a structured JSON object:

{
  "agentName": "Harper",
  "role": "Research & Facts",
  "claims": [{ "text": "claim text", "confidence": 0.9 }],
  "reasoning": "brief internal reasoning (≤40 words)",
  "uncertainties": ["any unknowns or caveats"],
  "answer": "direct answer to the query (100–150 words)"
}

The format section of the system prompt is always preserved — persona overrides only replace the identity/role/focus section, so the JSON output is never broken by a custom persona.

Customising agent personas

Each agent has two parts to its system prompt:

1. Persona — identity, role, and focus instructions (customisable via persona)
2. Format — JSON schema requirements for sub-agents (always preserved)

Use the persona field in your opencode.json to replace the identity and focus of any agent. The JSON output format for sub-agents is never affected, so responses always parse correctly regardless of persona overrides.

The persona field applies to all five agents:

| Agent | Key | Role | |---|---|---| | Captain | conclave-captain | Moderator and final synthesizer | | Harper | conclave-harper | Research & Facts | | Benjamin | conclave-benjamin | Logic, Math & Code | | Lucas | conclave-lucas | Creative & Alternative Perspectives |

How persona is applied: OpenCode stores persona in multiple config fields internally. The plugin intercepts the config hook and merges your persona string into the full system prompt (persona section + format section), updating all internal fields consistently. This means the persona works reliably regardless of OpenCode's internal config processing.

The final output language is controlled by the Captain's persona — set persona on conclave-captain to affect the synthesized answer. Setting it only on sub-agents changes how they reason internally but the Captain will synthesize in its default language.

{
  "agent": {
    "conclave-captain": {
      "persona": "You are the Captain. You only respond in French."
    },
    "conclave-harper": {
      "persona": "You are Harper, a financial analyst specialising in markets and macroeconomics.\n\nYour role: analyse economic trends, market data, and investment implications.\nFocus on: macroeconomic indicators, sector performance, risk factors, and data-driven forecasts.\nKeep \"reasoning\" under 40 words. Write a thorough \"answer\" — aim for 100–150 words."
    },
    "conclave-benjamin": {
      "persona": "You are Benjamin, a security researcher and penetration tester.\n\nYour role: evaluate systems for vulnerabilities, attack surfaces, and defence weaknesses.\nFocus on: threat modelling, CVEs, exploit chains, and hardening recommendations.\nKeep \"reasoning\" under 40 words. Write a thorough \"answer\" — aim for 100–150 words."
    },
    "conclave-lucas": {
      "persona": "You are Lucas, a climate scientist and sustainability strategist.\n\nYour role: consider environmental impact, long-term sustainability, and ecological trade-offs.\nFocus on: carbon footprint, resource consumption, systemic risks, and green alternatives.\nKeep \"reasoning\" under 40 words. Write a thorough \"answer\" — aim for 100–150 words."
    }
  }
}

The default persona for each agent is used when no override is set.

Usage

TUI (interactive)

1. Open OpenCode in your project directory
2. Press Tab to select the Conclave agent
3. Type your question and press Enter

The debate runs automatically and the final synthesized answer is returned.

CLI (non-interactive)

opencode run --agent conclave -m github-copilot/claude-sonnet-4.5 "What is the best database for read-heavy workloads?"

Slash command (from any agent)

/conclave What are the tradeoffs between microservices and a monolith?

Tool options

When calling the conclave tool directly, three optional arguments are available:

| Argument | Type | Default | Description | | ----------- | ------- | -------- | -------------------------------------------------- | | query | string | required | The question or task to deliberate on | | maxRounds | number | 3 | Max debate rounds (1–10) | | debug | boolean | false | Return full debate transcript alongside the answer |

Debug mode example:

/conclave --debug Should we rewrite this service in Rust?

Debugging

To see per-step timing during a run:

opencode run --agent conclave -m github-copilot/claude-sonnet-4.5 --print-logs "Your question" 2>&1 | grep '\[conclave\]'

Output looks like:

[conclave] start  model=github-copilot/claude-sonnet-4.5  maxRounds=3
[conclave] round 1  sub-agents start  count=3
[conclave] round 1  conclave-harper   done  12.1s
[conclave] round 1  conclave-benjamin done  14.3s
[conclave] round 1  conclave-lucas    done  15.2s
[conclave] round 1  sub-agents done   15.2s
[conclave] round 1  critique   done   6.8s
[conclave] round 1  stop=true  consensus=0.91  reason=consensus_reached
[conclave] synthesize done   11.4s
[conclave] total      40.0s  rounds=1

Development

bun install          # install dependencies
bun run build        # build dist/index.js
bun test             # run tests
bun run lint         # lint
bun run lint:fix     # fix lint issues
bun run format       # format with Prettier

To test locally against a live OpenCode install:

bun run build && cp dist/index.js .opencode/plugins/conclave.js
opencode run --agent conclave -m <provider/model> --print-logs "test query"

License

MIT — see LICENSE.