Files
faceclaw/notes/voice-assistant-design.md
T

22 KiB
Raw Blame History

Voice Assistant Architecture

Status: design draft, 2026-07-19; revised 2026-07-24 (tool availability ontology, misc fixes). Phase 1 is implemented (see the Implementation plan below); phases 2-4 are still design-only.

Goal

A default voice assistant reachable from anywhere in the shell: hold the button, speak, and the utterance goes to an agent that can answer in text on the lenses and act on the glasses through tools. The agent is pluggable:

  • Direct mode: the phone runs the agent loop itself against the Anthropic Messages API (tool-use loop, conversation state on-phone).
  • External mode: the utterance is routed to the user's own long-running agent (OpenClaw, Hermes, or anything similar running on their machine), and that agent gets access to the glasses' tools while it works.

Tools come from two sources, with four availability tiers (see "Tool availability ontology" below):

  • Glasses/system tools: provided by the shell, always available (show text, notifications, screen state, open app, ...).
  • App tools: contributed by apps. Each app tool declares how much of the app needs to exist for the tool to be live: merely installed ("start a 5 minute timer" must not require the Timer app to be open), open ("next track" needs a running Music window, foreground or not), or foreground ("scroll down" only makes sense aimed at the foreground app).

What we already have

  • PTT capture UI: VoiceInputLayer (app/ui/apps/voice-input.ts) handles push-to-talk, live transcripts, and a Send/Continue/Discard menu; Shell opens it on long-press and delivers Send to the foreground window's receiveTextInput (app/ui/shell/shell.ts).
  • Wakeword: On-device classifier listens for the wakeword "Hey Even", which sends a message to the phone which wakes the screen and starts voice input.
  • STT: voiceControlBridge with onboard Moonshine, Whisper, or ElevenLabs cloud.
  • LLM client: streamAnthropicMessage (app/native/anthropic.ts) — raw SSE streaming over FaceclawSseRequest. Text-only today; no tool use.
  • Worker app protocol: WorkerAppMessage/WorkerAppReply (app/ui/shell/worker-window.ts) — small-JSON postMessage protocol between shell and app workers; the natural place to add tool declarations.
  • Dial-out websocket precedent: G2MirrorClient — JSON over a websocket to a server on the user's machine (tailscale host/port/token settings). The external-agent bridge should copy this shape.

Architecture overview

                        main isolate (shell)
  ┌───────────────────────────────────────────────────────────┐
  │  Shell ──(PTT)──> VoiceInputLayer ──> AssistantSession    │
  │                                            │              │
  │  AssistantLayer (overlay UI) <──streaming──┤              │
  │                                            │              │
  │                                   AssistantBackend        │
  │                                   ┌────────┴─────────┐    │
  │                                   │ Direct  │ External│    │
  │                                   └────┬────┴────┬────┘    │
  │                                        │         │         │
  │  ToolRegistry <────────────────────────┴─────────┘         │
  │   ├─ system tools (in-process handlers)                    │
  │   ├─ app "installed" tools (shell-side or launch-on-call)  │
  │   └─ app "open"/"foreground" tools (proxied to app worker) │
  └──────────┬──────────────────────────────────────┬─────────┘
             │ postMessage                          │
       app worker (e.g. terminal)          Anthropic API (SSE)
       set-tools / tool-call /             or agent bridge (WS)
       tool-result                         on the user's machine

Everything assistant-side lives on the main isolate. Rationale: tool handlers need shell state (window list, focus, notifications, settings), the agent loop is I/O-bound not CPU-bound, and the overlay UI is a shell layer anyway. If conversation rendering ever gets expensive it can move to a worker later without changing the tool protocol, which is already async message passing.

Core pieces

  1. ToolRegistry (new, app/assistant/tool-registry.ts)

    • Holds ToolSpecs and handlers; answers listTools(scopeFilter) and callTool(name, args) -> Promise<ToolResult>.
    • Emits onToolsChanged (needed for MCP tools/list_changed in external mode, and consulted at each turn boundary in direct mode).
  2. AssistantBackend (interface) with two implementations:

    • DirectAnthropicBackend: tool-use loop over the Messages API.
    • ExternalAgentBackend: websocket bridge to the user's agent.
  3. AssistantSession: one conversation. Owns history/turn state, exposes sendUtterance(text, ctx), cancel(), and streaming callbacks (onTextDelta, onToolActivity, onTurnDone, onError). Both backends implement the same session surface so the UI is backend-agnostic.

  4. AssistantLayer: shell overlay showing streamed reply text, a status line for tool activity ("→ calendar.list_events"), and a small menu (Follow-up / Done). Double-click cancels the in-flight turn.

Entry points and routing

  • Wakeword Saying "Hey Even" wakes the screen if necessary and opens the dictation dialog. For one-off voice commands, this will be the main entry point.
  • Context menu The sidebar context menu and app context menu already have a "Voice input" option, which opens the dictation dialog whose Send goes to the foreground window. We'll split the Send option in this dialog into two options, "Send to Assistant" and "Type Into App". If the entry point was via context menu, "Type Into App" is highlighted by default; if the entry point was the wakeword, "Send to Assistant" is highlighted by default instead. (The capture UI is the existing VoiceInputLayer flow, with Send handing off to AssistantSession and morphing into AssistantLayer).

Tool model

type ToolAvailability =
  | "always"      // system tools; no app involved
  | "installed"   // app tool, live whenever the app is installed —
                  //   the app need not have any window open
  | "open"        // app tool, live while the app has a window open,
                  //   foreground or backgrounded
  | "foreground"; // app tool, live only while the app owns the
                  //   foreground window

type ToolSpec = {
  name: string;              // namespaced: "glasses.show_notification",
                             // "app.terminal.send_input"
  description: string;
  inputSchema: object;       // JSON Schema
  availability: ToolAvailability;
  /** May be invoked by the external agent outside an active voice turn. */
  proactive?: boolean;
  timeoutMs?: number;        // default 10s; registry enforces
};

type ToolResult =
  | { ok: true; content: string }           // text payload back to the model
  | { ok: false; error: string };

Tool availability ontology

Every tool sits in exactly one tier, and the tier answers two questions: when does the tool appear in listTools(), and who handles the call.

Tier Live when Handler Examples
always always shell, in-process glasses.get_state, notifications.list
installed app installed (launcher entry exists) shell-side handler registered with the app entry, or launch-the-app-then-proxy timer.set ("start a 5 minute timer" must work with no Timer window), music.play (cold-start playback)
open app has ≥1 open window, foreground or not app worker (or in-process window) via tool-call proxy music.next_track, terminal.send_input
foreground app owns the foreground window app worker via tool-call proxy teleprompt.scroll_down, anything meaning "act on what I'm looking at"

Mechanics per tier:

  • installed tools cannot be declared by the worker (it may not be running), so they are declared statically alongside the app's launcher registration. Two handler flavors: (a) a shell-side handler function registered with the entry — right for tools that don't actually need the app UI (timers, starting playback); (b) launch-on-call: the shell opens the app (as if launched from the launcher), waits for its set-tools declaration, then proxies the call. Start with (a) only; (b) adds a launch-latency/timeout dance that no phase-1/2 tool needs.
  • open and foreground tools are declared by the running app via set-tools (worker apps) or direct registration (in-process windows). The registry filters foreground tools by whether the declaring app owns the current foreground window at list time and again at call time — a foreground tool called after a foreground swap fails with a normal tool error rather than acting on a window the user is no longer looking at.
  • A tier is a floor, not a routing rule: an installed tool stays listed while the app is open or foreground. If an app wants different behavior when its window exists (e.g. timer.set also shows a countdown in the open Timer window), the shell-side handler checks for the window itself.

Flag: with multiple windows per app (the worker protocol is per-window), tool names are namespaced per app, so two windows of one app declaring the same tool name would collide. Rule: set-tools replaces the app's toolset per window, and a call routes to the declaring window; if two windows declare the same name, last declaration wins and earlier ones are dropped with a logged warning. Revisit only if a real app needs per-window tools with the same name (none of the planned ones do).

System tools (initial set)

Tool Notes
glasses.show_alert short text popup on the lenses; proactive-capable
glasses.list_windows / glasses.focus_window / glasses.open_app shell control
glasses.get_state screen on/off, foreground app, battery levels, time
glasses.read_screen text/description of the current foreground surface (start with window title + app-reported summary; pixel OCR is out of scope)
calendar.list_events wraps FaceclawCalendarProvider
media.play_pause / media.next / media.now_playing wraps FaceclawMediaController
notifications.list / notifications.dismiss wraps notification listener
timer.set / timer.cancel new, but trivial and high-value for an assistant

Keep the first release to ~8–12 tools; every tool costs prompt tokens and model attention in direct mode.

Note that calendar.*, media.*, and timer.* are conceptually installed-tier app tools that happen to be implemented by in-process shell code; listing them here just means the shell registers their handlers directly. If any of them later becomes a real app, its tools move to the app's launcher registration without the model-visible surface changing.

App tools

Running apps declare open/foreground tools per window over the existing worker protocol (installed tools are declared statically with the launcher entry, per the ontology above):

// WorkerAppReply additions (worker -> shell)
| { type: "set-tools"; windowId: string; tools: ToolSpec[] }   // replaces the set
| { type: "tool-result"; callId: string; result: ToolResult }

// WorkerAppMessage additions (shell -> worker)
| { type: "tool-call"; callId: string; windowId: string; name: string; args: unknown }

Registry rules:

  • Liveness follows the availability tier: open tools are removed when the app's last window closes, foreground tools additionally drop out of listTools() when the app loses the foreground. Either transition fires onToolsChanged; in-flight calls to a closed window resolve as errors, and calls to a foreground tool are re-checked at dispatch time.
  • Names are auto-prefixed app.<appId>. by the registry so apps can't shadow system tools or each other.
  • The shell enforces timeoutMs on the postMessage round-trip; a hung worker yields a tool error, not a hung turn (same philosophy as input handling — the shell must survive a stuck app).
  • In-process shell apps (calendar, music, ...) register handlers directly with the registry; the worker protocol is just the remote flavor of the same registration.
  • Mid-turn availability changes (foreground swap, window close): direct mode re-lists tools at each loop iteration (tool set is per-API-call anyway); external mode sends MCP notifications/tools/list_changed. A call to a tool that just vanished returns a normal tool error the model can react to.

First app to wire up: the terminal — app.terminal.send_input (type into the attached session; open, proactive: false — "rerun the build" should work while the terminal is backgrounded), app.terminal.read_screen (current grid contents; open), app.terminal.list_sessions (open). This immediately enables "tell the terminal to rerun the build" style commands and doubles as the reference implementation — deliberately all open-tier; the first foreground tool should come from an app where focus genuinely matters (teleprompt scrolling).

Direct mode: agent loop on the phone

Extend app/native/anthropic.ts:

  1. Streaming tool-use support: handle content_block_start with tool_use, accumulate input_json_delta, surface complete tool_use blocks alongside text deltas. (Today the client only reads text_delta.)
  2. runAgentTurn loop (new, app/assistant/direct-backend.ts): messages + tools from the registry → stream → if stop_reason == "tool_use", execute each call via the registry, append tool_result blocks, continue; until end_turn/refusal or a safety cap (max 8 loop iterations, max ~2 min per turn).
  3. Session state: history kept in memory per AssistantSession, trimmed from the head past a token budget. Sessions idle-expire (~10 min) so a new PTT starts fresh; "Follow-up" continues the session.
  4. System prompt: glasses context — output renders on a 576×288 monochrome HUD, so answer in 1–3 short sentences, no markdown, no lists unless asked; prefer acting via tools over describing; current time, foreground app, and screen state are injected per turn.
  5. Model claude-sonnet-5, effort: "low" (same latency posture as the dictation refiner). Model/effort become settings later if needed.

External mode: bridge to the user's agent

Topology

The phone dials out (g2mirror pattern — phones don't accept inbound connections reliably; tailscale host + port + token settings) to a small bridge server co-located with the user's agent. One websocket, JSON frames, three multiplexed channels:

{ v: 1, chan: "ctl" | "chat" | "mcp", ... }

ctl:  hello {deviceName, token, capabilities}, hello-ack, ping/pong, error
chat: utterance {turnId, text, ctx}        (phone -> agent)
      text-delta {turnId, text}            (agent -> phone)
      tool-activity {turnId, label}        (agent -> phone, optional status)
      turn-done {turnId, stopReason} / turn-error {turnId, message}
      cancel {turnId}                      (phone -> agent)
mcp:  { msg: <raw MCP JSON-RPC frame> }    (bidirectional)

On the mcp channel the phone is an MCP server (MCP roles come from the initialize handshake, not from who dialed). It serves tools/list from the ToolRegistry, tools/call into it, and emits tools/list_changed on foreground changes. This means the agent-side integration needs zero Faceclaw-specific tool code — the glasses appear as a normal MCP server.

Agent-side adapters

The bridge server is a small standalone package (faceclaw-agent-bridge, sibling of g2mirror) with two faces:

  1. MCP face: re-exposes the phone's MCP channel as a localhost streamable-HTTP MCP endpoint. The user registers http://localhost:<port>/mcp with their agent (OpenClaw, Hermes, Claude Code, anything MCP-capable) exactly like any other MCP server. When the phone is disconnected, tools/list returns empty rather than erroring.
  2. Chat face: pluggable per-agent adapter that forwards utterance frames into the agent and streams replies back:
    • OpenClaw: a plugin that injects the utterance as an inbound message on a dedicated "glasses" channel/session and relays the streamed reply.
    • Hermes: equivalent module against its session API.
    • Fallback adapter: spawn a configured CLI (e.g. claude -p --resume) per turn — lowest-fidelity but makes the bridge useful with no plugin.

Utterance ctx carries {foregroundApp, screenOn, localTime} so the remote agent has the same situational grounding the direct loop gets in its system prompt.

Proactive agent actions

Because the MCP channel is live whenever the bridge is connected, the external agent can call glasses tools outside a voice turn — e.g. push a notification to the lenses when a long job finishes. Gating:

  • Only tools marked proactive: true are callable outside an active turn (registry-enforced; enforced on the phone, never trusted to the bridge). Availability tiers apply to proactive calls unchanged: a proactive installed tool (e.g. timer.set) works anytime, while a proactive call to an open/foreground tool whose app isn't in the right state gets the same tool error a mid-turn call would.
  • Master setting assistant.allowProactive (default on — it's a marquee feature — but visible and easy to turn off).
  • Proactive display actions are rate-limited (e.g. 6/min) and never turn the screen on unless the user enabled that specifically.

Settings

New "Assistant" section in dashboard settings:

  • assistant.backend: anthropic | external (default anthropic if an API key is set, else external if a bridge host is set).
  • assistant.bridgeHost / assistant.bridgePort / assistant.bridgeToken (mirrors the terminal.* trio).
  • assistant.allowProactive: boolean.
  • voice.skipConfirmationAfterWakeword: Wakeword-triggered input sends to the assistant immediately at the end of an utterance, rather than waiting for a menu selection.
  • Reuses existing llm.anthropicApiKey for direct mode.

Security notes

  • Bridge auth: bearer token in the hello frame, TLS optional because the expected transport is tailscale (same stance as g2mirror); document that plainly.
  • The phone enforces all tool gating (proactive flags, rate limits, timeouts). The bridge and agent are treated as honest-but-fallible.
  • Tool results may contain private data (calendar, notifications); external mode ships them to the user's own machine only — no third-party hop. In direct mode they go to the Anthropic API like any other request content.
  • App tools execute inside the app's worker with the app's existing privileges; the registry adds no new capability beyond what the app could already do.

Implementation plan

Phase 1 — direct assistant, system tools only. ✅ Landed 2026-07-24 (builds + typechecks; not yet hardware-tested). ToolRegistry (app/assistant/tool-registry.ts) with the full availability-tier model but only always tools wired; 8 system tools (glasses.get_state, glasses.show_alert, calendar.list_events, media.now_playing/play_pause/ next, notifications.list/dismiss) in app/assistant/system-tools.ts; tool-use streaming in anthropic.ts (content-block assembly, tools param); DirectAnthropicBackend + AssistantSession; AssistantLayer overlay (app/ui/apps/assistant.ts); VoiceInputLayer reworked to route via a sendTargets list ("Send to Assistant" / "Type Into App") with entry-point default highlight and wakeword skip-confirmation auto-send; shell.showAlert popup; assistant.skipConfirmationAfterWakeword setting. Deliverable: speak wakeword, ask "what's on my calendar", get an answer on-lens.

Phase 2 — app tools. Worker protocol extension (set-tools / tool-call / tool-result), availability-tier filtering in the registry (open on window open/close, foreground on foreground change), installed-tier registration on launcher entries (shell-side handlers only; launch-on-call deferred), terminal app tools as the reference. Deliverable: "run the build again" typed into the terminal even while it's backgrounded.

Phase 3 — external bridge. Bridge wire protocol; ExternalAgentBackend (phone side, reusing the FaceclawWebSocket plumbing); faceclaw-agent-bridge server with the MCP face + fallback CLI chat adapter; settings; reconnect/offline behavior. Deliverable: utterances answered by the user's own agent, which can also call glasses tools mid-task.

Phase 4 — OpenClaw/Hermes adapters + proactive polish. Native chat adapters for both agents; proactive gating + rate limits; tools/list_changed on foreground swap; session persistence across reconnects.

Open questions (deferred, not blockers):

  • Continuous conversation: after a turn ends, listen for a follow-up utterance without requiring the wakeword/PTT again? There are some UI design questions here; maybe multi-turn conversations take the form of an Assistant app.
  • Launch-on-call for installed-tier tools that genuinely need the app UI (open the app, await set-tools, proxy the call) — deferred until a tool needs it.
  • glasses.read_screen fidelity beyond app-reported summaries? We probably don't want to do OCR, but we might do something like extending the draw-text draw call with an on-by-default option that also appends the drawn text to a string that gets submitted along with the frame.
  • Whether assistant conversations should be reviewable in the phone UI (transcript log) — probably yes, cheap once sessions are objects. The natural UI home for this would be inside an Assistant app window.