ikiwiki/doc/todo/Resolve_native_reStructuredText_links_to_ikiwiki_pages.mdwn

_NB! this page has been refactored, hopefully it is clearer now_  
_I propose putting discussion posts somewhere in the vincity of
the secttion Individual reStructuredText Issues_

## Design ##

**Goal**

To be able to use rst as a first-class markup language in ikiwiki. I think
most believe this is almost impossible (ikiwiki is built around markdown).

## Wikilinks ##

**WikiLinks**, first and foremost, are needed for a wiki. rST already allows
specifying absolue and relative URL links, and relative links can be used to
tie together wiki of rst documents.

1. Below are links to a small, working implementation for resolving
   undefined rST references using ikiwiki's mechanism. This is **Proposal 1**
   for rst WikiLinks.

2. Looking over at rST-using systems such as trac and MoinMoin; I think it
   would be wiser to implement wikilinks by the `:role:` mechanism, together
   with allowing a custom URL scheme to point to wiki links. This is
   **Proposal 2**.

        This is a simple wiki page, with :wiki:`WikiLinks` and other_ links
        
        .. _other: wiki:wikilink

        We can get rid of the role part as well for WikiLinks::
        
            .. default-role:: wiki
        
        Enables `WikiLinks` but does not impact references such as ``other``
        This can be made the default for ikiwiki.

Benefits of using a `:role:` and a `wiki: page/subpage` URL scheme are
following:

1. rST documents taken out of the context (the wiki) will not fail as bad as
   if they have lots of Proposal-1 links: They look just the same as valid
   references, and you have to edit them all.
   In contrast, should the `:wiki:` role disappear, one line is enough
   to redefined it and silence all the warnings for the document:

        .. role:: wiki (title)

### Implementation ###

Implementation of Proposal-2 wikilinks are in the branch
[rst-wikilinks][rst-wl]


        This is a simple wiki page, with :wiki:`WikiLinks` and |named| links
        
        .. |named| wiki:: Some Page

        We can get rid of the role part as well for WikiLinks::
        
            .. default-role:: wiki
        
        Enables `WikiLinks` but does not impact references such as ``named``
        This can be made the default for ikiwiki.

[rst-wl]: http://github.com/engla/ikiwiki/commits/rst-wikilinks

**rst-wikilinks** patch series includes changes at the end to use ikiwiki's
'htmllink' for the links (which is the only sane thing to do to work in all configurations).
This means a :wiki:`Link` should render just exactly like [[Link]] whether
the target exists or not.

On top of **rst-wikilinks** is [rst-customize][rst-custom] which adds two
power user features: Global (python) file to read in custom directives
(unsafe), and a wikifile as "header" file for all parsed .rst files (safe,
but disruptive since all .rst depend on it). Well, the customizations have
to be picked and chosen from this, but at least the global python file can
be very convenient.

> Did you consider just including the global rst header text into an item
> in the setup file? --[[Joey]] 
>
>> Then `rst_header` would not be much different from the python script
>> `rst_customize`. rst_header is as safe as other files (though disruptive
>> as noted), so it should/could be a editable file in the Wiki. A Python
>> script of course can not be. There is nothing you can do in the
>> rst_header (that you sensibly would do, I think) that couldn't be done in
>> the Python script. `rst_header` has very limited use, but it is another
>> possibility, mainly for the user-editable aspect. --[[ulrik]]
>>
>> (I foresaw only two things to be added to the rst_header: the default
>> role could be configured there (as with rst_wikirole), and if you have a
>> meta-role like :shortcut:, shortcuts could be defined there.)
>
> I have some discussion on the [docutils mailing list][dml], the developers
> of docutils seems to favor "Proposal 1", while I defend my ideas. They
> want all users of ReST to use only the basic featureset to remain
> compatible, of course. -- [[ulrik]]

[dml]: http://thread.gmane.org/gmane.text.docutils.user/5376

Some rst-custom [examples are here](http://kaizer.se/wiki/rst_examples/)

[rst-custom]: http://github.com/engla/ikiwiki/commits/rst-customize

## Directives ##

Now **Directives**: As it is now, ikiwiki goes though (roughly):
filter, preprocess, htmlize, format as major stages of content
transformation. rST has major problems to work with any HTML that enters the
picture before it.

1. Formatting rST in `htmlize` (as is done now): Raw html can be escaped by
   raw blocks:

        .. raw:: html
   	
	        \[[!inline and do stuff]]

   (This can be simplified to alias the above as `.. ikiwiki::`)
   This escape method works, if ikwiki can be persuaded to maintain the
   indent when inserting html, so that it stays inside the raw block.

2. Formatting rST in `filter` (idea)
   1. rST does not have to see any HTML (raw not needed)
   2. rST directives can alias ikiwiki syntax:
     
        ..ikiwiki:: inline pages= ...

   3. Using rST directives as ikiwiki directives can be complicated;
      but rST directives allow a direct line (after :: on first line),
      an option list, and a content block.

> You've done a lot of work already, but ...
> 
> The filter approach seems much simpler than the other approaches
> for users to understand, since they can just use identical ikiwiki
> markup on rst pages as they would use anywhere else. This is very desirable
> if the wiki allows rst in addition to mdwn, since then users don't have
> to learn two completly different ways of doing wikilinks and directives.
> I also wonder if even those familiar with rst would find entirely natural
> the ways you've found to shoehorn in wikilinks, named wikilinks, and ikiwiki
> directives?
> 
> Htmlize in filter avoids these problems. It also leaves open the possibility
> that ikiwiki could become smarter about the rendering chain later, and learn
> to use a better order for rst (ie, htmlize first). If that later happened,
> the htmlize in filter hack could go away. --[[Joey]] 

> (BTW, the [[plugins/txt]] plugin already does html formatting
> in filter, for similar reasons.) --[[Joey]]

>> Thank you for the comments! Forget the work, it's not so much.
>> I'd rank the :wiki: link addition pretty high, and the other changes way
>> behind that:
>>
>> The :wiki:`Wiki Link` syntax is *very* appropriate as rst syntax
>> since it fits well with other uses of roles (notice that :RFC:`822`
>> inserts a link to RFC822 etc, and that the default role is a *title* role
>> (title of some work); thus very appropriate for medium-specific links like
>> wiki links. So I'd rank :wiki: links a worthwhile addition regardless of
>> outcome here, since it's a very rst-like alternative for those who wish to
>> use more rst-like syntax (and documents degrades better outside the wiki as
>> noted).
>>
>>> Unsure about the degredation argument. It will work some of
>>> the time, but ikiwiki's [[ikiwiki/subpage/linkingrules]]
>>> are sufficiently different from normal html relative link
>>> rules that it often won't work. --[[Joey]]
>>>
>>>> With degradation I mean that if you take a file out of the wiki; the
>>>> links degrade to stylized text. If using default role, they degrade to
>>>> :title: which renders italicized text (which I find is exactly
>>>> appropriate). There is no way for them to degrade into links, except of
>>>> course if you reimplement the :wiki: role. You can also respecify
>>>> either the default role (the `wikilink` syntax) or the :wiki: role (the
>>>> :wiki:`wikilink` syntax) to any other markup, for example None.
>>>> --[[ulrik]]
>> 
>> The named link syntax (just like the :wiki: role) are inspired from
>> [trac][tracrst] and a good fit, but only if the wiki is committed to
>> using only rst, which I don't think is the case.
>>
>> The rst-customize changes are very useful for custom directive
>> installations (like the sourcecode directive, or shortcut roles I show
>> in the examples page), but there might be a way for the user to inject
>> docutils addons that I'm missing (one very ugly way would be to stick
>> them in sitecustomize.py which affects all Python programs).
>>
>> With the presented changes, I already have a working RestructuredText
>> wiki, but I'm admitting that using .. raw:: html around all directives is
>> very ugly (I use few directives: inline, toggle, meta, tag, map)
>>
>> On filter/htmlize: Well **rst** is clearly antisocial: It can't see HTML,
>> and ikiwiki directives are wrappend in paragraph tags. (For wikilinks
>> this is probably no problem). So the suggestion about `.. ikiwiki:` is
>> partly because it looks good in rst syntax, but also since it would emit
>> a div to wrap around the element instead of a paragraph.
>>
>> I don't know if you mean that rst could be reordered to do htmlize before
>> other phases? rst must be before any preprocess hook to avoid seeing any
>> HTML.
>>
>>> One of my long term goals is to refactor all the code in ikiwiki
>>> that manually runs the various stages of the render pipeline,
>>> into one centralized place. Once that's done, that place can get
>>> smart about what order to run the stages, and use a different
>>> order for rst. --[[Joey]]
>>
>> If I'm thinking right, processing to HTML already in filter means any
>> processing in scan can be reused directly (or skipped if it's legal to
>> emit 'add_link' in filter.)
>>
>> -- [[ulrik]] 

>>> Seems it could be, yes. --[[Joey]]
>>>
>>>> It is not clear how we can work around reST wrapping directives with
>>>> paragraph tags. Also, some escaping of xml characters & <> might
>>>> happen, but I can't imagine right now what breakage can come from that.
>>>> -- [[ulrik]]

[tracrst]: http://trac.edgewall.org/wiki/WikiRestructuredText

### Implementation ###

Preserving indents in the preprocessor are in branch [pproc-indent][ppi]

(These simple patches come with a warning: _Those are the first lines of
Perl I've ever written!_)

> This seems like a good idea, since it solves issues for eg, indented
> directives in mdwn as well. But, looking at the diff, I see a clear bug:
>
>	-                               return "[[!$command <span class=\"error\">".
>	+                               $result = "[[!$command <span class=\"error\">".
> 
> That makes it go on and parse an infinitely nested directive chain, instead
> of immediatly throwing an error.
> 
> Also, it seems that the "indent" matching in the regexps may be too broad,
> wouldn't it also match whitespace before a directive that was not at the beginning 
> of a line, and treat it as an indent? With some bad luck, that could cause mdwn
> to put the indented output in a pre block. --[[Joey]] 
>
>> You are probably right about the bug. I'm not quite sure what the nested
>> directives examples looks like, but I must have overlooked how the
>> recursion counter works; I thought simply changing if to elif the next
>> few lines would solve that. I'm sorry for that!
>>
>> We don't have to change the `$handle` function at all, if it is possible
>> to do the indent substitution all in one line instead of passing it to
>> handle, I don't know if it is possible to turn:
>>
>> 		$content =~ s{$regex}{$handle->($1, $2, $3, $4, $5)}eg;
>>
>> into
>>
>> 		$content =~ s{$regex}{s/^/$1/gm{$handle->($2, $3, $4, $5)}}eg;
>>
>> Well, no idea how that would be expressed, but I mean, replace the indent
>> directly in $handle's return value.
>>
>>> Yes, in effect just `indent($1, handle->($2,$,4))` --[[Joey]] 
>>
>> The indent-catching regex is wrong in the way you mention, it has been
>> nagigng my mind a bit as well; I think matching start of line + spaces
>> and tabs is the only thing we want.
>> -- [[ulrik]]
>> 
>>> Well, seems you want to match the indent at the start of the line containing
>>> the directive, even if the directive does not start the line. That would
>>> be quite hard to make a regexp do, though. --[[Joey]]
>>
>> I wasted a long time getting the simpler `indent($1, handle->($2,$,4))` to
>> work (remember, I don't know perl at all). Somehow `$1` does not arrive, I
>> made a simple testcase that worked, and I conclude something inside $handle
>> results in the value of $1 not arriving as it should!
>>
>> Anyway, instead a very simple incremental patch is in [pproc-indent][ppi]
>> where the indentation regex is `(^[ \t]+|)` instead, which seems to work
>> very well (and the regex is multiline now as well). I'm happy to rebase the
>> changes if you want or you can just squash the four patches 1+3 => 1+1
>> -- [[ulrik]]

[ppi]: http://github.com/engla/ikiwiki/commits/pproc-indent

## Discussion ##

I guess you (or someone) has been through this before and knows why it
simply won't work. But I hoped there was something original in the above;
and I know there are wiki installations where rST works. --ulrik

**Individual reStructuredText Issues**

* We resolve rST links without definition, we don't help resolving defined
  relative links, so we don't support specifying link name and target
  separately.
  
  * Resolved by |replacement| links with the wiki:: directive.

**A first implementation: Resolving unmatched links**

I have a working minimal implementation letting the rst renderer resolve
undefined native rST links to ikiwiki pages. I have posted it as one patch at:

Preview commit: http://github.com/engla/ikiwiki/commit/486fd79e520da1d462f00f40e7a90ab07e9c6fdf  
Repository: git://github.com/engla/ikiwiki.git  

Design issues of the patch:

The page is rST-parsed once in 'scan' and once in 'htmlize' (the first to generate backlinks). Can the parse output be safely reused?

> The page content fed to htmlize may be different than that fed to scan,
> as directives can change the content. If you cached the input and output
> at scan time, you could reuse the cached data at htmlize time for inputs
> that are the same -- but that could be a very big cache! --[[Joey]] 

>> I would propose using a simple heuristic: If you see \[[ anywhere on the
>> page, don't cache it. It would be an effective cache for pure-rst wikis
>> (without any ikiwiki directives or wikilinks).
>> However, I think that if the cache does not work for a big load, it should
>> not work at all; small loads are small so they don't matter. --ulrik