You miscast the trunk update. You are blasted with magical energy!!!!!!!

Hello crawlers. The 0.25 development cycle continues apace. Expect an announcement about a forthcoming release and tournament within the month.

This round of updates comes with an overhaul of the spell miscast system, to couple with the magic overhaul of this version. Miscasts have been thoroughly simplified, to the point that the effect of miscasting a spell can be included in its description! Read on for details and the scoop on the latest from the Iron City.

• Zot traps have had their effect flowchart replaced with an effect table with a more focused list of awful terrible things to do to the player.
• The Wu Jian Council met in executive session and tweaked some of their martial arts. Wall jumps move only to the ‘a’bility menu, movement triggering is no more. Whirlwind attacks no longer pin. Heavenly Storm has had its slaying mechanics simplified.
• With miscasts simplified Xom now finds them boring and has turned their attention to summoning more demons on low level players.
• As a prelude to the miscast changes several gods no longer use miscasts in their retribution. The new wraths are:
• Trog replaces fire miscasts with pure fireballs in her fiery rage.
• Lugonu exchanges translocations miscasts with a chance to banish, summon Abyssal things, or teleport the penitent to nearby monsters.
• Cheibriados replaces miscasts with a chime on their clock before putting the player to sleep at high tension, and stat damage at low tension.
• Fedhas’ elemental damage is now more focused: a corrosive bolt, a primal wave, or a thorn volley await those who would trample the flowers.
• Yredelemnul prefers the simplicity of a bolt of draining to necromancy miscasts.
• Kikubaaqudgha will still impose necromancy miscasts on those who attempt to use necromancy while under penance; for outright retribution they now apply the same ancient and terrible death curses used by mummies.
• The scythe of Curses also replaces necromancy miscasts with death curses.
• The distortion brand no longer teleports on hit; unwielding a distortion weapon has a chance to teleport the unwieldier to nearby monsters, contaminate, or banish.
• Regeneration items (troll armour, artifacts) must attune to a player before granting their effect. No more carrying TLA to swap while resting!
• Mummies have consulted the book of death and determined that necromancy miscasts no longer suit their vengeful needs. Death curses are now their own effect table, separate from miscasts. Kikubaaqudgha was involved as well, and still protects loyal followers from curses.
• New monster: nameless horror. Nameless horrors are brought into the world by summoning miscasts, have an antimagic attack and abjure other summons as a natural ability.
• Absolute Zero has been reduced to a fixed range of 5 (up to 6 with Vehumet), the noise now scales with power down to 0 at maximum power.
• Leda’s Liquefaction regains its slow movement penalty, but the caster remains immune to its effects on melee combat.
• Deflect Missiles has been removed. The tedious-but-optimal armour-wizardy-int dance has danced its last!
• Lesser Beckoning and Teleport Other now turn allies hostile.
• Tukima’s Dance cannot be used by Ru followers who have sacrificed Love.
• Two new Gauntlet maps are live: one presents a sequence of arenas, after each you can chose to press your luck or turn and face the minotaur. The other has a choke point in the sequence: there is only one middle arena in the three-arena-sequence to the minotaur.
• Miscast effects have been simplified. When miscasting a spell, the player receives magical contamination, and if the miscast is bad enough an additional effect that scales with the spell’s current failure chance. These effects vary by school as follows (compare to the old miscast effects, which could you even correctly lists all of them without a source dive?). For damaging effects the damage message is punctuated like other damage messages.
• Charms and Hexes: debuff and slow
• Summoning: durably summon a nameless horror
• Translocation: dimension anchor
• Transmutation: polymorph
• Conjuration: irresistible AC-ignoring damage
• Elemental schools: school flavored damage (Earth magic does fragmentation damage like Lee’s Rapid Destruction).
• Monsters on D:1 are now back asleep, after briefly getting into the coffee.

You mumbles some strange words. You chill the trunk update to absolute zero! The trunk update is frozen into a solid block of ice!

Greetings fellow crawlers, it’s time for another Trunk Update. The 0.25 development cycle is slowly winding to a close, but we have some changes for you to try out, with more to come! Expect a release and tournament in the next month or two, barring any unforeseen delays. As always, we’ll make an announcement once we have a date in mind. This update features a couple larger changes we’ll give some extra detail about, along with some smaller updates described further below.

First up is the new L9 Ice spell Absolute Zero, which replaces the L9 Ice/Conj Glaciate spell. Level 9 spells are tricky to get right since they cost so much XP, yet we want to give players a reason to learn them. We also want these spells to be both distinct from the other L9 spells and relevant to their school; the latter is especially a focus with our spell changes this version. Glaciate fits reasonably into the Ice theme we’ve established, being a “diffuse” AOE spell, but it has too many similarities with Fire Storm yet functions like a worse version of that spell! Where Fire Storm is one of the best ways to kill monsters safely, Glaciate requires close proximity to be effective, and is far less good at blocking monster movement and line-of-fire.

Instead of reworking Glaciate and likely ending up with something similar to Fire Storm, so we went for more distinctiveness with Absolute Zero. This spell kills the closest target, leaving behind an ice block. It chooses randomly among the closest monsters if there’s a tie by distance. This kill is unconditional, so it works on everything from orbs of fire to Cerebov! The positional aspect comes getting the right monsters close enough when multiple monsters are involved, which is usually the case for dangerous fights late in the game. Absolute Zero might feel too cheesy at its max range, which scales with spellpower up to 7. We may reduce the range to a fixed lower value of, say, 5, and have noise scale down with spellpower instead. The latter aspect was in fact part of hellmonk’s original implementation of this spell. Thanks to him for that patch as well as other Discord regulars for discussing the idea.

Next we have a round of item removals; that’s right, the dreaded r-word! On the chopping block are: lamp of fire, sack of spiders, fan of gales, crystal ball of energy, and staff of power. See the links for kate’s brief description of why each item was removed. The crystal ball and staff of power removals won’t surprise many long time crawl players. The former is notoriously useless in most situations and the latter is an annoying swap with an effect duplicated by other equipment. As an aside, we’d like all magical staves to have a good melee effect; see below regarding staves of poison and Olgreb, and look out for similar changes (or removals!) in the future.

Regarding the other removed items, we want the set of evocables be relevant reasonably often, as distinct as possible, and not too numerous. The fan of gales was an escape item you’d use very infrequently, and this game already provides so many escape consumables, spells, abilities, and god powers. The lamp was a strong item, yet its design was underwhelming, as kate says. Its notorious targeter was awful to use in open spaces, and while it was very good in corridors, we already have an evocable that makes clouds and which is good in both open and closed areas. The sack was a very strong item, like all summoning consumables, but the summons it created aren’t memorable, at least not until getting ghost moths at high evocations. We have a bunch of other consumable items that create interesting summons and that everyone has to carry around as things are. The sack’s net effect might show up on some future item or ability, but not attached to summons.

These removals make the remaining evocables feel more distinct, free up a lot of item slots, and reduce the cognitive load in a game with such a large number of items. Speaking of loadstones, if you’ve been using the new tin of tremorstones and found it underwhelming, have no fear. We’ll be taking a look at these and seeing how they can be improved before release. Now, for the rest of the changes:

• Phial of floods now applies a silencing “waterlogged” debuff to all monsters in the flooded area and no longer summons water elementals. So you can still get those cool silencing debuff on monsters, just without so much accompanying summoning cheese.
• D:1 level spawns can no longer generate within LOS distance of the player’s starting position. A surprising nerf to tiny arrival vaults, to be sure, but a welcome one.
• Potions of Brilliance now only provide a stronger universal spell enhancer and remove spell hunger. They no longer provide an Int bonus nor wizardry.
• Staves of poison now do resistable poison damage on hit like other staves instead of just having a chance to poison.
• The Staff of Olgreb now has a chance to deal poison-arrow flavored damage on hit, based on evocations skill. It no longer has an additional chance to cast Venom Bolt on top of casting OTR when evoked. Once again, I await kills from the Gauntlet minotaur wielding this.
• Potions of Might no longer provide a bonus to strength. Don’t worry, the 1d10 bonus damage remains, and that’s the really good part.
• The Dragonskin Cloak now provides rCorr instead of sticky flame resistance. Waiting for the complaints to roll in from the die-hard rSticky fans.
• The -Tele property no longer appears on artefact weapons and jewellery. Sorry, no more cheesing those Zot traps with your -Tele stick!
• Ziggurats now have a level set featuring many player ghosts. Zig ghosts are back! Accompanying them are a variety of undead assistants. This set replaces the “draining” monster set, since that idea is well covered by this new set and the existing ones.
• Frozen Ramparts has its radius of effect reduced to 2 and its damage was slightly reduced. Even with the removal of the monster slow movement debuff, I found this spell to be too strong, doing lots of damage over huge areas in many situations. With these nerfs, Ramparts feels like it has damage more appropriate to a level 3 spell.
• Many arrival vaults have been reworked to allow better player tactics. After all, tactics are important, or so the wiki guides say.
• All Evocable items can no longer be used by the player while confused. Target fuzzing, be gone!
• Trog now hates use of all magical staves and pain weapons. On the upside, Troglodytes won’t get offered these by scrolls of acquirement!
• Nemelex abilities can no longer be used while silenced. This was a loophole from when decks used to be items. Remember those long gone days?
• Call Imp no longer bases the type of imp summoned on spellpower. Don’t worry, this spell is still really good, crimson imps included!
• The WebTiles server has seen a major overhaul and now supports python 3 as well as Tornado 5+. Thanks to advil, aidanh, and chequers for all their work on this development and testing. More WebTiles updates will be rolled out after testing, with work to come from aidanh on making it easier for us to upgrade server chroots. That way we can have a modern development toolchain on every server.

Play-testing trunk is always appreciated, especially as the release gets closer. Please report any bugs you find on mantis or on the github issues page. I myself had great fun ascending a merfolk ice elementalist to experience all the changes to ice magic; looking forward to trying a fire elementalist or venom mage next. Look out for more trunk changes and likely another Trunk Updates before next release. So until next time, Happy Crawling!

You quaff a potion. It was a potion of stabbing. You feel ready to backstab. You open the trunk update like a pillowcase!

Greetings fellow crawlers. The 0.25 development cycle has proved to be a busy one, so we’re back with another Trunk Update. Many of us usual suspects have been hard at work bringing you some new toys to play with. I’m also happy to note that PleasingFungus has recently returned to drop some new crawl creations of his own. How he managed to escape from the dreaded Tower of Lemuel, we’ll never know!

Now, on with the update:

• Frozen Ramparts, a new L3 Ice spell that makes all walls within radius 3 become covered in icicles that release when a monster walks nearby. The icicles do light damage that checks AC but is only partly resistable. Affected monsters additionally receive a “frozen” debuff that slows their movement by 4 aut for a few turns. Available now for any Ice Elementalist in their starting book!
• Tin of Tremorstones, a new multi-use evocable. When used, it creates two radius 2 explosions at range 3 that are triple-affected by AC (like LRD), but the explosions are randomly skewed and may hit the user. Most useful for players with good armour. It has a 33% chance to fall apart with each use. It’s a tin of angry earth elementals waiting to explode!
• New Potion of Stabbing that gives a 50% chance to upgrade weak stabs (e.g.distracted monsters) to strong ones (e.g. sleeping monsters). I wonder what this potion would taste like?
• New unrand: Cigotuvi’s Embrace, a +4 leather armour with rN+, rRot and *Drain that automatically gathers corpses to increase AC. The AC decays slowly over time but decays more quickly as more corpses are added. That’s right, the former Cigotuvi’s Embrace spell has returned, but as an unrand!
• Ledas’ Liquefaction gets a boost: the player is now immune to the effects of liquefied ground made by casting it. With this spell, we can all pretend to be a centaur.
• Scrolls of Magic Mapping now reveal any floor traps on the level. Now you can feel the pain of finding of a double-zot-trapped Hall of Zot:5 right after you enter the level!
• The vorpal weapon ego no longer has distinct adjectives for each weapon class. Now all such weapons are described as ‘vorpal’. Vorpality is the new reality.
• Shields have gotten a reskin. Ordinary shields are renamed to kite shields and large shields have been renamed to tower shields. Sword-and-tower sounds cooler than sword-and-board, in my humble opinion.
• Confusing Touch is now level 3 and checks MR instead of monster HD. Additionally, the Confuse spell has been removed. Now there’s exactly one “confuse exactly one thing” spell and it’s not so overpowered.
• Boots of the Assassin are now called the hood of the Assassin and use the headgear slot. Imagine wearing this while quaffing a potion of stabbing and standing next to a distracted Rupert. Nothing personnel, kid.
• The demon blade Leech gains *Rage/+Rage and loses its AC-3 and EV-3 properties. Bloodbane has been removed, having been merged into Leech.
• Warlock’s Mirror can now reflect piercing ranged weapons. Is there anything this little gadget can’t do?!
• God conducts for harming allies now apply to spells and items that place harmful clouds. But don’t worry, demonspawn, your ignite blood mutation won’t anger Oka.
• The Staff of Olgreb now grants poison immunity to monsters wielding it. The gauntlet minotaur can’t wait to try this out when he next finds Olgreb in his loot pile.
• Potions of Agility have been removed. Among the less interesting of the many buff potions that fill our inventories, agility potions have been replaced with the more specialized potions of stabbing.
• Monsters such as spiders can no longer cling. Rest in peace, a myriad of clinging code bugs.

Some of the above items will continue to receive balance adjustments and tweaks over coming weeks. And we’re not done adding new stuff for 0.25, so stay tuned. Until next time, Happy Crawling!

Hello crawlers, and welcome to the first batch of trunk updates of 2020. The teams’ celebrations raised quite a bit of Pandemonium, but in the fray several trunk changes appeared. Here’s the latest!

• A new swamp worm tunnel vault offering a transporter puzzle for your death enjoyment
• A new Fedhas vs Yredelemnul themed swamp rune vault with a cameo by an undead Lernean Hydra
• On the back-end we’ve added catch2 testing to the build process
• Foxfire no longer damage the player when swapping, just dissipate
• A new abyss-escape themed ghost vault
• Regeneration has regenerated for the last time, and is removed. Agony replaces it in the book of Necromancy, and Ozocubu’s Armour in the Book of Battle
• A new unique Maggie enters the dungeon. Maggie is a younger version of Margery, and is a test of reaverb’s new multiple-unique-encounter storytelling approach. Maggie knows Bolt of Fire and Mesmerise, and comes equipped with a light dragon skin and a good weapon. Margery’s dragon armour is upgraded. Both Maggie and Margery have a chance of carrying Wyrmbane, which comes already slightly leveled up!
• dtsund can send us an “I told you so!”. Scrolls of acquirement now offer a choice of 4-5 items: food (if you eat), gold, and three random choices from the old acquirement categories. These items are fully identified, and you can inspect them before making your choice.
• However, the miscellaneous category of acquirement has been removed, so you won’t be seeing evocables on the list ever again!

That’s all for now! On the horizon, there are rumors of an upcoming crawl sudden death challenges “soon”, more magic changes, and who knows what manner of twisted sinister changes from the mind of Xom. Happy crawling!

Hello Crawlers, and welcome to a new batch of trunk updates. This sometimes-regular blog feature has seen the dark of Kikubaaqudgha and risen from the grave. There’s no telling whether or not it will survive the whole 0.25 development cycle. But while it’s here lets get the latest from Tar:7.

• Various bug fixes from the 0.24 tournament. The 0.24.1 bug fix release is coming soon.
• Monster clinging has finally lost its grip and is now removed
• Boris will now pursue you on the orb run
• While hostile guardian golems were great ways to make the game harder, summons now dismiss when attacked instead of becoming hostile. The exception is Greater Demons, who are none to pleased to be bothered with your mortal issues.
• Local tiles gains monster threat level indicators which can also be enabled in WebTiles with the tile_show_threat_levels option.
• Fedhas now protects your plants from collision damage.
• On the back end,
• aidanh’s long running ui overhaul has made significant strides
• Monster pathing code has received some optimizations, which should reduce lag when stepped from time. (Cheibriados regrets the need for haste).
• Carrying corpses has been removed again.

Looking to the future, my positional attack magic experiment has grown and twisted. It’s just been merged, details will appear in the next trunk updates post!

What is seeding?

The 0.23 release introduced a major and ongoing overhaul in how DCSS approaches random seeding from the player’s perspective. If you’re not familiar with this idea from other games with random procedural generation like Minecraft, Spelunky and brogue, here’s how it’s supposed to work: if you enter a specific code (the “seed”) into the game when starting it, you’ll get a repeatable form of randomness that encompasses dungeon layout, items, and monsters. Any dungeon generated with that seed should be identical. This has many applications: from Spelunky or Brogue (and many others) you might be familiar with the concept of a daily, weekly, or monthly run. In these, all players enter the same seed and over some time interval explore the same dungeon. Here are a few ways crawl community has been experimenting with this since the release:

• On the CPO server, run by chequers, there is an experimental weekly online challenge with a share dungeon.
• Ultraviolent4 has started running an offline Game of the Month challenge. (Edit: The next challenge begins on Apr 7, per uv4.)

I’m excited to see what else develops. But what does it take for this to actually work? This post goes into the background on what has been involved in doing this retrofitting for DCSS, where it is at, and what the remaining problems are.

Randomness and procedural generation

Procedural level generation happens via constrained randomness. Some process repeatedly draws on a random number generator according to a fixed, predefined set of constraints and rules, making choices that (ideally) generate a coherent, explorable, locale. At the heart of this is a random number generator, which usually takes the form of a function that the programmer can repeatedly call, generating a sequence of numbers.

Now, “true randomness” is more of an ideal than anything that can be implemented, so RNGs are really only “pseudo-random”. What this means is that they aim to (in addition to matching some overall distribution) be unpredictable, which a viewer or adversary will perceive as random. Unpredictability can be defined in precise mathematical terms that I won’t go into here, but it doesn’t actually mean complicated: some of the best rng algorithms are remarkably simple. They basically all work by initializing the rng state with a “seed”, and repeatedly transforming that state, and the heart of the one we use, PCG, is just 5 lines of code. For cases where there is an adversary, you want this initial seed to itself be unpredictable, and so common strategies are to use some external state of the program (e.g. exact time, external devices) to do this initialization: then, if the adversary doesn’t know the initial state, and the algorithm is unpredictable in the right way, they’re out of luck.

Video games flip this idea on its head. For some kinds of RNGs, if you initialize the generator in the same way, multiple instances will generate the same sequence of numbers in the same order, while still behaving unpredictably and well-distributed as long as the seed is unknown. This means that if the rules and constraints that draw on the RNG are implemented correctly (we’ll come back to this), the behavior given that seed acts random but is completely replicable. You can think of the RNG as, given a seed, providing a “code book” of integers that acts random, but as long as you have the code book and the correct rules/constraints (giving you position in the list), you can decode the behavior of the game.

There are, then, two main ways to make this go wrong: mess up the initial conditions, or cause the rng to get out of sync. This latter case is when two processes are using the same sequence, but make decisions starting from different points in the sequence. Because of unpredicatability, their behavior will diverge extremely rapidly (more on this later). In general, if RNG behavior diverges at all, it tends to diverge thoroughly and quickly.

DCSS has long used the PCG RNG which (like the Mersenne Twister, xorshift, and many others) supports this reproducibility; retired dev bh added this in 2015, replacing a hand-rolled implementation of an algorithm from the 80s that was more predictable. However, using a good RNG algorithm is only one piece of the picture, and turns out to be probably the least complicated part of this project.

Why player-facing seeding was (and is) hard for DCSS

Crawl wasn’t remotely designed with seeding as a player-facing feature in mind; though many roguelikes have used specifiable seeds for testing, the mainstream use as a player feature seems to be quite recent. Actually, I couldn’t find out much of the history of using seeding for completely repeatable procedural generation as a player-oriented feature, but it doesn’t seem to have been very mainstream before games like Minecraft and Spelunky; an earlier example, pointed out to me by Roguelikes discord user esc, is Civ 3. Seed-based pseudo-randomness, internal to the game, is much older, and for example Moria allowed choosing a seed for testing purposes. But getting seeding to work as a player feature is a whole different project.

Here are four challenges that DCSS faced (and still faces) when trying to get player-facing seeding to work well:

1. Mixing of gameplay / levelgen RNG
2. Arbitrary interactions in level generation between levels.
3. Extremely complex level generation with a huge diversity of procedural generation techniques, implemented in a mix of C++ and Lua code.
4. Diversity of build targets and settings: support for windows, mac, linux, with 32 or 64 bit builds. Builds on non-x86 architectures such as ARM.

When starting the project, I was aware of factors 3-4, but thought 1-2 would be the biggest challenges, and so it was there that I began. In retrospect, 4 is actually the biggest problem.

Before 0.23, generating a level and entering a level were completely intertwined: levels were not generated until you entered them. “Greedy” level generation goes back to Linley’s Dungeon Crawl as far as I’m aware, and on its own terms is a fairly simple and clean approach, spreading the CPU cost of level generation across the game in a way that hides it from the player. However, it can lead to a problem illustrated by the following toy example: suppose that you have a separate RNG for each level, the player is on level 2, and for both levels 3 and 4 Sigmund is poised to generate given the RNG state. If the player takes stairs to level 3, then he’ll generate there; if the player takes a shaft to level 4, then he’ll generate there. Since Sigmund is a unique monster, he cannot then regenerate if the player enters levels 4 and 3 respectively, and so the RNG diverges right away — the builder may place another unique, or do something different altogether, and after that initial decision will tend to be completely different.

Not only that, this first assumption wasn’t made: DCSS before 0.23 used a single RNG for gameplay and levelgen decisions, so any different action the player took might lead to completely different level generation.

Part 1: pregeneration and RNG separation

This problem has long been known, and comes up regularly in discussions of seeding. I adopted two solutions from these discussions: separate out levelgen RNG from gameplay, and allow pregeneration of the dungeon. The latter solves almost all of problems 1-2 on its own, and had several past implementations by community members. The most developed of these was due to tavern user bel in the above linked thread, and just needed some adjustments so that level generation was better separated from level entry. The initial seeding patch can be seen here.

Incidentally, if you’re designing a game from scratch, your life is going to be easier if you take a different strategy for these sorts of non-independent decisions. For example, you could place uniques independently of levelgen (something we do now only for one unique, Boris, who respawns), perhaps deciding where they go at the beginning of the game. However, it’s probably too late for DCSS in this respect.

Part 2: testing and fixing many, many seed stability bugs

This patch, it turns out, was only the beginning, opening things up to a whole world of ways in which a repeatable RNG can get out of sync based on problems 3-4.

To start with, I wrote a lua test that, given a seed, prints the vault list for that seed, re-runs generation, and compares the result. While true seed stability bugs on a single device have been fairly rare, this second part helped identify a tremendous amount of global state that can matter for seeding, scattered all over crawlcode. You can see an example of the output here if you scroll down to “Running test #31: ‘vault_catalog.lua’.”

The real challenge has been seed stability across devices. Here is a sampling of problems that have come up:

• If you’re drawing randomly from a list, and something permutes the underlying list, it doesn’t matter if your rng draw is the same! For example, if the OS provides a directory listing in a different order.
• If the order of some calls isn’t determined by the language/library standard, then it will probably be implementation dependent. This is a problem that has come up again and again:
• C++ (and C) doesn’t define the order of multiple function calls that make up an arithmetic expression, and in practice their order varies. You might think this would go by order of operations, but this would be wrong. They all have to be explicitly sequenced.
• Lua iteration order, both through the C interface, and via the “pairs” function, is not guaranteed to be stable.
• Anything in C++ that sorts may have implementation-specific behavior if a comparison operation indicates two objects are equal. This is fine if your comparison operation is fine-grained enough, but sometimes it wasn’t.
• A large set of bugs that were introduced with the initial pregeneration patch because the level builder (and related code) assumed the player was about to enter the level; something that could be designed around from scratch but was fairly baked into dcss as it stood.
• DCSS caches parsed maps, and some cases of randomness in the cache were causing divergence.
• Cases where the builder made use of player state: the most prominent of these were vaults that generated items using the acquirement code. (This still can happen in troves, but only there.)
• Floating point issues. First, Lua 5.1 uses different Number to int conversion techniques on different build targets, and these have different truncation behavior. Then, even controlling for this it turns out that e.g. SSE2 behavior for non-exactly-represented integers can lead to different ints than x87, for example, (1 / 100 * .7). I doubt all of these have been found yet, and for the 0.23.2 release I still needed to manually force the windows build to not use x87. I’m looking at you layout_delve…
• Cases where unexpected code was drawn upon by the builder. For example, levelgen tries to ensure connectivity, and it turned out the code that does this was being affected by certain rc options! Another example is that slime connectivity in particular uses a different algorithm (because of the wall damage) that led to an underspecification bug mentioned above.

In effectively all of these examples, the result of the bug was immediate loss of seed synchronization across devices — and many of these were things that wouldn’t be considered bugs at all without the goal of seed stability. Debugging them was uniformly extremely difficult, because it required both replicating the divergence and then finding the exact point of RNG divergence. Without VMs and docker, I would not have been able to do most of this.

Some of these problems can be programmed defensively against, with experience, but I doubt I’ve found all the seed stability bugs in crawlcode. At this point, I’m mostly reliant on people testing seeds on various platforms — Ultraviolent4′s seeded Game of The Month series has so far been invaluable for finding seed/pregen bugs. (Incidentally, the challenge of localizing seed bugs, together with the challenges of developing for seed stability, are why I would strongly recommend fellow roguelike devs against using seed-based save formats…)

Part 3: next steps

There are still many improvements to be made, and I expect a lot of my effort for 0.24 will go towards finding and quashing more seed stability bugs. Here are a few improvements I hope to include in the next version:

• Incremental pregeneration: generate the dungeon in a predetermined order, but only do as much as you need to do to play catchup on level changes. This will allow hopefully removing the distinction between pregenerated and non-pregenerated games, along with allowing seeded play online. (Right now, the main barrier is that up-front level generation is very hard on a server’s cpu.)
• More RNG improvements for repeatability: apply various techniques to make rng de-synchronization less impactful. Remove floating point code as much as possible, and improve some of the lua numeric issues further.
• Non-numeric (readable) seeds, a la cogmind.
• A seed catalog of some kind.
• What seed-related features would you like to see?

The DCSS community has been extremely helpful in finding and reporting bugs, and I doubt we are done. Please go play one of the seeding based challenges, and let us know if you find problems!