10. Glossary

Key terms used across this section and the rest of the documentation.

Term	Meaning
Agent	A mission, examples and the tools it may reach, written as data: an [[agents]] entry of the apollia.toml of a package, or an agent created in the application. The engine built into the runtime runs it in a ReAct loop, under the runtime's governance. No Python stands for an agent.
Package	A folder with an apollia.toml at its root, distributed as a .apo archive: any combination of agents, tools and workflows. A package with no agent is valid.
Tool	Something an agent may be granted and call: a native tool of the runtime, an MCP tool (mcp:<server>/<tool>), or a Python tool of a package (pkg:<package>/<tool>), an async function under the @tool decorator.
Workflow	An async Python function of a package under the @workflow decorator, whose steps are declared ahead and which may run agents, run other workflows and ask the user a question. Addressed as <package>/<workflow>.
Skill	A typed, invocable capability addressed by a skill_id on the A2A path. No decorator of the SDK declares one: the Python of a package brings tools and workflows.
A2A	Agent-to-agent invocation: an agent calling another agent's skill by skill_id, with guards for depth, self-call, timeout, and chain deadline.
ORIA	The autonomous execution engine (apollia-oria): the ReAct loop, planner, budget, resilience, verification, and context management.
Direct vs orchestrated	Two ORIA execution modes. Direct runs a single agent loop; orchestrated plans and drives governed tool steps with verification and re-planning.
ReAct	The reason-then-act loop an agent runs: think, call a tool, observe, repeat.
StepBudget	The non-bypassable ceiling on reasoning steps, tool calls, and wall-clock time the runtime enforces on every run.
ctx	The runtime context passed to every tool and every workflow of a package, exposing seventeen typed services. The contract is sdk/apollia/types.py. See the SDK reference.
AgentKit	The Python SDK (apollia): the two decorators, the schemas, the test doubles and the helpers the Python of a package is written against.
MCP	Model Context Protocol. Apollia is an MCP client that discovers and calls external tools, and it serves the agents and workflows the user exposed to an MCP client through a stdio bridge, off by default.
Autonomy tier	The per-run setting that governs the plan gate and the post-run verification pass. It governs no permission and no approval, so it changes how far a run goes unattended, not what an agent may touch.
HITL	Human-in-the-loop: the answer a person gives (once, always or refuse) when the governance decision asks about a tool call. The answer is recorded in the journal.
Audit journal	The append-only, hash-chained, signed record of governed actions, used for verification.
Verify	Checking a run's audit hash chain and signatures to confirm the record was not altered.
Rollback	Undoing filesystem changes made in a chat session. The journal format and the replay logic exist in the codebase, but nothing writes to the journal, so this is not a capability of this release.
Replay	Stepping offline through the trace the journal captured for a run and reporting it identical to its record, diverged or incomplete (apollia-os audit replay). It calls no model, runs no tool, and does not run the agent again.
Runner	The out-of-process sidecar (apollia-runner) that runs local speech-to-text (whisper). Local LLM inference is served by the embedded llama-server (upstream llama.cpp), which the daemon supervises.
GGUF	The single-file local model format the embedded llama-server loads.
Driving contract	The stable, versioned HTTP API plus generated host SDKs a host product uses to drive the runtime. See the HTTP API reference.
EventBus	The runtime's structured event stream, which the audit journal and observability share.
Sovereignty	The property that no user data leaves the machine without an explicit action, with local inference and storage by default.