This thread in the December archives remains unresolved.

See https://lists.w3.org/Archives/Member/w3c-xsl-wg/2013Dec/0028.html (member only) and subsequent thread.

We do not really have a clean way of doing multi-pass streaming in the current spec (other than relying on external frameworks like XProc to combine multiple stylesheets, of course).

The classic way of defining a multi-pass transformation in XSLT is to use variables:

<xsl:variable name="temp1">
 <xsl;stream href="input.xml">
   <some processing/>
 </xsl:stream>
</xsl;variable>
<xsl:apply-templates select="$temp1" mode="streamable"/>

But there is nothing here to say that $temp1 should be processed as a stream; indeed we encourage the idea that streamed nodes are never bound to variables.

I would be tempted to suggest something like

<xsl:pipeline>
  <xsl;stream href="input.xml">
    ...
  </xsl:stream>
  <xsl;stream>
    ...
  </xsl:stream>
</xsl:pipeline>

where each xsl:stream in the pipeline after the first takes its input from the result of the previous xsl:stream.

I'm trying to look around for a solution that doesn't involve new syntax, but it's hard to find one. 

* Adding "streamable="yes" to xsl;variable is a possibility but it's still new syntax, and the rules for what it means and how it can be used are potentially quite tricky.

* Some kind of coupling of xsl:result-document and xsl:stream might be possible:

  <xsl;result-document href="temp.xml" method="pipe">
    ...
  </xsl:result-document>
  <xsl;stream href="temp.xml">
    ...
  </xsl:stream>

* Or something like Saxon's next-in-chain:

<xsl:result-document next="mode2">
  <xsl:stream href="input.xml">
     ,,,
  </xsl:stream>
</xsl:result-document>

where the "next" attribute is a mode that is then used to process the result document.