@simplium/hive
v4.2.0
Published
Agent governance layer — 36 curated agent & skill definitions with portable permission translation, traceability and anti-fabrication discipline for Claude Code and OpenCode
Maintainers
Readme
HIVE Framework
Agent governance layer — 36 curated agent & skill definitions with a portable permission model, traceability, and evidence-based discipline. Installs into any Claude Code or OpenCode project.
What is HIVE
HIVE is an agent governance layer, not an AgentOS: the operating system is provided by the host runtime (Claude Code and OpenCode today; others via adapter generators). HIVE supplies the workforce and the policy — a capability pack of 36 definitions (18 agents + 18 skills) covering the full software development lifecycle, each with:
- an abstract permission model (
file_system/network/database/human_approval) translated to each runtime's native enforcement (Claude Code:disallowedTools; OpenCode: nativepermission: ask/allow/deny—human_approvalbecomes a real ask-gate) - tiers and
known_failure_modes— criticality-aware routing metadata - anti-fabrication discipline — evidence requirements and prompt-injection guards baked into every definition
- traceability — automatic invocation logging to
HIVE-LOG.md(ISO 42001 A.9) plus a JSONL event stream for external observability
Definitions target two stacks: Stack A (PHP 8.3 / Laravel / MySQL / Plesk) and Stack B (Next.js 15 / NestJS / Prisma / PostgreSQL).
The former Layer 2 NestJS runtime was retired per ADR-000 and removed in v4. It is preserved in the
legacy/layer2-runtimebranch.
Quickstart
From your project directory:
npx @simplium/hive .This installs 18 agents + 18 skills into the project's .claude/ and 18 OpenCode agents into .opencode/agents/ (with native permission translation and version stamps), plus the auto-logging hook (Claude Code). Open the project in Claude Code or OpenCode — everything is discovered natively (OpenCode reads skills from .claude/skills/). Re-run with @latest after a new HIVE release to update.
From source (contributors):
git clone https://github.com/marcablanca/hive-framework.git
cd hive-framework && npm install
npm run init-project -- /path/to/your-projectArchitecture
0*-*/AGENT.md · SKILL.md 36 source definitions (frontmatter + body)
│
├── npm run validate schema gate (Zod) — 0 errors required
├── npm run reindex AGENT_INDEX.yaml (JIT index, deterministic)
│
├── npm run adapters .claude/ + .opencode/agents (this repo)
└── npm run init-project same, into any consumer project
│
▼
Host runtime (Claude Code · OpenCode)
native discovery · permission enforcement · hooks (CC)
│
├── HIVE-LOG.md automatic invocation log (ISO 42001 A.9)
└── hive-events.jsonl event stream for external observabilityTier System
Agents and skills are organized into 8 tiers (00–07), from highest criticality to support roles:
| Tier | Name | Description | |------|------|-------------| | 00 | Orchestrator | Central coordinator; routes and manages all tasks | | 01 | Foundation | Architecture, product, legal/compliance, brand | | 02 | Core Development | Backend, frontend, API, database | | 03 | Quality & Security | Testing, security auditing, performance, accessibility | | 04 | Infrastructure | DevOps, cloud, billing, incident response, search | | 05 | Intelligence | AI/ML, data analysis, analytics, observability, voice | | 06 | Growth | SEO, marketing, CRO, email, mobile | | 07 | Support | Docs, releases, i18n, UX research, competitive intel |
7 agents in blocking tiers carry human_approval: true: in Claude Code this becomes disallowedTools: [Bash] (shell actions surface through the host approval flow); in OpenCode it becomes a native bash: ask / edit: ask gate.
Agents & Skills Catalog
36 total = 18 agents + 18 skills. Sourced from AGENT_INDEX.yaml.
| Tier | Focus | Agents | Skills | |------|-------|--------|--------| | 00 — Orchestrator | Task routing, coordination | orchestrator | — | | 01 — Foundation | Architecture, product, legal | architecture-planner, legal-compliance, product-manager | brand-design-system | | 02 — Core Development | Backend, frontend, API, DB | backend-developer, frontend-developer, api-designer, database-engineer | — | | 03 — Quality & Security | Tests, security, perf, a11y | test-engineer, security-auditor | web-performance, accessibility | | 04 — Infrastructure | DevOps, cloud, billing, search | billing-payments, incident-response | cloud-infrastructure, devops-engineer, search | | 05 — Intelligence | AI/ML, data, analytics, voice | ai-ml-engineer, data-analyst | analytics-implementation, observability, voice-ai | | 06 — Growth | SEO, marketing, CRO, email, mobile | — | email-deliverability, growth-analytics, landing-page-cro, marketing-communications, mobile-development, seo-aeo-geo | | 07 — Support | Docs, releases, i18n, UX, research | competitive-intelligence, cost-optimization, customer-success, ux-research | documentation-writer, release-manager, translator-i18n |
Model distribution: 7 opus / 28 sonnet / 1 haiku. JIT index reduces per-task context by 93–98% vs loading all definitions (npx tsx scripts/measure-jit.ts).
Scripts Reference
| Command | Description |
|---------|-------------|
| npm run init-project -- <dir> | Install/update Layer 1 (adapters + hook) in a consumer project |
| npm run adapters | Regenerate this repo's .claude/ adapters from sources |
| npm run validate | Validate all 36 definitions against the frontmatter schema |
| npm run reindex | Regenerate AGENT_INDEX.yaml (deterministic — no date churn) |
| npm run log-session -- --agent <n> --task "..." --outcome <O> | Manually append an invocation to HIVE-LOG.md |
| npm run cost-report | Estimated cost per agent from HIVE-LOG.md, with budget alerts |
| npm run secrets-audit | Scan definitions for hardcoded secrets |
| npm run eval | Layer 1 evals against the live API (needs ANTHROPIC_API_KEY; skips honestly without it) |
| npm test / npm run lint / npm run typecheck | Offline quality gates (also in CI) |
Golden Paths
Step-by-step walkthroughs for common workflows on each stack:
- Golden Path — Stack A (Laravel) — PHP / Laravel / Filament / MySQL
- Golden Path — Stack B (Next.js) — Next.js / NestJS / Prisma / PostgreSQL
Roadmap
| Phase | Status | Scope |
|-------|--------|-------|
| Phase 0 — Agent migration | Complete | 36 definitions in v3 YAML frontmatter format |
| Phase 1 — Core runtime | Retired | NestJS runtime built, then superseded by ADR-000; preserved in legacy/layer2-runtime |
| Phase 2 — Governance layer | Complete | Adapter generators, permission translation, traceability hook, Layer 1 evals, distribution without fork |
| Phase 3 — Ecosystem | In progress | npm package published (@simplium/hive); OpenCode adapter shipped 4.1.0 (verified live on opencode 1.17); public docs pending |
Documentation
License
Licensed under Apache 2.0.
Author
Built by José at Simplium.io.
