| FastBrain |
Historical name for the user-facing reasoning path; the realtime implementation calls this the FrontBrain |
| FrontBrain |
The realtime provider model filling the FastBrain role on the voice path |
| Proactive |
Proactive communication: discovers grounded proposals and selects optional updates; never executes tasks or controls the floor |
| Runtime spine |
The event loop that applies state and coordinates work |
| Memory |
Canonical per-channel observations, accepted handoffs, and revision-bound intake facts |
| Personal memory |
Cross-session facts stored by default in the unified local memory ledger; mem0 remains an explicit alternative; never execution authority |
| Knowledge corpus (K) |
Opt-in, separately admitted user documents retrieved through Knowledge MCP; stored apart from the personal-memory store; cloud embedding egress requires explicit disclosure and consent |
| Saved / applied |
Persisted configuration versus configuration activated by the backend; service changes can be saved but pending restart |
| Channel |
One append-only observation stream per capability: conversation, search, camera evidence, hidden Vision watch/guard, plus one per active executor |
| ContextView |
A bounded snapshot compiled for a model call |
| Floor |
Speaking-path arbitration with three verdicts: allow, preempt, defer |
| Priority |
Urgency bound to the triggering event, never chosen by the model: user 100, guard 90, active executors 50, ambient observations 40 |
| Preempt |
Floor verdict that interrupts Nova playback when allowed; it never interrupts user speech, and real audio cancellation exists only on the realtime path |
| Defer |
Floor verdict that sends the utterance to the suggestion pool instead of dropping it |
| Executor |
A manifest-declared capability behind the port contract |
| AgentController |
A host registry entry describing a model-facing controller and its owned hidden channels; separate from executor manifests |
| Direct MCP tool |
A consumer-projected ${name}__${op} operation, never an executor dispatch target or intake writer |
| Delegate |
One identity-bound unit of dispatched work; long-running coding work and bounded approval/model requests have separate lifetimes |
| Progress |
A non-terminal executor event bound to its delegate |
| Observation |
A non-terminal fact emitted by an active delegate outside the progress cadence (for example a monitoring hit) |
| Handoff |
A typed terminal executor result returned to Runtime |
| Suggestion |
A stored candidate that may or may not be spoken |
| Suggestion pool |
The suggestion lifecycle: pending → fired → cooldown + re-arm → pending, plus withdrawn and lazy expiry; re-arming requires both an elapsed cooldown and new evidence on the cited channel |
| Steering |
Appending a new instruction to an in-flight Codex turn (codex.steer over the app-server transport) without terminating or restarting it |
| App-server |
The native codex app-server JSON-RPC process driven by the live Codex backend (turn/start, turn/steer) |
| Wake reason |
Causal metadata describing why a model slot should run; distinct from an acoustic wake word |
| Wake word |
A local acoustic keyword that wakes the hidden desktop orb; never execution authorization |
| Auto-hide / sleep |
Idle desktop state that hides the orb and routes unmuted microphone frames to local wake detection |
| Frontend instructions |
Shared host-rendered context and speaking/tool rules in frontend-instructions.ts |
| Response origin |
Provider correlation evidence linking a response to a user turn or host request; not permission to execute |
| Workspace |
One isolated filesystem/Git project with its own Codex home; a Session never leaves its Workspace |
| Session |
One durable, resumable Codex thread inside exactly one Workspace |
| Proposal / structured confirmation |
Create, switch, and resume first produce a proposal; only the dedicated confirmation call with the exact proposal ID and a JSON boolean commits it — rejection, mismatch, or replay fails closed |
| Work order |
The single consolidated, bounded task statement handed to Codex; conversation history never crosses that boundary |
| Pipeline mode |
The top-level realtime shape: integrated (one realtime model) or cascaded (endpointing → ASR → LLM → TTS) |
| Endpointing |
The cascaded stage deciding when an utterance has ended (semantic turn detector or bounded silence) |
| Knowledge MCP |
Built-in mcp__nova_knowledge__recall evidence tool; optional Codex loopback additionally resolves digest-pinned chunks via get_chunk, never mutates the corpus |
| MyContext adapter |
An optional loopback-only, read-only evidence provider behind Nova's strict capability handshake; no adapter ships in this repository |
| Workbench |
The desktop main window: an icon rail beside personal-object pages, with Nova present as a collapsible chat pane |
| Rail |
The Workbench's icon navigation column, selecting Todos, Ideas, Goals, Feeds, Task and Profile |
| Task |
A durable, host-tracked unit of delegated work with a goal and acceptance criteria; distinct from an ad hoc executor delegate |
| Handback / takeover |
Transferring a Task's controller between Nova and the user; takeover lets the user message the executor directly and pauses Nova's automatic corrections, handback returns control to Nova |
| Acceptance criterion |
One statement a Task's goal must satisfy before Nova marks it complete; evidence is recorded per criterion |
| Candidate |
An unconfirmed Todo, Idea, Goal or Profile fact Nova notices in conversation; recorded once the user confirms it, except that an explicitly stated todo is recorded right away and can be undone |
| Source |
An authorized document, account or feed that personal memory and suggestions may draw on |
| Connector |
An integration (for example Google, macOS Calendar, Apple Mail or Feishu) that exposes a Source through a read scope the user grants |
