~/projects — claude /consensus-run

Multi-Agent Consensus · v0.2.0 · MIT · Hard-gate

Three agents must agree.
Or the gate stays shut.

Multi-Agent Consensus is a Claude Code plugin where Lead, Alpha, and Bravo independently validate every gate. Unanimity or nothing — 2-of-3 does not pass. Built from a real session where a single-agent reviewer explicitly approved a P2 streaming bug; three independent eyes catch what one set misses.

View on GitHub →Install ↓Read the launch post ↗

Agents

Commands

Skill

3/3

Quorum

Mocks

01What it does

One artifact. Three eyes. Zero majorities.

Multi-Agent Consensus is not majority voting. It is not weighted scoring. It is unanimity at every phase gate. The producer never reviews their own work, and a 2-of-3 PASS is treated as a failure to converge.

Independent verification

Three agents — Lead, Alpha, Bravo — each start fresh with no shared findings. They validate the same artifact in isolation, then submit votes. Independence is structural, not advisory.

Hard gates only

2/3 does not pass. Unanimity or nothing. The day a single-agent reviewer approved a P2 streaming bug is the day this plugin was born.

Re-validate after fixes

When a vote fails, the fix triggers all three agents to re-run — not just the dissenter. Drift between agents is the bug; closing the loop is the gate.

02Quick start

Add marketplace. Run. Read three votes.

Two distribution channels — a Claude Code plugin (interactive, native slash commands) and a pipx CLI (CI / non-interactive). Pick one; both produce identical evidence on the same target.

bashexample 1 · Claude Code plugin

# Install via the Claude Code marketplace.
$ claude plugin marketplace add krzemienski/multi-agent-consensus
$ claude plugin install multi-agent-consensus@multi-agent-consensus

# Run the full consensus pipeline against a project.
$ /consensus-run --target ./my-project --phases "explore,audit,fix,verify"
# Phase explore  → Lead, Alpha, Bravo each run independently
#                 evidence/explore/{lead,alpha,bravo}/*.{json,log}
# Phase audit    → 3 votes merged into evidence/audit/verdict.json
verdict: PASS   Lead: PASS · Alpha: PASS · Bravo: PASS
verdict: PASS   Lead: PASS · Alpha: PASS · Bravo: PASS
verdict: BLOCK  Lead: PASS · Alpha: PASS · Bravo: FAIL — "P2 streaming bug at line 42"

# Inspect the dissent. Fix the real system. Re-run.
$ /consensus-report --phase audit

bashexample 2 · pipx CLI for CI

# CI-friendly install — no interactive Claude Code session needed.
$ pipx install git+https://github.com/krzemienski/multi-agent-consensus.git

# Same model prompts, same evidence schema, exit code reflects verdict.
$ consensus run --target ./my-project --phases "explore,audit,fix,verify"

# Single-gate validation — useful when you only need to re-check one phase.
$ consensus validate --target ./my-project --phase audit

# View the JSON-schema-validated verdict from the last run.
$ consensus show-config
→ phases: 4   ·   gates: 4   ·   quorum: 3/3   ·   evidence_root: ./e2e-evidence

03Visual explainer

Three independent paths. One gate.

Each phase ends at a checkpoint where Lead, Alpha, and Bravo each cast a vote. The diagram below shows the structural shape: parallel independent validation, then merge. 2-of-3 does not unlock the gate — only 3-of-3.

04How it works

Three roles. One protocol.

Each role is a real subagent definition under ./agents/{lead,alpha,bravo}.md. Each owns its own evidence subdirectory and votes blind. The Lead authors the merged report only after Alpha and Bravo have submitted.

Lead

Orchestrates the run, owns the merged verdict, and authors the consensus report. Lead never breaks ties — only Lead+Alpha+Bravo agreeing breaks the gate.

Alpha

Independent validator. Reads the artifact cold, captures evidence to its own subdirectory, votes. Never sees Bravo's findings until after submission.

Bravo

Independent validator. Same protocol as Alpha; different fresh context. Two independent eyes on the same artifact catch what one set misses.

The five commands

Five slash commands cover the full lifecycle — from running a pipeline to inspecting a single gate’s verdict. Every command writes evidence to the configured root and emits a JSON verdict against the public schema.

Command

Purpose

/consensus-run

Run the full consensus pipeline across phases

/consensus-validate

Single-gate validation against current target

/consensus-roles

Print the active Lead/Alpha/Bravo agent roster

/consensus-report

Render the consensus verdict report from the last run

/consensus-show-config

Inspect the configured phases, gates, and evidence paths

05Feature matrix

What ships in v0.2.0.

Counts grounded in ~/Desktop/blog-series/multi-agent-consensus/ — the canonical plugin tree. Two distribution channels, identical evidence shape on the same target.

Two channels. Same evidence.

Run as a Claude Code plugin for interactive sessions, or as a pipx CLI for CI. Both produce the same JSON verdicts, the same per-agent evidence directories, and the same exit-code semantics.

bashinstall — choose one

# Channel A — Claude Code plugin (recommended for interactive use).
$ claude plugin marketplace add krzemienski/multi-agent-consensus
$ claude plugin install multi-agent-consensus@multi-agent-consensus

# Channel B — pipx CLI (recommended for CI / scripts).
$ pipx install git+https://github.com/krzemienski/multi-agent-consensus.git

# Verify install.
$ /consensus-roles
Lead    · agents/lead.md    loaded
Alpha   · agents/alpha.md   loaded
Bravo   · agents/bravo.md   loaded

# First run.
$ /consensus-run --target . --phases audit

07Documentation

Read the protocol. Then run it.

Every agent definition, command spec, vote schema, and example is open source. Below: the eight pages worth bookmarking on day one.

01 · Start

Install & First Run

02 · Run

/consensus-run end-to-end

03 · Validate

/consensus-validate single gate

consensus-pipeline skill

07 · Examples

Worked examples

08 · Schema

Vote/Verdict JSON schema

A 2-of-3 PASS is not a PASS. It is the shape of disagreement the gate is designed to surface.

08Changelog

From single-validator to 3-of-3 quorum.

Multi-Agent Consensus has shipped on monthly minor versions since launch. Major beats: independent evidence per validator (v0.1), plugin distribution (v0.2).

v0.2.0

Apr 2026

Plugin distribution. Native /consensus-* slash commands. Inherited Claude Code auth. Task-tool agent dispatch.

v0.1.0

Feb 2026

Initial release. Three-agent gate, vote-merge protocol, evidence isolation per agent, JSON verdict schema. Built from catching the P2 streaming bug.

Three agents must agree. Or the gate stays shut.