Appendix I: Glossary

This glossary provides precise definitions for the core terminology used throughout the Kelly handbook and OpenClaw ecosystem. It serves as a quick reference for the canonical meaning of terms as used in this knowledge base, eliminating ambiguity when terms might have different meanings in different contexts. Whether you're reading planning artifacts, debugging an issue, or writing new documentation, these definitions ensure consistent understanding.

The terms are deliberately concise — a glossary entry should clarify, not explain. For deeper treatment of any term, refer to the specific chapter or appendix that covers it in detail. The glossary is organized alphabetically rather than by category, making it easy to look up a specific term without knowing which domain it belongs to.

Key terms worth highlighting: TEA audit (Test, Evaluate, Assess) is the quality audit phase specific to the software factory pipeline — not a general term but important in this context. RALPH (Retry And Learn Protocol) is the escalation system triggered after repeated sub-agent failures, distinct from simple retry logic. Idempotent is a foundational concept — operations that produce the same result whether run once or multiple times — that underlies all reliable automation design. The distinction between Agent (an AI model instance with tool access in a session) and Sub-agent (an agent spawned by another agent for a specific delegated task) clarifies the multi-agent architecture.

The glossary excludes terms that are standard computing vocabulary (file, process, API, HTTP) and focuses exclusively on OpenClaw-specific or handbook-specific terminology where the meaning might not be immediately obvious from general knowledge.

Key Items

• **Agent** — An AI model instance with tool access, running within a session; can spawn sub-agents; the fundamental unit of autonomous action in OpenClaw
• **Context window** — The amount of text an AI model can consider at once, measured in tokens (~4 characters each); constrains how much context can be loaded and processed in a single session
• **Gate** — A quality checkpoint between pipeline phases; each gate has explicit PASS/FAIL criteria that must be met before work proceeds to the next phase
• **Handoff** — Structured transfer of work from one agent to another, including all relevant context and artifacts; explicit handoff format ensures critical information isn't lost between agents
• **Idempotent** — An operation that produces the same result whether run once or multiple times; idempotency is the foundation of reliable automation because cron jobs can run multiple times due to DST, system restarts, or retries
• **RALPH** — Retry And Learn Protocol; the escalation system triggered when sub-agents fail repeatedly (3 attempts); includes structured diagnostic information for human review
• **Skill** — A packaged behavior, available through ClawHub or defined in context files, that extends OpenClaw's capabilities beyond core tools; trigger-based activation with defined inputs, process, and outputs
• **TEA audit** — Test, Evaluate, Assess; the quality audit phase in the software factory pipeline where deliverables are verified against specification before proceeding

Related Concepts

• [[kelly-handbook-ch2-architecture]] for architecture and core concepts
• [[kelly-handbook-ch7-multi-agent]] for agent and sub-agent patterns
• [[kelly-handbook-ch6-scheduling-and-cron]] for cron and scheduling terminology
• [[karpathy-llm-wiki]] for the LLM wiki pattern this KB follows