Jarvis — AI Assistant on Reachy Mini

An embodied AI assistant inspired by Jarvis from Iron Man, running on the Reachy Mini robot with OpenAI Agents as its brain.

Design Principles

1. Latency is king. First audible response within 300-600ms (filler or real). Full answer streams after. People forgive dumb; they don't forgive slow.

2. Presence, not request/response. A continuous 30Hz "presence loop" runs independent of the LLM — breathing, micro-nods, gaze tracking. The robot feels alive even when silent.

3. Embodiment as policy, not library calls. The LLM outputs an "embodiment plan" (intent, prosody, motion primitives) with each response. A renderer maps those to physical behavior. No random "play happy" uncanny valley.

4. Barge-in. User can interrupt at any time. TTS stops immediately, new utterance is captured. This single behavior makes it feel 10x more real.

5. Guardrails. Destructive smart home actions use dry-run by default. Everything is audit-logged. Permissions model for sensitive operations.

6. Honest state broadcasting. It's obvious when Jarvis is listening vs idle vs muted, through posture and behavior, not just an LED.

Architecture

┌───────────────────────────────────────────────────────────────────┐
│                      PRESENCE LOOP (30Hz)                         │
│  Always running. Receives lightweight signals, outputs motion.    │
│                                                                   │
│  Signals in:              States:                                 │
│    vad_energy ──┐          IDLE     → breathing, drift            │
│    doa_angle  ──┤          LISTENING → orient, micro-nods, lean   │
│    face_pos   ──┼────────► THINKING → look away, processing anim │
│    llm_state  ──┤          SPEAKING → stable gaze, intent motion  │
│    embody_cmd ──┘          MUTED    → privacy posture             │
└───────────────────────────────────────────────────────────────────┘

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│   Audio Input    │     │   Agent Brain     │     │   Audio Output  │
│                  │     │                  │     │                  │
│  Mic → VAD ──────┼────►│  Agent SDK       │────►│  Stream TTS     │
│       ↓          │     │  + MCP tools:    │     │  (ElevenLabs)   │
│  Whisper STT ────┼────►│    embody/robot │     │                  │
│                  │     │    smart_home/* │     │  Barge-in:      │
│  Barge-in: ◄─────┼─────│    todoist/*    │◄────│  VAD interrupts │
│  stop TTS        │     │    memory/*     │     │  playback       │
└─────────────────┘     └──────────────────┘     └─────────────────┘

┌──────────────────┐     ┌──────────────────┐
│  Face Tracker    │     │   Audit Log      │
│  YOLOv8 → face   │────►│ ~/.jarvis/audit  │
│  position signal │     │   .jsonl         │
│                  │     │ (rotating)       │
└──────────────────┘     └──────────────────┘

Components

Setup

cd jarvis
uv sync
cp .env.example .env
# Fill in: OPENAI_API_KEY, ELEVENLABS_API_KEY
# Optional: HASS_URL, HASS_TOKEN for smart home
# Optional: HOME_PERMISSION_PROFILE=readonly (state only) or control (default)
# Optional: HOME_REQUIRE_CONFIRM_EXECUTE=true (require confirm=true on all executes)
# Optional: HOME_CONVERSATION_ENABLED=true (enable HA conversation intent tool)
# Optional: HOME_CONVERSATION_PERMISSION_PROFILE=readonly|control (default readonly)
# Optional: SAFE_MODE_ENABLED=true (force mutating actions into restricted/dry-run behavior)
# Optional: TODOIST_API_TOKEN / TODOIST_PROJECT_ID / TODOIST_PERMISSION_PROFILE
# Optional: NOTION_API_TOKEN / NOTION_DATABASE_ID (integration_hub notion notes backend)
# Optional: TODOIST_TIMEOUT_SEC=10.0 / PUSHOVER_TIMEOUT_SEC=10.0
# Optional: PUSHOVER_API_TOKEN / PUSHOVER_USER_KEY / NOTIFICATION_PERMISSION_PROFILE
# Optional: NUDGE_POLICY=interrupt|defer|adaptive / NUDGE_QUIET_HOURS_START / NUDGE_QUIET_HOURS_END
# Optional: EMAIL_SMTP_HOST / EMAIL_FROM / EMAIL_DEFAULT_TO / EMAIL_PERMISSION_PROFILE / EMAIL_TIMEOUT_SEC
# Optional: WEATHER_UNITS=metric|imperial / WEATHER_TIMEOUT_SEC
# Optional: WEBHOOK_ALLOWLIST=example.com,api.example.com / WEBHOOK_AUTH_TOKEN / WEBHOOK_TIMEOUT_SEC
# Optional: SLACK_WEBHOOK_URL / DISCORD_WEBHOOK_URL
# Optional: PERSONA_STYLE=terse|composed|friendly|jarvis / BACKCHANNEL_STYLE=quiet|balanced|expressive
# Optional: IDENTITY_ENFORCEMENT_ENABLED / IDENTITY_DEFAULT_USER / IDENTITY_DEFAULT_PROFILE
# Optional: IDENTITY_USER_PROFILES / IDENTITY_TRUSTED_USERS
# Optional: IDENTITY_REQUIRE_APPROVAL / IDENTITY_APPROVAL_CODE
# Optional: PLAN_PREVIEW_REQUIRE_ACK=true (require preview_token before risky execute tools)
# Optional: MEMORY_RETENTION_DAYS / AUDIT_RETENTION_DAYS (0 disables pruning)
# Optional: MEMORY_PII_GUARDRAILS_ENABLED=true|false
# Optional: MEMORY_ENCRYPTION_ENABLED / AUDIT_ENCRYPTION_ENABLED / JARVIS_DATA_KEY
# Optional: WAKE_MODE / WAKE_CALIBRATION_PROFILE / WAKE_WORDS / WAKE_WORD_SENSITIVITY / VOICE_TIMEOUT_PROFILE
# Optional: STT_FALLBACK_ENABLED / WHISPER_MODEL_FALLBACK / TTS_FALLBACK_TEXT_ONLY
# Optional: OPENAI_ROUTER_MODEL / ROUTER_TIMEOUT_SEC / POLICY_ROUTER_MIN_CONFIDENCE
# Optional: INTERRUPTION_ROUTER_TIMEOUT_SEC / INTERRUPTION_RESUME_MIN_CONFIDENCE
# Optional: SEMANTIC_TURN_ENABLED / SEMANTIC_TURN_ROUTER_TIMEOUT_SEC / SEMANTIC_TURN_MIN_CONFIDENCE
# Optional: SEMANTIC_TURN_EXTENSION_SEC / SEMANTIC_TURN_MAX_TRANSCRIPT_CHARS
# Optional: MODEL_FAILOVER_ENABLED / MODEL_SECONDARY_MODE / WATCHDOG_* / TURN_TIMEOUT_ACT_SEC / STARTUP_STRICT
# Optional: OPERATOR_SERVER_ENABLED / OPERATOR_SERVER_HOST / OPERATOR_SERVER_PORT / OPERATOR_AUTH_MODE / OPERATOR_AUTH_TOKEN
# Optional: WEBHOOK_INBOUND_ENABLED / WEBHOOK_INBOUND_TOKEN
# Optional: RECOVERY_JOURNAL_PATH / DEAD_LETTER_QUEUE_PATH (interrupted-action + failed-outbound journals)
# Optional: EXPANSION_STATE_PATH / RELEASE_CHANNEL_CONFIG_PATH (roadmap + release-channel persistence/check config)
# Optional: NOTES_CAPTURE_DIR / QUALITY_REPORT_DIR (integration capture + report artifact locations)
# Optional: OBSERVABILITY_* (DB/state/event paths, burst threshold, snapshot interval)
# Optional: SKILLS_ENABLED / SKILLS_DIR / SKILLS_ALLOWLIST / SKILLS_REQUIRE_SIGNATURE / SKILLS_SIGNATURE_KEY

Smart home safety defaults:

• Sensitive domains (lock, alarm_control_panel, cover, climate) require confirm=true when dry_run=false.
• HOME_PERMISSION_PROFILE=readonly disables mutating smart_home actions but keeps smart_home_state.
• HOME_REQUIRE_CONFIRM_EXECUTE=true enforces confirm=true for all non-dry-run smart_home actions.
• SAFE_MODE_ENABLED=true keeps mutating actions in restricted mode (dry-run where supported, blocked otherwise).
• PLAN_PREVIEW_REQUIRE_ACK=true enforces a two-step preview+ack flow (preview_token) before mutating medium/high-risk actions.
  • First call can pass preview_only=true to get a plan preview token.
  • Execute call must include matching preview_token=<token> before token expiry.
• NUDGE_POLICY controls due-reminder interrupts: interrupt, defer, or adaptive (quiet-window aware).
• Operational runbook: docs/operations/home-control-policy.md.
• Integration runbook: docs/operations/integration-policy.md.
• Trust/identity runbook: docs/operations/trust-policy.md.
• Dialogue response mode auto-switches by request context:
  • brief for urgent/short-answer requests,
  • deep for explicit detailed walkthrough requests,
  • normal otherwise.
• First-response strategy auto-selects per request:
  • answer for direct questions,
  • act for explicit action requests,
  • clarify when an action request is ambiguous (it/that/this targets).
• Confidence policy auto-calibrates language:
  • cautious for volatile/time-sensitive prompts (latest, today, right now),
  • calibrated for estimate/prediction prompts,
  • direct for stable factual prompts.
• Wake-word false-trigger suppression supports calibration profiles (default, quiet_room, noisy_room, tv_room, far_field):
  • profile tunes wake sensitivity, minimum post-wake phrase length, and adaptive suppression window after repeated wake-only triggers.
• Follow-up intent carryover preserves unresolved action context across short multi-turn replies:
  • short fragments like the bedroom or and in the office inherit prior unresolved action targets.
  • explicit new action phrasing (e.g. turn on the kitchen lights) remains a new request.
• Runtime STT confidence diagnostics are exposed in operator status:
  • voice_attention.stt_diagnostics reports confidence score/band, model source, fallback usage, and transcript quality signals.
• Low-confidence action requests trigger a lightweight repair loop:
  • Jarvis asks I may have misheard you as ... and accepts either confirm or an immediate corrected phrase.
• Semantic turn-end detection can defer utterance commit when speech appears incomplete:
  • an LLM router decides commit vs wait, then applies a short extension window before finalizing the turn.
• Barge-in follow-up handling is LLM-routed:
  • interruption turns are classified as replace, resume, or clarify with fail-closed fallback to replace.
  • resumed turns include continuity context from the interrupted answer.
• Per-user voice profiles can tune speaking behavior:
  • set_voice_profile / clear_voice_profile / list_voice_profiles manage per-user verbosity, confirmations, pace, and tone.
• Personality posture auto-switches by context:
  • social: allows one brief dry-wit line where appropriate,
  • task: stays precise and execution-focused,
  • safety: disables humor and favors explicit confirmation language.
• Home Assistant conversation tool requires both:
  • HOME_CONVERSATION_ENABLED=true
  • HOME_CONVERSATION_PERMISSION_PROFILE=control
  • and tool argument confirm=true
• Home Assistant helper tools:
  • home_assistant_todo (list|add|remove) for native HA to-do entities
  • home_assistant_timer (state|start|pause|cancel|finish) for HA timer entities
  • home_assistant_area_entities for area-aware entity resolution
  • media_control for simplified media_player actions (play, pause, volume_set, etc.)
• Automation consumers can use:
  • system_status (includes schema_version)
  • system_status.scorecard (unified latency/reliability/initiative/trust scoring)
  • system_status.observability.latency_dashboards (p50/p95/p99 total-turn latency with intent/tool-mix/wake-mode breakdowns)
  • system_status.observability.policy_decision_analytics (allow/deny reason counts by tool/user/status)
  • system_status.turn_timeouts (listen/think/speak/act timeout budgets)
  • system_status.integrations.*.circuit_breaker (open/remaining/failure state per integration)
  • system_status.recovery_journal (interrupted-action reconciliation summary)
  • system_status.dead_letter_queue (failed outbound delivery queue with replay status)
  • system_status.expansion (proactive, trust, orchestration, planner, quality, embodiment, integration roadmap feature snapshot)
  • jarvis_scorecard (standalone scorecard payload for dashboards and alerts)
  • system_status_contract (stable required-field contract)
• Memory retrieval now includes confidence/provenance details:
  • scoped classes: preferences, people, projects, household_rules (tagged as scope:<name>).
  • memory_search and memory_recent apply explicit scope policy (scopes=...) and expose scope=..., confidence=..., source=..., and trail=id/source/created_at.
  • memory_status includes confidence_model and scope_policy metadata for retrieval transparency.
• Audit logs now include readable authorization outcomes:
  • audit entries include decision_outcome, decision_reason, and decision_explanation.
  • this makes allow/deny/failure rationale machine-filterable and human-readable in /api/audit.
• Operator console/API security:
  • OPERATOR_AUTH_MODE=off|token|session controls operator auth strategy:
    • off: no auth challenge (highest risk)
    • token: per-request bearer/header token
    • session: login endpoint creates short-lived browser session cookie
    • if unset, mode auto-selects token when OPERATOR_AUTH_TOKEN is set, otherwise off
  • Set OPERATOR_AUTH_TOKEN when binding OPERATOR_SERVER_HOST to a non-loopback interface.
  • token mode protects /api/*, /metrics, and /events via X-Operator-Token or Authorization: Bearer <token>.
  • session mode protects the same endpoints via POST /api/session/login + jarvis_operator_session cookie.
  • The dashboard root (/) remains reachable and supports token entry for browser-based API calls.
  • GET /api/control-schema returns action/payload requirements for automation clients.
  • GET /api/conversation-trace returns live turn flow/tool/policy/latency trace rows used by the dashboard panel.
  • /api/status now includes episodic_timeline snapshots for recent important turns/actions.
  • /api/status now includes operator_controls with active_control_preset, available presets, and the current runtime profile snapshot.
  • /api/status now includes runtime_invariants (last check, total violations, auto-heals, recent entries).
  • Operator control actions support one-click presets/profile portability:
    • apply_control_preset (quiet_hours, demo_mode, maintenance_mode)
    • export_runtime_profile / import_runtime_profile
  • Control actions include explicit sleep/wake toggles via set_sleeping (sleeping=true|false).
  • Personality controls support live preview and rollback: preview_personality, commit_personality_preview, rollback_personality_preview.
  • /api/operator-actions now records tamper-evident chained signatures (previous_signature, signature, signature_alg).
• Release checklist: docs/operations/release-checklist.md.
• Security maintenance: docs/operations/security-maintenance.md.
• Error taxonomy: docs/operations/error-taxonomy.md.
• Observability runbook: docs/operations/observability-runbook.md.
• Personality research and tuning notes: docs/operations/personality-research.md.
• Proactive triage + preference loop: docs/operations/proactive-preference-loop.md.
• Fault resilience profiles:
  • local: make test-fault-profiles (runs quick, network, storage, contract)
  • CI: scheduled Fault Profiles workflow runs weekly with per-profile artifacts
• Skills developer guide: docs/operations/skills-development.md.
• Provenance verification: docs/operations/provenance-verification.md.
• Incident response: docs/operations/incident-response.md.
• Release acceptance: run ./scripts/release_acceptance.sh fast|full.
• Release channel checks: run ./scripts/check_release_channel.py --channel dev|beta|stable.
• Weekly quality artifact: run ./scripts/generate_quality_report.py --output-dir .artifacts/quality --markdown --compare-with .artifacts/quality/weekly-quality-<previous>.json.
• Deterministic eval dataset runner: run ./scripts/run_eval_dataset.py docs/evals/assistant-contract.json --strict --min-pass-rate 1.0 --max-failed 0.
  • Current dataset includes 250+ contract cases spanning home orchestration, planner/autonomy, integrations, comms, trust, and status contracts.
• Router policy eval dataset runner: run ./scripts/run_router_policy_eval.py docs/evals/router-policy-contract.json --strict --min-pass-rate 1.0 --max-failed 0.
  • Router dataset includes adversarial prompt-injection, identity-spoofing, escalation, and fail-closed routing contract cases.
• Interruption route eval dataset runner: run ./scripts/run_interruption_route_eval.py docs/evals/interruption-route-contract.json --strict --min-pass-rate 1.0 --max-failed 0.
  • Interruption dataset checks replace|resume|clarify routing, fallback behavior, and continuation metadata integrity.
• Trajectory trace grading runner: run ./scripts/run_trace_trajectory_eval.py docs/evals/trajectory-trace-contract.json --strict --min-pass-rate 1.0 --max-failed 0.
  • Trace grading scores full trajectories across completion success, response quality, interruption recovery, linkage integrity, and high-risk guardrail adherence.
• Autonomy cycle contract runner: run ./scripts/run_autonomy_cycle_eval.py docs/evals/autonomy-cycle-contract.json --strict --min-pass-rate 1.0 --max-failed 0.
  • Autonomy cycle dataset checks retry escalation, checkpoint gating, replan state transitions, and failure taxonomy accounting.
• Unified readiness gate: run ./scripts/jarvis_readiness.sh fast|full (or make readiness).
• One-command host bootstrap: run ./scripts/bootstrap.sh.
• Container profile: docker compose up --build (simulation/no-vision default).
• Home Assistant add-on starter path: deploy/home-assistant-addon.
• Todoist integration:
  • TODOIST_PERMISSION_PROFILE=readonly|control
  • readonly allows todoist_list_tasks and denies todoist_add_task
  • control allows both tools
  • TODOIST_TIMEOUT_SEC controls request timeout (default 10.0)
  • todoist_list_tasks supports format=short|verbose (default short)
• Pushover integration:
  • NOTIFICATION_PERMISSION_PROFILE=off|allow
  • off denies pushover_notify, slack_notify, and discord_notify
  • allow enables all channel notification tools
  • PUSHOVER_TIMEOUT_SEC controls request timeout (default 10.0)
• Email integration:
  • email_send requires confirm=true and EMAIL_PERMISSION_PROFILE=control
  • email_summary shows recent outbound email metadata
  • required SMTP env for send: EMAIL_SMTP_HOST, EMAIL_FROM, EMAIL_DEFAULT_TO
• Slack/Discord hooks:
  • slack_notify uses SLACK_WEBHOOK_URL
  • discord_notify uses DISCORD_WEBHOOK_URL
• Productivity tools:
  • timers: timer_create, timer_list, timer_cancel
  • reminders: reminder_create, reminder_list, reminder_complete
  • optional due reminder push dispatch: reminder_notify_due
  • calendar read helpers via Home Assistant: calendar_events, calendar_next_event
• Weather integration:
  • weather_lookup (Open-Meteo backend; WEATHER_UNITS=metric|imperial)
• Webhook integration:
  • webhook_trigger enforces https + WEBHOOK_ALLOWLIST domain policy
  • optional bearer token injection via WEBHOOK_AUTH_TOKEN
  • when identity enforcement is enabled, high-risk calls require approval_code or a trusted requester with approved=true
  • failed outbound webhook/channel/email/push attempts are queued for operator replay:
    • dead_letter_list to inspect queue state
    • dead_letter_replay to retry specific or filtered entries
• Proactive assistant workflows (proactive_assistant):
  • briefing, anomaly_scan, routine_suggestions, follow_through, event_digest
• Memory governance (memory_governance):
  • per-user partition overlays + duplication/contradiction/staleness audits + cleanup
• Identity and trust controls (identity_trust):
  • session confidence scoring, domain trust-policy management, guest-mode sessions, household profile admin
• Home orchestration (home_orchestrator):
  • intent-to-plan decomposition, preflighted multi-entity execution with partial failure reporting, area policy constraints, automation suggestions, long-running task tracking
  • automation pipeline: automation_create, automation_apply, automation_rollback, automation_status (supports dry-run diff previews)
• Skills governance (skills_governance):
  • capability negotiation, dependency health, quotas, harness runs, bundle signing metadata, sandbox templates
• Planning and autonomy (planner_engine):
  • planner/executor split output, task graphs with checkpoint/resume, deferred scheduling, self-critique
  • autonomy loop controls: autonomy_schedule, autonomy_checkpoint, autonomy_replan, autonomy_cycle, autonomy_status
  • autonomy cycle supports structured step contracts (precondition/postcondition), bounded retry with backoff, and automatic replan escalation with failure taxonomy in autonomy_status
• Quality and evaluation (quality_evaluator):
  • weekly report generation + deterministic dataset-runner summary
• Embodiment roadmap controls (embodiment_presence):
  • micro-expression library, user gaze calibration, adaptive gesture envelopes, privacy posture, motion safety envelope
• Integration workflows (integration_hub):
  • calendar CRUD policy flow, notes capture backends (including Notion when configured), messaging draft/review/send flow with channel dispatch, commute briefs, shopping orchestration, policy-gated research workflow
  • release-channel operations: release_channel_get, release_channel_set, release_channel_check

First-Time Operator Checklist

1. Copy .env.example to .env, then set required keys: OPENAI_API_KEY and ELEVENLABS_API_KEY.
2. If using integrations, set both values for each pair:
   • HASS_URL and HASS_TOKEN
   • PUSHOVER_API_TOKEN and PUSHOVER_USER_KEY
3. Choose explicit permission profiles before first run:
   • HOME_PERMISSION_PROFILE=readonly (recommended first boot)
   • TODOIST_PERMISSION_PROFILE=readonly
   • NOTIFICATION_PERMISSION_PROFILE=off
4. Run local validation gates:
   • make check
   • make test-faults
5. Start in simulation mode and confirm no startup warnings are emitted:
   • uv run python -m jarvis --sim --no-vision
6. If Home Assistant is enabled, run a dry_run=true smart-home request first before any live execute.

Usage

# Full Jarvis experience
uv run python -m jarvis

# Without face tracking (audio only)
uv run python -m jarvis --no-vision

# Text output instead of TTS (debugging)
uv run python -m jarvis --no-tts

# Simulation mode (no robot connected)
uv run python -m jarvis --sim

# Verbose logging
uv run python -m jarvis --debug

# Create a backup bundle (memory, audit logs, runtime state, operator settings)
uv run python -m jarvis --backup ~/.jarvis/backups/jarvis-$(date +%Y%m%d-%H%M%S).tar.gz

# Restore from a backup bundle (overwrite existing files)
uv run python -m jarvis --restore ~/.jarvis/backups/jarvis-20260227-120000.tar.gz --force

# Open operator console
open http://127.0.0.1:8765

Developer Checks

# Full lint + full test suite
make check

# Fast local regression pass
make test-fast

# Simulation-focused validation pass
make test-sim

# Fault-injection oriented subset (network, HTTP, summary, and storage taxonomy)
make test-faults

# Soak/stability subset
make test-soak

# Extended soak profile (simulation + fault profiles + checkpoint/retry validation)
make test-soak-extended

# Personality A/B drift checks (brevity + confirmation friction)
make test-personality

# Deployment/security gate (lint + tests + fault subset + workflow pin checks)
make security-gate

# Combined release-readiness gate (lint + acceptance + release checks + strict eval)
make readiness

# Marker-based subsets
uv run pytest -q -m fast
uv run pytest -q -m fault
uv run pytest -q -m slow

Equivalent scripts are available under scripts/:

• scripts/check.sh
• scripts/test_fast.sh
• scripts/test_sim.sh
• scripts/test_faults.sh
• scripts/test_soak.sh
• scripts/test_soak_extended.sh
• scripts/run_soak_profile.py
• scripts/test_personality.sh
• scripts/personality_ab_eval.py
• scripts/security_gate.sh
• scripts/jarvis_readiness.sh

CI runs the same lint + test gates on every push and pull request via ci.yml. Workflow linting and YAML hygiene run via workflow-sanity.yml. Nightly soak coverage is scheduled in nightly-soak.yml. Readiness-gate automation is scheduled/on-demand in jarvis-readiness.yml.

CI Workflow Intent and Failure Routing

Project Structure

jarvis/
├── pyproject.toml
├── .env.example
├── ~/.jarvis/audit.jsonl      # Auto-created audit log (runtime path)
├── src/
│   └── jarvis/
│       ├── __main__.py        # Entry point + conversation loop
│       ├── config.py          # Settings & env vars
│       ├── brain.py           # OpenAI Agents SDK orchestrator
│       ├── observability.py   # Telemetry store + metrics export
│       ├── operator_server.py # Local operator dashboard/API
│       ├── skills.py          # Local skill discovery + lifecycle
│       ├── presence.py        # 30Hz presence loop (the soul)
│       ├── tools/
│       │   ├── robot.py       # embody, play_emotion, play_dance
│       │   ├── services.py    # shared runtime/helpers + MCP tool registry
│       │   └── services_domains/
│       │       ├── home.py         # home_orchestrator domain handler
│       │       ├── planner.py      # planner_engine domain handler
│       │       ├── integrations.py # integration_hub domain handler
│       │       ├── comms.py        # channel/email/todoist/pushover handlers
│       │       ├── governance.py   # skills_governance + quality_evaluator + embodiment_presence
│       │       └── trust.py        # proactive_assistant + memory_governance + identity_trust
│       ├── audio/
│       │   ├── vad.py         # Silero voice activity detection
│       │   ├── stt.py         # faster-whisper transcription
│       │   └── tts.py         # ElevenLabs synthesis
│       ├── vision/
│       │   └── face_tracker.py  # YOLOv8 detection → presence signals
│       └── robot/
│           └── controller.py  # Reachy Mini SDK wrapper