perf(mcp): cache external tool discovery for search

[reachjalil]

Summary

• add a 60-second process-local cache for external MCP tool discovery used by search
• coalesce concurrent misses and avoid caching failures
• retain complete sanitized tool schemas so argumentsSchema, schemaDigest, and invocation metadata stay current
• key entries by organization, connection, principal, endpoint, auth and credential mode, connection version, and connected-account version
• bound retention to 256 entries and 16 MiB with LRU eviction; oversized catalogs are not retained

Why this version

The useful latency goal from #2570 is preserved without reviving its large persisted-manifest migration. Authorization and credential checks remain live, execution still performs live schema discovery, and only search discovery is cached.

Checks

• bun test test/external-mcp-search-cache.test.ts — 5 passed
• focused integration: repeated search variants reuse the full current external tool catalog — 1 passed
• benchmark fixture: cold 20.38 ms / 4 remote requests; warm 2.79 ms / 0 remote requests
• pnpm --filter @openwork-ee/den-api exec tsc --noEmit -p tsconfig.json
• git diff --check

Risk and gaps

• cache lifetime is process-local, so cold starts still discover tools remotely
• the benchmark is a deterministic local fixture, not a production latency claim
• backend-only change; no UI manual journey applies
• full Den API suite was not run in this branch

Supersedes #2570.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
openwork-landing	Ready	Preview, Comment, Open in v0	Jul 21, 2026 2:02pm

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
openwork-diagnostics	Skipped		Jul 21, 2026 2:02pm

[vercel]

@reachjalil is attempting to deploy a commit to the Different AI Team on Vercel.

A member of the Team first needs to authorize it.