Aivyx PA · ManualContents ↓

Glossary

Product terms first, then the codenames you’ll meet in commit messages, code comments and the deeper design docs. Links to those docs go to the repository on GitHub.

Product terms

Access level — How far the agent reaches on your machine: sandbox, workspace, home, full or custom. Set with aivyx-pa access. See Access and settings.

Audit log — A tamper-evident record of every tool call and decision, kept in the store. Each entry is chained to the previous one with an HMAC, so a change anywhere breaks the chain. Verify it with aivyx-pa --verify-only; export it with aivyx-pa audit export.

Autonomous loop — Works through a queue of stories (tasks) on its own until done or a limit is hit. Configured in [loop], driven with aivyx-pa loop.

Autonomy level — How much the agent may do without asking: manual, assisted, supervised, autonomous or unleashed. Set with aivyx-pa autonomy. See Autonomy and routines.

Capability (scope) — Permission for one kind of action, such as fs.read or email.send. Roles grant capabilities; trust tiers cap them.

Channel — A way of talking to the agent: the terminal, the Studio, the desktop app, Telegram, Discord, Slack, voice, or one you write.

Daemon — The one long-running Aivyx PA process. Every channel talks to it over a private local socket.

Default skills — SKILL.md capability packages bundled with Aivyx PA and shared with aivyx-coder, read through skill_defaults.* tools. Not the same as learned skills.

Escalation — With routing on, moving a call from a local model to a cloud one. Governed by [routing.escalation]; never happens for a sensitive conversation.

Gate — A pause for your approval before an action runs. You answer it in chat, the terminal or the Studio.

Instance — One complete, separate Aivyx PA: its own config, store, roles, memory and Studio. Everyone starts with default; add more with aivyx-pa instances create. See Named instances.

Integration — A connected service (Gmail, Calendar, Notion…) whose tools run as a separate, sandboxed tool process.

Learned skill — A procedure the agent wrote down (or you taught it), approved by you, which it can reuse. Lives in the store.

MCP server — An external Model Context Protocol server whose tools the agent can use, configured as [[mcp_server]].

Memory profile — lite or smart ([memory] profile): how much of the recall machinery is switched on. smart adds the knowledge wiki and graph, which cost model calls.

Mission — A long-running piece of work with its own record and status, often run by a team.

Notify target — A named place notifications can go: a Telegram chat, a webhook, an email address, or a desktop notification from the open Studio ([[notify_target]]).

Pack (vertical pack) — A signed bundle that adds a domain-specific team and tools, such as the kitchen example. Managed with aivyx-pa pack.

Passphrase — Unlocks the store. Aivyx PA can’t recover it for you.

Persona — The agent’s evolving character and what it has learned about working with you. Changes are proposed and wait for your approval.

Profile — What you tell the agent about yourself and how you like to work ([profile], aivyx-pa profile).

Proposal — A change the agent suggests to its own Persona, skills or behaviour. Nothing applies until you approve it (unless you set an auto-accept threshold).

Reflection — A scheduled pass in which the agent reviews recent turns and files proposals ([[reflection_schedule]]).

Role — A named bundle of system prompt, tools and capabilities ([[role]]). Roles can inherit from a parent.

Routine — Something that runs on its own: a schedule, a webhook or a file watch.

Routing — Choosing a model per call from the models you have ([routing]). Off by default.

Sensitive conversation — With routing on, one that has touched your own data (mail, files, memory…). It stays on local models.

Store — The encrypted database holding memory, Persona, skills, missions, the audit log and saved keys (store.redb).

Studio — The web interface the daemon serves at http://127.0.0.1:7843, also wrapped by the desktop app.

Team — Several specialist agents working one mission under a lead; the built-in team is called the Nonagon. See Teams.

Tool — One thing the agent can do, such as fs.read or gmail.send. See Tools.

Tool process — A tool that runs as its own program and talks to the daemon over a small protocol ([[tool_process]]). Integrations work this way, and you can write your own.

Trust tier — The ceiling on what a channel may ever do: Kernel, Trusted, SemiTrusted or Untrusted. A chat app gets less than your own terminal.

Workspace — The agent’s own directory for notes, plans and its journal, separate from your files.

Codenames

Features were built as named “chapters”. Each name below is what it refers to today; most have a design doc under docs/.

For the security names (Ward, Portcullis, Rampart, Bulwark, Picket, Gatehouse) in plain language, see Security and privacy.