CLAUDE - Z-M-Huang/openhive GitHub Wiki

OpenHive Wiki

Documentation for the OpenHive AI agent orchestration platform. Version 4.6.3.

Core Architecture Invariant

Main β†’ Orchestrator β†’ Subagents β†’ Skills β†’ Plugins.

  • Main agent routes user requests and delegates to child teams. No subagents, no skills, no direct execution.
  • Team orchestrator manages subagents. Never invokes skills directly (ADR-40).
  • Subagents follow skills and own learning/reflection cycles. Each brings a different perspective (ADR-40).
  • Skills orchestrate plugins via ## Required Tools. Pure orchestration β€” no raw API calls (ADR-39).
  • Plugins are executable TypeScript tools for single external operations.

Version

4.6.3. Do not reference v3 or prior versions in page content. Version references inside ADR historical text (Context/Decision sections) are preserved as-is.

Canonical Ownership Rule

Every concept has exactly one canonical page. Other pages use 1-2 line cross-references with [PageName#Section](/Z-M-Huang/openhive/wiki/PageName#Section) syntax. Never duplicate explanatory content.

Canonical Ownership Map

Concept Canonical Page
Skill adoption flow (12-step) Skill-Repository
Skill format, plugin/skill roles Skills
Subagent format, wiring, hierarchy Subagents
Memory schema, search, injection Memory-System
Isolation mechanisms Isolation-Model
Rule cascade, format Rules-Architecture
Three-tier data model, code inventory Architecture
Execution model (daily-ops / org-ops pool + mutex) Architecture#Execution Model
Learning cycle, reflection, governance Self-Evolution
Trigger engine, overlap, circuit breaker Triggers
window trigger type (continuous watch) Triggers#window Trigger Type
config.yaml, vault, data store tree Team-Configuration
Session lifecycle, tool assembly SDK-Integration
Message flow, trust gate Request-Processing
Topic classification Conversation-Threading
Dashboard views Admin-Dashboard
Browser sessions, SSRF Browser-Proxy
Organization tools, task queue Organization-Tools
query_teams parallel fan-out Organization-Tools#query_teams β€” Parallel Fan-out
enqueue_parent_task work handoff Organization-Tools#enqueue_parent_task β€” Work Handoff
Tool selection decision framework Tool-Guidelines
Activation Decision Framework (when→which trigger) Tool-Guidelines#Activation Decision Framework
End-to-end walkthroughs Scenarios
Design principles Design-Principles
ADR decisions & rationale Architecture-Decisions
Log levels, credential scrubbing Logging
Durable state, recovery Durability-Recovery
Channel adapters, WebSocket protocol Channel-Adapters

ADR Conventions

  • Fully superseded ADRs are deleted entirely (4.6.3 clean start).
  • Partially superseded ADRs use a **Superseded:** line after Status.
  • When deleting an ADR, update the intro text and any back-references in replacement ADRs.

Formatting

The wiki serves both humans and LLMs. LLMs read these pages at session bootstrap via the rule cascade, so structure-first formatting (tables, diagrams, explicit flows) is a hard requirement β€” not a stylistic preference.

  • Mermaid diagrams are required for every execution model, state machine, and flow (sequence, flowchart, stateDiagram). Narrative prose is acceptable for non-flow explanations (concepts, rationale, definitions).
  • No ASCII art for architectural diagrams. ASCII analogies are permitted inside prose when they clarify a concept an LLM will read (e.g., the guard-on-duty analogy in Tool-Guidelines#Why window ticks feel long-running).
  • Wiki links: [PageName](/Z-M-Huang/openhive/wiki/PageName) or [PageName#Section](/Z-M-Huang/openhive/wiki/PageName#Section). Prefer section-anchored links when cross-referencing a specific canonical diagram or table.
  • Tables: GitHub-flavored markdown pipe syntax.
  • No code snippets from actual codebase β€” reference source file names.
  • Every diagram has exactly one canonical home page; other pages cross-reference it rather than re-embedding. Canonical homes are listed in the Canonical Ownership Map above.