fix(cache): prevent Connection closed errors during SWR background revalidation

[tlgimenes]

What is this contribution about?

Fixes McpError: MCP error -32000: Connection closed errors that occur during background cache revalidation for HTTP/SSE MCP connections.

Root cause: fetchWithCache implements stale-while-revalidate (SWR) — on cache hit, it returns cached data immediately and fires a background fetchLive() to refresh the cache. For HTTP/SSE connections, fetchLive() creates clients via the per-request pool (ctx.getOrCreateClient). The pool's connections get closed before the background revalidation completes, causing the error.

Fix (two layers):

1. Keep pool alive: Add pendingRevalidations array to MeshContext. Background revalidation promises are registered via an onRevalidation callback. A Hono middleware fire-and-forgets awaiting these promises after the handler returns, keeping ctx (and its client pool) alive via closure while revalidations complete. 30s timeout prevents hung revalidations.
2. Defense-in-depth: Suppress "Connection closed" errors in the revalidation catch handler (consistent with existing isMethodNotFound handling), since SWR revalidation is best-effort.

No pool disposal is added — SSE/streaming proxy connections depend on pool clients remaining open after the handler returns.

Screenshots/Demonstration

N/A — backend-only change.

How to Test

1. Deploy to a staging environment with HTTP MCP connections configured
2. Make requests to a Virtual MCP endpoint that triggers SWR cache hits (connections with cached tool/resource/prompt lists)
3. Verify no more [fetchWithCache] ... background revalidation FAILED: McpError: MCP error -32000: Connection closed errors in logs
4. Run bun test apps/mesh/src/mcp-clients/mcp-list-cache.test.ts — all 27 tests should pass
5. Verify SSE/streaming proxy connections still work correctly

Migration Notes

N/A — no database or configuration changes.

Review Checklist

• PR title is clear and descriptive
• Changes are tested and working
• Documentation is updated (if needed)
• No breaking changes

---

Summary by cubic

Fixes MCP error -32000 "Connection closed" during SWR background revalidation for HTTP/SSE MCP connections by keeping the per-request client pool alive and treating revalidation as best-effort. Also skips redundant revalidations in the connections list to reduce noise and background work.

• Bug Fixes
  • Added pendingRevalidations to MeshContext and an onRevalidation callback in fetchWithCache; all call sites register revalidation promises, and hono middleware fire-and-forgets a 30s await after the handler returns, now wrapped in try/finally to keep the client pool alive.
  • fetchWithCache now suppresses "Connection closed" errors during revalidation and continues handling MethodNotFound by caching empty lists.
  • Skipped SWR revalidation in COLLECTION_CONNECTIONS_LIST; the GET tool already revalidates per-connection, preventing unnecessary fan-out and potential connection closures.

^{Written for commit be97752. Summary will update on new commits.}

[github-actions]

🧪 Benchmark

Should we run the Virtual MCP strategy benchmark for this PR?

React with 👍 to run the benchmark.

Reaction	Action
👍	Run quick benchmark (10 & 128 tools)

Benchmark will run on the next push after you react.

[github-actions]

Release Options

Suggested: Patch (2.204.4) — based on fix: prefix

React with an emoji to override the release type:

Reaction	Type	Next Version
👍	Prerelease	2.204.4-alpha.1
🎉	Patch	2.204.4
❤️	Minor	2.205.0
🚀	Major	3.0.0

Current version: 2.204.3

Note: If multiple reactions exist, the smallest bump wins. If no reactions, the suggested bump is used (default: minor).

[cubic-dev-ai]

1 issue found across 13 files

Prompt for AI agents (unresolved issues)

 Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately. <file name="apps/mesh/src/api/app.ts"> <violation number="1" location="apps/mesh/src/api/app.ts:989"> P1: Wrap `await next()` in `try/finally` so pending revalidation handling still runs when downstream handlers throw.</violation> </file> 

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}