senate

senate

Multi-agent debate skills for coding CLIs. Orchestrates codex, gemini, cursor, kimi, and claude through structured debate formats — parliament, court, red-team, peer-review, committee, brainstorm — to reach more robust answers than any single model.

Background

Multi-agent debate is a well-studied technique for improving LLM reasoning: independent agents propose, critique, and revise answers under a structured protocol. Results are protocol-dependent — strong single-agent prompting can match it on some benchmarks — but a substantial body of work reports gains in factuality, divergent thinking, evaluation quality, and truthfulness.

senate ports the protocols humans already use to coordinate disagreement — parliaments, courts, peer review, RFCs — and packages them as agent skills you can run across heterogeneous CLIs. See dev/PRODUCT.md for the full thesis.

Install

Prerequisite: the CLIs you want to put in debates must be installed and authenticated locally — senate shells out to them. Each skills/invoke-agent/references/<cli>.md has a paste-ready install check for codex, gemini, cursor, kimi, and claude.

This repo ships as a Claude Code plugin and as a cross-agent bundle via the skills CLI (works with most coding agents that load skills — Claude Code, Codex, Cursor, OpenCode, Gemini CLI, …).

# Claude Code plugin
/plugin marketplace add SebastianElvis/senate
/plugin install senate@senate

# Any host agent
npx skills add SebastianElvis/senate

Useful flags for the skills CLI:

Option	Description
-g, --global	Install globally (~/<agent>/skills/) instead of the current project
-a, --agent <name...>	Target specific agents (claude-code, codex, cursor, opencode, …)
-s, --skill <name...>	Install a subset (--skill senate)
-l, --list	List skills without installing
-y, --yes	Skip prompts

Other commands: npx skills list | find <q> | update senate | remove senate. Source forms accept GitHub shorthand, full URLs, git URLs, or local paths.

Usage

Ask your host agent for a debate in plain language:

• "Run a parliament between codex, gemini, and kimi on whether to migrate this service to Rust."
• "Hold a court debate — codex prosecutes my refactor, claude defends, gemini judges."
• "Committee of three models drafts this API design."
• "Red-team this deployment plan."
• "Peer-review this design doc."
• "Run a draft-review-finalize pipeline on this spec." (multi-stage)
• "Which format should I use?" (the planner recommends one without running)

Run artifacts land in <cwd>/.senate/runs/<id>/ — never in this skill repo. End-to-end walk-throughs of the headline cases live in examples/.

Architecture

Five skills compose one debate lifecycle:

              user request
                   │
                   ▼
         ┌───────────────────┐
         │      senate       │   mints .senate/runs/<id>/
         │   (orchestrator)  │
         └─────────┬─────────┘
                   ▼
         ┌───────────────────┐
         │   debate-agenda   │ ──▶ agenda.md
         │     (planner)     │
         └─────────┬─────────┘
                   ▼
         ┌───────────────────┐  dispatches   ┌──────────────────┐
         │  moderate-debate  │ ─────────────▶│ per-turn subagent│
         │    (moderator)    │ ◀─────────────│ + invoke-agent   │
         └─────────┬─────────┘  result       └────────┬─────────┘
                   │ appends                          │ shells out
                   │   ▶ transcript.jsonl             ▼
                   │   ▶ context.md          codex · gemini · cursor
                   │   ▶ agents/<cli>.md     kimi  · claude
                   ▼
         ┌───────────────────┐ ──▶ notes.md
         │   meeting-note    │
         │     (scribe)      │
         └───────────────────┘

moderate-debate dispatches every turn into a fresh per-turn subagent that loads the relevant invoke-agent playbook, shells out to the CLI, validates the contract, and returns only a structured result. Multi-stage pipelines are expanded once by the planner into a single agenda.md; the moderator then runs each stage under stages/<N>-<name>/, calling back to the planner only for clarification or mid-run re-planning. meeting-note consolidates after the final stage.