mirror of
https://github.com/simstudioai/sim.git
synced 2026-09-24 15:45:35 +08:00
WaleedandClaude Opus 4.7
0c96964f8f
* improvement(mcp): per-server tool queries + negative cache so one slow server can't block the workspace Move MCP tool discovery off the workspace-aggregated `Promise.all` fan-out and onto per-server React Query keys, matching how Cursor and Claude Code render remote MCP. `useMcpToolsQuery` is now a `useQueries` combiner: each server has its own cache entry, its own loading state, and a slow neighbor never gates the others. Public shape stays compatible with existing consumers. Add a short-TTL negative cache: when `listTools` fails (timeout, connection error, etc.) we mark the server unhealthy for 30s so subsequent discovery calls short-circuit instead of re-paying the timeout. OAuth-required errors are exempt so re-auth retries immediately. Drop `LIST_TOOLS_TIMEOUT_MS` from 30s to 10s to bound the worst-case first failure. Invalidations are per-server where the action is per-server (OAuth popup, per-server SSE event, refresh, update, delete). Bulk operations stay workspace-broad. Adds tests for the negative-cache behavior and the OAuth exemption. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * improvement(mcp): in-flight dedup + 2min negative TTL + no refetch on window focus Three small follow-ups on top of per-server tool queries: - Coalesce concurrent `discoverServerTools(userId, serverId, workspaceId)` calls into a single upstream `tools/list`. Races between OAuth-callback cache priming, post-OAuth UI refetch, and multi-tab loads no longer double-fetch the same server. - Bump negative-cache TTL from 30s to 2 minutes. Cleared on listChanged, OAuth completion, manual refresh, and the next successful discovery, so this floor only matters for genuinely dead servers — drops their floor traffic by 4x. - Disable `refetchOnWindowFocus` on per-server tool queries. listChanged SSE + mutation invalidations already cover real schema changes; alt-tab no longer triggers N parallel `tools/list` calls. * fix(mcp): address bugbot/greptile review on per-server tool discovery - Workspace-scoped `mcpKeys.serverToolsWorkspace(workspaceId)` prefix for bulk invalidations (create-server, refresh-all, SSE workspace fallback, OAuth fallback). The previous `mcpKeys.serverTools()` prefix was global and invalidated every workspace's tools cache. - `useMcpToolsQuery` folds `useMcpServers().isLoading` into the aggregate `isLoading` so mounting no longer flashes an "empty tools" state during the servers-list fetch. Aggregate `error` is suppressed when any per-server query already returned data so one slow server can't blank out the others. - `useForceRefreshMcpTools` invalidates the per-server query keys of servers whose force refresh failed, so stale tools don't linger. - `DiscoveryOutcome` error variant carries the original error, restoring the OAuth-exemption check that `getErrorMessage(...)` previously erased. - `discoverServerTools(userId, serverId, workspaceId, forceRefresh = false)` now consults the positive + negative cache by default. Per-server React Query refetches hit the cache instead of re-paying the listTools timeout; callers that explicitly bypass cache (refresh route, OAuth callback, bulk POST refresh) pass `forceRefresh: true`. Negative-cache hits throw a typed `McpConnectionError` so the route layer can surface a fast 503. * update icons * chore(mcp): remove dead query keys and trim verbose comments - Drop unused `mcpKeys.tools()` / `mcpKeys.toolsList()` — replaced by per-server keys, no remaining callers. - Trim narrative comments to keep only the non-obvious "why" notes. * fix(mcp): map negative-cache cooldown error to HTTP 503 `McpConnectionError` thrown when a server is in cooldown previously fell through `categorizeError` to a generic 500. Cooldown is a transient-unavailability condition, so route it to 503. * test(mcp): cover cooldown error → 503 categorization * fix(mcp): address second-round bugbot review - useMcpToolsQuery serverIds: filter on enabled + workspaceId match. Disabled rows no longer trigger discover calls that get negative-cached, and keepPreviousData on useMcpServers no longer races a workspace switch into cross-workspace discover requests. - Aggregate skips per-server data when that server's latest refetch errored, so a broken server's last-known tools no longer linger in the workspace view while its card shows an error. - discoverServerTools failure path drops the positive cache alongside writing the negative-cache marker. A cache-respecting follow-up now fails fast via cooldown instead of returning stale tools from a now-broken server. - useMcpTools.refreshTools drops the dead forceRefresh param — the per-server queryFn always sends refresh=false, so the flag was never effective. Callers wanting cache-bypass should use useForceRefreshMcpTools. * chore(mcp): trim verbose comments * fix(mcp): third-round bugbot review - discoverTools failure path now drops the per-server positive cache alongside writing the negative-cache marker, matching discoverServerTools' behavior so a workspace-aggregate failure doesn't leave stale tools cached. - useForceRefreshMcpTools filters disabled and out-of-workspace rows before fan-out so disabled servers don't 404 → negative-cache themselves. - Remove unused useMcpServerTools export — the aggregate goes through useQueries directly, no external consumer exists. --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>