Okay, I got annoyed trying to write parts of the automatic core updater utility, so I started digging through the server code and have discovered some things:

What happens when a connection is made to the LambdaMOO server

The another program connects to a port the server is listening on, is accepts the connection after some minimal amount of error checking or black list checking, and then creates a task with the following associates: a unique identifier for that connection (a negative object reference), and the object responsible for responding to that port. In the default case, this object will be #0 (the system object). To determine what objects are responsible for which ports, the built in function listeners() can be used.

After creating the previously mentioned association and task, the server pushes an empty string into the input queue of that task. The interesting thing at this point is that this empty string is what causes the welcome message in the LambdaCore database to be generated. What happens is this

If you don't believe this, start up a copy of your favorite MOO core, connect as a wizard, and @rmprop $login.welcome_message and then login in again. You should receive a fairly bizarre trace back error.

In the normal flow of events, one of the verbs called by "do_login_command" will return an object reference, or end the connection by calling boot_player. Assuming that an object reference is returned, that task becomes promoted to command task, and the rules for dealing with user input changes radically.

For input tasks, the flow of execution is as follows:

1. It is possible to completely redefine how the server deals with user input by defining a "do_command" verb on the object the player has logged into on. do_command is called before the normal processing of input lines into psuedo-english phrases. However, doing macro processing (as Mr. Fish was wondering about) would require doing all of the normal text processing functions in LambdaMOO code, since it is not possible since there are no destructive list maninpulation functions.

Each port the server is listening on is associated with an object. The port bound by the server on start up is associated with object #0 (The system object). do_command gets passed the input string after it has been broken up into white space delimited tokens. So, if for example #0:do_command looks like


  player:tell(@args);
  if (length(args)==3)
    return 1;
  else
    return 0;
  endif;

then a player connected through #0 who typed

 l at me

would see the text

 lame

echoed back to them, and that command would never receive normal processing. Interestingly enough, the command ".pr*ogram" in the form
 .program Object:verb
is hard coded into the server for programmer players before the check for "do_command". Interestingly enough, do_command isn't defined for either the Sell Game, or Planet Oit.

2. Server options

Yes, the server options to work. Essentially, the documentation is unclear about where the properties are supposed to be found. The server code expects them to be on $server_options, or rather that $server_options is expected to be an object reference which defines the properties.

3. Forking tasks

Granted that it looks like the server code for dealing with queued tasks is fairly complicated, it doesn't look like the overhead for forking a task is too much more that the overhead for invoking a verb on an object.

4. Server startup scripts

The verb #0:server_started is run when the server starts up, after the database is loaded and the network code initialized. This would be the place to put things like the http server initializations for the Sell Game and Planet Oit.

5. I've been looking through the server code for the MudOS LPC style mud server, and managed to locate the main loop of the server. My major reaction to the MudOS code is that it makes the LambdaMOO server code (even the confusing parts) look blindingly straight forward. Essentially, LPC muds implement a large amount of swapping objects out of memory, calling reset functions every so often, and maintaining a circular list of objects to call a heart_beat function periodically, all while trying to take up more disk space than memory. Of course, LambdaMOO servers try to do the exact opposite by trying to take up more virtual memory than disk space. In any event, I honestly don't see why (if periodic resets and heart_beat functions were desired) these two functions could be implemented by carefully written LambdaMOO code.

6. I still haven't found a good way of introducing a good destructive list append function into the Lambda server. At the moment, it appears that this would involve adding a built in function, and as such the naming convention would be static package bf_ItsName(...) and then call register_function("ItsName",?,?,bf_ItsName,ArgType,ArgType,ArgType,...); The notation in functions.c says

/*****************************************************************************
 * This is the table of procedures that register MOO built-in functions.  To
 * add new built-in functions to the server, add to the list below the name of
 * a C function that will register your new MOO built-ins; your C function will
 * be called exactly once, during server initialization.  Also add a
 * declaration of that C function to `bf_register.h' and add the necessary .c
 * files to the `CSRCS' line in the Makefile.
 ****************************************************************************/
Which is interesting, and I'll have to give it a try soon.

Conceivably, adding file io would not be necessarily very involved, once the appropriate interfaces were defined. For adding destructive set maninpulations, all that would be necessary would be to figure out how to dissasociate the previous copy of the variable from the reference counting scheme, modify it, and then reintroduce it to the ref counters.