it has been some years since the matching different kinds of links issue was tackled, but hardly a plugin is using it.

in order to enhance on the rel attribute for links and better bug tracking support issues and to provide a more general infrastructure, i'd like to propose a generic plugin for typed links. it can be also viewed of a way to have structured page data that consists of URLs inside the wiki.

following the use case i've developed it for, i'll call it blocks for the moment (but am open to better suggestions).

outline

the plugin has a configuration option called blocks_names, which consists of pairs of verbs; the typical example is blocks/blockedby, but other values could be next/prev up/down or owner/owns.

for each verb in the options, there is a directive which is used to state the relationship; relationships can be declared on both ends, so a page bugA with the contents [[!blocks bugB]] is semantically equivalent to a page bugB with the contents [[!blockedby bugA]].

for each verb, there is also a pagespec which matches all pages that are the origin of a relationship to a given page. if developerA [[!owns bug1]], then if bug1 contains [[!map pages="owns(.)"]], it will show the owning developer. these specs match both ways, ie. if bug1 [[!owner developerA]], the said map directive will still produce the same result.

details

  • single word relationships vs. symmetric relationships

    with some verbs, it is possible that a relationship is only used in one direction (eg index, even though one could declare it as index/isindexof).

    isindexof is not a very interesting relationship - it just clogs up the link-map, since the index is "the index of" all pages. I can't see any situation in which you'd want to do pagespec matching on it? --smcv

    that's why i used index as an example of a one-direction relationship.

    it wouldn't clog up the link map, though: in order to cleanly match both directions, when the "inverse" term of a relationship is used, the link in taggedlinks uses the "forward" term, but switches the objects.

    --chrysn

    other verbs are symmetric, eg. equivalent, which need different treatment.

  • "taglink" style directives

    the tag plugin would be a special case for this plugin (apart from the autotag and tagdir features). as there is a [[!taglink ...]] directive, there could be an analogous directive for every single directive.

    This is basically the traillink/trailitem duality, too. I'd be quite tempted to generalize to something like this:

    We can't fix [[!link  blocks="bug123" text="Bug 123"]] until we do this.
    
    [[!hiddenlink  owner="smcv"]]
    

    but perhaps that's too wordy?

    I think both trail and tag need their own special processing beyond the general case, but maybe not? --smcv

    i'd be all in favor of having this unified and deeper; there has been the idea of a [[!link ]] directive again and again.

    i like the [[!link text=""]] and [[!hiddenlink ]] conventions, but think that ${REL}="${TARGET}" isn't ideal because it implies that a single link can have more than one target. instead, i'd go for `[[!link to="bug123" rel="blocks" text="Bug 123"]]; as with the html rel parameter, rel would be a list of whitespace separated values.

    positional parameters ([[!link bug123 rel="blocks" text="Bug 123"]] or even [[!link Bug 123|bug123 rel="blocks"]]) would be possible, but i prefer explicit syntax and not joining stings back again with the whitespace that was split off it before.

    if the '|' character is not widespread in page names (which i assume it is not), instead of using positional parameters in [[!link ]] for shortcuts, we could extend the regular link syntax; the same relationship could then be declared as [[Bug 123|bug123|blocks]]; this would be an easy extension to the original link syntax. it would even work for hidden links ([[|smcv|owner]]), which previously made no sense because a link with neither a physicial representation nor metadat is of no use.

    --chrysn

  • implementation notes

    the way pagespec hooks are implemented required some nasty perl tricks, for which the people who showed me felt very bad for having spoilt me. indeed, no strict refs; and *$forward_name = $forward_match; are not exactly ideal. a change in the pagespec declaration api (why not just hook like everything else) would make the implementation cleaner.

    How about replacing blockedby(bug*) with linktype(blockedby bug*) or something? Then you'd only need one pseudo-hook. --smcv

    there has been the topic of pagespecs like typedlink(type glob) back in the matching different kinds of links discussion, but it was removed in favor of per-type matchers. --chrysn

    note to self: should use the inject function to avoid no strict refs. --chrysn

  • configuration location

    i aimed for static configuration of the block_names in the setup file. this could be made more general like in the shortcut plugin, but that would make things more complex.

  • no html links with rel= yet

    as there are no taglink style links between the articles so far, no htmllink gets rendered that could carry the relationship name in its rel field.

    having the inverse relationship description in backlinks (as in the link created by the map directive in the example above) would be hard to implement. (actually, i think it'd be easier to determine the rel values from the taggedlinks for every htmllink than to influence the backlinks in this plugin).

  • one direction also creates a normal link

    due to the way add_link treats relationships, the forward relationship is always going to be reflected in the links/backlinks. a section of matching different kinds of links was dismissed with "let's not worry about it", this plugin might be reason to worry about it again. (i'd consider what is in @links to be a representation of which hyperlinks are there, and in this case, none are generated).

    taglink and traillink already count as wikilinks without generating any visible HTML. --smcv

implementation

there is a working but slightly incomplete (basically where it comes to the details mentioned above) implementation in blocks.pm.

--chrysn