[rmorehig]

User-authenticated MCP requests previously discovered clusters only through the default deployment. Discover clusters across enabled deployments, resolve routing through each cluster’s region ID, and send requests to deployment-specific public API hosts while retaining routing headers. Authoritative regional records take precedence over gateway duplicates in either response order.

Match the UI’s regional-outage handling: each deployment lookup has a five-second deadline; successful deployments return healthy clusters with unavailable_regions containing failed region IDs and names. Authentication failures and total outages remain errors. Requests for clusters absent from a partial catalog, or gateway stubs lacking both a region reference and valid routing metadata, refresh discovery and fail safely if the cluster is still undiscovered, preventing default-gateway fallback and allowing immediate recovery. The list-clusters tool description and README explain partial results.

Supports current 8-character and legacy 32-character deployment IDs. Custom API origins and API-key routing retain their existing behavior. Merged current main to preserve the newer workflow and query-tool fixes; release notes are under Unreleased.

Validation: all 275 tests passed; pnpm lint, pnpm build, and git diff --check passed. Coverage includes direct EU requests, duplicate deployments and gateway records, custom endpoints, API keys, legacy IDs, partial/total outages, DNS failures, timeouts, authorization errors, empty healthy catalogs, recovery, cluster-ID mutations, and gateway stubs during outages in both catalog orders. An in-memory MCP tool check using the compiled server and a simulated stalled regional transport returned partial results after 5.02 seconds using the real deadline, then routed to the recovered region. Final structured autoreview reported no actionable findings.

Related Platform changes: https://github.com/rawtreedb/rawtree-platform/pull/1566, https://github.com/rawtreedb/rawtree-platform/pull/1575, and https://github.com/rawtreedb/rawtree-platform/pull/1637.

Companion CLI change: rawtreedb/rawtree-cli#85.

---

Note

Medium Risk
Changes core request routing and cluster discovery for OAuth/session users; mis-routing or stale cache could send operations to the wrong deployment, though extensive tests cover outages and recovery.

Overview
User-authenticated clients no longer list clusters only through the default API gateway. listClusters fans out to each enabled regional deployment (deduped by routing headers), merges catalogs with authoritative region_id metadata over gateway duplicates, and attaches unavailable_regions when some deployments fail within a 5s per-target deadline. rt_ API keys keep the single-gateway behavior.

Scoped cluster operations gain deploymentApiUrl, rewriting api.rawtree.com / api.rawtree.dev to deployment-specific public hosts while still sending X-RawTree-Deployment-Id / X-RawTree-Region. Custom API origins are unchanged. ClusterRouting refuses gateway-only stubs during partial discovery, invalidates cache for retry after recovery, and errors safely when a cluster is missing from a partial catalog. Legacy 32-character deployment IDs are accepted alongside 8-character IDs.

Docs and the list-clusters MCP tool description explain partial results and retry expectations.

^{Reviewed by Cursor Bugbot for commit a2ccc5e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4cc4c01. Configure here.}