Gateway kill switch
Fail-closed controls for RaksHex-routed traffic, with durable PostgreSQL state and Redis propagation.
Workspace, identity, project and agent scoped kill-switch state can be evaluated before a RaksHex-routed provider request. The gateway reconciles low-latency Redis state with durable PostgreSQL state so a cache miss must not silently clear an active durable switch.
Budgets
Hard gateway budgets apply to traffic routed through the RaksHex gateway. monitor_only budgets are visibility/alerting and must not be described as blocks. provider_native enforcement is attempted only for provider capabilities that genuinely support it.
Dynamic team AI pool
When enabled on the workspace gateway budget, identities can borrow unused capacity from teammates (personal → team shared → member shareable → emergency reserve). Borrowed amounts appear in x-rakshex-pool-borrowed-usd on allowed gateway responses.
Critical boundary
A RaksHex gateway kill switch cannot truthfully be described as disabling direct provider traffic that bypasses RaksHex. Provider-side revocation/limits require a supported, authorised provider-native control.
Runtime authorization of consequential agent actions is the separate Agent Firewall path.