16 December 2010

Stow meets design docs

Can the stow model help me write design documents?

The problem

I like to document my designs for my projects, such as Klink and Emtest. In theory writing down the things I decide or do should create a record of the design.

But I always face a problem. The issues, insights and decisions don't come in an organized manner. Quite the opposite. They trickle in a stream-of-consciousness manner, a few at a time, bearing no strong relation to the organization of the design. I want to bring order to this chaos without spending more time on it than it's worth.

I've tried several approaches:

  • Buckling down and putting every new thing into its proper, final place. After doing that for too long, my conclusion is that it's far, far too much work, especially navigating to the right spot all the time.
  • Separate files for major task areas: Design, decisions, progress. Before I switched to org mode, I wrote a meta-design of how to organize and some helper code and tried to stick to it. Again, after too much time and effort trying to keep it on track, I concluded that this just took far more work than it was worth.
  • Tried out John Wiegley's remember / planner / emacs-muse, but it really wasn't what I was looking for.
  • About this time I switched to Carsten Dominick's org-mode. It's wonderful for organizing text, but seemingly possesses no miracle big enough for this task.
  • Accumulating everything as notes and placing them in their final place in the outline later. Problems:
    • Later never comes.
    • Notes get lost in the big amorphous blob of notes.
    • Often I write a note that wants to modify the design, but that I might or might not ultimately accept. It isn't clear what should be done with it. I could move it into the design and maybe have to move it back out and try to restore the design to its old state, or leave it where it is, where it itself becomes an informal focus of later additions.

      I tend to choose the latter, so I am often finding recent notes and adding to them. This corrupts the file's organization too, so either way the design file degenerates.

  • Tagging each incoming note. I hoped that by doing this, I could filter on tags and thereby see all the items that should be moved to a place in the design section. However, that offers little for the pieces that are maybe in, maybe out.

    It's helpful up to a point. Theoretically I can filter on the tags, but in practice that hasn't done much for me. The tagset is simply nowhere near as fine-grained as my design outlines, and there's no reasonable way to make it so.

  • Splitting my design file template into information sections. "Design", "Issues", "Info", "Communication", "Progress", "Notes". Again, it's helpful up to a point, but doesn't solve my problem.

    One major problem is that items often evolve from one category to another. Eg, an issue will evolve into a task, which belongs in "Progress", or a design, which belongs in "Design". Progress items will spawn new issues, often as many as half a dozen. Decision items want to place one alternative into "Design" and archive the rest.

I'd really like to organize my designs even more strongly, separating out such facets as parts and behavior, and "cookbook" docs, etc. But without a much stronger approach, that's unlikely to happen.

What stow does

In order to describe my shiny new idea, I first have to describe how Stow works. Stow uses indirection. The parts are in one place (in subtrees of /stow/ or /usr/local/stow/) and the organized tree (/usr/local/bin/, /usr/local/man/man1/, etc) lives in a different place. Please ignore the fact that /usr/local/stow/ is really a location in the filetree - that's an inescapable bootstrap issue and poses little practical problem.

What stow does is tells the organized tree to hold the parts indirectly by symlinks. So the executable file in /usr/local/bin/my-exec really points at /usr/local/stow/this-package-0.1/bin/my-exec.

It figures out what to hold where from each package's internal organization. Ie, /usr/local/stow/this-package-0.1/bin/ will be symlinked from /usr/local/bin/, because after we get past the stow prefix /usr/local/stow/this-package-0.1/, what we see is bin/)

Stow's advantages include:

  • One can swap packages in and out very freely
  • Consequently nearly all of the pain of managing package versions disappears. You install them to use/local/stow and you stow the version you want. Upgrade? Unstow the old one, stow the new one. Didn't like the upgrade? Reverse the procedure. Don't like the whole package any more? Unstow it and delete the subdirectory.
  • Also, packages' install procedures can be much simplified, though most packages don't take advantage of this because they want to install on non-stow systems too.

A possible solution

Now, this could be another shiny idea that proves to be more work than it's worth. But it works neatly for stow, so maybe it will work here.


My approach would have these parts in analogy to stow:

/stow/ contents
Notes that mirror the existing structure of a document or subtree.
items that don't have their own contents but act as if they were other items.
Documents or outline items intended to contain the symlinks as well as some text themselves.
A command relating to a note, that causes appropriate "symlinks" to it to be created in /stow/ contents.

stowable notes

They are like org notes

These might be just ordinary notes, or they could be notes in a specific place. They should be creatable by org-remember, so they should live in a refile target.

Multiple stowables might live in a note

But it shouldn't be the case that each note corresponds to exactly one target location. A single coherent note may well want to add parts to different documents. Also, its target may be so deeply nested that mirroring the entire path to it would be very annoying.

So ISTM best for each stowable item to parameterize on a prefix. Presumably that prefix is stored in a property. ISTM the right interface is a command or commands to move singleton "directories" into and out of the prefix. Binding-wise, ISTM best to overload `org-shiftmetaright' and `org-shiftmetaleft' to do that, since their ordinary function is not useful here.

Having stowables not correspond to notes creates a problem that stow also has: How to recognize a parts directory when there can be more than one. We could just record "stowable", presumably as a tag, and not let new stowables encompass items tagged "stowable".

In stow, items don't need to know whether they are currently stowed or not. Here, for user convenience we'd probably want to at least tagged stowed items "stowed".

Indicating the mirror structure

My first thought is to just mirror the headlines. It works but it's fragile for split trees. If the user renames an item in the target document, suddenly we can't find it anymore. Let's also have a backup: store the org-id of any node that our mirror structure splits on.

Creating the mirror structure

But we still haven't talked about creating the mirror structure. It would be a PITA to do manually. Doing it supervised automatically doesn't seem difficult: It'd be much like the interaction like org-refile. Presumably we'd proceed in two stages:

  • Creating a stowable
    • Interactively find a root item
    • Make a fresh item with the same headline
    • Store the root's id.
    • Tag the fresh item "stowable"
  • Adding an item underneath a stowable
    • Choose a item in the respective target directory (whether our item is stowed or not)
    • Make a fresh item with the same headline
    • Store its id.

I'd also like support to repair the mirror structure, if it goes astray. It'd go something like

  • Visiting a possibly broken item:
    • Find where the org-id points.
      • No match, or no id? Uh-oh, that's not good.
        • Is there a headline match under the parent?
          • If so, fix the id.
          • Otherwise proceed as if creating a new item, just it will have old contents.
      • Match found: Fix the headline, if needed.

stow command

As with stow itself, there are a few issues under the surface:

  • Conflicts
  • Tree-splitting.

It's all fine to symlink /foo/bar/x into an empty directory, but what do we do when part of the path already exists in the target? If eg the target is /foo and not /.

Same thing that stow does: We split the tree. Sometimes to do so we must turn stowed "symlink" directories into real directories whose components are now symlinks. We may even have to do that recursively. Usually that suffices, because the user's intention was not to make competing, overlapping parts.

But sometimes it doesn't suffice. Two stowed parts might want to occupy the same position, and they can't both. Or maybe there is already text there.

Again we do as stow does: Refuse to stow, report the conflict(s), and asks the user to resolve them manually. That's acceptable. If the user's design document really has such a structural conflict, there's no magic here to fix it. Of course he will need to actually change whatever is making the conflict.

It might be handy to have a tool to pull existing text that participates in a conflict into a new note or notes. The user could then resolve the conflict just by unstowing that text.

Integration with org-choose

org-choose is my old project to "choose" among org items, where choosing one item makes its siblings non-selected. It would be very nice to automatically stow the item that is selected, and unstow the alternatives (but not in that order). This probably wants its own tag, perhaps "stow-choice".

But what about impure choices, where some items build on others? For instance, in this very post, the item "Do the same in place" builds on "Weave a separate read-only document". Pretend for a moment that they are full stow-choice items.

Stowing "Do the same in place" by itself wouldn't work; what's it the same as?. Stowing "Weave a separate read-only document" would be just wrong. If the two items have no structure, we're just out of luck. The second item will just have to be edited to be stand-alone.

So also pretend for a moment that the user has restructured both items cleanly: ie, each sub-item pertains entirely to one or entirely to both. Now there is something potentially stowable there. We just have to see how to do it right.

One thought is that the derived item points to be stowed in the base item, and if chosen, cascades to stow first the base and then itself. But this can't be entirely right. It implies that the derived item also overwrites the base item's text. That would be wrong, because what happens if we then choose the base item? Its text would be gone, replaced by that of the derived item. In any case, I don't contemplate using this mechanism to override text.

One could insist that the user rewrite further to remove conflicts, but the items are in fact alternatives. Sometimes we're lucky and one alternative builds on another in such a way that there is no conflict, but we can hardly rely on that. Making them never structurally conflict seems contrary to the spirit of alternative choices.

So instead, let this base/derived situation inform the choice of what to stow but not cause any base text to be overwritten. We'd recognize such situations just if a "stow-choice" alternative pointed into a "stow-choice" alternative in the same set.

No comments:

Post a Comment