fix(mcp): cache result of discoverServerTools to prevent post-OAuth refetch storm

[waleedlatif1]

Summary

• OAuth callback was calling discoverServerTools to populate tools immediately after auth, but the result was never written to the cache — so the UI's immediate refetch missed cache and tried to live-fetch again
• On servers with slow upstreams (PlanetScale taking 60s to respond to tools/list from prod ECS), this left the UI in permanent "Loading…" forever, refetching every few seconds and timing out each time
• Fix: have discoverServerTools write its successful result to the per-server cache key, matching what discoverTools already does in its fetched branch
• Removed the now-redundant clearCache(workspaceId) from the OAuth callback — it was wiping cached tools for OTHER servers in the workspace unnecessarily; the new cache write replaces just the one server we care about

Type of Change

• Bug fix

Testing

Tested manually by reproducing prod symptom on staging — confirmed OAuth flow leaves UI in "Loading…" indefinitely before the fix, and renders the tools list instantly after.
Added unit test: discoverServerTools primes per-server cache so follow-up discoverTools hits cache (no live re-fetch).
All 286 MCP tests pass.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		May 21, 2026 5:55pm

[cursor]

PR Summary

Medium Risk
Moderate risk: changes tool discovery side effects (cache writes and DB status updates) and introduces a new tools/list timeout, which could affect perceived server availability if mis-tuned.

Overview
Prevents the post-OAuth UI refetch loop by having discoverServerTools write successful tools/list results into the per-server cache (and update server status/tool count) so the next discoverTools call hits cache instead of re-fetching live.

Removes the OAuth callback’s workspace-wide cache clear, avoiding unintended invalidation of other servers’ cached tools, and caps McpClient.listTools with a dedicated LIST_TOOLS_TIMEOUT_MS to keep slow upstream metadata calls from hanging the UI. Adds unit coverage ensuring discoverServerTools primes cache, plus minor test-mock updates to support .limit() and the new timeout constant.

^{Reviewed by Cursor Bugbot for commit 4e6210c. Configure here.}

[greptile-apps]

Greptile Summary

This PR fixes a post-OAuth "Loading…" loop on MCP servers: discoverServerTools was live-fetching tools after auth but never writing to the cache, causing the UI's immediate follow-up discoverTools call to miss the cache and re-fetch — timing out repeatedly on slow upstreams. It also adds a 30 s hard timeout to listTools to prevent indefinite hangs.

• Cache priming: discoverServerTools now writes its result to the per-server cache key and calls updateServerStatus (marking the row connected with fresh toolCount/lastToolsRefresh), matching the behaviour already present in discoverTools.
• Targeted cache invalidation: The broad clearCache(workspaceId) in the OAuth callback (which wiped every server's cached tools) is replaced by the single-server cache write inside discoverServerTools.
• listTools timeout: A new LIST_TOOLS_TIMEOUT_MS: 30_000 constant caps the MCP SDK listTools call so a slow upstream cannot stall the entire discovery loop indefinitely.

Confidence Score: 4/5

Safe to merge; the core OAuth cache-miss fix is correct and well-tested, with one minor gap in error observability.

The fix is well-scoped and the new test validates end-to-end cache priming. The one remaining gap — no warning logged when updateServerStatus rejects inside allSettled — means a transient DB failure would leave the server row stale and produce no log signal, which could complicate future incident diagnosis.

apps/sim/lib/mcp/service.ts — the updateServerStatus arm in the new Promise.allSettled block has no .catch() logger, unlike the cache-write arm beside it.

Important Files Changed

Filename	Overview
apps/sim/lib/mcp/service.ts	Core fix: discoverServerTools now primes the per-server cache and calls updateServerStatus after a successful live fetch, resolving the post-OAuth "Loading…" loop and the previously-flagged missing DB status update.
apps/sim/lib/mcp/client.ts	Adds a per-request LIST_TOOLS_TIMEOUT_MS cap (30 s) to listTools so a slow upstream can no longer block the UI indefinitely.
apps/sim/lib/mcp/utils.ts	Adds LIST_TOOLS_TIMEOUT_MS: 30_000 to MCP_CLIENT_CONSTANTS for use in the new per-call timeout on listTools.
apps/sim/app/api/mcp/oauth/callback/route.ts	Removes the broad clearCache(workspaceId) call (which wiped tools for all servers) and relies on discoverServerTools to write only the authenticated server's cache entry.
apps/sim/lib/mcp/service.test.ts	Adds a regression test verifying that discoverServerTools primes the per-server cache so a follow-up discoverTools doesn't trigger a live re-fetch; also fixes the mock to support .limit() chains.
apps/sim/lib/mcp/client.test.ts	Adds DEFAULT_EXECUTION_TIMEOUT_MS to the execution-limits mock so the new constant import in client.ts doesn't break tests.

Sequence Diagram

sequenceDiagram
    participant UI
    participant OAuthCallback as OAuth Callback
    participant Service as McpService
    participant Cache as Per-Server Cache
    participant DB

    UI->>OAuthCallback: "GET /api/mcp/oauth/callback?code=..."
    OAuthCallback->>DB: Exchange code to store token
    OAuthCallback->>Service: discoverServerTools(userId, serverId, workspaceId)
    Service->>DB: getServerConfig(serverId)
    Service->>Service: resolveConfigEnvVars + createClient
    Service->>Service: client.listTools() timeout 30s
    Service->>Cache: set(serverCacheKey, tools, 5min)
    Service->>DB: updateServerStatus(connected, toolCount)
    Service-->>OAuthCallback: tools[]
    OAuthCallback-->>UI: htmlClose Connected

    UI->>Service: discoverTools(userId, workspaceId)
    Service->>Cache: get(serverCacheKey)
    Cache-->>Service: tools[] cache HIT
    Service-->>UI: tools[] no live re-fetch

Loading

_{Reviews (2): Last reviewed commit: "improvement(mcp): update server status f..." | Re-trigger Greptile}

[waleedlatif1]

@greptile

[waleedlatif1]

@cursor review

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 4e6210c. Configure here.}