Gemini Code Assist

Editions

Governance surfaces are identical across tiers.

Enable the service and grant roles

1. Cloud Console → Marketplace → enable cloudaicompanion.googleapis.com.
2. Cloud Console → Admin for Gemini (console.cloud.google.com/gemini-admin) → purchase and assign licenses.
3. IAM grants for end users: roles/cloudaicompanion.user + roles/serviceusage.serviceUsageConsumer. For auto-assignment on first use, grant cloudaicompanion.licenses.selfAssign.
4. Admin roles: roles/billing.admin or roles/consumerprocurement.orderAdmin for purchase; standard IAM admin for license management.
5. Optional custom-role permissions of interest: cloudaicompanion.companions.generateChat/generateCode, cloudaicompanion.instances.completeCode/generateCode/exportMetrics.

Runtime audit logging

Cloud Audit Logs under service cloudaicompanion.googleapis.com:

Data Access logs give per-user, per-method metadata (principal, timestamp, method, resource) but not prompt / response bodies. Enable via standard Cloud Logging Data access audit logs configuration on the log router.

Prompt-and-response capture — Configurable Code Assist logging

Google’s Security/Privacy page states Code Assist is stateless by default but offers optional logging of user input and responses to a Cloud Logging bucket. This is the closest analogue to Cursor’s admin dashboard export and is the only path for prompt/response fidelity in a Code Assist deployment.

Enable this from the Admin for Gemini page for your subscription. Route the resulting log entries through a sink to Arrakis:

1. Sink source: your Configurable Code Assist logging bucket.
2. Sink destination: BigQuery dataset (recommended for retention and JOIN with Data Access logs) or a Pub/Sub topic that a small OTLP shim converts to Arrakis coding-telemetry.
3. Grant the sink’s writer identity bigquery.dataEditor on the BigQuery dataset (or pubsub.publisher on the topic).

Arrakis reads from the sink destination — either via Data Warehouse Pull (BigQuery) or an OTLP push (Pub/Sub → OTLP shim). The exact shim configuration is provisioned per-tenant.

Preventive controls

• VPC Service Controls. Add both geminicloudassist.googleapis.com and cloudaicompanion.googleapis.com to the perimeter. Add ingress rules for developer identities.
• User Domain Restriction. Enforce at the Org policy layer to prevent OAuth sign-in from non-owned domains.
• .aiexclude files. Repo-scoped file filter that removes files/paths from Code Assist context. Client-side, committed to the repo.
• CMEK. Supported for the code-customization data plane.
• IP indemnification. Google’s Gemini indemnification covers Code Assist output.
• Recitation / citation filter. Long-form quotations from open-source cite the source; users can suppress citation matches from IDE settings. Not a preventive control per se — a licensing safety net.

Not first-class on Code Assist as of this writing: Vertex AI safety filters, Model Armor, and Sensitive Data Protection (DLP) integration. Code Assist does not run on the customer’s Vertex endpoint, so those Vertex-side controls don’t automatically apply. Customers who require DLP scanning of coding-assistant traffic need to attach it at the Cloud Logging sink layer (Arrakis-side DLP on the ingested content) rather than in the client path.

Endpoint-side reality check

Code Assist is delivered as an IDE plugin — VS Code Marketplace, JetBrains Marketplace, or bundled with Cloud Shell / Cloud Workstations. There is no native OpenTelemetry exporter and no documented managed-settings schema analogous to Claude Code or Gemini CLI. Client-side controls that exist:

• .aiexclude — repo-scoped file filter.
• IDE settings for citation-match suppression.
• IDE / IdE Marketplace policy: JetBrains supports pushed IDE settings and project settings; VS Code supports enterprise settings via --user-data-dir and workspace trust. Google does not publish a managed-settings schema for the plugin itself — these are IDE-layer controls, not Code-Assist-specific.

Arrakis integration path

• Metadata: Cloud Audit Log sink (Admin Activity + Data Access) → Arrakis ingest.
• Content: Configurable Code Assist logging → BigQuery or Pub/Sub → Arrakis ingest.
• Perimeter: VPC-SC + User Domain Restriction enforced on the customer’s Google Cloud project; no Arrakis action required beyond documenting the configuration.

All paths terminate at the same Arrakis coding-telemetry ingest as native-OTel sources.

Egress verification

For Cloud Logging sink destinations, verify the sink writer identity has permission on the destination and that a sample event flows end-to-end. For Pub/Sub-based paths, verify the OTLP shim is emitting to the tenant OTLP endpoint using the shared probe in OTel Collector & Arrakis Endpoint → Verifying connectivity.