Skip to main content
Feature documentation describes the checked-in implementation. Confirm the deployed commit and configuration before promising availability. Enable one vertical at a time in a non-production environment; the following is a conservative example, not a list of current Production values:

Interface gates

OpenAPI operations publish x-contact-button-required-abilities and x-contact-button-feature-flags from route middleware. Feature names in these extensions are configuration keys, such as multichannel_inbox, rather than environment-variable names. Service-level checks, workspace entitlements, and provider readiness still apply. For example, Agent streaming and action requests have additional runtime checks. A disabled route gate normally returns 404; it does not establish that a resource exists or is accessible. Inspect /v1, /mcp.json, the negotiated MCP capabilities and tools/list, and the current Page capabilities in the target environment. Global discovery lists Documentation and Workspace MCP; Page and direct Agent endpoints are selected from their resource public IDs.

Order

  1. Deploy additive migrations and API v1.
  2. Verify UUID backfills, tenant isolation, legacy embeds, and Page rendering.
  3. Enable public discovery and validate HTML, Markdown, manifest, sitemap, and Agent Card parity.
  4. Enable each engagement feature for internal canary workspaces. Enable surveys after testing dashboard publishing, public responses, and SMS/WhatsApp distribution preflight for a tenant. Enable agent form submissions only after testing a simple form and its OTP confirmation.
  5. Enable MCP, then A2A, only after authentication and confirmation smoke tests.
  6. Enable knowledge sync provider-by-provider.

Rollback

Disable the affected flag and clear config cache. Do not reverse a migration while newer application code may still write to it. Keep the additive schema through the observation period, then remove it in a later reviewed release only if the pivot is abandoned. Agent traces are sampled at 100% only when SENTRY_AGENT_MONITORING_ENABLED=true. Prompt and model-output capture is not implemented; enabling it requires a separate privacy approval.