A Copilot CLI + VS Code Copilot Chat package that implements the Multi-Agent Adversarial Convergence Protocol — a research-grade harness for multi-LLM collaboration.
A research-grade Copilot package that enforces rigorous multi-LLM collaboration. Real multi-turn debate, transparent assumptions, adversarial verification, and a "never settle" stance toward the project's north star.
New here? Read INSTALL.md for the full story: who this is for, how it works, and how to install in Copilot CLI or VS Code Copilot Chat.
| Surface | Trigger | What it does | Install path |
|---|---|---|---|
| Copilot CLI agent | /agent → convergence |
Orchestrates the full protocol | Plugin install or folder copy |
| Copilot CLI skill | Keywords: converge, align, consensus, brainstorm, discuss with other LLMs, adversarial review |
Auto-loaded by the model when these words appear | Plugin install or folder copy |
| Copilot CLI extension tools | convergence_protocol, convergence_check |
Loads the rules; checklist-validates each round | Folder copy to ~/.copilot/extensions/ (plugin install does not load extension.mjs today) |
| VS Code Copilot Chat | /converge |
Same protocol, one slash command | Copy .github/prompts/converge.prompt.md into your workspace's .github/prompts/ |
A complete install on a single machine is: clone or unzip the repo into
~/.copilot/extensions/multi-agent-convergence/ (one step gives you the agent,
skill, and extension tools), then copy the VS Code prompt into your workspace
if you also use VS Code. See INSTALL.md for details and
alternatives.
- Convergence = multi-turn adversarial discussion. Minimum 3 rounds. Never summarize and call it convergence.
- Three-agent setup. Parent + two persistent sub-agents from different model families. Reuse idle agents.
- Transparent assumptions. Every output answers: what does this mean, what was assumed, does it match the ask.
- Convergence protocol. Parent relays verbatim content between sub-agents, not summaries.
- Proposer implements, verifier tests. Never self-verify.
- Adversarial verification. "Looks good" is not acceptable. State what was checked and why you are satisfied.
- Research rigor. Check logic gaps, loopholes, whitespaces, biases, leaks, fitting, methodology.
- Goal-centered research. Define primary goal, success factors, north-star metrics, guardrails. Evaluate everything against them.
- Never settle. Push toward the north star; escalate honestly if the goal is unreachable.
- Per-decision Decision Cards. Every key decision, finding, or result ships with a Decision Card: assumptions (source + confidence + evidence), what would invalidate the result, alternatives rejected, and the exact adjustments the researcher can make.
multi-agent-convergence/
├── README.md ← this file
├── INSTALL.md ← research-grade overview + install guide
├── extension.mjs ← CLI extension (convergence_protocol, convergence_check)
├── .github/
│ ├── plugin/plugin.json ← Copilot CLI plugin manifest
│ └── prompts/converge.prompt.md ← VS Code Copilot Chat /converge slash command
├── agents/
│ └── convergence.md ← CLI agent (invokable via /agent)
└── skills/multi-agent-convergence/
├── SKILL.md ← Full protocol (10 rules)
└── templates/
├── project-goals-template.md ← North star + guardrails
├── convergence-log-template.md ← Per-round debate log
├── research-rigor-checklist.md ← 25-point rigor checklist
└── decision-assumptions-template.md ← Rule 10 Decision Card
Returns protocol text. Sections:
full, quick-reference (default), convergence-steps, research-rigor,
agent-setup, assumptions-template, decision-card.
Validates a finished round. Inputs: rounds_completed, agents_involved,
challenges_exchanged, disputes_documented, assumptions_listed,
decision_card_attached. Flags violations (e.g., fewer than 3 rounds, no Decision
Card, summarization instead of challenges).
MIT.