Hive: Hub & Spoke
AI Agent Orchestration for Open Source
From brainstorming to full autonomy and everything in-between.
A fully customizable fleet of AI agents covering every level of project maintenance.
hive.kubestellar.io
Hub & Spoke Architecture
Hub (hive.kubestellar.io) - central registry, cross-hive leaderboard, contributor directory
Spoke - independent hive instance with its own agents, repos, and GitHub App
| Hub | Spoke | |
|---|---|---|
| Role | Registry, federation, leaderboard | Runs agents, manages repos |
| Auth | Contributor directory | Own GitHub App + scoped tokens |
| Data | Aggregated metrics | Local beads, wiki, snapshots |
- Spokes register with hub, send heartbeats, share contributor metrics
- Each spoke is fully independent - hub coordinates, never controls
- Contributors can register with multiple spokes across the network
Agents & the Governor
10 built-in agents, each with a lane, a policy, and boundaries. Custom agents can be added easily.
| Agent | Role |
|---|---|
| Supervisor | Prioritizes, dispatches work |
| Scanner | Inbound triage, bug fixes |
| Quality | Test coverage, code quality |
| CI-Maintainer | Build health, workflows |
| Sec-Check | Security screening |
| Agent | Role |
|---|---|
| Architect | RFCs, refactors |
| Strategist | Cross-agent coordination |
| Outreach | Community, ADOPTERS |
| Guide | Repo setup advice (L1) |
| Brainstorm | Ideas to knowledge (L1) |
Governor adapts cadence based on queue depth (thresholds fully customizable):
| SURGE (> 20) | BUSY (10-20) | QUIET (2-10) | IDLE (<= 2) |
|---|---|---|---|
| Scanner only, all-hands triage | Scanner + Reviewer, reduced | All agents active | Architect + Outreach, deep work |
ACMM Levels & the Proxy
The AI-native Codebase Maturity Model - 6 levels from brainstorming to fully autonomous.
The ACMM Proxy sits between agents and GitHub, enforcing level constraints and preventing rogue behavior.
| Level | Name | Agents | Allowed Actions |
|---|---|---|---|
| L1 | Inception | 2 | Advisory only. Guide + Brainstorm. |
| L2 | Advisory | 5 | Advisory beads on dashboard. No GitHub writes. |
| L3 | Quality-Gated | 6 | Quality opens hold-gated PRs. CI gates begin. |
| L4 | Security-Aware | 7 | Agents open issues. Sec-check joins. |
| L5 | Semi-Autonomous | 9 | PRs with hold label. Human batch-review. |
| L6 | Autonomous | 10 | Auto-merge on green CI. Full fleet. |
Agents cannot bypass their level - the proxy blocks unauthorized PRs, issues, and merges.
Paper: The AI Codebase Maturity Model
Level 1: Inception (Brainstorming) & Spec-Kit
Inception - onboarding a project into Hive
| Phase | What Happens |
|---|---|
| Capture | Provide the raw idea or existing repo URL |
| Clarify | Brainstorm generates 5-7 questions; you answer |
| Structure | Creates knowledge base facts (vision, requirements, constraints) |
| Scaffold | Generates bootstrap files (README, CLAUDE.md, CI, tests) |
| Complete | Inception done, agents ready |
Two modes: greenfield (new idea, no repo) or brownfield (existing repo, extract facts).
Spec-Kit - structured artifacts generated at L1
specs/CONSTITUTION.md - immutable principles | specs/SPEC.md - requirements
specs/PLAN.md - implementation steps | specs/TASKS.md - task breakdown
These artifacts bootstrap the LLM-Wiki knowledge layer.
LLM-Wiki: The Knowledge Layer
A 4-layer wiki (powered by llm-wiki) giving agents shared memory.
Injected as markdown into every agent's work order from L2 onward.
| Layer | Scope | Source |
|---|---|---|
| Personal | Private | User edits |
| Project | Repo | Auto-extracted from merged PRs |
| Org | Team | Auto-promoted (confidence >= 0.9) |
| Community | Public | PR-based contributions |
Precedence: personal > project > org > community
Fact types evolve as the project matures
| Ideation (L1) | Operational (L2+) | Evolution |
|---|---|---|
| idea, vision, constitution | pattern, gotcha, decision | acceptance -> test_scaffold |
| requirement, constraint | regression, integration | constraint -> gotcha |
| stakeholder, acceptance | test_scaffold, coverage_rule | constitution -> decision |
Beads & Advisory Mode
Beads - the coordination ledger
A git-backed work tracker that keeps agents from stepping on each other.
| Field | Values |
|---|---|
| type | bug, feature, task, advisory |
| status | open, in_progress, blocked, done, closed |
| priority | 0 (critical) to 4 (minor) |
| actor | scanner, quality, architect, ... |
| metadata | pr_ref, test_status, contributor_engaged |
Agents claim beads before work. Skip what peers claimed. Dashboard shows real-time progress.
Advisory Mode (L1-L3)
- All findings go to beads, not GitHub - proxy blocks writes
- Governor posts a single Advisory Digest Issue consolidating all agent findings
- Operators review and decide what to act on - system proposes, humans dispose
Relay, Snapshots & Leaderboard
Relay - contributors donate compute to hives they care about
- Browse hives
- Authenticate with GitHub
- WebSocket relay connects CLI to hive
- Hive assigns tasks
- Report back, get next
Credentials never leave your machine. Relay is scoped to the contributor's GitHub permissions for each repo.
Trust tiers (auto-promote): Newcomer -> Contributor (5 tasks) -> Trusted (20 + vouched)
Dashboard Snapshots
- 15-min intervals, 30-day retention (2,880 data points)
- Governor mode, agent states, token usage, CI health
- Published as self-contained HTML pages
Cross-Hive Leaderboard
- Hub aggregates contributor metrics across all registered hives
- Tracks tasks completed, trust tier, last active - sorted by contribution
- Contributors build reputation across the entire Hive network
Hive: Enterprise Scale
Hive supports self-hosted LLM inference on Kubernetes GPU clusters.
- vLLM - high-throughput serving on GPU (or llama.cpp on CPU)
- llm-d - Kubernetes-native inference routing via InferencePool + EPP
- API Proxy - translates Anthropic Messages API to OpenAI Chat Completions
- ACP - agent-to-agent communication protocol
| Docker/Kubernetes | Kubernetes + vLLM + llm-d | |
|---|---|---|
| Inference | Cloud API | Self-hosted GPU clusters |
| Scale | 1 hive, ~10 repos | Hundreds of hives, thousands of repos |
| Cost | Per-token billing | Fixed GPU fleet, amortized |
| Privacy | Code sent to provider | Code stays on-prem |