Before the Shadow Play

Author: Joe McMahon

  • MVT under Hercules notes

    I needed a little mental relaxation this weekend, so I spent a while playing mainframe model trains by bringing up Hercules.

    I initially tried the MVS Turnkey system, but ran into some issues — mainly that there are no working (free) 3270 emulators for Big Sur. Since I couldn’t set up any consoles, and Homebrew x3270 didn’t seem to work under Xquartz, and I had no intention of spending $29 just to fool around with MVS for a bit, or multiple hours trying to get X11 builds working, I dropped back to Jay Maynard’s MVT installation instructions.

    They’re a bit out of date at the moment, and the Right Thing would probably be to make the fixes in both the instructions and the files, and move the corrected instructions over to a wiki somewhere. For now I’m leaving my notes here so I don’t forget what I did, and so I can do that later if I get the time. More fooling with the console and running jobs, less file twiddling.

    • You can get the OS/360 “CD-ROM” at http://www.jaymoseley.com/hercules/downloads/archives/os360mvt.tar.gz — this really ought to be on archive.org. It works with Maynard’s instructions and has these fixes:
      • dlibs/DN554.XMI was recovered. The srclibs/dn554 and related files in srclibs/TAPEFILE.ZIP have not been modified (recovered) to match, and should be at sone point.
      • srclib/fo520/IEYUNF.txt has been renamed to original.IEYUNF.txt and a version recovered from the MTS distribution has been added as IEYUNF.txt. The related file in srclibs/TAPEFILE.ZIP needs to be fixed as well.
      • These files are fine as they are to build MVT.
    • You need the JCL and HASP II tapes; they’re at http://www.conmicro.com/hercos360/os360ctl.tar.gz. Yes, I wish it was HASP IV, but I’m playing, so it’s not that big of a deal.
    • Maynard’s instructions are definitely of their time, when cutting and pasting commands was not a thing. The relevant commands are mentioned once, and one is expected to remember them. If I update these, I’ll inline the relevant command and note which console they get typed into. Switching back and forth between the Hercules “hardware” and the MVT console was a tad confusing at times. A significant omission: the devinit 00c foo.jcl command works okay for the MFT starter system and MVT without HASP, but once HASP is running the devinit must include eof at the end of the command or the reader hangs and the job never starts. Also, one of the HASP job filenames is called out by its right name in the section header, but by the wrong one in the text. Looks like a cut and paste error.
    • You should comment out or remove the 3270 definition at 0C0 in the mvt.cnf Hercules config file; if it’s there when you boot MVT, but no 3270 emulator is connected to it, the machine will hang with wait state 21. Took some googling to find that. TCAM will grumble about it when it starts up, but it doesn’t hurt anything.
    • You will need to install telnet on your Mac, since Big Sur removes it. telnet localhost 3270 to connect the virtual 3215 console. It might be worth trying to configure some 3215s for TSO and see if that works. We don’t really need to emulate real serial terminals.
    • Doing the mn jobnames, t and mn status commands will save you a lot of wondering whether anything is going on or not, even after HASP is up.
    • Be sure to copy the prt00e.txt virtual printer text file when:
      • You’ve finished doing both stages of the sysgen on the MFT system.
      • You’ve finished installing HASP.
      • You’ve finished installing TCAM and TSO

    Otherwise you lose all those useful assembler outputs of the HASP hooks and the TSO interfaces to it. The sysgen and TCAM build output isn’t critical, but it’s nice to have.

    Other things — http://hercstudio.sourceforge.net/ is supposed to be a hardware console emulator (lights and dials and stuff) interface for OS X and Linux; it’s written in Qt, which is fine, but the Makefile that qmake builds works to build it, but wants to install things in /usr/bin and Big Sur will not let it go there even with sudo. I might consider just writing a native iOS one instead.

    I’m up a couple hours later than I intended, but I have my notes, and as Adam Savage says, “the difference between science and screwing around is writing it down!”

  • Praise and a warning about using Gemini II and CleanMyMac X to clean up old merged backups

    Time to clean up!

    Earlier this year, my company, in its push to get things squared away for an IPO at some point (note to the SEC: I know nothing about IPO plans, I am not suggesting anyone invest in anything, I’m just this guy, you know?), installed a remote management tool for MacOS. Initially, I was concerned that we might end up being monitored as to what was on our machines, and non-work use might be frowned upon – plus I learned the hard way at WhiteHat that if you’re going to get laid off or fired, no one’s going to give you a day or two to back up anything personal on your machine. (I lost, and later managed to partially recover, all the patches for my Radio Free Krakatau album.)

    IT and upper management, after a couple of days of general consternation and concern about keylogging, etc., formally told us, “no, we don’t care what you do on your laptop, just don’t do anything illegal,” but by that point I’d scoured off the personal files and data and moved them to iCloud, Dropbox, or a spare 2012 MacBook Pro I had.

    The 2012 MBP was the last one that allowed upgrading by the end user. It could accept up to 16GB of memory, had a lot of ports (including a DisplayPort, native Ethernet, and FireWire), and had an internal disk that could be swapped to an SSD.

    I picked up a 2TB Crucial SSD, pulled the old disk, installed the new one, and used Carbon Copy Cloner to copy the old internal disk back on to the SSD. I also pulled in several older backup spinny disks into a folder called “Backups to Clean Up”. This was a superfund site of duplicates, junk, and accumulated files. I took a first cut at cleaning it up right away — deleting old stuff I knew I didn’t care about anymore, like partial iPhoto/Aperture/Photos libraries and old iTunes folders — but I was left with a considerable stash of data that I knew contained duplicates. At the time I was busy and decided I’d work out the rest later. I had removed my Adobe apps and music apps and data from my work laptop, and at the time I just didn’t have any time to work on those.

    Diving in with Gemini II

    Last weekend, I decided it was time to do the cleanup. I had bought Gemini II a couple years ago in a MacHeist bundle, and had tried it a little, but found it too slow on a spinny disk to to be useful. I decided that the job was big enough that I really needed to have some help, so I tried it again. I fired it up on Wednesday afternoon, and pointed it at my home directory on the MBP, and said go get ’em.

    Friday morning (I neglected to exclude Dropbox from the duplicate check, resulting in a lot of “download the file, check it” for the 200 or so GB or data in there, slowing things down considerably), I had a complete comparison. I spent the better part of Friday evening and Saturday and a chunk of Sunday looking at the recommendations and clearing duplicates. In general Gemini had made good choices as to which files to keep and which were duplicates, and this got rid of almost 200GB of duplicates. I did a couple rescans and found another hundred or so that I could clean up.

    Gemini recommended I try CleanMyMac X to help with getting rid of extra junk on the disk, and being in a cleaning mood, I decided to try it. I signed up for a month, and only after I’d done that did I see a “30% off if you own one of our other products”, despite being signed in. MacPaw was very kind and extended my CleanMyMac X subscription for three months to compensate.

    On the initial run, CleanMyMac X was very useful. It cleared a bunch of old caches, got rid of unused languages, etc., and helped me cleanly delete some old apps that were cluttering up ~/Library. It installed a very attractive cleanup and virus checking monitor, and I thought nothing of it at the time.

    Problems surface

    I continued working with Gemini II, and the monitor was solicitously clearing the trash when it got full, and so on. I then tried to use Gemini to just dedup my Music folder, and here’s where the fun started.

    It ran for an hour or so and then I got a “You are out of memory” warning; Gemini II apparently had 69GB of memory allocated. I shut some stuff down, but I ran out of memory again. And again. And again. Quit Gemini. Tried to run Ableton Live; the cursor was sluggish, sound was breaking up, and trying to select a patch in plugins was causing outright crashes. And the laptop was so hot I couldn’t leave it on my legs.

    This was not going to do at all. I wanted to use this machine for music, and it wasn’t able to handle it anymore. Was I going to need a new laptop? It was late. I went to bed.

    The solution

    In the morning, after some time spent with Activity Monitor, I twigged to the problem: CleanMyMac X had installed a lot of startup items. Like four. And Gemini had installed some too. This was not going to get any better with those hanging around. I decided that I was going to have to remove them, and the easiest way was to have CleanMyMac X do it. All credit to MacPaw: it simply warned me that it would shut down all the monitoring if I removed CleanMyMac X, did I want to do that? I did.

    And now the machine is running fine. I’m able to keep a couple of instances of Arturia’s 2600 emulator open and running with Live actively generating sound, and I can tweak the settings without significant effort or the sound breaking up. I ended up using Song Sergeant to do the Music Library cleanup; I can recommend it as doing a good job of finding duplicates, even in different formats.

    Conclusions

    The machine is slimmed down by about 250GB total and running fine; if I decide to do a similar cleanup again, I will probably use both Gemini II and CleanMyMac X to get the cleanup work done, but without being able to easily say no to the monitoring they install, I’ll probably delete them again as soon as I finish. MacPaw, if you’re reading this: make it optional to install the startup items, and give us an easy way to turn them off. If I had those I’d leave the two apps installed, but I just can’t and get any work done.

  • obliquebot returns

    Some time back, when beepboop.com was still around, I wrote a little Slack bot that listened for “oblique” or “strategy” in the channels it had been invited to, and popped out one of Eno’s Oblique Strategies when it heard its keywords or was addressed directly.

    It worked fine up until the day that BeepBoop announced that they were going away, and eventually obliquebot stopped working.

    This month, I decided that I would stop ignoring the “you have a security issue in your code” notifications from GitHub, and try catching obliquebot up with the new version of the SLAPP library that I’d used to get Spud, the RadioSpiral.net “who’s on and what’s playing” robot back online.

    I went through all the package upgrades and then copied the code from Spud over to the obliquebot checkout. The code was substantially the same; both are bots that listen to channels and respond, without doing any complex interaction. I needed to add the code to load the strategies from a YAML file and to select and print one, but the code was mostly the same.

    I also needed to update the authentication page to show the obliquebot icon instead of the RadioSpiral one, and to set the OAuth callback link to the one supplied by Slack.

    Once I had all that in place, I spent a good two or three hours trying to figure out why I could build the code on Heroku, but not get it to run. I finally figured out that I had physically turned off the dyno, and that it wasn’t going to do anything until I tuned it back on again.

    obliquebot is now running again at RadioSpiral and the Disquiet Junto Slack, and I’ve updated the README at the code’s GitHub page to outline all the steps one needs to take it and build one’s own simple request-response bot.

  • Show report: 2020-10-31 “Pharoah Nuff” at radiospiral.net

    My last performance was not as smooth as I hoped, so this time I decided that I would find a way to streamline it even further.

    I decided to go further in the direction I’d taken with the Wizard of Hz show, and strip down even more. I decided to try to perform as much as possible of the set on the iPad, and use the laptop solely for streaming and Second Life. This freed me from hassles in switching setups in VCVRack, Live, and the other software I’d been using, but it also meant that I wouldn’t be using either of my favorite synths for this performance (the Arturia 2600 and Music Easel).

    Having had some time between performances to really experiment with AUM and I felt comfortable using it to lay out my performance. I decided that I wanted to keep Scape as my background/comping program, and that I’d set up a series of light-handed scapes to give me a through-line. I then sat down with MIRack and Ripplemaker to create multiple Krell textures that I could bring in and out, and also discovered a couple of lovely lead patches for Ripplemaker that I paired with a Kosmonaut looper. I also brought in a couple public-domain samples from old sci-fi movies, heavily processed with Kosmonaut again, and felt like I had enough material to do an hour’s performance.

    I used the iConnect Audio4+, which I now finally have the hang of, and set it up so that I had two stereo channels from the iPad and one mono channel routed to the iPad through Kosmonaut (again!) for some subtle reverb when I was doing my intro and outro. With the setup I used, the iConnect kept the iPad fully charged through the whole set.

    I used Loopback to connect the multiple outs from the iConnect to the stereo ins on my Mac, and monitored on headphones. I pulled up Audio Hijack, entered the stream setup, and was ready to broadcast.

    I got up early on the day, started up AUM, and ran a soundcheck to make sure everything was working. All sounded good, and I was good to go.

    Mostly.

    I didn’t stop AUM, and as a result, it ran for several hours before I tried to start using it. This apparently triggered some kind of a memory shortage, and when I started streaming, I was completely mute. Fortunately, I’d cued up a prerecorded VCVRack texture, and started that while I was trying to figure out what was wrong. I gave up and restarted the iPad, and AUM came up like a champ.

    After that it was pretty smooth. I was able to fade the various patches in and out, play the sci-fi samples, and improvise over the Scape-provided background. Once it was off the ground, the performance was very easy to do. I did forget and leave the audio feed from Second Life enabled, so as a result this was a very sparse performance, but the sparseness worked out very well.

    Overall this was a great way to do a performance and I plan to refine this further. Of particular note is that AUM saves things so well that it will be trivially easy to do this performance again, should I decide to; this is probably the first time I’ve had a performance setup I felt was robust enough to say that!

  • RadioSpiral Wizard of Hz Performance Notes

    Last time I did a live streaming performance for an audience, it did not go well. I had long pauses, the mic didn’t work, and miscommunication over Slack to the remote venue resulted in my getting cut off before my set was finished. And this was even after a good bit of practice.

    So when I signed up for the Wizard of Hz concert on RadioSpiral, decided that I needed to have as much backstop as possible in place so that no matter how tangled up I got mentally, I’d have a fallback to something that sounded good and would be a nice navigable arc from point A to point B. Ideally, I should have something that would sound great even if I got called away for the entire set!

    My go-to process for this is Scape. I’ve had it since it first came out, and it meshes very well with what I enjoy hearing and enjoy playing. I started off with the Scape playlist that I often use to relax and get to sleep; this is a seven-scene playlist, with the transition time at max, with the per-scene time adjusted to be just a bit over an hour. This gives me a fallback for the whole hour; I can pull everything else back and lean on Scape while I decide what the next section should be.

    In addition, Scape provides a very nice backdrop to improvise over, so I can be playing something while Scape gives me a framework.

    I then put together a couple of Ableton Live sets: one built on the Arturia ARP 2600 and Buchla Music Easel emulations, and another built on Live’s really nice grand piano and the open-source OB-Xa emulator, the OB-Xd. I finally figured out how to change patches on the OB-Xd about 20 minutes before showtime.

    I had set up a piano with a nice looping effect from Valhalla Supermassive (Supermassive and Eventide Blackhole figured heavily in the effects), but ended up not using it, and doing a small Launchpad set instead using the Neon Lights soundpack.

    I was also able to open and close with the large singing bowl, played live and processed through the Vortex, which was a nice real analog performance touch.

    Overall, I strove for a set that sounded played-through, but that had enough breathing room that I could fall back on Scape while making changes (switching Live sets, etc.), and I think I achieved that.

    I did have Audio Hijack recording the set, so if it sounds OK, I’ll be releasing it on Bandcamp. (Followup: it came out pretty well! Definitely at least an EP.)

    Only real issue was a partially-shorted cable between my iPhone and the mixer that I didn’t figure out until most of the way through the set.

  • Squaring numbers and a forgotten book

    I happened on a demonstration of a mental math trick on Reddit for squaring numbers in your head and was immediately reminded of a technique I learned in 1972 from a great book on speed arithmetic that I have unfortunately forgotten the name of.

    The video’s formulation uses the identity n^2 = (n^2 - a^2) + a^2 to make the multiplication simpler, but the book had an extremely elegant way to notate a different identity that works nicely for doing the squares of two-digit numbers in one’s head, and rapidly doing multi-digit squares on paper.

    The Reddit example squared 32 by changing it to 32 * 32 = 30 * 34 + 4 = 1024, which is clever, but check this out!

    Start with the identity (a + b)^2= a^2 + 2ab + b^2 and treat 32 * 32 as (30 + 2)^2.

    Visualize this in your head:

    0904 
 12

    That’s the a^2 + b^2 on the first line, and 2ab on the second. Now just add it up normally, with blank spaces equal to zeroes, and you get 0, 10, then 102, then 1024.

    The left-to-right add means you never have to remember the carry value, just the changed result. Let’s try 47.

    1649
 56

    1, 21, 220, 2209. Simple.

    The Wikipedia page on mental arithmetic is a great resource that has this technique, but lacks the notation visualization shown here which honestly is what makes it easy. The same technique works for larger numbers too. There’s more to remember, which may make it too hard to do in your head, but it makes squaring large numbers on paper trivial.

    Let’s say we want to square 123:

    010409 
 0412 
  06

    1, 14, 151, 1512, 15129. (a^2 + b^2 + c^2 + 2ab + 2bc + 2ac). Squares on the top row, 2ab on the left in the middle, 2bc on the right in the middle, 2ac on the bottom.

    I will admit that I didn’t properly get how to do the multi-digit notation right 45 years ago, but I hadn’t really understood the mapping of the identity to the positions on the page and was doing it by rote. The notation is the slickest part of this, as it automatically handles the proper number of multiplications by 10 for you.

    The left-to-right addition and a trick of doing mental addition by repeating the current total to oneself when adding the next number to keep from losing one’s place (ex. 45 + 37 + 62 – 45, 75, 75…82, 142, 144 and cast out 9s — 0, 1, 9, and 1+4+4 =9) were all also in that same book. I really wish I could remember what it was!

  • Belloq fail: Roblox

    In the category of “we can’t handle email right” again, or at least, they haven’t convinced me they can: the email that is this blog’s domain name plus .me.com is apparently on someone’s list of “valid emails you can put in forms”, or there’s a tool that exists somewhere to grab an email off one of the numerous breaches that included it, because it gets used by random people around the world to sign up for stuff.

    This is definitely an “I’m doing this on purpose” because the name is unusual for anyone who doesn’t speak Bahasa Indonesia, and I have never yet had a fraudulent sign-up from Indonesia.

    As I do for my other email, I usually punish them by resetting the password and locking them out of the account. For dating apps I add a really savage profile about how dumb they are.

    But every once in a while there’s one I can’t do this for — Capital One, for instance, allowed ROBIN JEAN (yep, it was all caps) to supply the me.com address as their email for a credit card without verifying that it was accessible by their customer. Their password reset requires, if I recall, the account number to do a reset, so there’s nothing I can do about that one except complain every month when the balance email shows up. (We’re three months in; hasn’t helped, though they keep swearing they’ll fix it.)

    The one I’m writing about today, however,  is one that leaves me gobsmacked. And somewhat alarmed.

    On July 1, I got a purchase confirmation from Roblox that read like this (please note that I do not have a Roblox account):

    Thank you for your purchase on Roblox, the online gaming platform that is powering imagination globally!

Please contact us at roblox.com/support, or call us at +1-855-333-4734 if you have any questions about this charge.

Your 6/28/2020 3:11:10 AM order:
Item Purchased: Roblox Premium 2200
Item Price: CAD25.99
Next Renewal Date: 7/28/2020
Total: CAD25.99

Billing Information:
sdf sdf
pemungkah@me.com
Visa ending in 1563
sd
sd
fsd v6e
United States
Username: 45dfgerdfwerewr
Sale ID: 543250908

You will be charged CAD25.99 per month for this service until you cancel. You can cancel at any time by going to the billing tab of the account settings page and clicking cancel membership. If you cancel, you still will be charged for the current billing period. We hope you enjoy your membership!

    Let’s just luxuriate in the utterly transparent fakery of that address and username for a minute.

    It is blatantly obvious that whoever is using this credit card is not on the up-and-up. So I immediately tried to reset the password. Nope. No password reset email. Well, they allow several other authentication schemes, maybe I can’t reset it this way . I’ll make sure that Roblox Support knows about this; possibly unauthorized, fraudulent charges are most certainly going to be a serious issue for Roblox, and they’ll want to be sure that they’ve protected whoever this actually was, and they’ll take quick action to fix this.

    Ha. No.

    I spent the next eleven days simply trying to communicate that someone was very possibly committing fraud, that I had evidence, and that maybe they should do something.

    Roblox “support” spent that time sending me their form emails about unauthorized charges. Once I battered my way past that, I said, fine, you can’t tell me anything. Please make sure my email is removed from your system.

    They couldn’t find it.

    I supplied the email with full headers.

    Still couldn’t find it.

    Do you have any explanation as to how this order ended up in my mailbox, then? Because it certainly was not me or anyone in my household. I would think this would be an issue, that there are orders going out to emails that you don’t have any record of.

    Time passes. Crickets.

    Then I get the automated “you haven’t replied and we want to close this ticket so our KPIs look good” email. All right, I will explain it carefully so we can perhaps get an understanding going here.

    Hi. Look. This should not be as hard to understand as it seems to be.

I forwarded you an email I got. 

It came to my email address, and had my email address in the purchase record.

The data in the purchase record is obviously random typing on the keyboard.

It’s not my credit card.

It is, however, my email.

SOMETHING must have created this purchase. There has to be an audit trail that points back to some account that this purchase order is associated with, and some transaction that initiated it.

Whatever account it is. Whatever purchase it was.

NONE OF IT should be associated with my email.

Have I made it clear?

    Reply:

    To assist with or provide information about any account, we must first verify account ownership. Unfortunately, there is no email address or purchase information associated with the account. Without this information, we are unable to verify ownership or assist further with the account.

Please make sure that with any account you create, you add and verify your email address. This will allow us to verify ownership and also allow you to use the reset password feature.

    What did I just send you, other than the complete email, with all the headers, containing the account name, the email address, the literal transaction ID for the possibly fraudulent sale…? So I gave up.

    I’m guessing that they may actually have caught that it was bogus right away, and immediately deleted the account, and the stonewalling is to prevent me trying to social-engineer my way into, I don’t know, getting them to confirm the credit card is good or something.

    I’m guessing that there is a  record that this account was deleted because of fraud, but because of policy they can’t tell me that.

    But we’ll never know. To whoever owns the credit card, sorry, I did my best. I hope they did protect you, or that you catch the charge and dispute it.

    I’ll just say that I don’t feel warm and fuzzy about the whole thing.

  • The Harp of New Albion’s Tuning for Logic

    The Disquiet Junto is doing an alternate tunings prompt for week 0440 (very apropos!).

    I’ve done several pieces before using Balinese slendro and pelog tuning, most notably Pemungkah, for which this site is named. I wanted to do something different this time, using Terry Riley’s tuning from The Harp of New Albion, using Logic Pro’s project tuning option.

    The original version was a retuning of a Bosedorfer grand to a modified 5-limit tuning:

    However, Logic’s tuning feature needs two things to use a tuning with it:

    • Logic’s  tuning needs to be based on C, not C#
    • The tuning has to be expressed as cents of detuning from the equal-tempered equivalent note.

    This leads one to have to do quite a number of calculations to put this in a format that Logic will like.

    (more…)

  • Life in the fast lane / Surely makes you lose your mind

    I came back to the Radiospiral iOS app after some time away (we’re trying to dope out what’s going on with metadata from various broadcast setups appearing in the wrong positions on the “now playing” screen, and we need a new beta with the test streams enabled to try things), only to discover that Fastlane had gotten broken in a very unintuituve manner. Whenever I tried to use it, it took a crack at building things, then told me I needed to update the snapshotting Swift file.

    Okay, so I do that, and the error persists. Tried a half-dozen suggestions from Stack Overflow. Error persists. I realized I was going to need to do some major surgery and eliminate all the variables if I was going to be able to make this work.

    What finally fixed it was cleaning up multiple Ruby installs and getting down to just one known location, and then using Bundler to manage the Fastlane dependencies. The actual steps were:

    1. removing rvm
    2. removing rbenv
    3. brew install ruby to get one known Ruby install
    4. making the Homebrew Ruby my default ( export PATH=/usr/local/Cellar/ruby/2.7.0/bin:$PATH)
    5. rm -rf fastlane to clear out any assumptions
    6. rm Gemfile* to clean up any assumptions by the current, broken Fastlane
    7. bundle install fastlane (not gem install!) to get a clean one and limit the install to just my project
    8. bundle exec fastlane init to get things set up again

    After all that, fastlane was back to working, albeit only via bundle exec, which in hindsight is actually smarter.

    The actual amount of time spent trying to fix it before giving up and removing every Ruby in existence was ~2 hours, so take my advice and make sure you are absolutely sure which Ruby you are running, and don’t install fastlane into your Ruby install; use bundler. Trying to fix it with things going who knows where…well, there’s always an applicable xkcd.

    You are in a maze of Python installations, all different

  • Broken iframes and HTML::TreeBuilder

    We had a situation last week where someone had entered a broken <iframe> tag in a job description and our cleanup code didn’t properly remove it. This caused the text after the <iframe> to render as escaped HTML.

    We needed to prefilter the HTML and just remove the <iframe>s. The most difficult part of this was figuring out what HTML::TreeBuilder was emitting and what I needed to do with it to do the cleanup. It was obvious that this would have to be recursive, since HTML is recursive (there could be nested, or multiple uncosed iframes!) and several tries at it failed until I finally dumped out the data structure in the debugger and spotted that HTML::TreeBuilder was adding “implicit” nodes. These essentially help it do bookkeeping, but don’t contain anything that has to be re-examined to properly do the cleanup. Worse, the first node contains all th text for the current level, so recursing on them was leading me off into infinite depths, as I kept looking for iframes in the content of the leftmost node, finding them, and uselessly recursing again on the same HTML.

    The other interesting twist is that once I dropped the implicit nodes with a grep, I still needed to handle the HTML in the non-implicit nodes two different ways: if it had one or more iframe tags, then I needed to use the content method to take the node apart and process the pieces. There might be one or more non-iframes there, which end up getting returned untouched via as_HTML. If there are iframes, the recursion un-nests them and lets us clean up individual subtrees.

    Lastly, any text returned from content comes back as an array of strings, so I needed to check for that case and recurse on all the items in the array to be sure I’ve filtered everything properly. My initial case checks for the trivial “no input so no output”, and “not a reference” to handle the starting string.

    We do end up doing multiple invocations of HTML::TreeBuilder on the text as we recurse, but we don’t recurse at all unless there’s an iframe, and it’s unusual to have more than one.

    Here’s the code:

    +sub _filter_iframe_content {
  my($input) = @_;
  return '' unless $input;

  my $root;
  # We've received a string. Build the tree.
  if (!ref $input) {
    # Build a tree to process recursively.
    $root = HTML::TreeBuilder->new_from_content($input);
    # There are no iframe tags, so we're done with this segment of the HTML.
    return $input unless $root->look_down(_tag=>'iframe');
  } elsif (ref $input eq 'ARRAY') {
    # We got multiple strings from a content call; handle each one in order, and
    # return them, concatenated, to finish them up.
    return join '', map { _filter_iframe_content($_) } @$input;
  } else {
    # The input was a node, so make that the root of the (sub)tree we're processing.
    $root = $input;
  }

  # The 'implicit' nodes contain the wrapping HTML created by
  # TreeBuilder. Discard that.
  my @descendants = grep { ! $_->implicit } $root->descendants;

  # If there is not an iframe below the content of the node, return
  # it as HTML. Else recurse on the content to filter it.
  my @results;
  for my $node (@descendants) {
    # Is there an iframe in here?
    my $tree = HTML::TreeBuilder->new_from_content($node->as_HTML);
    if ($tree->look_down(_tag=>'iframe')) {
      # Yes. Recurse on the node, taking it apart.
      push @results, _filter_iframe_content($node->content);
    } else {
      # No, just return the whole thing as HTML, and we're done with this subtree.
      push @results, $node->as_HTML;
    }
  }
  return join '', @results;
}