fix(mcp): repair stdio startup diagnostics

[joestump]

Two independent bugs that leave maybeStdioErr unable to do its job, so a stdio MCP server that dies during startup reports EOF and nothing else. Split out of #3522 at a reviewer's convenience — neither has anything to do with that PR's feature.

The symptom

Misconfigure a stdio server — a package name that doesn't exist, or node not on PATH:

{ "mcp": { "demo": { "type": "stdio", "command": "npx", "args": ["-y", "not-a-real-package"] } } }

The server shows as errored with EOF. The child's actual complaint is discarded, which is the exact case stdioCheck was written for.

Bug 1 — the diagnostic re-run duplicates argv0

exec.Cmd.Args already carries argv0 as its first element, and exec.CommandContext prepends Path as argv0 itself. Passing Args through whole therefore re-runs npx npx -y pkg:

cmd := exec.CommandContext(ctx, old.Path, old.Args...) // "npx" + ["npx", "-y", "pkg"]

So on the occasions the diagnostic did run, it reported that malformed command's failure — typically cannot execute binary file, the shell trying to exec itself as a script — instead of the child's real output.

Bug 2 — the transport is never unwrapped

createSession wraps every transport in a channelTransport before Connect, and the error path then hands that wrapper to maybeStdioErr:

transport = &channelTransport{inner: transport, name: name, gate: channelGate}
// ...
session, err := client.Connect(mcpCtx, transport, nil)
if err != nil {
    err = maybeStdioErr(err, transport)   // always the wrapper, never the CommandTransport

The transport.(*mcp.CommandTransport) assertion inside can therefore never match, and the function returns the bare EOF unchanged. Every stdio startup diagnostic has been silently dead since transports started being decorated.

The unwrap goes through a small transportWrapper interface rather than a type switch on channelTransport, so a future decorator cannot quietly reintroduce this — failing to implement the interface is the only way back to the bug.

Notes

• Commits are ordered so each is green on its own: the argv0 fix lands first, because the unwrap's tests assert on the child's real output and would fail without it.
• Three tests, each verified to fail with only its own fix reverted: TestStdioCheck_DoesNotDuplicateArgv0, TestMaybeStdioErr_UnwrapsChannelTransport, and TestMaybeStdioErr_UnwrapsEveryWrapper (which stacks a second, test-local decorator to pin that the unwrap peels the whole chain rather than a fixed number of layers).
• go test -race -failfast ./... and golangci-lint run are green on this branch.
• Both fixes have been running in a downstream fork; this is the upstream-adapted version, cut fresh from main.

🤖 Posted on behalf of @joestump by claude-opus-5 using Claude Code.

[meowgorithm]

Nice!