Single-assistant rollout

Operator guidance for developers deploying a new version of the account-mode feature. User instructions are in Single assistant and ensembles.

Release and rollback runbook

Single-assistant mode changes durable profile state, so treat code rollout, account switching, and changing the new-account default as separate actions.

Before rollout:

Build and test one exact candidate hash outside the live source checkout.
Take and verify a complete data snapshot, including user profiles, agents, sessions, assignments, memories, tasks, credentials, and configuration.
Rehearse upgrade, single → ensemble → single round-trip, and snapshot restoration against an isolated copy of representative profile data.
Keep the previous known-good build and the pre-upgrade snapshot available.

For a canary, start with one adult account. Confirm the mode round-trip, primary selection, browser refresh, voice and scheduled entry points, worker completion, permissions, and the account’s real integrations before expanding. Do not use a child account as the first canary.

The fastest behavioral rollback is to switch the canary account to ensemble. For a full rollback, restore the previous build and the matching pre-upgrade data snapshot during an approved maintenance window. Do not restore only the code after the new build has written profile state.

Monitor policy fallback warnings, unknown-agent errors, permission denials, tool-recovery rate, worker completion and duplicate-effect signals, bound-MCP refusals, token use, and p50/p95 latency throughout the canary.

The Run Inspector records the exact provider-native tool field prepared for each logical model round. A cost result is promotable only when the inspector marks both token totals and request/completion/usage cardinality complete. OpenAI Responses currently supplies that exact cardinality. Other provider adapters retain their schema evidence but do not yet qualify for an exact cost pass; LM Studio can additionally report estimated token totals, which are shown as estimates and never counted as exact evidence.