Gemini Runtime Configuration — AI Access & Model Governance
How Hussh One governs AI access and Gemini model execution: managed Vertex AI vs. BYOK (Google AI Studio / Google Cloud Vertex), PCHP consent boundaries, workload tiering, and privacy-preserving credential probing.
TL;DR: Hussh One decouples AI access into a governed runtime configuration layer beneath One, supporting managed Vertex AI and BYOK credentials. Typed turns use canonicalized Gemini 3.7 Flash with bounded reasoning, while live voice uses Gemini Live native audio without leaking keys, data, or authority.
Status as of 2026-08-14: see body.
Relations
Overview & AI Access Boundary
AI access in Hussh One is an explicit private runtime configuration boundary rather than an additional specialist subagent. It establishes the reasoning and execution posture for the user's private agent across typed chat, specialist orchestration, and bidirectional live voice.
Configuration appears during initial setup and can be reopened in connection settings. It operates strictly as infrastructure governance: it does not publish public actions, does not appear as a standalone dashboard tile, and never bypasses Personal Consent Handshake Protocol (PCHP) authority gates.
Runtime Access Modes
The runtime provides two distinct modes for model intelligence:
| Mode | User Input | Credential Storage | Operational Scope |
|---|---|---|---|
hushh_managed_vertex | None | No user secret | Default typed and live private-agent experience through managed workload identity. |
byok | Google AI Studio Gemini API key or Google Cloud Vertex API key (with project & location) | Encrypted PKM only | Typed private-agent turns; live voice is an explicit, separately gated compatibility path. |
Explicit Key Transport
Key transport is strictly typed and explicit:
- Developer API (
developer_api): Uses the Google AI Studio endpoint for direct personal keys. - Vertex API Key (
vertex_api_key): Uses Google Cloud Vertex endpoints, requiring explicit project and region bindings.
A credential is never guessed or classified by string shape, and service account JSON files are never accepted in client setup.
Lifecycle & Zero-Leakage Credential Probing
- User Mode Selection: A user selects managed Gemini or BYOK in AI access settings.
- In-Memory Staging: During setup, a BYOK credential is held transiently in process memory. Any page refresh, vault lock, sign-out, or account deletion wipes it immediately.
- Readiness Probe: The backend executes a bounded generation probe against the selected Google endpoint using the canonical
gemini-3.7-flashmodel before persisting the key. A successful probe verifies authentication, quota availability, and minimal generation capabilities. - Encrypted PKM Storage: Finishing setup unlocks the canonical user vault and writes the verified key into encrypted Personal Knowledge Model (PKM) storage. Raw keys never touch URLs, relay tickets, browser storage, native app preferences, relational database tables, application logs, or model prompts.
- Turn-Local Resolution: Typed private-agent turns resolve the unlocked key strictly for that individual request.
- Session-Local Live Voice: Live voice relays receive credentials only in the initial authenticated WebSocket frame to instantiate a connection-local runner, dropping raw references immediately.
Model Capability Matrix & Canonicalization
The agent runtime standardizes model tiers to optimize latency, cost, and reasoning fidelity:
- Primary Reasoning & Specialists (
gemini-3.7-flash): Serves as the GA reasoning engine across typed Agent Chat, manifest-guided action selection, and all domain specialist agents. Requests run over global Vertex endpoints with legacy sampling controls omitted andthinking_levelretained for bounded reasoning. - Bounded Reduction & Probes (
gemini-3.1-flash-lite): Handles lightweight summary reduction, receipt extraction, and background readiness probes at minimal latency. - Bidirectional Live Audio (
gemini-live-2.5-flash-native-audio): Powers real-time, full-duplex speech in the persistent Agent Bar, served regionally with native audio streaming.
Privacy & Security Principles
- Consent Isolation: Audio transport and model access never grant implicit access to personal records. PKM memories, connected system integrations, and personal data exports remain strictly gated by PCHP consent receipts and runtime verification.
- Privacy by Omission: When optional services or private operational split pages are unauthenticated or unavailable, references are omitted entirely to prevent dead ends and unauthorized leakage.