leonardo

Arbiter

A2A-native agent consensus framework — 5 packs, 26 skills, zero inherited bloat.

Arbiter is an A2A-native agent consensus framework delivered as a curated collection of Claude Code skill packs. Agents design, reconcile, ratify, and audit — with filesystem durability as the floor and JSON-RPC/SSE as the live wire.

Originally derived from an MIT-licensed upstream, inherited plugins were removed on 2026-05-16. Arbiter now ships only original contributions.

---

Plugin Packs

Pack	Category	Skills	Description
autonomous-ai-agents v0.4.0	ai-agency	3 + 2 MCP bridges	Fleet coordination: identity, hermes-bridge, openclaw-bridge
grill-each-other v1.3.1	skill-enhancers	10	Dialectic claim discipline: grill-me, peer-grill, fleet-ratify, permutation, etc.
athenaeum v0.2.0	skill-enhancers	4	Streamlined dialectic: design, reconcile, ratify, audit
leonardo v1.1.0	ai-agency	1	Protected-string decoder with audit trail
pocock-engineering v1.0.0	skill-enhancers	8	SDLC skills forked from Matt Pocock's framework

---

Skills by Pack

autonomous-ai-agents (3 skills + 2 MCP bridges)

Skill	What it does
agent-identity	Declares who an agent is — name, role, capabilities, boundaries
hermes-bridge	MCP bridge to Hermes messaging relay
openclaw-bridge	MCP bridge to OpenClaw dispatch system

grill-each-other (10 skills)

Skill	What it does
grill-me	Interview the user about a plan until shared understanding
grill-me-agents	Grill multi-agent designs — roles, handoffs, failure modes
grill-me-with-agents	Code-aware variant that cross-references existing agent definitions
grill-with-docs	Grill against domain model and documented decisions
peer-grill	Two agents interrogate each other to converge on shared state
agent-show-and-tell	Agents write status reports; one reader collates a roundup
fleet-ratify	Ratify fleet decisions with SHA-256 attestation
permutation	NxN fleet topology ratification with visual diagrams
caveman	Ultra-compressed communication mode (75% token reduction)
find-skills	Discover and install agent skills

athenaeum (4 skills)

Skill	What it does
design	Propose a design claim for peer review
reconcile	Resolve conflicting claims into a merged design
ratify	Sign off on a finalized design with attestation
audit	13-branch audit protocol with confidence rules

leonardo (1 skill)

Skill	What it does
protected-string-decoder	Decode protected strings with audit trail

pocock-engineering (8 skills)

Skill	What it does
zoom-out	Step back for broader architectural context
diagnose	Reproduce → minimize → hypothesize → instrument → fix → regression-test
triage	State-machine issue triage with role-based workflows
tdd	Test-driven development with red-green-refactor
to-issues	Break plans into tracer-bullet vertical-slice issues
to-prd	Turn conversation context into a PRD on the issue tracker
improve-codebase-architecture	Find refactoring opportunities informed by domain language
setup-matt-pocock-skills	Bootstrap AGENTS.md with issue tracker, triage labels, domain docs

---

Install

# Install a pack
/plugin install grill-each-other@arbiter
/plugin install athenaeum@arbiter

Or install individual skills by copying skills/<name>/SKILL.md into your own project.

---

Architecture

All skills follow the 2025 schema — name + description as the portable minimum, with allowed-tools, version, and package manifests as optional adapter fields.

See AGENTS.md for the full developer guide, schema details, and GitNexus integration.

---

A2A Native

Arbiter speaks the Agent-to-Agent Protocol natively:

• Agent Cards — agent_card.py generates capability descriptors with Athenaeum extensions
• A2A Tasks — a2a_task.py serializes design/reconcile/ratify/audit workflows to JSON
• JSON-RPC endpoint — tasks/send, tasks/get, tasks/cancel on port 18765
• SSE streaming — tasks/sendSubscribe pushes live task updates until terminal state
• Filesystem floor — athenaeum poll <topic> for zero-server async status checks

A2A is opt-in. Default mode is filesystem-only: