Glossary

Short definitions. Each links to the reference, the single source of truth.

Term	Meaning
A2A	Agent2Agent protocol. Agents served by a provider gateway such as LiteLLM, referenced as a2a:<provider>:<agent>. See agents.
Agent	A role-based AI agent with a model alias, instructions, skills, tools and guardrails. Local agents are Markdown files and are referenced as local:<id>.
Agent catalogue	The list of agents a workflow can use: local agents plus one section per A2A provider.
Alias (model)	A name such as agent-chat that agents use instead of a model tag. LiteLLM maps it to an upstream model.
Approval gate	A human decision point inside a run: accept, deny or accept with comments. See approval gates.
Audit trail	Recorded policy decisions and approval decisions, stored as files in a Docker volume.
Checkpoint	A commit made on the run's branch after each agent finishes.
Fail closed	When a control cannot decide, such as OPA unreachable or a missing guardrail, the action is blocked rather than allowed.
Guardrail	A check on LLM input and output, such as pii-protection.
LiteLLM	The gateway between agents and models. It also serves agents over A2A and fronts the Git MCP server.
MCP	Model Context Protocol. How agents call tools such as the Git tools.
OPA	Open Policy Agent. Evaluates the Rego policy for workflow starts, agents and every tool call.
Policy	OPA rules in infrastructure/opa/policies/. See policy enforcement.
Project	A software project with its Git repository and credentials.
Skill	A reusable instruction package for agents.
Spec	Markdown specification with requirements and acceptance criteria, stored in the project's specs/ folder. See spec-driven development.
Spec queue	A per-project queue that runs specs one at a time.
Tool	A capability an agent may call, granted server-side and checked on every call.
Workflow	An ordered chain of agents, optionally with approval gates.