[roomote-roomote]

​Opened on behalf of @daniel-lxs. Follow up by mentioning @roomote-roomote, in the web UI, or in Telegram.

Related issue

Internal product improvement; no separate issue.

Why this PR exists

• A maintainer explicitly invited this PR in the linked issue or discussion
• I am a maintainer / this is internal Roomote work

What changed

Add Test connection inside a saved custom remote MCP server's Manage tools dialog. It explicitly fetches fresh tool metadata and reports named success/failure, including a valid empty catalog. Previously there was no successful-dialog retest action, and reopening could reuse cache.

Reuse the existing authorized, credential-aware, SSRF-guarded endpoint. Pending requests coalesce; testing does not execute tools or change enablement/approval choices. No new endpoint, persistent state or dependency. Public docs explain the remote-metadata scope and stdio limitation.

How it was tested

• Baseline develop fb5e231d: genuine public DeepWiki catalog loads three schemas but has no Test action; the fresh-cache regression fails for that missing capability.
• Final head 16df38eebd20e795784579e96d98f474ccb33b4c repairs the review-found observer race by awaiting the captured server-keyed fetch. Both opposite-outcome regressions fail before repair and pass after it. 37 client + 21 server command tests, web types, docs build/links/render and actual commit/pre-push gates pass. Full local suite not run.
• Native desktop1440×1000/mobile390×844: real public metadata success; request abort → Test failure → remove cause → Test recovery; existing Retry recovery. Final-head private placement repeats Test/failure/recovery/Retry, with database enable/tool-policy values unchanged. Four synthetic fixture UUIDs removed; all transport overrides restored.
• Final-head race: a browser fetch timing gate delays an aborted A request's original rejection until B's genuine public catalog loads. Releasing A correctly reports A failed while B stays successful. No successful response or UI state is fabricated; this is controlled scheduling/transport proof, not an upstream outage. No tool execution or external user messages.
• Latest three images opened/inspected and independently checked; full committed patch matches the proof snapshot. Mobile diagnostic/Retry and complete catalog are unobstructed. Stills prove final appearance, not timing or freshness. No real-device, screen-reader, dark-theme or live OAuth/header-auth provider claim.
• Implementation CI is green, exact-head automated re-review clean, and the review thread resolved. Normal human review remains required.

Screenshots

Genuine local app and public schemas, with disclosed delayed/aborted transport for negative cases. Latest final-head evidence replaces earlier captures.

Related PRs

• #3651: independent pending custom-skill edit protection.
• #3652: independent contextual automation/settings Edit names.

Checklist

• The PR title follows the repo convention: [Fix], [Feat], [Improve], [Refactor], [Docs], or [Chore] followed by a user-facing description
• This PR is small and scoped to one change
• pnpm lint and pnpm check-types pass locally — normal broad fast equivalents and actual hooks passed
• I added tests or included a clear manual validation note above
• I removed secrets, tokens, private keys, and customer data from code, logs, and screenshots
• If this change should appear in the changelog, I ran pnpm changeset — no changeset added

[roomote-community]

No new code issues found. The prior finding is fixed: connection tests now await the server-keyed fetch, and regression coverage checks both opposite-outcome switches. An isolated TanStack Query probe confirmed that each fetch retains its original outcome after switching observers. Lint, type-check, docs, Knip and CodeQL checks passed; tests and the app build were still running when checked. See session

• P2: Keep connection-test outcomes scoped to the tested server — apps/web/src/components/settings/CustomMcpServers.tsx:790–794: switching servers during a pending test can report the second server's outcome under the first server's name, including a false success for a failed test.

_{Reviewed 16df38e}

[daniel-lxs]

Reviewed head 16df38e. The connection-test race is addressed by awaiting the original server-keyed fetch, with regression coverage for both opposite outcomes. All review feedback is addressed and required checks pass.