The prompt handler in registerPromptsFromBackend duplicated the connect → send → error-check → unmarshal pattern that executeBackendToolCall was introduced to centralize, creating two diverging implementations with inconsistent error messages.

Changes

• internal/server/unified.go: Adds executeBackendRequest[T any] — a generic helper that owns the full backend RPC lifecycle for any method/result type. Refactors executeBackendToolCall into a thin wrapper delegating to it:

func executeBackendRequest[T any](ctx context.Context, l *launcher.Launcher, serverID, sessionID, method string, params map[string]interface{}) (T, error) {
    var zero T
    conn, err := launcher.GetOrLaunchForSession(l, serverID, sessionID)
    // ... connect → send → error-check → unmarshal
    return result, nil
}

func executeBackendToolCall(...) (interface{}, error) {
    logUnified.Printf("executeBackendToolCall: serverID=%s, toolName=%s", serverID, toolName)
    return executeBackendRequest[interface{}](ctx, l, serverID, sessionID, "tools/call", map[string]interface{}{
        "name": toolName, "arguments": args,
    })
}

• internal/server/tool_registry.go: Replaces the ~20-line inline duplicate in the prompt handler closure with a single call:

result, err := executeBackendRequest[sdk.GetPromptResult](ctx, us.launcher, serverIDCopy, sessionID, "prompts/get", params)

Future backend capability types (e.g. resources/read) can now reuse executeBackendRequest with zero boilerplate, and error handling improvements apply uniformly across all backend call sites.