Broken
Names are text, not entities
Whisper hears “Rivvl” as ribbon or ripple. Exact matching fails, then the last project wins or clarification loops forever.
Architecture and acceptance briefing · 14 August 2026
Jarvis needs a task desk, not a smarter regex
The reliable path to FRIDAY-like interaction is stable entity identity, explicit conversation state, constrained resolution, and repair. An LLM can be a language adapter later; it must not own thread IDs, approvals, or execution.
DecisionDeterministic core
Optional semantic fallback
Typed execution only
Current failure modes
The rough edges share one cause
Broken
One last task is not memory
Opening a second conversation overwrites the only reference. “Go back” has no cursor, history, candidate set, or stable target to operate on.
Weak
Prompts are strings, not frames
“Which project?” is spoken, but the system does not preserve a typed slot and candidate IDs robustly enough to repair the next turn.
Weak
Approval sees shell syntax
A wrapper command hides its actual read-only purposes. The agent says “something” because no operation-set adapter explains the inner work.
Deep module
A server-owned task desk
| State | What it means | Voice examples |
|---|---|---|
focusedThreadId | The exact task receiving referential commands. | “Stop that”, “add tests to this.” |
backStack / forwardStack | Stable navigation history, independent of the visible T3 screen. | “Go back”, “return to the newer task.” |
recentTasks | Bounded entities: IDs, project, title, objective, state, aliases. | “Switch to the Rivvl pull-request review.” |
pendingFrame | A typed clarification or confirmation with candidate IDs. | “The second one”, “no, the review task.” |
newConversationArmed | A one-turn instruction to create independent work while preserving focus history. | “Start another conversation.” |
Resolution order
Language proposes; state proves
- Repair first. If a clarification frame exists, resolve only against its candidates.
- Exact controls next. Back, forward, stop, status, new conversation, approval, and ordinals are closed grammar.
- Entity resolution. Search real project/task entities using exact aliases, relationships, lifecycle, recency, and conservative phonetics.
- Confidence gate. One clear winner proceeds; ties remain pending and ask a short discriminating question.
- Optional semantic adapter. An LLM may propose an action and entity text only when deterministic parsing rejects the utterance.
- Typed authorization. The host resolves stable IDs, validates state, and emits ordinary T3 orchestration commands.
Release acceptance
Test the experience as a system
A green unit test is not enough. Every release should survive these user journeys on a real laptop host and Windows Companion.
InstallShipped
Install once, update thereafter
Install the NSIS build once; verify tray version, background differential download, restart-to-install, and preservation of pairing/defaults.
CaptureShipped
Hold, speak, review, release
First words survive immediate speech, every VAD segment is retained, capture lasts for the full hold, canonical terms are repaired, and the surface dismisses.
ProjectsShipped
Typed project intelligence
“What projects are there?” reads T3 state without Codex. Explicit, phonetic, ordinal, ambiguous, and unknown project cases never silently fall back.
ControlShipped
Exact task control
Start, steer, queue, status, stop, reroute, continuation, restart recovery, blocked replies, and cross-project ownership all retain stable IDs.
ApprovalShipped
Explain purpose and risk
Reads, writes, tests, builds, installs, pushes, migrations, network, elevated, destructive, compound, and unknown commands are distinguished; questions never approve.
SpeechShipped
Brief, useful reports
Completion speech keeps outcome, verification, blockers, and next action while dropping code, paths, hashes, headings, generic “Done,” and changelog noise.
ResilienceShipped
Failure is actionable
Exercise offline host, revoked session, stale provider/default, duplicate surfaces, rapid reports, host restart, and the three supported connection modes.
Task deskPlanned
Multiple conversations
New conversation, back, forward, named-task switching, durable candidate frames, and task relationship search remain the next vertical slices.
Tracer bullets
Build it in five vertical slices
Task desk projection
Typed per-client focus, history, recent task entities, restart durability, HTTP and WebSocket parity.
Back / forward / new
Instant voice navigation with integration proof that the next turn reaches the selected thread.
Named task search
Title, project, objective, relationship, state, ordinal, recency, and learned aliases.
Conversation repair
Persisted clarification frames; “second one” fills the open slot instead of becoming a new task.
Semantic fallback
Optional small model only for rejected paraphrases, evaluated against deterministic baselines and never trusted with authority.
Sources: Apple entity queries, Alexa dialog schema, Rasa dialogue management, and Rasa dialogue frames. Accessed 13 August 2026.