Glossary

Shared vocabulary for Stella and agentic coding. Keep entries short — one or two sentences — and alphabetized. Link to the pillar page for depth.

Term	Definition
ADR	Architecture Decision Record — see Decision Records.
Agent loop	The orchestrator cycle of model call → tool execution → observation, repeated until the task completes.
Context window	The token budget available to the model per call; managed by the context manager (Agent Design).
Eval	An automated test of end-to-end agent behavior against a graded task suite (Agent Design).
Sandbox	The isolated environment where Stella executes tools (shell, file edits) safely.
Tool	A structured capability exposed to the model (read file, run command, search); defined in Agent Design.
Turn	One user message plus the full agent loop it triggers, ending in a response.

Add new terms as they enter team usage — if you had to explain a word twice, it belongs here.