vMCP: a backend's protocol revision is probed on every Legacy connect, but a cached Modern backend that stops speaking Modern is never reclassified

未关闭
#7,163 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
48/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
go

调研方向

Start by reading vMCP's probeRevision, dispatch, and isRevisionMismatch paths, plus mcpcompat's Initialize and the revision cache behavior. Trace how Legacy and Modern requests handle protocol-version errors, then check the relevant health-check and client connection tests. Done means cached Legacy backends avoid repeated Modern probes and cached Modern backends can recover when a backend serves only Legacy revisions.

由索引模型根据 Issue 内容生成。

描述

needs-triage

Both observations are from vMCP v0.63.0 (toolhive-core v0.0.58, go-sdk v1.8.0), with a backend that runs the TypeScript MCP SDK's streamable HTTP server: firecrawl-mcp 3.37.0, whose CLI serves only the 2025 protocol revisions.

1. A backend cached as Legacy still gets a server/discover on every connect

probeRevision classifies such a backend as Legacy once and caches that. The Legacy path then builds a client through mcpcompat/client.Client.Initialize, which calls gosdk.Client.Connect(ctx, tr, nil). With nil ClientSessionOptions, go-sdk uses its latest version, 2026-07-28, so Connect first POSTs server/discover with MCP-Protocol-Version: 2026-07-28 and falls back to initialize only after that fails. InitializeParams.ProtocolVersion, which vMCP sets, is not passed on to Connect.

So every Legacy connect, including the health check (ListCapabilities every 30 s by default) and every session's backend connect, sends a request the backend is already known to reject. The TypeScript SDK backend answers each one with HTTP 400 (-32000 Bad Request: Unsupported protocol version: 2026-07-28 ...) and logs it at ERROR with a stack trace. Tool calls succeed after the fallback, but one such backend produced several thousand ERROR log entries a day from health checks and connects alone.

Suggestion: when the cached revision is Legacy, connect with the Legacy version directly, for example by passing &gosdk.ClientSessionOptions{ProtocolVersion: request.Params.ProtocolVersion} from mcpcompat Initialize and having vMCP request 2025-11-25 for a backend cached as Legacy. The legacyInit cache flip that turns a Legacy backend into a Modern one when initialize negotiates 2026-07-28 would then need another trigger, such as an occasional re-probe, bounded the way #6185 bounds the refuted-Modern-hint memory with a TTL.

2. A cached Modern backend that answers Modern requests with a generic JSON-RPC error is never reclassified

The reverse direction. Once a backend is cached Modern, dispatch reclassifies it only when isRevisionMismatch holds: errWrongEra, errLegacyResponseBody or errModernNegotiatedDown. A Legacy-only TypeScript SDK backend does not produce any of those. It answers a Modern request with HTTP 400 and a well-formed JSON-RPC error body (code: -32000, message Bad Request: Unsupported protocol version: 2026-07-28 ...), which readModernEnvelope decodes and interpretModernResult returns as a generic modern <method>: rpc error -32000. That is not a mismatch, so the error goes to the caller, nothing re-probes, and the revision cache has no TTL.

We hit the setup for this when enabling 2026-07-28 on such a backend, and it matters in two ordinary operations:

  • Rolling back the backend to a build that serves only the 2025 revisions. Every tool call and health check to it fails until the vMCP pod restarts.
  • Rolling out a backend with more than one replica. Once one new pod answers server/discover, vMCP caches Modern, and requests that reach a pod still on the old build fail until that pod is replaced.

The same 400 on the first probe of an uncached backend is classified Legacy correctly (probeRevision's default branch), so only the cached path is affected.

Suggestion: treat an HTTP 400 on a Modern request whose JSON-RPC error is not one of the Modern codes (-32020 to -32022, -32601) as a wrong-era signal, as a bare 400 already is. server/discover has no side effects and a 400 from a Legacy transport means the request was rejected before any handler ran, so retrying under Legacy is safe in that case. Alternatively, re-probe on a failed health check of a cached Modern backend.

主要语言
Go
星标
2.3k
派生
332
平均合并
19 小时 39 分钟
30 天内合并 PR
174

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

stacklok/toolhive 的其他 Issue

查看 stacklok/toolhive 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。