Release log#
Note
3.1.1 was the first release, tagged on 18 August 2026 — a macOS disk image and a source archive. The most recent is 3.111.3, and all of them are on the Releases page. Every release so far is marked a pre-release.
What follows is the development log, grouped by the milestones the work is organised into, newest first, with everything done since the last tag under Unreleased.
This page covers ZangbandTK only. For Angband’s own long history, which this game is built on, see Version Information.
Development is tracked by milestone. M0 to M10 are complete: the balance scalars, the bestiary and treasure, the effects and vaults, the wilderness, the towns, the quests, the races and classes, the mutations and virtues, and all seven realms of magic. The Warrior-Mage and the High-Mage, which M7 could not finish because they are defined by their choice of realm, arrived with M9.
M10, pets, is done. All nine phases — allegiance, the AI, orders, upkeep, player-side safety, the sources, Trump, following you downstairs, and the stable that costs you one in twenty at each level change — and, in 3.92.0, the last two of PLR-28’s five ways to acquire one: a Mindcrafter’s domination and a Chaos Lord’s gift. Both turned out to be data rather than the code hook they were recorded as needing. Two of Zangband’s routes remain and want objects that do not exist here: a thrown figurine, and a wand of charm monster.
M11, nightmare mode, is half done. Stage 1 — every behaviour Zangband’s own nightmare mode has, sixteen of them — is complete. Stage 2 is the menu of things its spoiler describes and Zangband never built, and that is a scheduled decision rather than a closed one (DEC-84). See Features for what all of this adds up to in the game, and for what is left.
Version numbers move with the work — patch for a fix, minor for a feature, bumped in the commit that does it — so a build can be identified from its title bar. They begin at 3.0.0, continuing Zangband’s own line from 2.7.5-pre1 rather than the Angband 4.2.6 the code sits on.
Unreleased#
The borg starved holding a meal — 20 September 2026#
3.124.2 — The borg counts every food as food, not only the ones Angband named (BRG-13, DEC-89).
The symptom looked like shop selection: walk into a shop, note “would prefer ‘2’”, leave the level. Traced, that is not it. The borg reaches the General Store perfectly well and sells its food there, then gives “5 Food” as the reason it will not dive.
borg_notice()classifiedTV_FOODby a list of svals, and the list is Angband’s — apple, ration, waybread and eight more. This game imports one food of its own, Strips of Venison, and it is not on the list, so it counted for nothing: the dive gate saw an empty larder and the sell logic dumped the venison as worthless on the way past.The mushroom branch three lines above already asked the right question — whether the item’s effect feeds. The food branch now asks it too, as a floor under the named list rather than a replacement: the names keep upstream’s calibration, and anything else that feeds counts as low-value food instead of nothing. Only
INC_BYandINC_TOcount;DEC_BYandSET_TOare what a mushroom of Purging does.Measured over twelve runs of 60,000 turns: food in the blocked tally 4 → 3, sum-depth 21 → 23, deepest 3 → 4, best character level 6 → 7. Warrior seed 7 went from allowed depth 1 on “5 Food” to allowed depth 4 on “restock recall”. Casting drifts 20 → 18 and shops 17 → 16 because the runs diverge once behaviour changes — one Priest now dives further and dies sooner.
The three runs held on “2 cure” are a different problem: there are no imported healing potions, so nothing is mis-classified there. That one is affordability, not naming.
borg/preparednow walks every food kind in the game and asserts that what feeds is recognised and what empties the stomach is not, so the next imported food cannot reintroduce this quietly.
Two borg diagnoses, both wrong — 19 September 2026#
3.124.1 — A shops-reached counter for the borg, and the two fixes it disproved (DEC-88). Only the counter shipped.
There was no figure for how many shops the borg had actually been inside, so one was added. Its first reading was zero across all twelve runs, which looked like confirmation of the reach bug. It was confirming nothing: the counter was zeroed in the status reporter, so it was cleared immediately before being printed — runs whose own notes read
# Currently in store '8'reportedshops=0. A measurement that always reads zero looks exactly like a behaviour that never happens, which is the defect this counter exists to catch, committed while building it.Fixed, the baseline is 17 shops across twelve runs. The borg reaches shops, and that invalidates the premise both fixes were built on.
The reach fix moved nothing and its diagnosis is false. Moving the
GOAL_DARKresumption below the world crossing left every total and every individual run byte-identical. Traced,borg_flow_world()is entered 32 times in 60,000 turns with and without the change — it was never being pre-empted. It declines becauseborg_choose_town()finds the borg already near a town stocking what it wants. There is nothing to cross to. Reverted.What is actually wrong is a shop-selection defect: the borg walks to a shop, records “would prefer ‘2’”, and then leaves the level rather than walking to store 2 — while blocked on food, in a town with a general store.
BRG-11a is a rewrite, and the numbers say so. Its
CLEVEL <= 8list matches monster names, and the names are the starting town’s. Skipping those checks on the surface cost the fleet nine character levels, a death and four spell casts for one shop fewer and no change in depth — the fleeing is protecting a level-one character from town monsters that can kill it. Upstream still carries its own!FIXon this at 4.2.6, so there is nothing to inherit. Not attempted.
A Draconian Mage breathes disenchantment — 19 September 2026#
3.124.0 — The Draconian’s class-substituted breath (PLR-01, DEC-87), which closes the last open half of DEC-80. From level 15, with a chance equal to your character level in a hundred, the fire or cold is replaced by a pair belonging to your class.
The mechanism is general, because it will be wanted again. A power band can now carry two optional qualifiers:
power-when-class:restricts it to named classes, andpower-chance:makes it an alternative — bands with a chance are tried in order and the first whose roll succeeds is the only one that runs, while bands without one always run. That is what lets one power hold seven class tables and a fallback and fire exactly one of them. The mutation firing loop reads the same two fields, because it is the same struct.Class names are matched at use rather than resolved at parse time, because races are parsed before classes — which makes a misspelling a band that silently never fires.
player/racenow asserts that every class name used by any race resolves to a real class.My earlier summary of the archive’s table was wrong in three ways, which is the reason for re-reading it. It missed that the Ranger shares the Warrior’s branch, the Warrior-Mage and High-Mage the Mage’s, and the Paladin the Priest’s. It called
GF_MISSILE“raw force” where the archive calls it “the elements”. And it implied even splits: the archive writesone_in_(3)for some branches and!one_in_(3)for others, so a Warrior breathes shards twice as often as the elements and a Chaos-Warrior breathes confusion twice as often as chaos.Two elements had to be built to carry it. Zangband’s confusion is a damage type that also confuses, where 4.2’s
MON_CONFis a status effect dealing nothing — mapping one to the other would have turned two thirds of three classes’ breath into a spell that does no damage, which is valid and silently wrong. And holy fire is not hell fire: hell fire doubles against evil and isHOLY_ORBexactly, while holy fire also makes good creatures immune and everything else resist, so takingHOLY_ORBfor both would have collapsed a Priest’s pair into one element.Druid, Necromancer and Blackguard breathe fire and cold, and that is a judgement rather than a lookup. Zangband has exactly eleven classes and its switch covers all eleven, so the missing
default:is not a designed fallback — it is a table that never needed one. Those three are Angband’s own, and falling through is the only option that invents nothing.The shape stays a cone, as ruled: 4.2’s breath is what every dragon in the game uses and what a player reads as breathing, where the archive throws a ball. Recorded again in DEC-87 so it reads as decided rather than forgotten.
Four tests, four falsifications. One of them was void the first time — the build failed on an unused variable under
-Werrorand the stale binary passed, which is exactly the false verdict the falsification harness now greps for. And a test that had been passing was found stale after a restore landed inside the same timestamp second: the binary had not been relinked against the restored library, and 298 of 600 iterations were reporting the falsified behaviour.
The Mindcrafter drinks what it hurts — 19 September 2026#
3.123.0 — Psychic Drain, the eleventh psionic power and the only one that pays you. It shipped as a psi ball and nothing else, on a note in
class.txtsaying neither half of it could be expressed here. Both can.First, a correction to the premise. This was asked for on the understanding that a Mindcrafter cannot recover mana at all — which I had said, and which is wrong.
calc_mana()gives a power-list class a pool (PLR-06) andplayer_regen_mana()runs for every class with no gate on any of them, so a Mindcrafter regenerates like every other caster and a night at an inn fills it. Zangband’sregenmana()is equally unconditional, so the archive’s Mindcrafter was not stranded either. Psychic Drain is the fast, active route — it turns a fight into mana instead of making you stop. The regeneration is now asserted in a test so it cannot be mislaid a third time.What is true, and is now in the manual: no shop sells anything that restores mana. Clear Mind and Debility are mushrooms and Restore Mana is a potion from 750 feet; all three are floor finds. The class is not stranded, but it cannot buy its way out of an empty pool.
Neither end needed a decision. The conversion is
damroll(5, dam) / 4capped at the maximum, and what cannot roll dice against damage another effect just dealt is a data chain — a projection handler has the damage in front of it, which is where the archive puts this too. The extra cost isp_ptr->energy -= randint1(150)on top of the turn the power already spends, and 4.2’s player energy is the same signed counter drained by the same subtraction. “No effect that spends energy” was true of the effect vocabulary and said nothing about the field.The branch order is the behaviour. Zangband’s conversion is an
else ifafter the resist test, so a mind that resists gives no mana rather than a third of it, while still taking a third of the damage. An empty mind gives nothing and takes nothing. And an undead or demon above your level turns the drain around one time in two, costing you mana and hit points unless you save.One deliberate divergence, recorded in DEC-86: the archive charges the extra time only when the ball was seen, so draining something invisible was free. That is an artefact of reading a return value as “did it hit”, and the energy is now charged whenever the drain finds a mind — and still not at all when it finds nothing.
No other class has the same hole. All twelve psionic powers are present at the archive’s levels and only this one was a shell; the other Zangband mana-gain routes are natural regeneration, the inn, the EAT_MAGIC mutation (built, as
TAP_DEVICE) and Omnicide’s soul absorption, which is a separately recorded realm deferral.Nine tests, seven falsifications. Two of them changed the tests rather than confirming them. Doubling the conversion rate did not fail the test meant to pin it, because a level 25 Mindcrafter’s whole pool is 26 points and the ceiling was clamping every sample — the test was measuring the cap, not the rate. And the suite flaked one run in eight on level generation dropping the player somewhere with no free neighbour, which presented as the spell failing; the setup now rebuilds the level until there is somewhere to stand a monster.
You’re dead, and that’s the risk — 18 September 2026#
3.122.1 — The wraith-form death is attributable, and the manual says plainly what the mutation can do to you.
The project owner ruled on the expiry case independently of the archive and landed in the same place: “so what happens when it expires while you’re inside rock. You’re dead. That’s the risk.” Good agreement, and the code was re-checked against both rather than against either alone — the floor is lifted by
!porous, so the damage applies at any hit point total and kills.What it did not do was say so.
died_fromis what the tomb prints after “by” and what the character dump prints after “Killed by”, and it read “solid rock”. A player who walked into a wall on purpose and died of generic crushing has been handed something that reads like a bug rather than like the risk they took. A character holding the mutation now dies of “an expired wraith form”, and the message is “Your form solidifies, and the rock closes!”The generic branch stays for a character who has never held the mutation, because that case is a bug and should read like one.
A Spectre is safe, and it is the race doing it. The crushing branch reads
OF_PASS_WALL, not the timed effect, so a race holding the flag permanently keeps it when the form ends and drops back to the floored density damage. Asserted on the race directly: a Spectre at depth 30 pays for the rock, pays from one hit point, and does not die. The same is true of anyone carrying an item that grants the flag. Had it gone the other way it would have been a bug — a mutation killing a Spectre for doing the thing its race is for.The manual leads with the danger instead of mentioning it. The warning is now the first thing in the section, and it carries the numbers that make the risk calculable rather than a surprise: the form lasts between half your level and your level in turns, rolled on arrival, so a level 20 character gets between 10 and 20 and must assume 10. The crushing is one point a turn on the surface and six at two thousand five hundred feet. Both ways it can end early — light, including light you are immune to and light you cast yourself — are stated, along with who is exempt.
Two tests added, both falsified: making the crushing branch ignore the permanent flag fails four, and reporting the generic killer fails one.
The deferral that stopped being true — 18 September 2026#
3.122.0 — Wraith form, the one deferred mutation whose stated reason had since expired (PLR-16). It was written off in 3.45.0 with the note that “4.2 has no wraith form and no player-passes-walls state at all”, and both halves of that stopped being true when the Spectre work built
OF_PASS_WALLandplayer_can_pass_walls()(DEC-74). The timed effect grants the flag throughflag-synonym, so it inherits the movement, the permanent-wall exception and the wall damage rather than reimplementing any of them.The deferral note was also wrong about what wraith form does, which is the reason for reading the archive rather than the note. It said wraiths “pass through walls and take damage from doing it”. They do not:
dungeon.c:1202tests!p_ptr->tim.wraith_formbefore charging anything, so the form suppresses the damage a Spectre pays. And it carries far more than passing walls — a hundred points of armour and reflection (xtra1.c:2795), nine tenths off every blow (effects.c:3285), darkness healing instead of hurting and light cutting the form short wherever you are, resisted or not (spells1.c:3632,:3595). A one-line summary of a nine-hundred-line function had lost all of it.What happens when it runs out inside a wall was the thing to settle before building, and the archive is not silent: you are crushed. The guard is
(chp > depth / 10) || !PASS_WALL, which lifts the hit-point floor for exactly that case — the comment beside it says so outright, that everyone else in a wall dies and a pass-wall character does not. That branch existed here and was deliberately left out, on the grounds that an ordinary character in a wall grid was a bug rather than a Spectre. Wraith form makes it a reachable state, so it is now built, and it is the only thing standing between that character and being entombed alive with nothing killing them.Ten tests, each asserting the same thing with the form and without it, and each falsified by removing the behaviour it covers and watching exactly one fail. Two of them were falsified against the plausible wrong build as well: cancelling wraith form only on unresisted light, and writing the nine tenths as
perc_dam_red = 90, which rounds the other way and turns a five-point blow into one point rather than none.One of those tests was rewritten because the falsification caught it. “A wraith pays nothing for the rock” asserted a single turn at depth 10, where the charge is two points — and the wraith’s own nine tenths turns two into nothing nine times in ten by itself, so the test passed against a build with the suppression removed. It now runs two hundred turns.
Four deferred random mutations remain, from six when the count was last stated: the warning, the two hit-point/spell-point exchanges, and the chaos gift, which fires on gaining a level rather than on a turn passing.
The manual catches up with the build — 18 September 2026#
3.122.0 — Five things the manual had wrong, found by checking it against the data rather than against itself.
“Races we decided against” listed eight races that are in the game. Barbarian, Klackon, Nibelung, Imp and the four undead were described as “not rejected, only waiting”, twenty lines below their own full entries and a hundred lines above their racial powers. They arrived in 3.117.0. The count in the same paragraph said nine of Zangband’s twenty-one extra races were built; it is seventeen.
The appendix tables listed twenty of twenty-eight races and ten of fourteen classes. Barbarian, Klackon, Nibelung, Imp, Skeleton, Zombie, Spectre and Ghoul were missing from both race tables, and the Mindcrafter, Chaos-Warrior, Warrior-Mage and High-Mage from both class tables. All twenty-four rows are generated from
p_race.txtandclass.txt, and the rows that were already there reproduce unchanged — which is what makes the five that did not worth reporting.Five classes had their experience penalty listed as “none” when they have one: Mage +30%, Priest +20%, Paladin +35%, Rogue +25%, Ranger +30%.
expfactisrace->r_exp + class->c_exp, so the column is real and was wrong for a third of the table. Inherited from upstream rather than introduced here, and wrong either way.The Golem’s stun immunity was nowhere in the manual. It has
PROT_STUNat birth and it is one of the better things about being one — a stunned character fights and casts worse and can be knocked out. Its entry listed the paralysis immunity beside it and stopped.Not bleeding was stated only in the Skeleton’s entry, which mentions in passing that a Golem and a Spectre are the same and a Zombie becomes so at 12. True, and not where anyone reading about a Golem would find it. Now in all four.
The mode that never said it was on — 15 September 2026#
3.121.0 — A character in nightmare mode says so, as
Nightmarein red on the status line, for as long as it lives. It sits besideCheat, which was added in 3.43.0 for the same reason and after the same kind of confusion.Zangband showed the mode in exactly one place, the character dump (
files.c:2766), and a dump is a file you write after the character is dead. In play the mode is sixteen changes and every one of them looks like the ordinary game going badly — a monster with twice the hit points looks like a monster, a sustain that failed looks like a sustain that was never bought, a wall that looks like floor looks like floor. The option is set once at birth and can never be turned off, so there was no later moment at which the game would have mentioned it.Ours, and a display rather than a behaviour. It changes nothing about how the mode plays, which is why it is not one of stage 2’s additions and needs no ruling as one (DEC-84).
The manual chapter also now says where the option is, because it is not on the race and class menus:
=at any of them, last of the sixteen birth options.
The manual stops saying M11 has not been started — 14 September 2026#
3.120.1 — Three status claims in the documentation had gone stale, and all three would have been wrong on the face of a release tagged today. The feature table said nightmare mode was not started; the release log said the same, and also that the most recent release was 3.49.5 when two later ones are tagged; and the feature summary said nine races when the game has twenty-eight.
No code. Worth its own entry because a release is partly a claim about what is in it, and these were the claims.
The bell, the walls that are not floor, and a recall that lies — 14 September 2026#
3.120.0 — M11 stage 1 is complete. The last three of nightmare mode’s sixteen behaviours are in, and every one of the sixteen is now either built or closed by a decision with its reasoning written down.
A bell tolls four times in the hour before midnight, and on the stroke the Ancient and Foul Curse falls on you wherever you are. Nothing in Zangband’s own documentation mentions the bell; it is the only warning the mode gives about anything, and there is time to drink something or be elsewhere.
Some walls look like open floor. A few on every level and one door in several hundred. You cannot walk through them, you cannot see past them, and nothing about how they are drawn will tell you.
Word of Recall occasionally arrives deeper than you asked — about one in six hundred and sixty-six. It will not take you past the bottom of the dungeon you are recalling into, which is a deliberate departure: this game has thirteen dungeons with their own floors, where Zangband had one. In a shallow dungeon that makes the corruption mild; in a deep one it can end a character who was going shopping.
Stage 2 — the menu of things Zangband’s spoiler describes and Zangband never built — is deferred until after a release, and is a scheduled decision rather than a closed one.
The wrong monsters, on the wrong floor — 14 September 2026#
3.119.2 — Two more of M11’s sixteen, and they are the two that change what you meet rather than what it does to you. The game’s ordinary rule for generating something out of its depth adds a few levels at most; under nightmare it multiplies, uncapped, so the fifth floor can hold something from the eighty-fifth. And the monsters normally pinned to their own depth — the ones the game will never generate early whatever else happens — are no longer pinned. Only quest monsters still are.
These are the two I pulled out on Friday, saying the tests did not discriminate. The tests were fine. What was wrong was the harness measuring them: restoring a source file and rebuilding within the same second does not recompile it, so every “restored” run was still testing the mutated binary. Two working guards were deleted on the strength of a stale build. The harness now asserts the file actually recompiled, and
check-build’s notes say why.
Nightmare mode takes your sustains and your stairs — 14 September 2026#
3.119.1 — Three more of M11’s sixteen. A sustain now fails one time in thirteen, and when a stat drain does land it is permanent twelve times in thirteen rather than wearing off. Nothing in Zangband’s own documentation mentions sustains at all, which makes this probably the cruellest thing in the mode and certainly the least expected: a sustain is a property you buy once and stop thinking about.
And no stairs can be made. Stair creation does nothing — no scumming for a convenient descent, no escape hatch out of a level that has gone wrong.
Two of the sixteen are also now closed rather than pending. Sleeping at the inn is unchanged under nightmare, and a sleep attack cannot give you nightmares because nothing in this game can put you to sleep — monsters paralyse, which is a different thing. The second is deferred rather than refused: if a sleep state ever arrives for another reason, the nightmare can hang off it.
The Intel Mac comes back — 14 September 2026#
3.120.0 — macOS on Intel is supported again, and released. Every release from now on carries four Mac files instead of two:
ZangbandTK-<version>-osx-intel.dmgandZangbandTK-<version>-osx-terminal-intel.tar.gzbeside the Apple Silicon pair, which keep the names they have always had.This reverses half of DEC-22, which ruled Intel Macs out on the grounds that they reach legacy status this month. The other half of that decision stands and is now the whole of it: no universal binaries. A fat image would make every download carry a second copy of the game that the machine downloading it can never run, and a release page can name two files as easily as one. The makefiles take
ARCHS=x86_64and append-intelto the file name by themselves, so the two builds cannot overwrite one another.Both Intel builds — the Cocoa application and the curses one — are built on
macos-15-intel, a real Intel runner, rather than cross-compiled on Apple Silicon. That matters for the checks rather than for the compiler: clang cross-compiles either way without complaint, but the signature check, theotoolcheck for a Homebrew library that would not exist on a player’s machine, and the smoke test that starts the game and walks a new character into the dungeon all have to run somewhere the executable can actually start. On the terminal build the Homebrew check had to learn a second path — Homebrew lives in/usr/localon Intel and/opt/homebrewon Apple Silicon.One difference that is not cosmetic: the Intel terminal build targets macOS 10.9 rather than 11.0. The 11.0 floor was never a choice, only the first release that ran on arm64 at all, so it does not apply to a processor that has been running macOS since 2006.
Download now opens its macOS section by asking which Mac you have, and both READMEs inside the downloads name the two files so that somebody who took the wrong one can see it.
Nightmare mode arrives, and the monsters are worse — 12 September 2026#
3.119.0 — M11 has started. Nightmare mode is a birth option now, off by default and irreversible once chosen, and the first six of its behaviours are in: monsters have twice the hit points, five more points of speed, arrive awake, arrive with twice the energy, give no free first move however dangerous they are, and notice you twice as fast while asleep.
Zangband’s own text for the option is “this isn’t even remotely fair!” and its documentation says the mode is not meant to be winnable. The manual has a chapter saying what it changes, what it does not change yet, and what it will never change — because the mode’s spoiler describes about three times as much as Zangband ever built, and a player comparing the two deserves to know which is which.
Two of those never-changes are decisions rather than gaps. Your score is unaffected: Zangband gives nightmare characters a bonus on a score multiplier this game has never had for any option, so adding one would be inventing a system rather than porting a number. And a Golem keeps its immunity to stunning: Zangband takes it away, but not for nightmare alone.
A Draconian breathes cold as well, and for the right damage — 11 September 2026#
3.118.3 — A Draconian’s breath was fire only and did three-quarters of the damage the original gives. It is now twice your character level, and fire two times in three with cold the third, which is where every Draconian breath starts in Zangband. At level 30 that is 60 rather than 45.
Still not built, and the manual says so: in Zangband the breath sometimes becomes something belonging to your class instead — shards for a Warrior, disenchantment for a Mage, hellfire for a Priest — on a chance equal to your character level in a hundred. A racial power here cannot ask which class is holding it.
The bloodless stop bleeding — 11 September 2026#
3.118.2 — A Golem, Skeleton and Spectre do not bleed, and a Zombie stops at level 12. Nothing that cuts anyone else leaves a wound on them, which is what Zangband does and what the manual says being made of bone or stone ought to mean.
This wanted a protection the game did not have — it had protection from fear, blindness, confusion and stunning, and none from bleeding. And it turned up something older. Of the four places the game can cut you, only a monster’s claws consulted that kind of protection at all; spell exertion, shard breath and ice storms went straight round it. So a character in stone form has been bleeding from shards this whole time too. All four now check.
A Mindcrafter grows into four things it never had — 11 September 2026#
3.118.1 — Four classes gain intrinsics on a level threshold in Zangband and gained nothing here. A Mindcrafter now resists fear at 10, sustains wisdom at 20, resists confusion at 30 and reads minds at 40. A Chaos-Warrior resists chaos at 30 and fear at 40. A Paladin resists fear at 40.
And a Monk speeds up. One point for every ten levels from level 10, and free action from 25 — both of them only while it is carrying little enough armour to fight bare-handed, which is the one place in the game where an intrinsic depends on what you are wearing. Put plate on a Monk and it loses the speed along with the blows.
The mechanism is the one the races already had, extended rather than copied, with a named condition on the band. The manual has a table of what each class grows into, because nothing in the game tells you it is coming.
A Dwarf can find the stairs again — 11 September 2026#
3.118.0 — The six races Angband and Zangband share had lost their racial powers. A Dwarf, Hobbit, Gnome, Half-Orc, Half-Troll and Kobold each have one in Zangband, and the import took the shared races’ stats and skills and stopped there. A Dwarf arrived as Angband’s Dwarf: the same race minus the one button it can press.
A dwarf now examines its surroundings from level 5 — traps, doors and stairs. A half-orc plays tough from level 3, which is the earliest racial power in the game. A gnome blinks, a half-troll works itself into a frenzy, a kobold throws a poison dart, and a hobbit cooks a ration of food onto the floor at its feet.
The manual’s racial-power table had also gone stale twice over: it listed nine powers where the game has twenty-three, having never been updated when the eight deferred races landed. It is now complete and ordered by level, and each of the six races has a line in its own entry.
A Sprite and a Half-Titan stop being the same size — 11 September 2026#
3.117.5 — Nine imported races shared one placeholder height, weight and age, so a Sprite, a Golem and a Half-Titan were all born seventy inches and a hundred and fifty pounds. They now have Zangband’s own figures: a Sprite is thirty inches and seventy pounds, a Half-Titan is eight foot nine and a quarter of a ton, and a Golem is one year old plus however many the dice give.
Only flavour — the character sheet and nothing else reads these — but the manual has always said height and weight are “randomly determined according to the race of the character”, and for nine races that was not true.
The Sword of Chaos stops being easy to clean up — 11 September 2026#
3.117.4 — The object-side BAL-08 sweep, fourth stage: curse severity. Zangband has three tiers of curse — one a common scroll removes, one that needs the greater scroll, and one nothing removes at all. Every imported cursed item arrived at the first tier.
The Sword of Chaos is permanently cursed in the original; here a Scroll of Remove Curse dealt with it. Stormbringer, Twilight, the Rod of Thanos and the Amulet of Destruction all need the greater scroll there and did not here.
The three tiers now line up with this game’s own scrolls: an ordinary curse is within reach of Remove Curse, a heavy one needs *Remove Curse*, and a permanent one cannot be removed. Heavier curses also make an item worth less to a shopkeeper, in both games, so the prices move with them.
A range that was one out, in six places — 11 September 2026#
3.117.3 — The object-side BAL-08 sweep, third stage. Zangband’s
rand_range(a, b)includes both ends. Half the places that expression was converted by hand wrote it as a range of the right width shifted up by one, and one value short.A Scroll of Logrus did 151 to 301 damage to its reader where the original does 150 to 300; a Scroll of Fire did 101 to 201 where it does 100 to 200. Seven artifacts and one ego could not recharge in the shortest time the original allows. The other three conversions were right, and nothing distinguished them.
No player could ever have noticed any of this, which is exactly why it survived: a number wrong by one has no symptom. It is written out once now, and the converter refuses to run if a hand-written effect does not say what its own source says.
Two rods stop recharging instantly — 11 September 2026#
3.117.2 — The object-side BAL-08 sweep, second stage: object kinds. Zangband keeps a rod’s recharge time in the same field it keeps a ring’s bonus in, and the converter only ever read it as the second. Both imported rods therefore arrived with no recharge at all — and a rod with no recharge is a wand with infinite charges. The Rod of Havoc, found around depth 95, throws a 150-point elemental ball; it could be zapped every single turn, for ever.
Havoc now takes 250 turns to recharge and Pesticide 3, which are Zangband’s own numbers. They cross unchanged: eleven of the twelve rods that exist in both games already carry the identical figure.
And a trap worth knowing about if you ever hand-edit the data. An object’s
values:line is read by a different parser from itscombat:andtime:lines, and the two disagree about a minus sign. Incombat:,-d50means “1 to 50 off”. Invalues:, the same text means a positive bonus of up to +5 — quietly, with no error. Trying it on the Amulet of Destruction produced an amulet of destruction that improved all five of your statistics. Its penalty stays flat, and a test now fails if anyone writes the tempting version.
Two egos get the half of themselves that was missing — 11 September 2026#
3.117.1 — The object-side BAL-08 sweep, first stage: ego items. Three faults, all of the same kind — the item parses, generates, and reads like an ego while a property it is meant to have is quietly absent.
A Sword of Sharpness now digs. Zangband gives it VORPAL and TUNNEL together, and the size of the digging bonus is not on the line the converter read: it is computed at creation from the depth, in a hook. With no size, the flag was filed under “modifier with no value” and dropped. It is +1 near the surface and up to +6 deep, which is what the original gives.
A Trump Weapon can be activated again. Its Lua teleports the wielder on a 50–100 turn timer, and its separate random-teleportation flag is the drawback that power pays for. Only the drawback had been imported. The flag table said activations were “handled separately”; nothing handled them.
Four egos rolled for their bonus and were shipping the best case. A Lantern of Vision gave +6 searching where Zangband averages +3.5. The Pattern Weapon’s +STR and +CON no longer move together, which Zangband’s single value made them do — 4.2 has no way to say it, and holding the average matters more.
Also: the eleven artifacts that had lost their activations, and the Amulet of Hunger that had invented one, from earlier the same day. A
--checkpass now compares the committed ego and artifact data against what the converter produces, so the next one of these fails a build instead of shipping.
The nightly says why a run stopped — 10 September 2026#
3.117.0 — The borg has always known why it will not go deeper, and nothing wrote it down.
borg_prepared()returns a reason in one word —5 Food,30 hp,2 cure— and the nightly recorded only the depth reached, so a shallow fleet looked like bad play.It is not bad play. Asking the borg directly, ten of the twelve runs are forbidden from leaving the first floor rather than choosing to stay: the reasons split roughly half supplies (food, cure potions, a recall scroll) and half level or hit points. Establishing that took an afternoon of tracing depth over time; it now takes a glance.
Each run’s line is followed by
allowed to depth N, held back by: ..., and the summary tallies the causes across the fleet, which is the part that separates one shared cause from twelve unrelated ones.ready=in the status line is gone. It reportedcharacter_dungeon— “a level exists” — so it was true of every run that had started and told nobody anything.allowed=andblocked=replace it and say something.
The nightly stops gating on its luckiest run — 9 September 2026#
3.116.0 — Totals replace maxima, and the per-run rows became the message. The nightly went red for best depth falling 8 to 3, on a night when the fleet’s character levels went 41 to 39 and its casting 15 to 14. Nothing was broken: twelve runs, twelve deaths, none capped.
Best depth is a maximum over twelve runs, so it reports whichever run got luckiest. Pulling the per-run tables out of four nightlies made that concrete: eight of the twelve never leave depth 1 on any night, so the whole depth signal lives in four runs and best depth lived in one. When the pass-wall commit shifted that run’s random stream, the headline moved wholesale.
Also visible in that history, and worth recording: the twelve races added on 7 September left all twelve rows byte-identical, and 7 to 8 September moved no rows at all — the runs are deterministic on CI. The runs set no race, so every character is a Human, which is why the new races could not have touched them.
Both directions were evaluated against that real history plus deliberate regressions — casting collapsing the way
BORG_SPELL_UNKNOWNmade it (4 cast against 15), nothing learned, the fleet ceasing to level, the fleet ceasing to descend, every character level halved. The new rules raise no false alarm and miss none. The old rules missed none either; they cried wolf twice, which is its own failure, because a job that goes red for nothing gets ignored.So the gate is now four totals — cast and learned halving, total character levels down more than a third, total depth down more than two fifths — and the twelve
run:rows in the baseline decide nothing. They exist so a failure says “Warrior seed 1: depth 8 to 3” instead of leaving a reader to diff two files, which matters because a scheduled task reads this output cold. The report states in as many words that rows moving is not a regression and that the totals decide.Depth is the loosest threshold on purpose: with two thirds of the runs stuck on the first floor it is close to a dead metric. It wants tightening when the borg reliably gets below depth 1, and until then the character levels carry the weight — which is also the honest reason the old depth rule was fragile rather than merely unlucky.
scripts/check-build’s header gained two notes while this was verified, both from mistakes made verifying it: the gate takes some twenty-five minutes and has to be waited on rather than backgrounded and checked for a phrase that may never appear; and nothing else may run against the build while it does, because a hand-run suite collides where the harness does not. It also records the staged-data trap — the copy into each build tree compares timestamps, so a data file edited, built, and reverted within the same second keeps the edit in every staged tree, which presents as unrelated suites segfaulting.The baseline is reshaped in the same commit and retaken from the 9 September nightly, which measured this exact commit. Its totals are that night’s: depth 18, levels 39, learned 20, cast 14. Updating the rows is now free, since they gate nothing; changing a total still wants a reason. The cross-platform guard is untouched — it caught the first false alarm this file ever produced.
Mountains charge like the walls they are — 9 September 2026#
3.115.3 — The surface exemption is reversed. 3.115.0 shipped the in-rock damage underground only, on the archive-faithful ground that Zangband’s mountains were not
FF_BLOCKgrids, so a Spectre there was never inside a wall and paid nothing for a crossing. The project owner overruled it on the same consistency ground as the pass-wall ruling itself: “Our mountains are walls so the Spector takes damage.” A mountain here is a wall, so being inside one costs what being inside a wall costs.DEC-74 records the reversal above the reading it overruled rather than in place of it, because the next person to compare us against Zangband will find the difference and should be able to see it was considered and taken deliberately.
Measured before it was asserted, since the formula at depth 0 is not obviously survivable: the depth term is zero, so it is one point a turn, and the guard
chp > depth / 10becomeschp > 0— its tightest possible form. It holds, because death ischp < 0. Measured, a Spectre in a mountain settles hovering between nothing and one hit point, taking a point back off whenever it regenerates one, and survives a thousand turns of it. Underground at fifty it takes six a turn and floors at one.So a crossing is not lethal in itself and it is not free either. A range is many blocks wide and a first-level Spectre has sixteen hit points, so it arrives at the far side empty — and what kills it will be the first thing it meets, not the rock. That is a real price for a race that pays 180 per cent, and it is what the ruling intends.
The falsification that used to prove “damage on the surface fails” now proves the opposite: reinstating the old exemption fails the suite, and removing the cannot-kill guard fails it twice over, since it now breaks the surface and the dungeon both. The manual says the cost applies to mountains, and roughly what it comes to.
The Spectre walks through walls, and the view was never the problem — 9 September 2026#
3.115.0 — Pass-wall, the last outstanding piece of any race. The project owner ruled on it after going and checking: “I tested the Spectre in the old zangband and he can walk through walls. If he can do that he can walk through anything.” Mountains included. Recorded as DEC-74, with the note that this is a consistency argument rather than an archive one — and that it lands where the archive was anyway, because Zangband’s mountains were passable to everyone in the first place, so the original never had to decide it.
The view, which was the whole reason this waited, needed no work at all.
los()never inspects its own origin: every scanner in every branch starts one step away fromgrid1, so line of sight from inside a wall is already well defined. A character in rock sees its eight neighbours by the adjacency short-circuit and nothing beyond, because rock is not projectable.add_light()goes through the same function, and no assertion anywhere requires the player’s grid to be passable. One afternoon of reading retired the fear rather than confirming it.The rest is what was sized: a branch in
do_cmd_walk_test(), a condition inmove_player(), and per-turn damage of1 + depth / 10inprocess_world()that cannot kill — the archive applies it only whilechp > depth / 10, so a Spectre can always walk back out of a mountain. Attacking monsters inside walls needed nothing:do_cmd_walk_test()already handles the monster case before it tests passability. No energy cost, as in the archive.A test caught a real hazard. The rule is permanence, and the obvious helper for that —
square_isperm()— tests for permanent rock, requiringTF_ROCKas well. The edge of the world is permanent open sea and carries no rock flag, so the first version let a Spectre walk off the edge of the world. It now testsTF_PERMANENTdirectly, and the reason is written where someone would otherwise simplify it back.Two consequences checked rather than assumed. Nothing depended on mountains being solid: danger comes from a block’s
lawand not its terrain, mountains exclude town placement and cost 25 in the road pathfinder — roads route around them rather than being stopped — and a mountain block is 65 per cent rock at grid level with the rest dirt and grass, so it was always porous. A Spectre crossing a range saves time, not access. And the borg, which knows nothing about any of this: five seeds at 60,000 turns each, no crash, no wedge, no oops. It simply never uses the ability.no_regenreplacesburninginprocess_world(), since two different things now stop a character healing — sunlight on a Vampire, rock around a Spectre. Zangband called itcave_no_regenand used it for exactly those two.The damage is underground only, and that is a judgement. Zangband keys it on
FF_BLOCKand its mountain does not carry that flag — “rock face” isHALF_LOS | USE_TRANS | ICKY | OBJECT— so a Spectre there crossed a range without ever being inside a wall. Ours are walls, and a range is many blocks wide, so the archive’s own test would charge for a crossing the original never charged for. The surface is free; the dungeon is not. That reproduces Zangband’s outcome rather than its code, and the manual entry now states the cost as plainly as the ability — a hit point a turn, six at the bottom, no healing while you are in there, and it stops rather than killing you.3.115.1 — An audit of all twenty-eight race entries, which found five things wrong. Written up piecemeal across a dozen commits, and the eleven inherited from 4.2 had never been reviewed at all.
Human: “go up levels faster than any other race” — the Yeek matches them at 100 per cent. A tie described as a lead, which is the third time today that exact shape has turned up.
Skeleton, Zombie, Spectre, Ghoul: five thresholds written “above level N” where the data says
gain-at:N, so each was off by one. My own text, hours old. The Yeek’s “above level 19” was correct againstgain-at:20but used the archive’s idiom rather than ours; everything now reads “from level N”, matching the directive.Spectre: “the best saving throw, magic device skill and stealth of any race” — it ties with a High-Elf on saves and a Mindflayer on devices. Only the stealth is a lead.
Imp: “feeble in every way except constitution” — its dexterity is positive too.
And the coverage question: all twenty-eight have entries in Creating a Character. The eleven shared races have no row in Races and Classes, which is correct rather than missing — that page is explicitly what differs from Angband.
Two inherited claims were checked and are true, which is worth recording so nobody re-checks them: the Half-Troll’s regeneration really does cost it extra food (
OF_REGENdoubles consumption ingame-world.c), and the Vampire’s “as far as anything sees” is tie-safe wording for a genuine five-way tie on infravision.3.115.2 — A test wrote to a fixed temp filename.
game/savesused"saves-roundtrip.tmp"for every run, so any two concurrent runs of that suite raced on one file — reliably, eight times out of eight in parallel. Nothing in the harness collides, becausecheck-flakesand the CMake runner both go suite by suite; a person running a suite by hand while a verification job is going does, and it presents as an intermittent savefile failure, which is about the most alarming thing it could look like.test_savefile_name()already existed for this and three other suites were already using it.
Every race Zangband had has now been considered — 7 September 2026#
3.114.0 — The four undead: Skeleton, Zombie, Spectre and Ghoul. With these, all twenty-one races Zangband has and Angband does not have been decided: seventeen are in, four were turned down under DEC-35 as near-duplicates. Twenty-eight races in the game.
They share one shape, which is why they arrived together: none can eat, all wake just after midnight, and each takes scrolls of Remove Hunger in place of rations. The kit substitution and the
UNDEADflag both already existed — built for the Vampire in 3.111.0 — so all four needed was oneequip-insteadline each. That is what building the Vampire’s mechanism generally rather than specially bought.The Skeleton has the hardest bargain of the four: it cannot eat and digests at the ordinary rate, where the other three digest slowly. The Spectre pays 180 a level for the best saving throw, device skill and stealth in the game on seven hit dice. The Zombie is the second-strongest thing you can be and very nearly mindless.
The Ghoul’s paralysing touch was already built. The audit called it “Zangband-specific, needs designing”; in fact
OF_GHOUL_TOUCHand its hook inplayer-attack.chave been there since CNT-09, written for items that carry the flag — bare-handed only,MON_TMD_SLEEP, using 4.2’s own save so it honoursNO_SLEEP. Declaring the flag on the race was the whole of the work. Its other power, eating corpses, has nothing to work on: 4.2 has no corpse objects.The Spectre ships without pass-wall, deliberately, and its manual entry says so in the same terms as the Draconian’s breath and the Sprite’s old missing speed. Three finished races should not wait on one unknown. What was established while sizing it is recorded as a follow-up rather than left in a transcript: four small parts (a movement branch, per-turn damage shaped like the Vampire’s sunlight, no energy cost, attacking monsters in walls), with the genuine unknown being what a character standing inside solid rock can see, since 4.2’s view algorithm assumes the player’s grid is passable.
That follow-up also carries a question that is not about races at all. Zangband’s mountains were never walls —
FEAT_MOUNTAINis passable to everyone at an energy penalty (cmd1.c:2382), slow terrain rather than an obstacle. Ours areROCK | WALLand notPERMANENT. So a faithful pass-wall would let a Spectre cross our mountain ranges, which the original never had to consider because it had no such walls to cross. Anything that reads “wall” reads our mountains as one, so it may bear on more than one race.Also recorded: all nine races of the first wave share a placeholder age, height and weight, so a Half-Titan is currently the same size as a Yeek. It is cosmetic — those fields reach the character sheet and nothing else — and correcting it changes existing characters, which is a different kind of change from filling a gap. The eight of the second wave carry the archive’s own figures.
Four more races, and the mapping proved rather than trusted — 7 September 2026#
3.113.0 — Barbarian, Klackon, Nibelung and Imp. The four of the eight deferred races that are essentially data, taken first so the two that are not do not hold them up.
Before writing any of it, the field mapping DEC-35 recorded — stats and hit dice copied, disarm/device/save/stealth copied,
searchfrom Zangband’sr_sns, melee and shoot doubled, Zangband’s experience kept, charisma dropped — was run backwards against all nine existing imports and reproduces every one exactly. Only then were the new four generated from the archive’s own table, and they round-trip the same way. Eight races’ worth of numbers is too many to transcribe by hand and hope.What they bring: the Barbarian resists fear from birth, which no other race does. The Klackon is the second user of the
speed:field built an hour earlier —lev / 10, exactly as the Sprite — which is what building it as a general mechanism was for. The Nibelung is the only race in the game that resists disenchantment from birth. The Imp sees the invisible from level 10. All four have a racial power; the bolt-becomes-a-ball ones usepower-whenbands, which the Mindcrafter already established.The Golem’s stun immunity is folded in — one of the three grants found yesterday outside
player_flags(), and a one-line fix sinceOF_PROT_STUNexists.Two things checked because of yesterday’s lesson that
player_flags()is only one function. Every one of the four was swept for non-flag grants: Klackon and Nibelung have none; the Barbarian’s starting ring of resist fear is swapped for sustain strength (birth.c:672), which is moot because no class in this game starts with a ring; and the Imp’s exemptions fromsanity_blast()andhave_nightmare_aux()are moot because DEC-32 dropped that whole path. Nothing outstanding.Also found while sweeping: all nine existing imports share a placeholder age, height and weight — 20:20, 70:6, 150:20 — where the archive gives each its own. A Half-Titan should not be the same size as a Yeek. The four new races take the archive’s figures (age verbatim, height and weight the midpoint of its separate male and female rows, which is how 4.2 derived its own). Correcting the nine is cosmetic and left out of this commit, because it changes existing character sheets.
The Sprite flies faster and the Yeek stops corroding — 7 September 2026#
3.112.0 — Two gaps the audit found, and a sweep to see whether there were more.
A Sprite had no speed at all. Zangband sets
TR_SPEEDon it above level 9 and cashes the flag aspspeed += lev / 10(xtra1.c:2550) — a point at 10 rising to five at 50, which is worth more than most races’ entire list of resistances and is most of what 175 experience a level buys. It is now aspeed:base:scalefield, sibling to thelight:andarmour:fields the Vampire and Golem needed, because the Klackon is blocked behind exactly the same thing and this makes it a data line when it arrives.Worth recording about the arithmetic: the archive’s level-10 threshold is redundant with its own formula, since
9 / 10is 0. Sospeed:0:10reproduces every level exactly with nogain-atband, and the test checks the boundary rather than the endpoint — a flat bonus, or a scale of five, both fail it.A Yeek was resistant to acid but never immune, which the archive grants above 19. One data line now the gates exist. That exposed a real flaw in the generic gate test: it required a gated resist to be absent below its band, which is true of the Draconian and false of the Yeek, who upgrades one it already has. It now requires the band to raise the resist.
And the set is closed, which was worth checking rather than assuming. Sweeping every race-conditional grant in the archive’s
player_flags()against our data: 10 level gates, all now present, and no birth flag missing from any of the nine imported races.That sweep covers one function, though, and the Golem’s armour was never in it. Sweeping outside it turned up three more, all on the Golem: it cannot be stunned (
effects.c:1875), cannot bleed (effects.c:2064), and a monster draining life from it — or from a Vampire — does not heal itself (melee1.c:1327).OF_PROT_STUNexists, so the first is one data line; the others need thought. Not built here; recorded so they are not found again by accident.Three documentation corrections while in the area. The Draconian’s breath was described as changing “every five levels” in Zangband — it changes by class, with a chance equal to your level in a hundred (
racial.c:380), and our damage is one and a half times level against the archive’s twice. The Sprite’s entry said Zangband quickened it at level nine, where the archive is above nine, meaning ten. Andp_race.txt’s own header saidmutation-ratewas a one-in-N chance; it is a percentage, so a Beastman’s1:20is one chance in five per level, which is what the manual has always said and what the header got wrong.
The release stopped one step short of the zip — 7 September 2026#
3.111.3 — The Windows release built the game and then could not package it. All 223 objects compiled and
angband.exelinked;scripts/pkg_winthen stopped onlib/tiles/neon/graf-neo.prf: No such file or directory, which names a file that exists. Line 28 is theawkinsidecp_unix2dosand the name in the message is its output: there was nolib/tiles/neonin the staging tree to write into.The script built its tileset directories from a list written out by hand, and the neon set had been added to
list.txtand not to that list. Thefindthat gathers prf files goes by pattern, so it picked the new set up happily, had nowhere to put it, anderrexitstopped the release on the spot.pkg_winnow takes the tilesets from whatlib/tilesactually holds, the same way it already mirrors the built manual a few lines above.Makefile.stdandMakefile.wasmeach kept their own copy of the same stale list, so the source tarball and the web build would have shipped without the set and said nothing about it; both are explicit lists on purpose, so both get the name added rather than a glob.Makefile.osxalready globbed the subdirectories, which is why the Mac app has had the set all along.And the version had come adrift, in the same way the headings had at 3.108.2:
buildid.hstill read 3.108.2 while this log had reached 3.111.2. Six entries — 3.109.0 through 3.111.2 — were written without the bump the top of this page asks for, so every build sincefee2b4cbehas told the player it was 3.108.2. The header is back in step with the log.
A use-after-free at birth, and the gate that was not watching — 7 September 2026#
3.111.2 — msys2 went red on the new race suite, and it was a real defect. The suite died rather than failing, and the sanitizer report never reached the log — the test runner captured the child’s output and printed it on every path except the one where it is the only evidence there is. Fixed first, so the next one is diagnosable.
The fault: the level and the player’s memory of it are a pair, and a level object’s
knowncounterpart lives inp->cave.player_init()releasedp->caveand left the globalcaveholding objects whoseknownpointers now dangled — and the next character’s birth walked exactly those, throughplayer_learn_innate()intoupdate_player_object_knowledge(), which dereferencesobj->knownon every level object. A real game never notices, because a birth reset happens before any level exists. The chunks are now released together, the global one first so thatcave_free()can still unhook the known copies while its pair is alive.Fixing that exposed a second, older one underneath:
update_stuff()guards its map work withcharacter_generated, which says a character has existed rather than that one is standing on a level, soupdate_monsters()could walk a nullcaveduring birth. It needsPU_MONSTERSraised to get there, and what raises it at birth is a character whose light differs from the previous one’s — which before the Vampire no race could do.It was never a Windows quirk. The suites seed from the clock and the pid, and the fault needs a level that happens to have a visible object on the floor: 3 seeds in 30. msys2 drew a bad one. Both fixes then held over 80 runs across 40 seeds.
And the local gate had stopped reproducing the job it claims to: it created
build-asanonly when the directory was missing, so a directory made before-fno-sanitize-recover=undefinedwas added kept the old flags for ever — reporting undefined behaviour and carrying on where msys2 aborts. It now configures every time.
The nightly was measuring the wrong machine — 7 September 2026#
3.111.1 — The first red night was the baseline, not the game. The nightly reported best character level down from 13 to 9, spells learned 30 to 22, cast 24 to 15, and called it a regression.
Two things in the shape of it said otherwise. Every number had moved, where a real regression usually moves one — and best depth had gone up, 7 to 8. And the Warrior rows were byte-identical to the previous night, to the digit, across all the race and class data that landed in between.
Both turned out to be true and innocent. The runs play a Human, and everything in 3.110.0 was scoped to four other races, so the Warrior’s random stream genuinely never moved. And the numbers all differed because the baseline was taken on a different computer from the one measuring it.
At the same commit, same seed, same flags, Warrior seed 13 played 280,358 turns to character level 13 on Darwin arm64 and 5,252 turns to level 1 on Linux x86_64. Both machines are internally exact and reproduce their own figures every time; neither is wrong. Build type was ruled out — the local figure reproduces identically at
-O0and underRelWithDebInfo, which is what CI builds — and the RNG is fixed-width and deterministically seeded, so the divergence is somewhere in the sequence of calls rather than the dice. What exactly has not been chased down, because it does not change the remedy.The baseline is re-taken from the nightly’s own log and now records
platform:.borg-progressprints the platform it ran on and skips the comparison when the two differ, rather than comparing figures that cannot be compared, so this fails loudly instead of looking like a bug in the game. Running a sweep locally is still useful for the table; it just no longer judges your laptop against CI.The lesson is one line, and it is now the first line of the update instructions: a baseline has to be generated the same way it will be measured.
The Vampire becomes playable — 7 September 2026#
3.111.0 — The Vampire can shelter from the sun, and packs for it. From a playability question: does a Vampire start taking damage the moment it goes outdoors?
Everything about the damage turned out to match the archive exactly — one point per world tick, both games ticking every ten game turns, both with a ten-thousand-turn day, both lighting every wilderness grid unconditionally at dawn, and resist-light gear cancelling it in both. The midnight start works and buys about 2,499 moves. What was missing was the shelter.
Zangband gives a Vampire two to five scrolls of Darkness at birth, and its darkening has no guard at all (
spells1.c:552) — so a Vampire caught out at dawn puts out the ground it is standing on and waits. 4.2 refuses to darken the surface in daylight, which is right for a game where nobody minds the sun and left ours with no daytime shelter but a staircase. We had inherited a rule written for a game with no race that burns.Both halves are now in. The guard passes for a caster the light hurts — the caster, not the player, so a monster casting darkness near a Vampire does not put the daylight out on its behalf. Zangband removes the guard for everyone; that was rejected as more than is needed, and every other character’s play is unchanged. Recorded as DEC-73, because it looks exactly like a bug to anyone who finds it later.
Race starting-kit substitution is new (
equip-insteadinp_race.txt). Zangband decides food and light by race before it looks at the class: the undead-or-built races take Satisfy Hunger scrolls instead of rations, and the Vampire takes Darkness instead of torches. The grant is unconditional rather than an edit of the class kit, because a Vampire Necromancer — the one class that carries no light — must still get its scrolls. Of twenty races only the Vampire and Golem qualify today; four of Zangband’s six are among the deferred races and will need one line each.Two things worth recording from the verification. A first falsification run reported the darkening tests passing with the change reverted; they were reading stale binaries, because reverting left the helper unused and the
-Werrorbuild failed silently behind a redirect. Falsifying through the predicate body instead gave real answers, and turned up a genuine gap — theSRC_PLAYERhalf was untested, and is now. And the kit suite needed a cave for the first time:calc_light()raisesPU_MONSTERSwhen a character’s light differs from the last one’s, and the Vampire is the first race to carry a light of its own, so the birth after a Vampire’s was the first that ever needed one.
The four half-imported races are finished — 6 September 2026#
3.110.1 — The manual now describes the four races that were just finished. Worth saying what this was not: the pattern elsewhere has been documentation describing intended behaviour rather than built behaviour — the Monk’s page promised armour class the class did not grant. That is not what happened here. The manual was accurate to what was built every time. A Draconian really did resist fire, a Golem really did hold its life, a Vampire really did digest slowly. The build was wrong and the manual faithfully described the wrong build, which is the harder failure to notice, because nothing disagrees with anything.
The four descriptions in Creating a Character are rewritten around what the races now do, and the level thresholds are stated in the text — a short note before the list explains the convention, since half of Zangband’s races are shaped by it and Angband has almost nothing of the kind.
The Vampire’s entry is the one that mattered most: it is a race a player picks for the flavour, and the flavour is now mechanical. Sunlight, the tenth of a meal, and the midnight start are all things you would otherwise discover by dying at dawn. Writing it up turned up a claim worth checking rather than reasoning about — dawn calls
cave_illuminate()over the whole surface, wilderness included, so a Vampire burns anywhere outdoors and not only on a town street. There is now a test for that, run with and without the grid lit by hand.Three overstatements caught in review before they landed: a Golem can eat, it just gets a twentieth of the good out of it (
cmd6.c:135); the Vampire’s infravision ties with the Dwarf and the Kobold rather than leading; and the Amberite was described as the second dearest race when at 225 per cent it is the third, behind the Half-Titan and the Draconian. That last one predates this work.3.110.0 — A race can now grow into what it is. Zangband grants most racial properties on a level threshold rather than at birth, and 4.2 has nowhere to put one:
struct player_racecarries a single set of flags, applied at birth. The import flattened every threshold to never, which is why a Draconian had wings and no scales.gain-at/gain-obj-flags/gain-valuesdeclare the bands inp_race.txt, and they are applied in the two places 4.2 already assembles a character —player_flags()beside 4.2’s own single hard-coded threshold, andcalc_bonuses()after the innate resists. A gained property is then indistinguishable from an innate one. Eight more races are queued behind this and are now a data change (DEC-72).What the four were missing:
Draconian — all five resistances, at 5, 10, 15, 20 and 35.
Mindflayer — sustained intelligence and wisdom at birth, see invisible at 15, telepathy at 30. Its racial blast fired a bolt of
MANA; the archive firesGF_PSI(racial.c:492). A conversion error, not a recorded substitution —MON_PSIhas existed since DEC-37 added it for exactly this, and the Mindcrafter uses it.Golem — twenty points of armour and a point every fifth level (
xtra1.c:2670), which is thirty at level 50 and the largest single omission of the nine imported races. It was missed because it is granted incalc_bonuses(), not in the flag switch the import read.Vampire — its own dim light (
TR_LITE), on the one race that light hurts. It also hadSLOW_DIGEST, which is not in the archive and was the wrong sign twice over: a benefit, on the race whose defining problem is that food does not work. Zangband gives it a tenth of a meal instead, burns it for a point a turn on a lit town grid by day, and starts it just after midnight so the sun is not already up.
src/tests/player/race.cis new, and is where race work belongs from now on — the race assertions had accumulated ingame/wild.c, a 6,295-line wilderness suite, because that is where somebody happened to be working. Eight tests, falsified ten ways. One of them walks everygain-atband in the data rather than a list written by hand, so a race added later cannot declare a threshold that never opens.Finding this also turned up real test pollution:
player/mutation.cset the race to Vampire in one test and never put it back, so the beak test — which measures what a beak costs you at mealtimes — had been measuring it on the one race whose own nutrition rule takes precedence.
The README stops describing an older game — 6 September 2026#
3.109.0 — The nightly says what the borg managed, and goes red when it drops. Pass and fail cannot carry this: every run dies, which is what unaided play looks like at these levels, and a run that hits the clock has not failed either — so the job went red only for a crash. Best depth could have halved overnight and nobody would have known.
tests/borg/BASELINErecords what the borg currently manages, on the same pattern astests/saves/EXPECTED-FAILURES: a statement of what is true now that somebody has to change on purpose. Seeded from a sweep atfee2b4cbe— best depth 7, best level 13, 30 spells learned, 24 cast, over twelve unaided runs.The thresholds are deliberate: depth down 3, level down 4, or either spell count halved. Not tighter, because although the runs are exact — the same seed gives the same turn count every time, checked three times over — any change to game data reshuffles the random stream and moves which seed happens to do well. The same seed that died at level 1 one night reached 13 the next, on a change to starting equipment.
A run that hit the clock is excluded from the comparison and the job says so, since a capped run stopped where the clock landed rather than where the borg did.
The results now render on the run page as a job summary rather than sixty lines of log to expand, and the failure names the metric, both values, which run reached each, and the command to reproduce it — enough for someone arriving cold without reading the log.
And the artifact step is fixed rather than left pretending: it globbed
build/game/*.txt, which matched nothing, andif-no-files-found: ignoremade it pass silently — the first red night uploaded zero bytes. It now keeps the sweep’s own output and the borg’s death log, for ninety days.3.108.2 — The front page was out of date and the release log had drifted. Both found by taking stock rather than by anything failing.
The README still said “Not yet: magic realms, pets, and nightmare mode — and the two classes defined by which realm they choose”. All of that has landed except nightmare mode: the seven realms, the pets, the Warrior-Mage and the High-Mage. It also said 17 ego types where the data and three pages of the manual say 16, and “nine races and four classes” while listing three. The status line called the project early with ten of eleven milestones done.
Every number on that page is now checked against the data files rather than remembered: 16 ego types, 387 imported monsters, 51 artifacts, ninety-six mutations, thirteen dungeons, nine Lords of the Courts.
And the release log’s headings had come adrift. Entries are prepended, and seven versions’ worth had been filed under whichever heading happened to be at the top: everything from 3.102.2 to 3.108.1 sat under “Gervais’ unexplored squares go black too”, and five versions of borg work sat under “An unexplored square is black everywhere”. Each group has a heading that describes it and carries the date its work was actually done.
Every caster has a book, and M8 closes — 5 September 2026#
3.106.0 — Every caster begins with a spellbook again. Six classes started the game unable to cast: the Mage and the Priest lost theirs to the realm renames —
d1eadd951and0847955a7each fixed one class’s book and deleted another’s in the same edit — and the Monk, Chaos-Warrior, Warrior-Mage and High-Mage never had one. A level 1 Mage readStudy (1)on the status bar with nothing to study from.The book is resolved from the realm chosen at birth, which is the only way it can be right. Zangband does the same:
player_init[]holds a marker andplayer_outfit()turns it into the book of the realm picked moments earlier. A fixedequip:line names one object kind and the realm is not known until the player picks it, so the Rogue’s and the Paladin’s fixed books were wrong in general too — a Rogue who took Death began with an arcane primer it could not open. Those two are fixed by the same change.No new data directive was needed: the class already declares which realm each of its books belongs to, so the book to give is the first one of a realm this character studies.
And a realm already taken is no longer offered twice. A Mage defaulted to Arcane in both slots — thirty-two spells where it should have sixty-four, and one book where Zangband gives two — because the function answering “what may I pick” only ever saw the class, never the character. Zangband excludes the first realm from the second list; now so does this. Mage, Priest, Ranger and Warrior-Mage start with two books, matching the archive exactly.
New suite
player/kit, because nothing pinned starting equipment and that is how two classes lost a book without anybody noticing. It asserts that every class starts with something, that every caster holds a book of each realm it studies, that no class holds a book it cannot open, that the two-slot classes take two different realms, and that the bookless classes stay bookless.Not changed, and recorded instead: five classes shared with Angband carry no experience penalty where Zangband gave them one. That is BAL-04 and DEC-36 working as written — 4.2’s classes keep 4.2’s conventions — but it leaves the five imported classes as the only ones paying for their power. Changing it is a balance judgement about Angband’s classes rather than a conversion one.
3.105.0 — Virtues are on the character sheet, and M8 is finished. PLR-20 asks that virtues be displayed to the player. They reached the character dump — a file you have to ask for and open in something else — and nothing on screen. The requirement had been recorded as met for fifty releases, its status note describing “a
[Virtues]section in the character sheet”, because that is what the code calls the dump.It is the identical failure PLR-17 was caught on, one requirement later.
The sheet has a fourth page now, reached with
Cand space, after mutations: eight lines on Zangband’s thirteen-rung ladder, from the polar opposite of to the living embodiment of. No numbers, because Zangband showed none and the rung is the part that means anything. The page and the dump both build their line withvirtue_line(), so they cannot say different things again.Folded in, all in the same code and all found by the M8 audit:
Mutation blows name what they hit. They said “You hit it with your tail” whatever was in front of you, with the monster in scope and the rest of the game using
monster_desc()throughout — unnoticed because until 3.104.0 the first swing of any melee mutation crashed.A body enveloped in flames gives light. Its description promised “+1 light” and its data carried
LIGHT[1], and nothing read it:calc_light()opens by zeroing the total and then walks equipment slots.Good luck does something; bad luck is recorded as doing nothing. Both were inert and neither was written down anywhere. Good luck’s object- generation half ports cleanly — one time in thirteen an object rolls against a level twenty deeper — and its other half, and the whole of bad luck, hang on an identification moment 4.2 does not have. So one is built and one is deferred with a reason, and the count of inert mutations goes from a claimed two to a measured three.
The mutation converter emitted a directive the parser refuses. It wrote
power-effect:msg:whereinit.cregisterspower-effect-msg, so regeneratingmutation.txtproduced a file the game would not load. The committed data had been hand-corrected months ago and the converter never was.zconv mutations --checknow exists and the gate runs it; four of the six converters still have none.And the stale counts are recounted from the data rather than carried forward: ten mutations do nothing, not the “six activatable and four random” the plan had;
mutmap.tomlcarries ninedeferkeys, not thirteen; Tame Monster has left the deferred table it had sat in since it was built, which drops the object deferrals from fourteen to thirteen; and the decision log no longer says Mindcrafter domination and Chaos patron gifts are open eleven releases after they shipped.3.104.0 — Two crashes in mutation melee, found by giving the borg mutations. Five mutations grow the character an extra limb – horns, a beak, a scorpion tail, tentacles, a trunk – and
player_mutation_blows()swings it after the weapon. It never completed a swing. Not once, for any of the five, since the code was written.critical_melee()was handedNULLwhere the monster goes, and it asksis_debuffed()whether the target is confused, held, afraid or stunned – which reads the monster’s timers. And it writes its message type through an out-param the caller may not want, unconditionally, so passingNULLfor that crashed too. Two dereferences, one after the other, on the first blow.The monster is passed now,
is_debuffed()treats “no monster” as “not debuffed”, and the message type is genuinely optional rather than optional-looking.Nothing had ever made a character with horns hit something. The borg did, on its first run with mutations granted, which is the argument for granting them:
borg-granthands over mutations rolled the way the game rolls them and every book of every realm the character studies, because the deeper books are dungeon drops and a granted level 25 caster with its starting book knows the same four spells it knew at level 5.With the grants, one Mage run held 21 spells learned, 4 mutations and a pet – against 5, 0 and 0 without them.
3.103.0 — The borg plays every night. The project owner set the cadence: “Agreed we should not run the borg too often. Once a night for now, then we can drop to once a week when development stabilizes.” Dropping to weekly is one line – the cron in
.github/workflows/borg.yaml– and nothing else needs touching, which is the point of it being one line.Scheduled, not on push. A borg run is minutes of play, and a check that slows every push is a check people route around.
It reports the three numbers – best depth, best character level, and how each run ended – and, per run, what was actually exercised: spells learned and cast, pets held, mutations. Depth alone would hide a borg that arrived deep having done none of the things M8 to M10 are about.
Deaths and time caps are data, not failures. A borg that dies at character level 3 is playing badly, which is worth knowing and is not a broken build; a nightly that goes red for that is a nightly everyone learns to ignore. Only a crash, an abort or a wedge exits non-zero, and each prints its seed so it can be replayed with
scripts/borg-progress -c <class> <seed>.Both build switches are named explicitly rather than assumed, because both are off by default and both fail quietly: without
SUPPORT_BORGthe borg compiles away to nothing and the binary plays a game nobody is driving, and withoutSUPPORT_TEST_FRONTENDthere is no-mtestmodule at all. The first draft of this job had the second wrong and died on “Unable to prepare any ‘display module’!”.
Gervais’ unexplored squares go black too — 5 September 2026#
3.102.2 — The same fault the generated tilesets had, reported again from play: Angband points
FEAT_NONEat cell (0,0), which in Gervais’ sheet is fully transparent, so what an unwalked square looks like is decided by whatever the front end leaves behind a transparent tile. On macOS that is white — a field of snow with the map drawn in it, which reads as a fault rather than as unexplored ground.Fixed as a mapping, not a modification. Of the sheet’s 3840 cells exactly one is fully opaque and pure black —
0x80:0x82— and nothing else refers to it.feat:NONEpoints there now.The three tilesets this game generates terrain into each carry a black square of their own, appended to the sheet. Gervais gets none, because
LICENSE.mdandlib/tiles/READMEboth say in those words that its32x32.pngis byte-identical to Angband’s. Drawing one in would have been easier and would have made that claim false.
Looking at a monster stops making you forget the trees — 5 September 2026#
3.102.1 — The disappearing squares were real, and they were trees. Reported from play as a graphical fault: squares going unexplored and coming back as the character walked, in every tileset. It was never graphical. It was in the ASCII build too — as a blank among the grass, where nobody had noticed it in the months it has been there.
path_analyse()inmon-util.cwalks the projection path from the player to every monster in view and forgets any grid along it that the player’s memory says blocks line of sight. In 4.2 that inference is airtight: nothing there lets a projection through while blocking sight, so such a grid is by definition misremembered.A tree is the exception, and the only one in this game —
PROJECTandPASSABLEwithoutLOS, which is the whole of “you can push through, but you cannot see far”. A bolt crosses it and sight does not, solos()andproject_path()disagree about it: a monster can be plainly in view around a tree while the projection path to it goes straight through one. Every such tree was forgotten, every turn.And it compounded. A forgotten grid remembers as
FEAT_NONE, which carries noLOSflag — sosquare_allowslos(player->cave, ...)stays false for it and the same grid is re-forgotten on every later pass. One tree could hold a line of country open indefinitely, which is why it looked worse than one square at a time.The memory is now dropped only when it actually disagrees with the ground: remembered as blocking, really not. On any level without trees this cannot change what 4.2 did.
Measured against the reported savefile, driven headlessly through the test front end and paced left and right as it was reported: 38 lost squares before, 0 after.
Pinned by
game/treesight, which has a suite to itself. The test has to stand the character in woodland, remember the whole map and put monsters round them; done insidegame/wildthat is three pieces of shared state to put back, and putting them back is where it went wrong three times — it wiped the town’s residents, then left the character somewhere else, then crashed one run in twelve restoring a knowledge array against a level that had moved under it. A suite of its own gets a level of its own and owes nothing to anybody.An earlier version of the test also passed against a deliberately broken build, because it never refreshed the view and
path_analyse()only runs for a monster whose grid is flagged in view — so it walked straight past the thing it was testing.And it took the sanitizer pass to find the last fault in it: the test placed its monsters on a fixed ring three grids out, and a tree is passable but is not floor, so
square_isempty()refuses one. Standing in woodland — which is the entire point — that ring can be trees the whole way round, and on about one seed in a hundred there was nowhere to put a monster at all. It searches outward in rings now. Forty flake passes clean, and it still fails on the seed that caught it when the fix is removed.
The borg learns to play — 5 September 2026#
3.102.0 — The borg can go shopping in another town. The starting village keeps four trades by design (WLD-11a) and the borg’s own restock rule wants a Staff of Teleportation from depth 10, which is a magic shop item. A borg that can only shop where it was born stops at depth 9 with two hundred thousand gold, whatever else is fixed.
Crossing the world to a town is the same walk as crossing to a dungeon mouth, so it is the same machinery pointed at a different landmark: the goal now carries a kind as well as an index, and only the choosing and the arrival differ – a mouth is descended, a town is stood in and the ordinary shopping takes over.
Three things had to be right, and the first two were wrong first:
What the borg wants is read from traits, not from
borg_restock()’s message and not by askingborg_prepared()where the wall is. The obvious version does ask, and it is not a pure query – it settlesborg.ready_morgothon the way past and returns a pointer into a shared static buffer. Speculating with it once a turn from the flow wedged a cheated Warrior on an unexpected direction prompt at turn 787,049.What the borg already has is the trades of the town it is standing in or beside, and nothing at all out in the country. “The nearest town” reads better and is wrong: a borg 334 grids from Avalon was credited with Avalon’s magic shop, concluded it wanted for nothing, and stood in a field.
Where the shops are had never survived the world moving.
track_shop_x/yhold window coordinates and the surface is a 144-grid porthole that re-anchors as the character walks, so a borg that walked to Avalon arrived still holding four sets of coordinates that now named open country. They are translated with the window now, and dropped when they scroll out of it.Measured: the borg picks Avalon – the nearest town with a magic shop, about 180 grids – walks there and arrives. What it does not yet do is settle and shop once it gets there; it wanders off again, because
borg_flow_dark()will explore a scrolling wilderness for ever and there is always more of it. That is the next thing, and it is a bigger problem than this one was.3.101.0 — The borg can learn spells, and cast them. Five hypotheses about the study loop failed, including memory corruption – ASAN ran clean straight through it. Reading the borg’s study path against the game’s, line by line, found it in one pass.
The borg matched a spellbook in the pack to a class book on tval and sval. That is not what makes a book readable here.
player_object_to_book()applies a second test the borg had no notion of: the book must belong to a realm this character studies (PLR-08). A class carries the books of every realm it may study, because a book has nowhere else to live, and the choice made at birth decides which of them can be opened – so a Warrior-Mage’s list holds twenty-eight books and the character can read four.A Warrior-Mage carrying a Chaos Book it cannot open recorded chaos book 1 as held, chose Magic Missile from it, and sent
G a aforty-one times.obj_can_study()filters the pack by the same realm test, so the book was never offered, the letter never meant anything, and the game refused before any prompt appeared – no message, no failed command, nothing in any log. The borg was asking for a book that does not exist as far as the game is concerned.It now asks
player_object_to_book()instead of repeating its rules, so the two cannot drift apart again. It is the function the study command itself goes through.Measured across eight classes, two minutes each, from
learned=0 cast=0:class
learned
cast
Mage
3
1
Priest
2
0
Warrior-Mage
3
0
Ranger
1
0
Rogue
2
1
Monk
5
1
Paladin
1
1
Mindcrafter
0
0
Mindcrafter has no spells in the data yet, which is the closed not-a-defect. The Warrior-Mage that used to wedge now plays until something kills it, which is data rather than a failure.
This is M9 finally being exercised rather than asserted. Nothing had ever cast a spell in a borg run before.
3.100.0 — The borg was not choosing to stay shallow. It had no legal move. A Warrior granted level 30 and 200,000 gold, placed at depth 25, climbed out and ground at depth 2 while losing a character level. That looks like a borg with bad judgement about danger. It was three separate faults, none of them about danger.
It was forbidden to descend and never sent home.
borg_prepared()refuses every depth below 4 without a Word of Recall.borg_restock()– the rule that sends the borg to town – has its matching line commented out upstream. So the borg may not go deeper and is not required to go anywhere else, and it grinds where it stands. It sat at depth 5 for thirty thousand turns holding forty Phase Door scrolls, no recall, and enough gold to buy the Alchemist’s entire stock of it. The same gap exists at every band boundary, because restock answers about one depth at a time: phase door becomes required at 6 and teleportation at 10, so a borg at 5 or at 9 is told it is fine where it is and refused the next step down.borg_must_return_to_town()now asks about the next level as well as the current one, and restock demands the recall scroll that prepared already demands. It cannot send the borg home for something town could not supply: restock only counts consumables, and the requirements town cannot sell – character level, hit points – live elsewhere and are not consulted.It threw away its own progress. Answering yes to “Set recall depth to current depth?” does not redirect a recall, it sets
max_depthto wherever the character is standing. The borg answered yes at depth 5 with a best of 9, discarded the 9, and spent three quarters of the following run re-earning depth 6.Two faults behind that.
unique_depths[]keeps the three shallowest live uniques by comparing against the deepest kept so far, but starts at zero, so the comparison islevel < 0and nothing is ever inserted – the borg’s idea of what it is hunting reads as depth zero for the whole game, and every depth looks far below it. And the reset itself had a catch-all yes. It now refuses when the borg is shallower than its best and still equipped for that best, which is the case where the depth is progress rather than a mistake.Measured on one seed, same budget: deepest reached 9 to 10, best depth held at the end 6 to 8, armour class 66 to 72.
And the ceiling that is left is not a borg fault at all. The borg is now stopped at depth 10 by
restock tele + tele staff < 2. A Staff of Teleportation is sold in the magic shop; the starting village deliberately has no magic shop (WLD-11a – the armoury, the weaponsmith, the magic shop and the black market are what make a larger town worth the walk). The borg has never needed to travel between towns, because before the wilderness there was only one. Recorded as BRG-25; it is the next thing standing between the borg and depth 30.New:
src/tests/borg/prepared.c, the first unit test to call into the borg – unit test targets now compile withALLOW_BORG, without which the borg headers preprocess away and a borg test fails to compile rather than to link. New harness commandsborg-prepared?(the depth ladder and the first rung the borg cannot climb) and aborg-maxdepth-dropwatch, since a record that ratchets downwards is invisible in any single reading.3.99.0 — A borg run is time boxed, and says whether it finished. The project owner: “The borg also needs to be time boxed on the nightly run not that it runs for 25 hours.” A turn budget does not bound wall clock — a character at depth computes far more per turn than one at depth 1, and a wedged run consumes no turns at all while running for ever.
ZTK_BORG_MINUTESdefaults to twenty, chosen to fit a nightly window with room for the gate and the suites beside it, and meant to be retuned once there is data on how long a descent to depth 30 takes.The two limits fail differently and the report distinguishes them: reaching the turn budget is
result=ok reason=budget spentand means the run finished; reaching the clock isresult=capped reason=time cap -- run did NOT finish. Neither is a crash, for the same reason a death is not — a nightly that goes red for running out of time is a nightly everyone learns to ignore. A capped run still emits its status line and exercise report, so it yields its numbers rather than nothing.The first capped run was itself informative: a cheated Warrior bought armour from 9 to 56, upgraded its weapon and earned money, which retires the suspicion that the borg could not value equipment — it never had any to spend.
3.98.0 — The study loop is named, and it is not what four fixes assumed. The borg has never successfully studied a spell in any run this project has made, which is why every exercise report reads
learned=0 cast=0 of 224.Tracing the key queue settled what three hypotheses had got wrong. The borg queues
G a d— study, book, spell — all three keys are consumed cleanly, and the game answers “You cannot learn any new spells from the books you have.” There is no desynchronisation and no stray key; the object context menu chased through three earlier fixes was a red herring. A Warrior-Mage is born with food, a torch, a sword, armour and a Word of Recall — no spellbook — so the game is right and the borg is wrong.One real defect found on the way and fixed: an empty pack slot is a zeroed
borg_item, so its tval and sval are both 0 and it matched any book entry reading zero, recording that book as being in inventory slot 0.The wedge itself is not fixed, and it is now specified precisely rather than guessed at:
book_idx[16]reads −1 immediately after the scan that sets it and 0 when the study code reads it, with only two writers in the whole subsystem and both inside that scan. That is a memory-consistency contradiction rather than a logic error. An ASAN build runs the first 400 turns clean.
An unexplored square is black everywhere — 4 September 2026#
3.97.1 — The generated tilesets give the unseen grid an opaque tile. Angband points
FEAT_NONEat cell (0,0), which in two of the four sheets is fully transparent — so what an unexplored square looks like was decided by whatever the front end leaves behind a transparent tile, and that is not the same everywhere. On macOS it came out white in Nomad’s set: a field of snow with the map drawn in it, which reads as a fault rather than as ground nobody has walked on.Reported from play, and the cause was ours.
make-terrain.pyrewrites the three sheets it extends, and Nomad was the only one stored as a palette PNG; writing it back as RGBA changes whatCGImageGetAlphaInforeports, which is whatmain-cocoa.mreads to decide whether to blend. Adam Bolt’s and the 8×8 set were already RGBA, which is exactly why Nomad alone changed.The fix does not argue with the front end: the generated row now carries a thirty-first tile that is opaque black, and
feat:NONEpoints at it. An opaque tile composites the same either way. Black because black is what “you have not been here” has meant in this game since it was ASCII, and one lighting variant because an unseen square is not lit three different ways.Left as it is, and said out loud rather than buried: the generator still converts a palette sheet to RGBA. Keeping Nomad’s format would mean quantising the generated tiles to its existing palette, which is a better answer and is not one to start at the end of a long session.
Terrain for the other three tilesets, drawn by a script — 4 September 2026#
3.97.0 — Adam Bolt’s, Nomad’s and the original 8×8 get a wilderness too. Gervais’ set had one hiding in it and was mapped; these three had nothing to map, so
scripts/tiles/make-terrain.pydraws it: ten features in three lighting variants, appended as one new row of thirty tiles to each sheet, with the matchingfeat-ztk.prfwritten beside it.Generated rather than shrunk out of Gervais, for three reasons. Nine of the ten are textures — grass is mottled greens, deep water is a blue that barely varies — and a texture is a rule, which survives being asked for at 16×16 and again at 8×8 where a downscaled 32×32 tile is mush. The palette can then be taken from the sheet being extended, so the grass is that tileset’s own green rather than one this project chose. And the art is ours, so no attribution follows it into three more files.
Nothing already in those sheets is touched. The row goes below the existing art, so every original cell keeps its coordinates and every original pixel its value. The script is deterministic and idempotent — re-running reproduces the same bytes and reclaims the row it wrote last time rather than adding another, which the first version got wrong.
What the palette sampler had to learn, each fault found by dumping the three numbers rather than squinting at the thousand pixels:
r >= g >= badmits pure red, so the first generated earth was scarlet in two tilesets; and merely clearing blue was not enough either, because(110, 45, 35)passes that and is brick. Earth wants green a quarter of the way up from blue towards red.g >= r and g >= badmits yellow, where red and green are equal — Nomad’s brightest “green” came back(248, 248, 8)— and teal, where green and blue are: Adam Bolt’s came back(88, 136, 136), which is why its first tree canopy was blue.Brightening by multiplication drains the colour out: the first channel to reach 255 stops while the others climb, so the hue slides towards white.
3.97.0 — What we have done to the tilesets is written down. A new
lib/tiles/README, and the same inLICENSE.mdand Copying and licence information, recording per tileset whether it was mapped (a lookup table onto art already in the sheet, altering no image) or modified (pixels written into it).Gervais is mapped only and its
32x32.pngis byte-identical to Angband’s. The other three are modified. Adam Bolt’s licence permits that in terms; the original 8×8 and Nomad’s carry no licence statement anywhere and stand on the project licence like the rest of the tree, which is a reading of a silence rather than a permission granted and is recorded as exactly that.Still missing, deliberately: the inn, magesmith, chaos tower and recharger everywhere, and the dungeon mouth and secret door in the three generated sets. Those are buildings and openings — shapes rather than rules — and a borrowed picture for one of them is actively wrong rather than merely absent.
The wilderness gets its tiles — 4 September 2026#
3.96.0 — Thirteen of the eighteen missing terrain tiles, and not one of them drawn. The wilderness is this game’s flagship feature and it was the only thing on screen with no art: grass, earth, sand, water, roads and trees all rendered as ASCII letters inside a tiled game.
They were in the sheet the whole time. Angband maps 1,151 of David Gervais’ 3,840 cells; another 1,239 unmapped cells hold art, and a contiguous block of it — sheet rows 23 to 25 — is a wilderness terrain set. Gervais drew for a variant audience; Angband has no wilderness, so nobody wired them up. So this is a mapping and not an art commission: no tile was made, resized or altered.
The lighting convention was measured rather than guessed. Gervais gives a feature three consecutive columns, and mean luminance runs
torch > lit > darkin every one of the fifteen featuresgraf-dvg.prfalready maps three ways. Candidate families were then found by that rule — a run of three unmapped, fully-covered cells whose luminance descends — which narrowed 3,840 cells to 75 to choose from.And then each one was looked at, which is the part that mattered. Colour alone puts lava and bare earth in the same bucket, and the first pick for bare earth was molten rock; the first open sea had a snowflake painted on it. Both were caught by rendering the picks before writing them down.
Water is three distinct looks rather than one recoloured three times — pale ripples, strong blue, and a smooth featureless open sea — because wadeable against drownable is a decision the player makes off the map.
SECRETis mapped to granite, closing a gap that is Angband’s own: unmapped, a secret door was the only ASCII#on screen and gave itself away by looking different.Four features are deliberately still letters: the inn, the magesmith, the chaos tower and the recharger. Gervais’ row of shop-front arches had exactly two unclaimed, and both fitted something — a cross for the healer, a crown for the magetower. The remaining four are not given a borrowed arch, because two buildings sharing a picture is worse than a building that is honestly a letter: the first is wrong silently.
3.96.0 — The stand-in monster tiles were never being installed. Found while adding the file above. Each tileset’s main prf has a
%:graf-ztk.prfline, and that file — which gives 234 imported monsters a tile so they are not letters among the pictures — was in none of the four hand-maintainedDATAlists. Somake installshipped four tilesets asking for a file that is not there, and every build that installs rather than runs from the tree lost the feature entirely.Nothing said so, because the only thing that would notice is an install and the developer machine does not do one.
scripts/check-build-listsalready checkslib/gamedata/Makefileagainst disk and stopped there; it now does the same for every tileset, which was verified by removing an entry and watching it fail.
Shockbolt’s tiles are removed — 4 September 2026#
3.95.0 — The 64×64 tileset is gone, on its licence. Angband ships it; this game may not, and had been shipping it anyway.
The terms are plain once read: permission is granted to use the tileset “with in-development and released versions of Angband”, and is expressly not granted to “use or distribute the tileset with other games or projects”. This is another project — separately named, separately released, with a savefile format of its own — so the two clauses point the same way. The author offers permission on request; nobody asked. Shipping it was an assumption, and it was ours to correct rather than to keep quiet about.
Removed:
lib/tiles/shockbolt/(17MB), graphics modes 5 and 6, the entries in every build and packaging list, and the licence grant that never covered us. Four tilesets remain — the original 8×8, Adam Bolt’s, Nomad’s and David Gervais’ — and Gervais at 32×32 is now the largest.If you were playing in one of those two modes, the game will start in ASCII and you should pick another tileset from the graphics menu. Nothing crashes:
grafmode_is_installed()already refuses a mode whose image is absent, and a saved preference naming 5 or 6 now finds no such mode and falls back. The two ids are deliberately not reused, so a stale preference cannot be handed a tileset nobody chose.Two things follow from this that are worth saying out loud. The web build now carries every tileset it lists — the only reason it left any out was Shockbolt’s 17MB. And the wilderness terrain-tile problem got smaller: the one tileset whose licence forbade us from adding the missing tiles is the one that has gone, and Gervais, which already contains an unmapped wilderness set, is CC-BY and may be both mapped and modified.
Also recorded while reading the terms: the original 8×8 and Nomad’s tiles carry no licence statement, here or in Angband’s own copying page. That is an inherited gap rather than a settled position, and it is now written down in
LICENSE.mdinstead of assumed.3.95.0 — Two dead tile mappings, and the check that was not being run. Found by accident while removing the tileset above:
graf-ztk.prfis generated, committed, and had gone stale.hoboandraving lunaticwere dropped from the import in77b340f50for being Angband’s own monsters wearing Zangband names, and their tile lines sat in four prf files for weeks afterwards, pointing at monsters that no longer exist.borrow-tiles.pyhas had a--checkmode for exactly this the whole time, and it catches both immediately. Nothing ran it. It is a pass inscripts/check-buildnow, beside the realm content check and the build-input lists, both of which are there for the same reason: a check nobody runs is a check that does not exist.
A polymorphed pet is still yours — 4 September 2026#
3.94.3 — And polymorph stopped deleting things (PLR-22). The second half of the same bug, found by asking what happens when the placing fails.
4.2 polymorphs by deleting the monster and placing a new one, and never looks at whether the placing worked. When it did not the monster was simply gone — which is a spell that occasionally works as a deletion, and since PLR-22 an occasional way to lose a pet with no message at all.
It was not rare. Measured at 3.5 per cent from depth thirty down — about one polymorph in twenty-eight — 1.2 per cent at fifteen and 0.2 per cent in the town. The curve is this game’s own bestiary: two dozen races are fish,
place_new_monster_one()will not put a fish on dry land (a ZangbandTK rule — 4.2 has no water to need it), andpoly_race()drew from the whole allocation table without asking where the monster was standing. The deep fish are krakens, whales and Charybdis, which sit squarely in the band a deep polymorph draws from.Fixed at the draw rather than at the failure. The preconditions
place_new_monster_one()applies that do not depend on who is standing there — terrain, water, glyphs, unique counts and depth — are now one function,monster_race_fits_grid(), which that function obeys andpoly_race()consults. So the shape drawn is one that can stand where the monster stands, and the spell works instead of fizzling. Zangband’s fallback is kept behind it — “Placing the new monster failed - use the old” — because the cost of being wrong about “unreachable” is a creature vanishing.Both halves are pinned separately, and each regression fails its own assertion: drop the fallback as well and monsters vanish; keep the fallback and drop the filter and nothing vanishes but the spell silently does nothing eight times in six hundred.
One incidental finding, recorded because it cost an hour: the level generator will sometimes start a character in a doorway with seven walls round it, and the test helper that places a monster beside the player wants an empty floor grid — an open door is passable and is not floor. One level in twenty has nowhere to put anything. That was a 12 per cent flake with nothing to do with polymorph.
3.94.2 — Polymorph no longer takes your pet off you (PLR-22, PLR-33). Reported from a reading of what players said they liked about the original, which names “recruiting an army of summoned pets that can polymorph into incredibly powerful dragons”. Ours did the opposite: it handed the dragon to the enemy.
A polymorph is one creature to the player and two to the code. 4.2 deletes the monster and places a new one, and everything not carried across that seam is lost — which, since PLR-22, includes whose side it was on. So a pet you polymorphed came back hostile, standing next to you.
And silently.
MON_POLYzeroes the damage, so the one site that turns an ally against you —mon_take_hit(), PLR-33’s choke point — was never reached, and nothing said “gets angry!”. The player was left with an enemy where their ally had been and no message accounting for it. Zangband read the attitude before the delete and handed it back after (polymorph_monster(),spells3.c:4368); this now does the same, at the same two points.This is the mechanic the strategy rests on, not merely a tidy-up.
poly_race()aims at(depth + monster level) / 2 + 5, so polymorphing your own animal deep in a dungeon is how a weak pet becomes a strong one. PLR-30’s upkeep charges the sum of pet levels against mana regeneration, so the reward already carries a price that scales with it and no cap was wanted.PLR-33 is not bypassed. A chaos attack that polymorphs also deals damage, and damage runs before the side effects, so the pet is hostile before its shape changes — dropping a chaos ball on your own animal does not become free by virtue of it polymorphing. Three tests in
monster/angerpin all of it, and the two that matter were checked against a deliberate regression: with the restore removed, the pet and the friendly monster both fail and the chaos case still passes, which is exactly the split the rule describes.
M10: the last two pet sources — 4 September 2026#
3.94.1 — PLR-28’s last two paths (PLR-28, DEC-63). M10 was declared complete carrying three of the five documented ways to acquire a pet, and Pets and Allies said so at the bottom of the chapter — an honest gap, not a closed one. Both turned out to be data, where the plan had them down as needing a code hook.
A Mindcrafter’s domination existed at level 9 and held minds: paralysis under the name of domination. Zangband’s is two powers in one. Below level 30 it fires
GF_DOMINATIONat a single target, and that projection cannot make a pet until its damage passes 29 — which, the damage being the character’s level, it cannot yet be. From 30 it ischarm_monsters()on everything in sight. So the low band is a mind attack and only the high band charms, and that is what it does now. 4.2 has no projection that picks among stun, confusion and fear, so the low band is confusion alone — the one Zangband’s own switch reaches for most often.A Chaos Lord’s gift. Zangband has three servant rewards; this game carried none of them. All nine Lords now have one — a servant, a demonic servant or an undead one — at rungs 8 to 17 of the twenty, which is the band Zangband’s own sixteen patrons place them in, each replacing a reward that Lord already gave twice, so nothing it could do was lost.
Both are pinned by tests, and both tests were checked against a deliberate regression: reverting the projection, or one ladder entry, fails them.
Tidying — 4 September 2026#
3.94.0 — Four prompt failures turn out to be one mistake, and it is the harness’s. Sampling a wedged process settled a question three hypotheses had got wrong.
A prompt reads its input with
inkey_ex(), which consults the borg’sinkey_hackbefore polling the terminal — so while the borg is active the borg answers every prompt in the game. A key pushed withTerm_keypress()is never read at all;internal_borg_inkey()only peeks at the terminal for the user-abort check, which a headless run ignores by design.That unifies four failures previously treated as separate: the borg’s own initialisation, level generation, a store, and 4.2’s object context menu. The first three looked fixed because in those cases the borg was not yet active and the terminal layer genuinely was the reader — which is why the fourth resisted three separate fixes.
It also retires a suspected game defect. “ESCAPE does not clear that context menu” was a fact about the harness, not the menu:
menu_dynamic_select()honours ESCAPE and returns −1, so a player is not trapped there.The whole of it is written up in Borg beside the harness commands, because it is the sort of thing that costs the next person a day.
The context-menu wedge is not fixed. Delivering ESCAPE through the borg’s own queue is the right layer and still does not clear it, and it reproduces on every seed tried — turns 725, 755, 1063 and 3765. So the scoped route cannot yet complete a verification run.
3.93.0 — A borg run says where it got to, and breaks out of prompts it cannot answer. Two instruments, and the second turned a class of half-hour investigations into a line of log.
Status was printed only when a run finished, which is exactly wrong for a run that may not: three separate questions this session were unanswerable for that reason, the last costing half an hour to establish that a process was at 0% CPU. A progress line now goes to the log every 500 game turns.
It localised the standing hang immediately — depth 25 at turn 1, depth 1 by turn 504, stopped between 1004 and 1504 — and sampling the wedged process named it: the borg opens 4.2’s object context menu through
do_cmd_equip()and cannot get out, spinning inside it while the game turn never advances.That was the fourth prompt the harness could not answer, after the borg’s own initialisation, level generation and a store. So the fix is general rather than a fourth special case: when the game keeps asking for input while the game turn does not advance — a borg that is playing moves the clock, one trapped in a menu does not — the harness sends ESCAPE and logs it, and fails the run as wedged if forty attempts do not clear it.
It fired correctly and ESCAPE did not clear the context menu, which is a second finding and possibly a game-side one rather than a borg one.
3.92.0 — The scoped route: a Warrior-Mage, four cheats, and depth 30 as the bar. Standard play does not reach the depths this game’s content lives at, and the early-game grind is not what the borg exists to verify — so the route is one character taken deep with the cheat infrastructure this project already built.
Warrior-Mage, because it has all seven realms: 224 of 224 spells are reachable by one character, and that includes every source of pets in the game. It leaves the Chaos patron and the Mindcrafter’s psionics untested, which no single route covers.
Four levers, all pre-existing:
borg-cheatgrants character level and gold,borg-jumpplaces the character at a dungeon whose band contains the wanted depth,BORG_CHEAT_DEATHremoves the attrition, andcheat_powersstops mutation powers failing. The cheats remove attrition, not decisions — grants happen at the start, never in reaction to danger, and the jump goes to a dungeon rather than to the target depth, so the intervening levels are generated and played. Deaths are counted and reported even though they are cheated.Gold rather than granted equipment, deliberately, so the borg shops — and it does: dropped at depth 25 it walks to town for Phase Door scrolls, which is exactly the decision the route is meant to preserve.
It does not reach depth 30 yet, and the honest reason is that a shopping round trip from depth 25 takes more turns than the runs so far have allowed. Also fixed here: granting experience raises maximum hit points without healing, so the first cheated character woke at depth 25 with fourteen hit points of 244 and fled to depth 1 — which looked like the borg refusing to descend and was the harness handing it an invalid character.
Housekeeping:
docs/hacking/borg.rstnow describes this variant before upstream’s text rather than only Angband and Morgoth — discharging BRG-21, which had gone unmet across five borg releases. (The duplicateDEC-65was fixed independently and better in 3.91.0, which also caught a second duplicate atDEC-31and added a check for it.)3.91.0 — The SDL 1.2 front end is retired (DEC-70).
main-sdl.cwas 6,169 lines built by CI and shipped by nothing — reachable only by configuring it by hand, and not testing what its label said. Debian and Ubuntu replacedlibsdl1.2-devwith sdl12-compat, which implements the SDL 1.2 API by dlopening SDL 2, so the job named SDL was building an SDL 1.2 front end against SDL 2 through a translation layer. It went green and verified a configuration nobody ships.Nothing is lost, and that was checked rather than assumed. Our only two local changes to the file were the About box printing every credit line and a guard against a saved configuration naming an uninstalled tileset. The first is already in
main-sdl2.c— the same commit changed both. The second does not apply there: the SDL 2 front end validates a graphics mode when it parses the config, so it never had the hazard.Gone with it:
SUPPORT_SDL_FRONTENDandSUPPORT_SDL_SOUND, the two CMake macros,--enable-sdland--enable-sdl-mixer, theAM_PATH_SDLdetection,SDLMAINFILES,SYS_sdl,SOUND_sdl, the desktop file, the now-unreachablefont-sdl.prf, and the CI job. The two statistics builds kept their sound — the point was to compile more infrastructure — and now get it through SDL2.snd-sdl.cstays: it includes plainSDL.handSDL_mixer.h, so the include path decides which SDL it gets, and SDL2 sound uses the same file.3.90.0 — A cheat that hands you a pet, on the
=xscreen and on the debug menu as^AthenN. It asks which monster and puts one beside you, already yours.Every route the game offers to a pet is a roll — a Trump summon, a charm that has to beat a saving throw, a patron in a good mood — so seeing what pets do meant playing until one arrived and hoping it was the kind you wanted to look at. The placement is a function rather than part of the command, because the borg’s test harness wants exactly the same thing, and a test can call it.
The decision log had two numbers that each meant two things. DEC-31 was both the dungeons are Amber’s places and Zangband’s six town inhabitant types were three unused constants; DEC-65 was both the SDL 2 front end stays on SDL 2 and what a permanent stable costs. Both collisions were being cited from live files and commit messages before anyone noticed — the file parses and reads perfectly well with a duplicate in it.
The cited meaning keeps the number in each case, and the other is renumbered with a note saying which commit cites it the old way: the dungeons become DEC-69, SDL becomes DEC-70.
scripts/check-build-listsnow fails the build on a repeated number, which is the only thing that would have caught it.
Pets follow — 3 September 2026#
3.89.0 — The borg follows the road, and stops teleporting itself to its death. Two changes, one of which lowers a number this log has been reporting.
It follows the road. The project owner: “The borg should follow the road, that’s what most players will do.” And the game was already built for it —
wild_place_roads()lays a spanning tree over the towns, short hops between neighbours, and a spur to every dungeon mouth, its own comment recording that before that pass existed six of thirteen mouths happened to sit on a road and the rest were up to a thousand grids of open country away. Verified: every mouth in both seeds tested is on a road, Rebma included. A road is passable by construction, which retires the terrain question rather than sampling around it; it routes around mountains, so following one is safer; and it is legible in a log. The measured clear-line test is kept as the fallback for a mouth with no road.And it no longer reads Scrolls of Deep Descent it cannot survive. Deep Descent drops the character
(4 / stair-skip) + 1levels — five as the data ships — and the borg’s preparedness rules gate walking down one level and said nothing about this. A Warrior at character level 4 read one at depth 6, arrived at depth 11, and died there.That correction lowers the best depth this log has reported since 3.85.0 from 11 to 6. Depth 11 was never earned: it was the scroll accident, and the borg died as a direct result. Depth 6 is what it walks to, and it matches its character level exactly, which is what its own rule intends.
3.88.0 — The borg can walk across the world to a dungeon it chose. There is no route to depth 30 without this: the Vaults of Amber, which the town staircase leads into, ends at depth 15; a town staircase always leads to the shallowest dungeon there is; and none of the thirteen dungeon mouths is inside the starting 144×144 surface window — the nearest reaching past 15 is 576 grids away, across a world roughly fourteen windows square.
It is a bearing rather than a route, and a measurement licenses that. Sampling the straight line to every mouth at block resolution across four seeds, at least one dungeon whose band reaches past 15 always has a completely clear line, usually several — and Rebma, whose band of 25–50 contains the target depth, was clear in all four. Where a line is blocked it is mountainside, never open sea. So the borg prefers a target it does not have to route around, which turns a pathfinding problem into a target-selection one.
The destination is held in world coordinates, because the surface window is rebuilt and re-anchored as the character crosses it, and a level grid stops meaning anything the moment it moves. Inside the window the borg flows to a known grid; outside it walks the bearing and lets the world scroll, choosing among all eight directions by how much each closes the world distance so a step round an obstacle is a step sideways rather than a stop.
It refuses to swim, and it gives up rather than thrashing: the step budget is spent on failing to close the distance rather than on steps taken, and an abandoned mouth is remembered as hopeless so the next choice is a different one. And it only crosses when actually stuck — it must have reached the bottom of the dungeon it has been in. Without that gate it walked 831 grids on its first turn to the mouth of the very dungeon whose staircase stood ten grids away, and died at character level 1 in open country for its trouble.
3.87.0 — The borg learns that a dungeon has a bottom, and that every other one is a long walk away. Two array overruns and a wall the borg has not reached yet.
dungeon_get_next_level()clamps a descent to the dungeon’s own floor rather than refusing it, so taking down stairs at the bottom of the Vaults of Amber returns you to depth 15. The borg’s preparedness rules knew nothing of dungeons and would have read each clamp as a successful dive, looping for ever with no crash and no message. It now reports “no deeper in this dungeon” and goes up, which is its own existing answer to being unready for deeper.Alongside that, two more arrays sized for Angband’s widest caster:
num_bookis indexed by an object’s sval, which nothing bounds, and had nine slots. That is the fourth instance of the same shape after the map arrays, the spell table andbook_idx— recorded as a pattern rather than four separate fixes, because none of the four had a guard and the one added here refuses to start rather than corrupting memory.And a measurement that sizes what remains: of the thirteen dungeon mouths in the world, none is inside the 144×144 surface window — not even the Vaults of Amber’s own. The nearest dungeon reaching past depth 15 is 576 grids away. Since the town staircase always leads to the shallowest dungeon, there is no route deeper than 15 that does not cross the world.
3.86.0 — Every spell the borg can cast now has a rating, and it made no difference yet. BRG-08’s derivation, done properly now that the rating rubric exists: 224 distinct realm-and-name pairs across the seven realms, 49 already rated upstream, 175 needing a judgement. Rated by effect category against the rubric rather than one spell at a time, so they stay consistent with each other and a spell added later is rated the day it is added.
The rubric leads with survival, and so does the mapping: escape and healing at 90–95, above the best attack spell at 85, so a cornered borg runs rather than swings. Sleeping or scaring a whole room rates 88 — survival wearing an offensive coat. Summoning a pet rates 40, deliberately low, because the borg cannot reason about the mana upkeep that would leave it unable to cast.
Measured, and it did not move the casters: over three seeds at 40,000 turns a Mage still dies at character level 1 and a Priest at 1–2, where a Warrior reaches level 6 and depth 11. A level 1 Mage has four hit points; it dies before spell selection means anything. Recorded as a result rather than presented as an improvement.
3.85.0 — The borg reaches depth 11, up from 2, and it was food. Two thresholds inside the borg disagreed with each other and with this game’s data.
borg_restock()refuses to descend past depth 1 with fewer than three rations;borg_choose_shop()only went shopping for food at zero. Between one and two the borg would neither dive nor buy — and our classes start with one to three rations where Angband’s start with three to five, so two rolls in three begin below the line. It rose to the surface, bought nothing, went back down, and reported “unable to dive: restock food < 3” for ever.Fixing the shop threshold alone was not enough, because the borg’s power curve valued food only up to seven rations and then almost stopped, so it set off with barely more than the minimum and starved at depth — where it could not path back to an up-staircase, would not dive without food, and escalated its own “bravery” from 3 to 30000 until it ignored all danger and died to a grid bug. The carry target is now twelve, named separately from the dive threshold so the two cannot drift together again.
Measured over six seeds at 40,000 turns each: best depth 2 → 11, median character level 3 → 5. Deaths rose from three runs in six to four, which is the honest half of the result — it goes deeper and dies more, and survival is now what limits it rather than starvation.
Also new:
scripts/borg-progress, which records best depth and character level per class from fixed seeds (BRG-18, pulled forward), because that sweep had been hand-rolled four times and slightly differently each time.3.84.0 — The borg stops treating your pets as targets.
src/borg/held no concept of allegiance, so a pet was simply a monster to it: something to target, to be frightened of, and to walk across a level to kill. Sinceborg_kills[]is what danger is computed over and what the borg walks toward, one check at the single point where entries are created covers targeting, fear and pursuit together; a second drops any creature that changes sides while the borg is watching, which a Wand of Tame Monster can do.Implemented but not demonstrated, and the reason is the interesting part. An instrument was built for it —
borg-kills?, which reports how many tracked entries point at something that is not hostile, a number that must be zero. It reports no tracked monsters at all, with the fix and without, after sixty turns on a level holding twenty-four hostiles: the borg never accumulates a target list because it never fights, and its list is wiped at every level change. So this is gated on the same blocker as 3.83.0 — the borg cannot restock, so it stair-scums instead of playing. Landed regardless, because the only failure mode of the guard would be the borg ignoring a monster, which its condition makes impossible.3.83.0 — The borg stops believing the surface is the outside wall of the world. Four files read Angband’s
DUNGEON_HGT— 66 — as the current level’s height, and this game’s wilderness surface is 144 rows, with a character starting at about row 81. So the borg’s own position was off the bottom edge of its world model: the grid test it uses to decide whether anywhere is worth standing returned false for every grid the player ever occupied. With that fixed it registers the starting town’s shops, which sit at rows 83 to 87 and were previously outside its universe entirely.It also visits shops in a branch that ran after the decision to leave the level, making it dead code in town — the borg arrives from the dungeon standing on the town staircase and always leaves before it gets there. Angband never shows this, because there the borg’s first town visit happens before it has ever descended.
The borg still does not get past depth 2, and that is now measured rather than assumed: 3000 turns for each of the twelve playable classes, best depth 2 and best character level 2, before and after. The remaining blocker is a food deadlock — it refuses to dive below three rations and only shops for food at zero — plus a bootstrap where it cannot want to buy from a shop whose stock it has never seen. Aligning the two thresholds did not fix it and was reverted rather than kept on a hunch.
And a correction to the harness worth more than the rest: a dead character is no longer a failure. Three of the twelve classes die inside 3000 turns at character level 1, and the borg reports death down the same channel it uses for defects, so a nightly job would have been permanently red.
3.82.0 — Twelve of fourteen classes can be played by the borg, up from two. B1’s robustness half, taken before its ratings half because nothing crashing matters more than playing well. With the harness from 3.81.0 the failure could be measured rather than assumed, and the measurement corrected the plan: twelve classes failed rather than thirteen, and ten of them killed the process rather than reporting a startup failure.
The borg matched each class’s spell list against a hand-written table positionally, by name, and M9 replaced every spell list — so a Mage’s first spell is now
Zapwhere the table expectsMagic Missile, and the loop then read 194 entries past the end of a 30-entry array. Ratings are now looked up by name within the table’s real length, an unrated spell is merely unloved rather than fatal, and the five classes this variant added no longer fall through to an unallocated array. A spell with no effect chain no longer dereferences NULL either.The two remaining classes, the Necromancer and the Blackguard, loop trying to study their first Death spell. They now fail with
reason=wedgedand the loop visible in the notes, rather than hanging — which is the step budget from 3.81.0 earning its place a day after it was added.3.81.0 — The borg plays the game, and says so when it breaks. B0 of the borg plan. Nothing in this repository played the game: 112 unit suites test rules in isolation and none of them walks a character out of a town, into a dungeon and back.
scripts/borg-smokedoes, from a seed, and exits non-zero on a crash, an abort or a wedge with the seed printed beside it.The borg segfaulted on the first turn of every game because its map arrays are sized for Angband’s 66×198 dungeon and this game’s depth 0 is a 144×144 wilderness surface. They are now sized by a ceiling that the game’s own largest level is checked against, so growing
wild:cache-blocksproduces a startup failure rather than memory corruption.Two build defects had to be fixed first, both single-target option leaks that failed silently:
ALLOW_BORGreached the executable only in the Windows branch, so on macOS and Linux no front end could name the borg, andUSE_TESTnever reached the core library, so guarded code there compiled to nothing. A headless birth was needed too — the four existing front-end tests all quit before the game starts, so a live character had never existed in the test front end.Also new because the phase needed them: a decision budget, so a borg waiting for a prompt it cannot see fails instead of hanging a CI job with no diagnosis;
borg-notes?, which produced every diagnosis in the phase; and a mid-run savefile round-trip, pulled forward from BRG-19. The borg reaches depth 1 in 300 turns and the same seed gives the same run.3.80.1 — Every class is asked whether its books and its spells agree. The borg reported the Mindcrafter as having zero spells, which turned out to be correct and by design — PLR-06’s psionics are twelve
power:entries and no realm, andgame/wildalready asserted it. Next door to it was a real gap: the Monk casts twelve books of ninety-six spells and its spell count was pinned nowhere, becauseplayer/realm’s counts table names classes one at a time and nobody had added a row. Fixed with the row, and with an invariant asked ofclassesitself — books and spells present together, the total equal to the sum of the books, every book one to eight spells of a named realm, and every choosable realm backed by books. Falsified by perturbing the loaded data three ways, and it catches an empty book that the existing entitlement test passes. No game change.3.80.0 — Four pets, and one in twenty walks away. Two decisions from the project owner, both about how the game feels to play rather than about the numbers. The carry cap drops from 24 to four: it was written as a safety valve against a case that would never arise, and it becomes a deliberate limit, because “more then that at it becomes unmanageable as the screen will be a mess of pets”. And each pet now has a flat 5% chance, at each level change, of leaving you for good — “He does a runner. Gone”, “Looking for a new owner” — removed from the game, not left standing on the level behind you, and not recoverable under persistent levels either. Read next to what it replaces: Zangband deleted every pet at every staircase, which is the project owner’s own justification — “That’s fair enough given the original zangband was 100%”. Compounded, a pet lasts a median of thirteen or fourteen level changes, and a full stable of four loses one about one descent in five.
The four ways of not coming now read as four messages, and the distinction they carry is whether the pet is still yours: one ran off and is gone, the other three are standing where you left them.
pets:leave-chanceis data and 0 turns it off. Recorded as DEC-68, and the pets chapter is rewritten with the whole picture in it — following, the four, the 5% and what it compounds to, which losses are permanent, and the upkeep table showing the third young red dragon costing 93% of a caster’s regeneration.The rule is off inside
game/carryand switched on by the tests that are about it, because a suite that changes levels constantly and asserts exact counts cannot also be losing a pet one descent in five. Getting that right took a correction thatscripts/check-flakesfound and nothing else would have:a-carry-then-a-save-round-tripsre-initialises the game in the middle of the suite, which re-readslib/gamedataand quietly turned the rule back on for every test after it. One failure in twenty whole-set passes, and about one run in fourteen for the test that actually broke.3.79.2 — The gate checks the build input lists. Three consecutive pushes went red on Linux CI for the same reason — a new
lib/gamedatafile missing from the install list, which would have shipped a packaged build without it — becausescripts/check-build-listswas a separate script you had to remember andscripts/check-builddid not run it. It does now. That is the fourth pass added to the local gate because CI caught something the gate did not ask about, after-Werror, the missing test-frontend option and the GCC pass; the answer has been the same every time.3.79.1 — A charm test stops assuming a charm is certain. The saving throw is
level > randint1(power * 3), so a level 30 creature against a power of 200 rolls againstrandint1(600)and wins about one time in twenty. Four assertions inmonster/charmcalled the charm once and required it to take — a five per cent flake apiece, and the comment on one of them claimed power 200 “beats the saving throw essentially always”, which the arithmetic does not support. Raising the power only makes it rarer, so the tests now ask whether a creature can be charmed rather than whether it was charmed first try. The assertions that a charm is refused need no such loop: those are certainties. Found bycheck-flakesreporting the seed.3.79.0 — What a permanent stable costs (PLR-26 phase F; DEC-65). Pets following the player raised a question: the mana upkeep had gone from being the second limit on a stable to being the only one, and the worry was that this made Trump straightforwardly the strongest realm. Measured, that worry was wrong.
A level 30 Mage regains 36 spell points per hundred turns and keeps two pets free. Two young red dragons cost nothing; the third costs 93 per cent, taking the same Mage to 2 per hundred turns. The upkeep is charged on the sum of pet levels, so for anything worth summoning it is a cliff rather than a tax, and a caster with an army of deep pets cannot cast at all.
What following actually changes is narrower: a caster may keep their free allowance permanently — two pets at level 30, four at 48, five for a High-Mage. A few standing companions rather than a horde, and everything past the allowance now costs continuously instead of being rebuilt, which makes the large stable a worse proposition than before rather than a better one.
No change to the upkeep. The figures are printed by
player/pet-upkeep, so a change to the regeneration formula or to monster levels moves them where somebody can see it.3.78.0 — Told which ones, and by name (PLR-26 phase E). The player is told their pets followed, and every pet that stays behind is named individually.
Two defects in the obvious version of that, both found by the tests.
The messages were said where the decision was made, which is before the old level is torn down — so they landed on a screen the player never sees. The names are now held and said on arrival, beside whatever else is being reported.
And they said “It cannot follow you.”
MDESC_IND_VISgives the indefinite form for a monster the player cannot see, and nothing is visible during a level transition — so every pet got the same useless line, and since the message log collapses identical lines, three animals declined produced one. Named by race now (MDESC_SHOW): a player knows their own animals whether or not they can see them. The pet roster and the dismissal prompt had the same fault and are fixed with it.Nothing is said when the player has no pets, so the line means something when it appears.
3.77.0 — The awkward cases (PLR-26 phase D). Uniques, groups, stored levels and the arena.
A real bug, found by the test written for it. The racial counter was released when a pet was lifted off the old level — which looks tidier and is wrong, because the new level is generated between the collection and the placement. A carried unique therefore became available to the generator, and the player could arrive to find a second copy of their own pet standing there. Two Grips, reproducibly. The count is now held through generation and handed over at placement.
Groups. A carried pet’s old group index is cleared, so it joins a fresh group of its own.
monster_group_assign()self-heals a dangling index, which is why a test that only asked for a valid group passed against a build with the clearing removed; the case that is not harmless is a collision with a real group on the new level, which quietly enlists the pet among strangers. The test forces that collision and asks that the pet’s group contains the pet and nobody else.Stored levels. With persistent levels the old chunk is kept rather than freed, so the collection has to happen before
cave_store()— otherwise the stored level holds a monster that is also standing on the new one, and coming back up hands the player a second copy.The arena carries nothing in and nothing out, and that is tested rather than merely implemented: an arena that let a stable in would be found by a player long before it was found by us. Writing that test also turned up that
arena_gen()readshealth_whowithout checking it, so an arena entered without an opponent named segfaults — unreachable in play, since only the effect can get there, but worth knowing.3.76.0 — And they bring what they are carrying (PLR-26 phase C). A held object belongs to a chunk twice over —
obj->oidxis a slot in the real object array and the player’s knowledge of it sits in the same slot of the known array — andcave_free()deletes anything listed with no grid, which is every held object. So a carried pet’s belongings are released from both of the old chunk’s lists and re-listed on the new one at matching indices, which ismonster_carry()’s idiom run backwards and then forwards.Falsified both ways: skipping the release crashes on freed memory, skipping the re-listing is caught by the integrity checker and loses the object.
A mimic still does not follow, and that is not a gap to be closed. What it pretends to be is an object on the floor, with a grid and a place in a pile: it belongs to the level. Carrying the monster and leaving the disguise gives a creature imitating something that is not there.
One diagnosis worth keeping. A test reported the carried object duplicated — two apples where one was given — which is precisely the failure this phase risks. It was not: the new level had generated a monster carrying an apple of its own, and the test was counting the whole object list rather than the pet’s inventory. Chasing it down rather than adjusting the assertion is the only way to tell those two apart.
3.75.0 — Pets follow you downstairs (PLR-26 as reversed; DEC-67, reversing half of DEC-60). Every pet on the level comes with you, put down within five squares of where you arrive, nearest-follower first. Anything that will not fit is left behind and named individually, however many lines that takes — arriving with four when you left with six and not knowing which two you lost is worse than losing them.
Zangband deleted every pet at a level change, and that was its region model deciding rather than a design: no pet-carrying code anywhere in its source, no mention in its documentation, nothing in nineteen changelog entries about pets. The recommendation here was to keep that behaviour, and it lost: “Pets are one element of zangband that people like. Usually people don’t leave their pets behind.”
Two constants, both data:
pets:max-carried(24, a safety valve rather than a balance limit) andpets:carry-radius(5, measured — a mean of 6.6 empty grids within three of an arrival against 16.9 within five, and the player lands on a staircase, which is usually in a corridor).A pet carrying objects is not carried yet; it is left behind and told why. Its held objects belong to the old chunk’s object list, and moving them is the next phase. The alternative was destroying a player’s property to make a transition tidy.
Nothing follows the player into an arena or out of one.
3.74.0 — The chunk’s bookkeeping is checked (M10 phase A of PLR-26’s reversal).
cave_check_integrity()and agame/integritysuite: object indices agree with the slots the chunk filed them under, held objects name the monster holding them and are listed, monsters agree with the map about who is standing where, mimicked objects point back, and group indices name groups that exist.4.2 ships
object_lists_check_integrity()and nothing calls it in tests — it is built onassert(), which aborts rather than reporting and says nothing about monsters. This one reports, covers both lists, and is safe to call from a test.scripts/check-flakesalso reports the seed of a failing pass now. That is the whole diagnosis of a rare flake: without it a one-in-three-hundred failure is unreproducible and the harness can only say “intermittent”. It earned that on the run it was added for, naming three failures inmonster/ally-aithat six, forty and two hundred passes had not found — all of them the same mistake in the test rather than the code, assuming a randomly generated level would leave a free grid where one was wanted.It is here ahead of pets following the player downstairs, which moves live monsters and their held objects between chunks. That failure mode does not crash where the mistake is: it surfaces as a corrupt savefile or a duplicated artifact several levels later. An instrument that fails at the corruption is worth more than care taken while writing the thing that might cause it — and these invariants have never been checked at all, so it earns its place either way.
Pets — 3 September 2026#
3.73.0 — Trump (M10 phase 7; DEC-64, discharging DEC-54). The seventh realm, deferred whole in 3.55.0 because fourteen of its thirty-two spells summon a creature that serves you and nothing could. Every realm now has books behind it.
Twenty-seven of the thirty-two are built, across four decks — Conjurings & Tricks, Deck of Many Things, Trumps of Doom, Five Aces — for the six classes Zangband entitles: Mage, Priest, Rogue, Ranger, Warrior-Mage, High-Mage. The Mage, Priest, Warrior-Mage and High-Mage now carry 28 books and 224 spells apiece.
Zangband writes
bool pet = successin every Trump summon, with a comment recording that it had been a dice roll and was deliberately made certain. So a Trump summon that goes off is a pet; the angry version is a failed casting, which reaches the miscast machinery. A test pins zero plain summons in the realm, because one mapped the old way would look right everywhere else.Five are deferred, each for its own reason and each named in a test so a sixth cannot join them quietly. Shuffle is the interesting one: a d120 read off twenty unequal bands, and 4.2’s
RANDOMpicks uniformly — the outcomes could be listed and the weights could not, and the weights are the spell.Two things turned up on the way. The converter’s book-line pattern assumed every realm’s books were called “books”; Trump’s are a deck, so its books were invisible and the realm before it in each class appeared to run on through them. And
TransferLib, which stages the game data for the unit tests, compares timestamps and can decline to re-copy a file reverted within the same second — two falsifications read a stale copy before that was understood.scripts/check-flakesnow stages it every run.3.72.0 — Where pets come from (M10 phase 6; PLR-28, PLR-32, DEC-63). Three charm projections and a pet-summoning effect, and with them the queue this milestone has been holding since M8 comes off.
Built: Life’s Day of the Dove, Nature’s Animal Taming, Animal Friendship and Summon Animal, Death’s Enslave Undead, the
GROW_MOLDmutation, and the Wand of Tame Monster. Restored: Chaos’s Summon Demon to one demon in three serving you, which DEC-53 recorded as a loss; theHYPN_GAZEmutation from a timed command to a real charm; andATT_DEMONto one in six arriving friendly. Spells shipping without an effect fell from 120 to 83.A summon arrives on its summoner’s side, which is why the upkeep is charged on the sum: a pet that summons builds you a stable you are paying for.
Two things stayed deferred with their reasons corrected rather than carried. Raise the Dead animates the corpses actually lying on the floor, and 4.2 has no corpse objects — “summon undead as pets” would be a different spell wearing its name. The magical figurine summons the monster named in it, and 4.2 has no object whose name interpolates a monster race; it is re-filed with the ten statues, since a figurine is a statue that does something.
Still open from PLR-28: Mindcrafter domination and Chaos patron gifts.
3.71.0 — Turning on your own (M10 phase 5; PLR-24, PLR-33, DEC-62). Anything the player does that hurts an ally turns it hostile, and carrying something that aggravates turns them all. It costs virtue: Individualism up, Honour, Justice and Compassion down — Zangband’s own four writes, dead numbers there and live ones here.
The anger goes in one place.
mon_take_hit()is the player-caused damage entry point in 4.2 — melee, missiles and every projection reach it, while monster-caused damage goes throughmon_take_nonplayer_hit()— so a pet cannot blame the player for something else’s fireball. Zangband had to write its call at seven sites and missed some.PLR-24 asked for a confirmation prompt; Zangband does something better. Walking into an ally changes places with it. The danger was never that a player decides to punch their own animal, it is that the animal steps into the doorway on the turn the player was walking through it — and since pets follow you, a prompt on that step would appear constantly and be answered unread. The exceptions are Zangband’s: confused, hallucinating, stunned, berserk, or unable to see it.
Aggravation is checked in the monster’s turn rather than in
monster_reduce_sleep(), where 4.2 reads the flag — no ally is ever asleep since 3.68.0, so the rule would have been unreachable for exactly the monsters it is about.3.70.0 — Pets cost mana, and earn you nothing (M10 phase 4; PLR-30, PLR-31, DEC-61). The two halves of pet balance, which Zangband’s own documentation says are the whole of it.
A count is free —
1 + level/pet-upkeep-divpets, the divider now class data at Zangband’s three values (20, 15 for the Mage, 12 for the High-Mage; the three 4.2 classes take their DEC-55 donor’s). Past that count the sum of the pets’ levels is the percentage of mana regeneration withheld, clamped to 5..95 — for the whole stable, not for the excess. That cliff is the balancing pressure: two pets inside the allowance are free and a third is charged for all three.The weight is the monster’s level, and that was measured rather than assumed. Zangband’s code adds
hdice * 2where its documentation says “the sum of the levels”; across its 883 monsters those are equal for 48% exactly, median difference zero, within two for 96% — the same number written twice.A Blackguard is exempt, deliberately. Its mana regeneration is negative, so applying the factor would make a stable of pets slow the burn down and pay the player for what the mechanism charges for. Measured: 15 spell points burnt over a hundred turns alone, 1 with the factor applied. Zangband had no such class, so this is our call and DEC-61 records it.
PLR-31 turned out to be already true and is now proved: monster-versus-monster kills award no experience and leave uniques alive, so a pet can neither farm nor clear a vault.
The sanitizer pass found a real use-after-free on the way through, of the same shape as the four fixed in 3.60.x: the stacked monster-message queue holds race pointers,
cleanup_angband()frees the races, and nothing emptied the queue — so a message left over from one character was read after the races were gone when the next character flushed it. Cleared on cleanup now.3.69.0 — Pets take orders (M10 phase 3; PLR-25, PLR-26, DEC-60). The pet menu on
A: five leash lengths, two switches, a roster and a dismissal. Zangband usedp, which is auto-explore here.The requirement calls these “nine command modes with per-mode distance behaviour”; they are nine entries of which five are distance modes. All nine are built. The orders are one policy for every pet at once, which is how Zangband holds them, and they survive a save in player block version 6 — the leash written signed, because its sign is its meaning and reading it unsigned turns “stay twenty-five squares away” into a leash of 65511.
Pets do not follow you downstairs, and that is deliberate. PLR-26 asks for it; Zangband does not do it. There is no pet-carrying code anywhere in 2.7.5 — the mechanism Hengband later added for this is absent from both archived lineages — and its documentation, which covers the upkeep, the killing-blow rule and every way of obtaining a pet, never mentions taking one with you. A pet is a per-level asset, which is also most of the answer to whether a summoner trivialises the game. A test pins it, so reversing it is a deliberate act. Levels kept under the persistent-levels option do keep their pets, which has been true since 3.67.0.
3.68.0 — Allies fight (M10 phase 2; PLR-23, DEC-59). A monster on the player’s side now picks its own enemy, stays awake while it is nowhere near the player, casts at what it is fighting, and comes back when it drifts off the leash.
Most of the machinery was already in the tree. 4.2 carries a complete monster-versus-monster path built for the Necromancer’s command power —
monster_attack_monster()for blows,mon_take_nonplayer_hit()for the damage, and monster-aware targeting in the spell code — so phase 2 reuses it and writes none of it. That also means PLR-31 is nearly already true:mon_take_nonplayer_hit()awards no experience and leaves uniques at one hit point, so a pet can neither farm experience nor finish a unique.Allies get their own branch of
get_move()rather than a swapped target, because everything that function does afterwards is about the player: the noise and scent heatmaps flow from the player’s grid, the flee logic runs away from it, and the pack AI works to surround it.One ordering choice is load-bearing: the enemy check comes before the push-past check, because
monster_can_kill()lets aKILL_BODYmonster walk over a weaker one and delete it. Without that, a pet standing between the player and something large would vanish with no blows struck.The three standing orders pets will obey — leash length, doors, picking things up — arrive with the player block at version 6. Older characters load with Zangband’s defaults: follow me, doors shut, hands off.
3.67.1 — A wilderness test stops depending on the world it drew.
game/wild’sa-towns-services-are-builtwalked the generated world and ended withrequire(checked > 1)— a floor on how many towns the world gave services to, which is a random draw. Measured across thirty worlds it ran from two towns to twelve, and a world with one or none is legal, so the suite failed about one run in twenty and said nothing about the code when it did. Found byscripts/check-flakes, one release after it was written.Replaced with two deterministic tests rather than a looser floor.
town_gen_wild()takes the service mask as an argument, so the masks can be chosen: each of the six services alone, then all six together, then none — that last being the case a random world almost never supplies and the only one that catches a generator building something it was not asked for. Falsified both ways, by making the generator ignore the mask and by making it build nothing.3.67.0 — A monster can be on your side (M10 phase 1; PLR-22, PLR-27, PLR-29, DEC-58). Angband 4.2 has no such notion — searching its source for friendly finds a comment and a shopkeeper’s greeting — so this is a change to an invariant the monster subsystem is built on rather than a feature on top of it, and it lands on its own before anything depends on it.
Three states, not a flag. Hostile, friendly and pet: a friendly monster will not attack you and promises nothing else, where a pet takes orders and will cost mana upkeep. Zangband carried this as two bits inside its smart-learn bitfield, and neither setter cleared the other, so a monster could be both and behaved as whichever test ran first — a released pet went on taking orders. One enum field cannot reach that state, and hostile is zero so no creation path has to remember anything.
Monsters can be enemies of each other. Alignment is checked before sides, which means two of your own pets will fight if one is good and the other evil. That needed
RF_GOOD, which 4.2 does not have: added, with the seven imported races Zangband flags. It also turned up a contradiction — 4.2’sdragonbase template carriesEVIL, so the law drake and the Great Wyrm of Law came out good and evil, which is Balance rather than Law. The balance drake and the Great Wyrm of Balance keep both flags, because Zangband gives them both on purpose.Looking at a monster says whose side it is on, first — before its wounds, where Zangband put it last and after the recall prompt. The monster list counts pets the way it counts sleepers. The glyph is not recoloured, which is also what Zangband did: its pet colours belonged to the Tk client.
Saved games keep it. The monster record gained a byte, so the
monstersandchunksblocks both went to version 2 with version 1 loaders; older characters open with every monster hostile, which is what they were. A test reads the block headers out of a saved file and insists on the new version, because the failure that matters is a forgotten version number rather than a lost byte.Nothing makes a pet yet. That is phase 6; for now the states are reachable through a wizard command added with them.
Balance — 2 September 2026#
3.66.1 — A harness for flakes, taking its list from the build rather than from disk.
scripts/check-flakes -n 12runs every suite twelve times and reports anything that was not the same every time — the suite, the pass, and the individual test, obtained by asking the short suite again with-v. A single green run says the tests passed once; it does not say they are deterministic, andgame/wild’sa-refused-power-is-freefailed one run in sixteen for two releases while every gate stayed green.The list of suites comes from
ninja -t commands allunittests. The throwaway script this replaces globbedbuild-*/unittests/*/*/*, which picked up a diagnostic suite that had been written, run and deleted — its binary outlived its build rule, and its three tests were counted into two totals reported here. Anything on disk the build does not know about is now named as an orphan and left out of the count. Current figure: 1246 tests, and twelve consecutive whole-suite passes clean, 14,952 test runs.3.66.0 — Every entitlement Zangband gives is carried over (CNT-10, PLR-08, DEC-57). Zangband’s
realm_choices1/2[]were transcribed verbatim from the start — a test has asserted that since 3.53.0 and never failed — but the books for a realm only ever reached classes that already carried that kind of book. So five classes could choose realms with nothing behind them: a Mage was allowed six realms and had three, a Priest likewise, and the Rogue, Ranger and Paladin were each short of Death or Arcane.Ten class-realm pairs, forty books, 305 spells. It was an artefact of the rollout rather than a decision — the three new realms were appended to their entitled classes and the four mapped ones replaced in place, and replacement only touches a class that already has those books. The tell was the asymmetry: the Warrior-Mage and High-Mage, emitted after all six realms existed, had twenty-four books, and the Mage with the same entitlement had twelve. The project owner chose fidelity to the tables.
A Mage that can study Death is not Angband’s Mage. Healing at one end and Hellfire at the other; a Priest may take Arcane. It is a larger change to how the game plays than any single realm import, taken for fidelity rather than for balance, and BAL-14’s dials are the answer if playtest disagrees.
The new books go on the end of each class’s list, so every spell index that already existed still means what it meant — the Mage’s Chaos books are still at flat index 64. The savefile fingerprint covers the whole list, so it changes: a saved caster with spells recorded is refused rather than silently rebound, and one with none loads unaffected. The corpus needed no change, because all thirty-five files already fail a step earlier on the prayer-book kind DEC-50 deleted.
Three things checked and found not to move: the chosen-versus-random gate is keyed on the class and never on the realm, so a Priest gaining Arcane still learns at random while a Mage gaining Life chooses its prayers; no deferral is class-specific, so widening who carries a realm widened the deferrals with it, 87 to 120; and seven spells sit above level 50 on the Rogue’s and Ranger’s own figures, which is
usable()doing Zangband’s arithmetic.One consequence beyond the spell lists. A Mage’s pack now sorts a nature book and a prayer book as readable, because it can read them — and since readable books come first and then sort by decreasing tval, the magic book comes last of the three.
player/calc-inventorycaught it; the new order was derived from the rule before the test was re-run.3.65.1 — A power paid for in blood is no longer also penalised for it (DEC-56). Racial and mutation powers charged 5% failure per point of mana the character was short — but the condition that triggered it,
csp < cost, is character for character the same condition that decides the price comes out of hit points. So the surcharge fell on exactly the people already paying blood, and on nobody else. They were charged twice for one shortfall.A Draconian’s breath, cost 25, at the level it arrives: 95% to fail on an empty pool, now 7% — the same as with a full one. Worth saying that the old number was not a permanent 95: the level term ate into it, so the power came good around level 30. An earlier note of mine claimed a Warrior “still cannot breathe”, which was that one figure generalised too far.
Spells keep their penalty, and that is not an inconsistency. A spell is not paid for in blood; it is paid for in unconsciousness —
over_exert()faints the caster and may take a point of constitution, after a confirmation prompt. Blood is a price, fainting is a risk, and neither needs a surcharge on top.The blood price itself is unchanged: 13 to 25 hit points for a cost-25 power, and still refused outright below that. The two classes with no mana at all — the Warrior and the Mindcrafter — are the ones this was blocking.
M9: magic realms — 2 September 2026#
3.65.0 — The Monk casts, and a gate the field names lie about (PLR-03, PLR-08, DEC-36). M7 built the Monk and deferred its casting: realm selection did not exist yet. It does, so the Monk now takes one of Life, Nature or Death — twelve books, ninety-six spells, on its own row of
magic_info[], so DEC-55’s donor rule never touches it. DEC-36’s prose said “choosing between Life and Nature” and was one realm short;realm_choices1[]includes Death, andrealm_choices2[]gives the Monk nothing, so it picks one of three and has no second slot.Punching and casting coexist, and they meet in the armour. A Monk is the only class that is both
MARTIAL_ARTSand a caster, and Zangband weighs the same six slots twice: at100 + level * 4for the martial penalty and atspell_weightfor the mana one. The Monk’sspell_weightis 300 — the Mage’s figure, not the Priest’s 350 — so a Monk in heavy armour loses its unarmed bonuses and its spell points together. Both halves are asserted, because a class where one quietly disables the other is the failure worth testing for.A Monk learns its spells at random, like a Priest — and that is the thing no field name advertises. Zangband has a
spell_typefield that looks exactly like the chosen-versus-random switch and nothing reads it; the real gate is the class’s ownspell_bookconstant, which staysTV_LIFE_BOOKwhichever realm the Monk picks. So the Monk does not getCHOOSE_SPELLS, which is the opposite of what treating it as “just another caster” would have done. Applied to the six classes both games share, the rule agrees with Angband’s own convention six times out of six.It also found a defect in a class shipped six versions ago. The Chaos-Warrior’s constant is
TV_SORCERY_BOOK, so it chooses its spells, and 3.58.0 gave it books without the flag — it has been learning Chaos at random ever since. Fixed here.Three things checked and found not to apply: Zangband’s glove penalty is gated on
TV_SORCERY_BOOKso a Monk never had one (and 4.2 has none at all); none of the twenty-nineCLASS_MONKsites touches mana or failure; and Zangband’s own study command filters books by the class constant, which means a Zangband Monk of Nature cannot select its own books — a bug in the original, not carried across, because books here are matched by realm.Documentation and plan corrections from the 2 September audit: four doc claims that had gone stale (three realms described as still holding Angband’s spells; a chapter note contradicting its own chapter; a mutation counted as broken that works), two plan lines still reading as open decisions after they were settled, CNT-10 phrased honestly as met for six realms of seven with Trump deferred rather than as a tick, and the mana-shortfall penalty written into the plan as the one thing outstanding for the project owner — it had been recorded only in the options chapter since 3.49.3, which is why it kept being reported as “nothing outstanding”.
3.64.1 — The Windows CI failure was a missing directory, and the gate now builds the way that job builds.
game/savesfailed two Windows CI runs, 0/1, while Linux and macOS were green — and it was neither Windows nor the sanitizers.The savefile corpus is staged into the unit tests’ working directory, and that staging hung off
SUPPORT_TEST_FRONTEND— an option about the end-to-end test front end, which has nothing to do with whether the unit tests can find their data. It defaults to OFF;linux.yamlpasses it andmsys2.yamldoes not, in all three of its jobs that run the tests. So the suite was looking fortests/savesin a directory that did not exist, and failing on the assertion written for exactly that (“absent rather than empty is a real answer”). Staging now hangs offallunittests, so any configuration that can run the unit tests can feed them.Direction of the failure: neither. No save loaded that should not have, and none failed that should not have. Reproduced on macOS by configuring as that job does, and confirmed by staging the corpus into that same build, where the whole suite passes 1242/1242 under ASAN+UBSAN with no sanitizer reports at all.
The corpus is now one test per savefile — 38 rather than 1. A tally of “0/1” was why two CI runs could not say which file had broken or in which direction;
game/saves 37/38names it, and an absent corpus says so in words and reports the path it looked in. Two invariants came out of writing it: a manifest entry naming a file that is not in the corpus used to pass silently, and nothing checked that every file had actually been offered to a test.scripts/check-buildgained a sixth pass that reproduces that CI job — Debug, ASAN+UBSAN, and no-DSUPPORT_*options at all. The missing options are the pass, not an oversight; the sanitizers are worth having on their own but would not have caught this. And.gitattributesnow pins the corpus as binary, which git’s heuristic already got right, because the corpus is the one place a silent byte-for-byte change would go unnoticed until a character failed to load.3.64.0 — The Warrior-Mage and the High-Mage, and PLR-03 is closed (PLR-03, DEC-36). M7 built three of Zangband’s five extra classes and deferred these two, because each is defined by which realms it may choose and realm choice did not exist yet. Every class in PLR-03 is now playable.
A Warrior-Mage studies Arcane in its first slot, always, and anything it likes in the second — the generalist realm plus one specialism, on a class that fights respectably and levels slowly because it pays for both halves. A High-Mage chooses one realm out of seven and gets no second slot at all, and Zangband’s table pays it for that: it reaches a realm’s deepest spells earlier and for less mana than any other caster, on the worst hit dice in the game. That difference lives entirely in two
realm-choice:lines, so that is whatplayer/realmasserts, both directions.Twenty-four books apiece, over the six realms that have content. Both are entitled to Trump and neither is offered it, because it has no books (DEC-54).
The skill conversion is DEC-36’s, re-run rather than re-judged. That mapping was measured against the six classes both games share and recorded as a table of rules — save increment × 0.42, search base × 0.62, melee × 3.1 and its increment × 0.56, shoot × 3.45 — and applying it here reproduces the Monk, the Mindcrafter and the Chaos-Warrior exactly, field for field, which is what made it safe to apply to two classes nobody had converted.
3.63.0 — Death is Zangband’s, and DEC-50 is finished (CNT-10, DEC-50, DEC-55). Thirty-two spells in four books — Black Prayers, Black Mass, Black Channels, the Necronomicon — replacing the five shadow books the Necromancer carried and the three the Blackguard had. All four of the realms Angband already had now hold Zangband’s spells, and every realm in the game is four books of eight, which is Zangband’s shape throughout.
Both of Death’s classes cast on borrowed figures, because Zangband had neither: the Necromancer on Zangband’s Mage, the Blackguard on its Paladin (DEC-55).
A miscast Death spell hurts, which 3.59.0 left out on purpose — it changes the balance of a class whose content had not been replaced yet, and now it has.
(book + 2)d6on a roll against the spell’s depth in the realm, plus a one-in-six chance of losing experience from the realm’s second half unless you have Hold Life. Out of the Necronomicon it is worse half the time: a saving throw against confusion and hallucination, and failing the next one instead costs a point of intelligence and a point of wisdom. It is the only spell failure in the game that can permanently cost a statistic.Four deferred: Enslave Undead and Raise the Dead need a creature able to take your side; Wraithform needs an incorporeal player, which this game has no notion of; and Omnicide is a whole-level sweep with a running mana cost that would otherwise duplicate the realm’s own Mass Genocide.
GF_HELL_FIREis the last of DEC-53’s missing elements to matter. Zangband’s hell fire does double damage to good monsters, and 4.2 has a projection that favours evil and none that favours good, so Malediction and Hellfire useMANA: the property that decides how they play is that nothing resists them.NETHERwas the closer flavour and the worse spell — undead resist nether, and it would have handed the realm’s own quarry immunity to its biggest attacks.3.62.0 — Nature is Zangband’s, and the Druid casts on borrowed figures (CNT-10, DEC-50, DEC-55). Thirty-two spells in four books — Call of the Wild, Nature Mastery, Nature’s Gifts, Nature’s Wrath — replacing the five nature books the Druid carried and the two the Ranger borrowed from it. The Ranger now gets the whole realm and goes from ten books to twelve.
The Druid is the first class to cast on another class’s figures. Zangband never had a Druid, so
magic_info[]has no row for it, and DEC-50 replaces its realm anyway. DEC-55 settles that: a class Zangband never had takes the figures of the one it matches onspell_first,spell_weightand casting stat — the three constants that six of Angband’s nine casting classes match exactly and by name, which is what makes the key evidence rather than a guess. The Druid’s are 1 / 350 / WIS, which among the six Zangband classes carrying Nature is the Priest’s and only the Priest’s.Recorded as an obstacle “wanting deciding”; it turned out to be a derivation. The same rule gives the Necromancer Zangband’s Mage and the Blackguard its Paladin, which Death will use.
Five of the thirty-two are deferred, and three are one wall: Animal Taming, Summon Animal and Animal Friendship all need a creature able to take your side, which is what deferred Trump whole. The other two are Stone Tell, whole-object identification in a game that replaced it with runes, and Protect from Corrosion, which is an object property here rather than anything an effect reaches.
Two tool faults surfaced. A realm’s book region can hold a directive that is not about books — the Druid’s
equip:nature book:[Lesser Charms]sits between its first book’s properties and that book’s first spell, and cannot be moved out because anequip:line naming a book must follow thebook:directive that creates the object kind. Replacing the region wholesale deleted it, and the game then refused to start. Andextract_realm_blockran to the end of the class block, swallowing the banner comment that separates one class from the next and moving the Priest’s heading into the middle of the Druid’s spells.3.61.0 — Arcane is Zangband’s (CNT-10, DEC-50). Thirty-two spells in four books — Cantrips for Beginners, Minor Arcana, Major Arcana, the Manual of Mastery — replacing the five magic books the Mage and the Rogue carried. The second of the four realms Angband already had, after Life.
All four are sold in town, for 100, 250, 1000 and 2500 gold, which is the realm’s whole bargain and the half of the spoiler’s sentence that pays for the other half: Arcane “has no ultra-powerful high level spells” and “all Arcane spellbooks can be bought in town”. Emitted under the generic two-town rule, half the realm was unbuyable and the reason to choose it was quietly gone. The flag is not cosmetic —
store.cstocks every town book permanently and never a dungeon one.A Mage no longer learns its attacks from Arcane. One ball, one bolt and one beam in thirty-two spells. That is Zangband’s design, where a Mage takes two realms and the second is where the damage lives; Arcane plus Chaos is the classic pair. DEC-49’s consequence is therefore inverted, and its addendum says so rather than being edited to match.
Two deferred: Phlogiston, because refuelling a light is a command in 4.2 and not an effect, and Detect Enchantment, for the same reason Sorcery’s is.
Along the way,
apply_realm.py— the tool that swaps a realm’s books rather than appending them — learned to fix starting equipment. A class can begin the game holding a book by name, and the Rogue’sequip:magic book:[First Spells]outlived the book: the game refused to start with “unrecognized sval”. Nature and Death will each hit the same thing.3.60.2 — The gate checks the realm content, and a flake is fixed.
scripts/check-buildgained a fifth pass:zconv realms --checkasks whether every spell name inrealmmap.tomlmatches Zangband’s own table, in both directions, and whether every realm block inclass.txtis still what the converter produces for that class. This is the guard against the failure that put Sorcery’s Teleport into the game doing nothing — a name keyed wrongly yields a spell with a level, a mana cost, a failure rate and no effect, which parses, appears in the book, casts, and does nothing. With three realms and about 128 more spells still to import, noticing is not a plan. Proved by reintroducing the mis-key (caught) and by hand-editing one spell’s experience value inclass.txt(caught).The pass picks its interpreter rather than assuming one: the converter reads TOML with
tomllib, which is Python 3.11 and later, andpython3on the development machine is 3.9 — so taking whatever is onPATHmade the pass die on an import and say nothing about the realms. A missing interpreter is a failure with a recipe, not a skip.And ``game/wild``’s ``a-refused-power-is-free`` was failing about one seed in sixteen, on master before this work. It hand-set a spell-point pool and asserted an upper bound on what a power charged. Neither holds: the power’s effect projects,
project()recalculates when anything is pending, and a mind blast aimed nowhere in particular can come back on the caster — psi damage drains experience, which drops the level and shrinks both pools. On the failing seed the character had 43 spell points before a 12-point power and 2 after. The test now takes the pool the game computes, checks it can hold the price, and asserts only the lower bound, which survives because nothing gives mana back. The Draconian test beside it had already reached that conclusion for the same reason.3.60.1 — The savefile fence is live again, and three latent faults are fixed. DEC-50’s Life realm left every one of the thirty-five historical savefiles refused, so nothing proved a character could be saved and loaded at all — and the spell-list fingerprint, which is exactly the guard the three remaining realm replacements need, had stopped being exercised by anything: every corpus file now fails a step earlier on a missing object kind, so the fingerprint could have been broken outright with the suite still green.
game/roundtripbuilds its own characters instead. A Priest with realms chosen and prayers learned survives a save and load with the same spells at the same indices; the fingerprint refuses that character when a spell in the class list has been renamed underneath it, and accepts it when it has not. The rename is deliberate — it leavestotal_spellsalone, so the count check cannot catch it and only the fingerprint can.Making that possible needed three fixes, all the same defect — memory freed and left live — and all invisible while a process only ever starts the game once:
cleanup_angband()freed the data directory strings without clearing them, so the nextinit_file_paths()double-freed and libmalloc aborted.vformat_kill()freed theformat()buffer and left it live, so the nextformat()wrote into freed memory and the abort landed on some unrelatedfree()much later.calc_spells()read the spell arrays before they existed. Birth chooses a class and only initialises the arrays at the end, so between the two a literate class has none — and every class change recalculates bonuses. Upstream survived it by reading freed memory; clearing those pointers in 3.58.0 turned that into an honest NULL dereference.
Together they are why a unit test could not build a second character in one process, which was recorded in 3.60.0 as its own piece of work.
3.60.0 — Life is Zangband’s realm now (CNT-10, DEC-50). Thirty-two prayers in four books — Book of Common Prayer, High Mass, Book of the Unicorn, Blessings of the Grail — replacing the five Angband prayer books the Priest and the Paladin carried. The first replacement rather than an addition, and the first realm where 4.2 already had content.
Twelve of the thirty-two chains came from spells 4.2 already had, because Life is 4.2’s divine under DEC-49. Two are deferred: Day of the Dove charms, which needs monster allegiance and waits for M10 with Trump, and 4.2 has no effect that blesses a weapon. Two of 4.2’s own shortfalls are recorded rather than hidden — there is no DISP_DEMON projection, so Exorcism and Dispel Undead & Demons dispel undead by name and demons by being evil, which is wider than asked; and GLYPH places one glyph where Warding True laid a ring.
The Paladin gains the most: 4.2 gave it three of the Priest’s books and sixteen prayers, and Zangband’s table gives it all thirty-two, at much higher levels.
Every savefile in the corpus is now refused, and that is the accepted cost. A book’s object kind is created by its book: line, and a savefile names objects by tval and sval as text — so deleting [Novice’s Handbook] made every save holding one unreadable, and every save holds one, because every save carries the town temple’s stock. Measured before the change: the two town prayer books appear in 35 of 35 files, the three dungeon books in 0, 1 and 1. The choice between Zangband’s titles and keeping the two town ones was put to the project owner with those numbers; Zangband’s titles were chosen. The thirty-five files are listed in
tests/saves/EXPECTED-FAILURESwith the reason, and the suite still fails if one of them loads or if anything crashes rather than refusing. It no longer proves a character can be saved and loaded at all — a fresh corpus of played characters is needed for that, and is the next thing this suite wants.
M9: magic realms — 1 September 2026#
3.59.0 — Chaos spells backfire (CNT-10). A failed Chaos spell may produce a chaotic effect instead of merely failing, and how bad it is depends on how deep the spell was: Magic Missile never backfires, Call the Void almost always does. Zangband’s forty-band table, boundary for boundary, from a short teleport at the top to the Ancient and Foul Curse at the bottom. Only three bands are substituted — the six bizarre summon groups and the Cyberdemon do not exist here, and
wall_breaker()becomes aKILL_WALLsphere. Everything else was already an effect the game had,GAIN_MUTATIONfrom M8 andANCIENT_CURSEfrom M3 among them.The scaling is the part that could have been quietly wrong. A spell’s index counts across its whole class, so Chaos’s first spell is 0 for a Chaos-Warrior and 62 for a Mage — scaling by that would have looked correct for the one class and made every Chaos spell a Mage owns backfire as though it were the deepest in the game.
player/realmwalks all 218 spells of three classes asserting each realm restarts at zero.Death’s miscast penalty is deliberately not built with this: it would change an existing class’s balance before that class’s spell content is replaced, and belongs in the Death realm’s own commit. See DEC-53’s addendum.
3.58.0 — The Chaos-Warrior casts (PLR-03, CNT-10). Zangband is blunt about the class — “trained in Chaos magic. They are not interested in any other form of magic. They can learn every Chaos spell” — but until the Chaos realm existed there was nothing to give it, so it shipped declaring itself unable to hold spell points. It now has all four Chaos books and all thirty-two spells from level 2, and it is the only class in the game with exactly one realm and no choice about it. Its figures are its own out of
magic_info[], not the Mage’s.The savefile corpus is unchanged at 31 loaded and 4 refused: the version-2 spell block only checks its fingerprint when the character actually has spells recorded, and the Chaos-Warrior in the corpus has none.
Two latent faults surfaced doing it. Swapping
player->classin place without re-sizing the spell arrays segfaulted as soon as the class had spells — a patron test had been doing exactly that harmlessly for as long as the Chaos-Warrior had no books. Andplayer_spells_free()freed without clearing, so a second free aborted; upstream leaves those pointers dangling and it is harmless only while nothing frees twice.3.57.0 — Swap position and Sterilize work (PLR-16). The two mutation powers M8 deferred with the note “deferred to M9, not open-ended”, built where they said they would be. Eleven mutations did nothing; nine do now.
Swap position had to be an exchange rather than a move. 4.2’s nearest existing effect brings the monster to you and leaves you where you are, and
player_place()asserts the square it puts you in is empty — so the naive version does not play wrong, it aborts.monster_swap()is 4.2’s own primitive, run on every monster turn, and already handles one of the two grids holding the player. A monster that resists teleportation refuses outright, which is Zangband’s behaviour and is not how the same flag works against teleport-other.Sterilize pushes the level’s breeder count past the ceiling that gates breeding, which is what Zangband does and is temporary in both: breeding resumes once enough breeders die. The seventeen-to-thirty-four hit points it costs are a separate effect in the chain, so an effect that stops breeding is only that — which needed a
power-effect-msgdirective the power parsers had never grown, since without it a power that hurts its own user could only report “yourself”.3.56.1 — A realm you can choose is a realm you can read (PLR-08, CNT-10). Zangband’s entitlement table says a Mage may take Trump, and it is imported whole because it is the record of what Zangband permits. But this game gains a realm’s content one commit at a time, and the birth step was offering realms with nothing in them: choose one and the character is made, the realm is fixed, and the spell menu is empty for the rest of that character’s life with no way back. It was live for Sorcery between 3.54.2 and 3.55.0. The birth step now skips a realm the class has no book in.
Trump is deferred whole (DEC-54). Fifteen of its thirty-two spells are summons, and in Zangband every one of them turns on whether the creature that arrives is yours. 4.2 has no monster allegiance at all, so all fifteen would collapse onto the failure branch — Trump Cyberdemon would spend 60 mana to put a hostile Cyberdemon next to you. That is the spell inverted, not a weaker version of it, so the realm waits for pets rather than shipping half a book.
3.56.0 — Chaos is playable, and a spell that did nothing is fixed (CNT-10). Thirty-two more workings in four more books — the Sign of Chaos, Chaos Mastery, Chaos Channels and the Armageddon Tome — for the three classes entitled to the realm: Mage, Priest and Ranger. The Rogue is not entitled and did not grow, which is the entitlement table doing its job.
Nothing was deferred, which is not the same as nothing changed. 4.2 has no disintegration, no rocket and no radiation, and a spell’s blast radius is a fixed number here rather than one that grows with the caster. Six spells do something measurably different as a result, each recorded next to its own entry; the two worth knowing are that Call the Void no longer punishes you for casting it beside a wall, and that Summon Demon always calls a hostile one. See DEC-53.
A found match rather than a translation: 4.2’s
WONDEReffect is Zangband’s Wonder table band for band, down to theplev / 5added to the roll and the comment explaining it.Sorcery’s *Teleport* did nothing in 3.55.0 and now works. Its effect chain was keyed under the wrong name, so the converter emitted the spell with its level, mana and failure and no effect at all — indistinguishable from a deliberate deferral. Two things now catch that:
zconv realmsfails if any name inrealmmap.tomldoes not match the spell table, in either direction, andplayer/realmasserts that every effectless spell carries the deferral wording and that there are exactly sixteen of them.And a failing test now fails the build. The CMake runner set its exit code only when a suite died or its output was malformed, so a suite reporting 10/11 printed in red and exited 0 — and
scripts/check-build,make alltestsand CI all called that a pass. Found by a real failure sailing through the gate during this work. Upstream Angband has the same hole.3.55.0 — Sorcery is playable (CNT-10). Thirty-two workings in four books — the Beginner’s Handbook, the Master Sorcerer’s Handbook, Pattern Sorcery and the Grimoire of Power — emitted by
zconv realmsinto the four classes entitled to the realm: Mage, Priest, Rogue and Ranger. The numbers come frommagic_info[]per class, the effect chains from a newrealmmap.toml, and the two are kept apart on purpose so the Mage’s Sorcery and the Rogue’s are the same spells at different levels rather than two transcriptions that can drift. All four classes’ blocks reproduce from the converter.The books are appended, never inserted. A Mage’s five Arcane books are still books one to five with Sorcery’s as six to nine, so every spell index a saved character holds still means what it meant — and the corpus is unchanged at 31 loaded, 4 refused. Inserting would have shifted them, the game would have loaded, and the sheet would have looked reasonable.
Four of the thirty-two do nothing and say so: Identify True, Detect Enchantment, Self Knowledge and Explosive Rune, each needing a mechanism 4.2 removed or never had. They are in their books with their levels and costs, described as beyond what the game can express. A book with a hole in it and a note saying so is honest; a spell absent without explanation is not.
This is also the realm filter’s first real test. Until now no class carried books from two realms, so removing the filter failed nothing; a Priest of Life can now read prayer books and not Sorcery ones, and a Priest of Sorcery the other way round.
3.54.2 — Realm choice gates which books you can open, which is what makes the choice mean anything. A class carries the books of every realm it may study — a spell lives in a book which lives in a class, and there is nowhere else to put one — so the choice made at birth is what decides which of them the character can actually read.
The gate asks how many realms the class has books in, not how many it is entitled to, and that distinction was a live regression rather than a nicety: a Priest may study Life or Death and every Priest book in the game is Life, so filtering on the entitlement left a Priest who chose Death unable to read anything at all. The gate is therefore inert until a class carries books from two realms, which is exactly when Sorcery arrives.
Honest about what the tests cover: a Priest who chose Death is proven to still read their prayer books, and gating on the entitlement instead fails that. But removing the filter fails nothing yet, because no class in the data carries two realms’ books for it to sort. That test comes with Sorcery’s content.
3.54.1 — FETCH, built once for three callers, and telekinesis works. Zangband’s
fetch()is called by the telekinesis mutation, by Sorcery’s Telekinesis spell and by the Trump realm, and the three differ in exactly two ways — so both are parameters rather than being levelled away. The mutation liftslevel × 10and needs line of sight; Sorcery’s liftslevel × 15and does not, which is what makes Sorcery’s the better spell.The rest of the rules are Zangband’s and each of them is there because leaving one out turns a spell with texture into one that just moves things: it refuses while you are standing on something, refuses artifacts — “The object seems to have a will of its own!” — has a weight limit, and given a direction rather than a target walks outward until it finds an object and stops at a wall. One rule is not carried across and is recorded rather than dropped quietly: Zangband refuses to fetch out of a vault, and 4.2 marks vault squares differently enough that the equivalent test would be an invention.
Ten mutations now do nothing, six activatable and four random.
3.54.0 — The Midas touch comes back (DEC-52, reversing DEC-48). Turning an object into gold was dropped because the mechanic “would have exactly one consumer and no prospect of a second”. It has three:
alchemy()is called by the Midas touch, by Sorcery’s Alchemy spell, and by a random-artifact activation Zangband priced at 10,000. So it is built once, for all three, the same wayFETCHwill be.The numbers are Zangband’s, taken from the source. A third of the object’s real value — what a sale prices at, not what the character believes it is worth. Divided before the quantity multiplies it, so ten objects worth two gold each pay nothing rather than six; that is integer division doing what it does. Capped at 30,000. Artifacts refuse, because one-of-a-kind items are not currency. And a worthless object becomes fool’s gold — destroyed, nothing paid.
One deliberate divergence: Zangband multiplies in a signed 32-bit value and caps afterwards, so a stack valuable enough wraps negative and pays nothing. It needs a cost above about sixty-five million to reach and Zangband’s data has none, so the bug is latent there rather than live — but a cap that can be jumped is not a cap. Found by a test written to pin the cap, which failed on it.
Eleven mutations now do nothing, all deferred and none refused, and the activatable split is 25 working to 7. DEC-48 is kept in the record with a reversal notice rather than deleted, because the reason it gave turned out to be wrong and that is the part worth being able to read.
3.53.1 — Books for the three new realms, and a stop on DEC-52. A
sorcery book, achaos bookand adeck— Trump’s book-noun is a deck, because that is what Trump magic is. An object records its base by name rather than by the tval list’s numbering, so adding these mid-list is safe. A book’s kind is created byclass.txtitself, so no object needed writing.The spell import stops on DEC-52, which is open. Sorcery’s Alchemy needs the object-to-gold mechanic that DEC-48 refused for the Midas touch, and DEC-48’s stated reason — that the mechanic would have exactly one consumer — is wrong:
alchemy()has three callers in Zangband, one of them a realm spell inside this milestone. Building it would reinstate something explicitly refused; skipping it would leave Sorcery a spell short with no record. Either way it is not a decision to make in passing.Also recorded: the three new realms have to be imported before the four mapped ones. DEC-50 replaces Arcane, Life, Nature and Death outright, and a spell with no effect chain does nothing — so replacing a working realm before its nineteen-or-so hand-written chains are finished would leave every existing caster holding books that cast nothing. A new realm costs only whoever chose it.
3.53.0 — Realms are chosen at birth (PLR-08, PLR-11). A fourth step after race and class, offering the realms the class allows — which is what makes the combination mean anything. The entitlements are Zangband’s own, out of
realm_choices1[]andrealm_choices2[], andplayer/realmchecks the shipped data against that table across all 8 classes × 2 slots × 7 realms rather than against transcribed values.A slot offering fewer than two realms is not asked about, because a list of one is an entitlement and not a choice: a Ranger’s first realm is Nature and the question would have one answer. Angband’s Druid, Necromancer and Blackguard have no Zangband counterpart, so each is entitled to the one realm it already studies — which keeps what they can cast exactly as it was.
The choice is written to the savefile by name and by count, as the patron and the virtues are, so inserting a realm cannot rebind a saved character to a different one. The
playerblock goes to version 5 and the version 4 reader is kept; a save written before realms existed comes back studying whatever its class studies, which for every caster in the corpus is what it was studying anyway. The corpus is unchanged at 31 loaded, 4 refused, so the version bump broke nothing incidentally.Two things worth recording. The realm list arrives back-to-front from the parser, so
ridx0 was Trump and every list shown to the player would have run backwards; it is reversed into file order at load. And the character panel asserts rather than growing when given one line too many, so adding the realms line aborted the game — caught by the birth frontend test, which is the pass that exists for exactly that.3.52.1 — The spell converter, and a correction to the order M9 goes in. No game change.
zconv.py realmsslices Zangband’smagic_info[]by position — 2,464{level, mana}pairs, exactly 11 classes × 7 realms × 32 spells — and derives the two fields 4.2 stores and Zangband does not:sfailfrom the base Zangband computes at cast time,sexpas5 × book²(DEC-51). It prints ready-madespell:lines, and re-derives the per-class realm entitlements from the table as its own cross-check. 1,376 lines across every class and realm, none breaking either rule.PLR-08 has to come before CNT-10, which is the reverse of the order they are listed in. A spell in 4.2 lives inside a
book:inside a class — there is nowhere else to put one. Sorcery, Chaos and Trump are realms no existing class carries, and DEC-49 deliberately pinned every existing class’s progression, so until a character can choose a realm at birth an imported Sorcery spell has no class to belong to. Realm choice first, then the import.3.52.0 — The savefile guard for the realm import (DEC-50). No game change yet; this is the fence built before the work that needs it.
spell_flags[]andspell_order[]are recorded by flat index across all of a class’s books, so what index 7 means is decided entirely by the data files. DEC-50 replaces the spell content of four realms outright — which means an old character would load successfully, with the right number of known spells and the wrong spells. The sheet would look reasonable and the game would play wrong.The
player spellsblock goes to version 2 and now carries a fingerprint of the class’s spell list. On load it is compared, and a character whose spells were learned against a different list is refused with a message saying so. The version 1 reader refuses too — but only when there is something to get wrong: a character with no spells recorded, which is every Warrior and every caster who has not learned one, is unaffected and loads as before.The fingerprint is not written for this one occasion. Every later change to a spell list is caught by the game rather than remembered by whoever makes it.
Of the 35 characters in the test corpus, four refuse and thirty-one load. The four are listed in
tests/saves/EXPECTED-FAILURESwith their reason, and the suite fails both if an unlisted file breaks and if a listed one starts loading again — a manifest that outlives its reason would silently excuse the next real break.
M9: magic realms — 1 September 2026#
3.51.1 — M9 stops on two questions, with the groundwork recorded. No game change. Phase 2 of the realms — importing the spellbooks — is not built, because CNT-10 and PLR-12 cannot both be taken literally once four of the seven realms already hold Angband’s own spells. Both readings are written up as DEC-50, with what each costs: 96 spells and no savefile risk, against ~224 spells and a migration for every casting class.
DEC-51 records the spell-experience question and corrects an earlier report of it. It was flagged as an eight-fold scale gap on the basis that Angband’s
sexpruns 2 to 10; it runs 2 to 300, and the earlier figure came from looking only at low-level spells. Measured properly against Angband’s median at the same spell levels, Zangband’s values cluster on 1.0 and diverge in both directions — 0.33 to 5.6 — because Zangband ties the award to the book and Angband to the spell. A balance note, not a conversion problem.Phase 2’s conversion rules are recorded so they are not derived twice. Both fields Angband needs and Zangband does not store come out exactly:
sfailbecause the two failure formulas are structurally identical after the base, andsexpbecause both games multiply by the spell’s level. The source table is parsed positionally — 2,464{level, mana}pairs, exactly 11 classes × 7 realms × 32 — and that slicing is cross-checked against the independentrealm_choicestables with zero mismatches across all 77 class-realm combinations.3.51.0 — Seven realms (PLR-09, PLR-10, DEC-49). Angband has four realms and Zangband has seven, and the two lists overlap by name without overlapping by meaning. The four are mapped onto four of the seven by content —
arcanekeeps its name,divinebecomes Life,shadowbecomes Death,naturekeeps its name — and Sorcery, Chaos and Trump are added. Exactly seven, with all of Angband’s metadata kept: casting stat, verb, spell noun and book noun, which Zangband held as scattered per-class constants.Nothing an existing character can cast has changed. Realms are not written to the savefile — books are matched by tval and sval — so the rename cost sixteen
book:lines and nothing else. Every class keeps the same books in the same order with the same spells at the same levels; the only difference is the internal name of the realm two book sets belong to, which no player ever sees.player/realmpins the shape of all eight casting classes so that no later phase of this milestone can disturb one quietly: inserting a book anywhere but the end of a class’s list shifts every spell a saved character knows one place along, and the game would load and look reasonable.This game’s Arcane is stronger than Zangband’s, and that is the one place the mapping is not clean. Zangband’s Arcane was the deliberate weak generalist with every book buyable in town and no high-level power; Angband’s is the Mage’s realm with two books of attacks and a dungeon-only Tome of Power. Having exactly seven realms requires folding them, and folding keeps the stronger.
Also here: the third flake of the session, identified and derived rather than nudged.
a-borrowed-lord-is-kindercompared how often a sworn and a borrowed Chaos patron reach the cruel end of the reward ladder, over four thousand rolls, and required a ratio above 1.5. The bottom quarter of the ladder is rare enough that four thousand rolls give only 74 to 99 events, so the ratio ranged 1.475 to 2.284 and the bound sat one and a half standard deviations out. It failed about one whole-suite run in fifteen, and the run that caught it named its own seed — 1604416127 — which is what the runner change in 3.50.3 was for. Ten times the rolls puts the bound six standard deviations clear, and the arithmetic is in the comment so the next reader does not have to redo it.
Testing — 31 August 2026#
3.50.4 — One bad crawl, and two wrong accounts of it. Every archived citation in the plans and docs re-checked one at a time — thirty-six of them — before M9 begins, because DEC-16 makes those documents a primary source and the realm spell lists are M9’s immediate input.
Twelve were dead, and they have one thing in common. All twelve carried the timestamp
20220527225941, and all twelve are spoilers. The 2022-04-20 crawl ofwww.zangband.orgcaptured everything; the 2022-05-27 crawl of the bare domain captured the homepage and almost nothing under it. All twelve are whole at20220420164xxxand have been repointed.There is no parking page, and
spoilers/life.txtis not unarchived. Both were recorded here and neither is true: fetched withid_, the 2022-05-27 root is the genuine zangband.org homepage, and life.txt returns 13,854 bytes — the complete Life realm spell list — at the timestamp already cited for it. It had been written off on a bad availability check.No requirement rests on a document nobody read. The four that cite a spoiler as their basis were checked against the recovered text and all four match: nightmare mode is the irreversible birth option BAL-15…17 describes, the Ancient and Foul Curse is CNT-15’s, random object abilities are CNT-16’s, and DEC-38 names all sixteen of Zangband’s patrons correctly before replacing them with nine Lords of the Courts. The citations were stale; the reading was not.
A note on method, because the first pass got it wrong. A throttled fetch from archive.org returns zero bytes and looks exactly like an absent document: the first sweep declared twenty-one citations dead, and retrying each failure three times reduced that to twelve. Nine documents were nearly rewritten out of the plan because the archive was busy.
3.50.3 — The same mistake, in the test written to replace it. The fourteen-run check on 3.50.2 flaked twice, and it was neither the code nor the doorstep: it was
a-character-at-sea-still-finds-ground, added one release earlier, assertingrestored > 200. Eight rings around the character are 289 squares only when the character is not within eight squares of the edge of the level; nearer the edge the out-of-bounds squares are skipped and the count comes in under the floor for a reason that has nothing wrong with it. A constant standing in for an invariant — exactly what 3.50.2 removed from the doorstep — put back by the test that removed it. It now asserts that everything flooded was restored, which is the actual invariant and has no number in it.And the seed the runner threw away. 3.50.2 gave every sampling suite a seed line so a flaky failure could be replayed, and it did not work: a suite returns zero whether or not its tests passed, so
run_tests.cmaketook the success branch and discarded the output — seed included — unlessVERBOSEhappened to be set. That is precisely the case the line was added for. The runner now echoes a suite’s output whenever it reports fewer passes than tests, and the next failure named its seed on the first try: 936496483.Sixteen consecutive whole-suite runs clean, which is the claim 3.50.2 should have waited for.
3.50.2 — A bound set from eight worlds, and the seed that found the ninth.
the-doorstep-is-survivablerequired the mean wilderness danger within six blocks of the starting town to be at most 12 — a figure taken from eight worlds where it came out between 1 and 6. Eight was not enough: on seed 1829551357 it is 14, and the test failed about one whole-suite run in eight for six releases.The bound was not wrong by a little, it was measuring the wrong thing. On that same seed the country beyond twelve blocks averages 27, so the doorstep is less than half as dangerous as the wilderness proper and the world is behaving exactly as intended. What the requirement says is that danger climbs as you walk out — a comparison between two parts of one world, true by construction because towns are placed in lawful country and law is what danger reads. It has no constant in it to be set from too small a sample. Near 14/7/1/1/1 against far 27 every time, across five worlds including the one that broke the old bound; and measuring around an arbitrary block instead of the town fails it on most worlds, so it still discriminates.
Every sampling suite now reports its seed. Only
game/wilddid, which is why its flake took one command to reproduce once the seed was in hand and the one inobject/importedwas never identified at all. Seven suites gainedtest_seed_rng_reported(), printed unconditionally rather than under-v, because a flaky failure in CI is exactly the case where nobody thought to ask for verbose output first.The rest of the audit found no other test at risk.
cave/wild’s terrain composition looks like the same shape and is not — it aggregates twenty fixed seeds and cannot vary. The remaining sampled bounds are all far from their means: weird luck’s out-of-depth rate is 8.7 standard deviations clear, a Vampire’s hypnotic gaze 6.5, the patron notice rate 11.9, and the cross-habitat rate measures 93–95 against a bound of 70. Those were derived rather than guessed, and two of them say so in their own comments.3.50.1 — The Midas touch is dropped (DEC-48). Turning objects into gold needs an effect that destroys an object for money, which 4.2 has not got — and building one means settling what fraction, of what valuation, and what that does to the shops pricing against the same numbers. That is an economy mechanic with exactly one consumer, for a mutation that saves the player a walk back to town.
Recorded as a rejection and not a deferral, so it is not picked up again by someone reading the deferred list as a queue.
mutmap.tomlgained arejectkey besidedeferfor it — the same splitobjmap.tomlalready had — and the conversion report lists it under REJECTED.The consequence reaches the screen: the power menu now shows six of its unusable entries as not yet and this one as dropped, and says so when it is chosen. Not yet is a promise, and it is not one this game intends to keep. The mutation is still gained, described and saved, on DEC-44’s reasoning that a mutation vanishing from a savefile is worse than one that does nothing.
3.50.0 — Closing M8: the clearing form, and six brackets that lied. Four items the audit turned up after the milestone was declared complete.
The converter read only half of Zangband’s flag syntax. SET_FLAG(of_ptr, TR_X) has a counterpart written flags[n] &= ~(TRn_X), with the word index baked into both the subscript and the macro name so it matches nothing the parser looked for. Three mutations use it and all three were kinder here than in Zangband: rotting flesh is meant to stop a character regenerating, and the panic-hit and warning mutations to stop them resisting fear — whether that regeneration or resistance came from another mutation or from a ring they are wearing. Reading the clearing form across all seven parsed regions of the Zangband source turns up nothing else: it appears in
player_flags()and in one line its own authors had already commented out.Fixing the pattern exposed a second bug of the same kind. Both block regexes closed on a fixed indentation and were blind to brace depth, so the chaos-gift test — which sits one tab shallower than the mutation blocks around it — ran past its own closing brace and claimed a flag belonging to vampirism. Both now match the closing brace to the opening one. This is the third brace-blind read of this source; the shape is now shared whether or not a region happens to be uniform.
Three mutations stop ordinary food feeding you, which reached no data file and no code at all. A beak, a mouth that eats rock and a taste for blood all set Zangband’s
TR_CANT_EAT, whose eat command leaves such a character a twentieth of what they swallow. Scoped to edible things, which is Zangband’s own scope — a potion of Cure Light Wounds still nourishes as much as it ever did, and the test asserts that half too, because dropping the scope passes every other check.A mutation’s bracket is now generated from its effects. Six descriptions named some of what they do and not the rest, and the worst of them — a living computer brain — advertised four points of intelligence and wisdom and said nothing about the vulnerability to electricity. Curating six strings would have left the seventh to be found the same way, so the text is derived: Zangband’s sentence, and a bracket built from the entry’s own fields. It costs the occasional redundancy (completely fearless (immune to fear)) and buys never shipping a description that hides a downside.
And the flake this suite carried for six releases.
find_open_grid_near()ingame/wildsearched eight rings for a square that is empty and not damaging, andsquare_isdamaging()is true of deep water — of which this wilderness holds some 330,000 grids. A character generated in open ocean has no acceptable square in the 289 that eight rings cover, so two tests failed about one whole-suite run in eight and never once in two hundred runs of the suite alone. It falls back to the whole level now, which cannot fail on a level with one open square on it. A new test floods the surroundings deliberately, so the case is exercised every run rather than one in eight; removing the fallback fails it and both historic offenders.3.49.6 — The mutations page drew over the stat table, so
WISread asISandDEXvanished. The stat panel starts at column 42 and a mutation’s description runs to 57 characters; the two cannot share a screen. The page takes the whole screen now, which it needs anyway — a character can carry far more mutations than the sheet has rows.Found from a screenshot, not from a test. The check that passed had grepped for the description being present, which it was, on top of the stat labels.
With it: the page is in Screenshots, captured from the running game like the others (
ZTK_SHOT_MUTATE), and Features no longer lists mutations under Not yet — M8 landed.3.49.5 — The character sheet shows mutations. A third page, reached with
hfrom the sheet like the other two, listing what chaos has made of the character — or “Chaos has not touched you.”Reported from play: a Sprite Mindcrafter whose message log was announcing “You mutate!” and “Electricity starts running through you!” while the sheet said nothing at all. Mutations reached the character dump and nowhere on screen; the power menu lists the activatable ones, and the passive ones — most of the ninety-six — could only be seen by writing a file and reading it.
Two things went with it. Going backwards through the sheet used
(mode - 1) % INFO_SCREENS, which is-1in C rather than the last page, and worked by accident with two screens; with three it would have made the new page unreachable in that direction. And the dump listed mutations by walking the linked list, which is reverse file order despite the comment above it claiming file order — the list is built backwards and renumbered. The dump, the new page and the wizard menu now all read inmutation.txtorder, so they can be compared against each other.3.49.4 — Mutations can be granted by hand.
^AthenU, under Debug → Player: all ninety-six listed, the character’s own marked[*], select to toggle.Chaos hands out mutations and the player does not choose them, which makes everything that touches them hard to test on purpose. Polymorph Self is the worst case — you cannot ask for the mutation that grants it, you cannot see the passive ones you have (the power menu lists only activatables, and the full set reaches the character dump and nothing on screen), and the effect itself does nothing about 60% of the time by design.
Urather thanMbecause everyDbg*list shares one key namespace, andMis Write map in another submenu. That is an assert incmd_init(), so the clash is a crash before the title screen rather than a menu that quietly misbehaves — which is how it was found.3.49.3 — A cheat for powers that will not fire.
cheat_powers, on the=xscreen: every racial and mutation power succeeds, and the listing shows0% to failbecause the menu and the roll ask the same function.Reported from play: a level 21 Beastman Chaos-Warrior looking at Polymorph at
20 hp, 95% to fail. Worth recording why it is 95, because it is neither the difficulty (18) nor the level (21 against a power level of 18) — a character short of mana is charged 5% per point of the shortfall, and a class with no spell points is short by the entire cost. 20 × 5 = 100, which buries every other term and clamps to the ceiling.So the blood price that exists precisely so a spell-less class can use a power at all is cancelled by a penalty for needing it:
player_use_power()lets a Warrior pay 20 hit points, andplayer_power_chance()then makes it fail 95 times in 100. Whether that penalty should apply to a character paying in blood is a balance decision and has not been made — the test records the present behaviour and will notice when it changes.
M8: mutations — 31 August 2026#
3.49.2 — A shared helper that overflowed on its twenty-seventh call.
menu_dynamic_add_label()writes a row’s shortcut key into the caller’s label buffer at the row’s own index, and did so without knowing how long that buffer was. Every dynamic menu in the tree allocated it withstring_make(lower_case)— twenty-six letters and a terminator — so the twenty-seventh row overwrote the terminator and the twenty-eighth ran off the end of the allocation.This is an Angband defect, not one this port introduced, and it is latent upstream rather than live: an audit of all twenty dynamic menus found the largest upstream one is eleven rows, and every row of every fixed list supplies its own key, so the pre-filled letters are entirely overwritten and the buffer’s length never matters. Four menus in this game are driven by data instead — the magetower’s destinations (up to 48), the Chaos Tower’s and the power list’s mutations (up to 89), and the quest giver’s (capped at 16) — and three of those can pass twenty-six. The magetower had already hit it and fixed it locally; the other two were fixed by clamping in 3.49.0, which was the wrong place to be defensive.
The bound now lives on the helper.
struct menucarries how many bytes of its label buffer may be written,menu_dynamic_labels()is the one way to get such a buffer, and all twenty call sites use it. A menu that never asks for one has a bound of zero and cannot be written into at all, so the unsafe thing a future caller might do is now the thing that does nothing. Fifty-one keys are available rather than twenty-six, fromall_letterslessq— skipped because these menus end with a “no thank you” row keyedqand the selection scan takes the first match, which is how the magetower’s seventeenth destination came to be unvisitable. Rows past the bound keep their place and lose only their key; refusing the row would silently shorten a long menu, which is the more dangerous failure of the two.Long dynamic menus still draw past the bottom of the terminal —
menu_dynamic_calc_location()setspage_rowsto the row count and nothing scrolls. That is a separate upstream limitation, unchanged here, and the magetower has lived with it at 48 destinations since M5.3.49.1 — The manual is part of the gate.
scripts/check-buildbuilds it as a fourth pass with-W, so a broken directive or a bad cross reference fails locally instead of failing the job that publishes zangbandtk.com. Missing Sphinx stops the gate and prints the two commands that fix it, rather than skipping the pass.The virtualenv had not gone anywhere.
docs/Makefiletooksphinx-buildfromPATH, nothing ever put.venv-docsthere, andmake -C docs htmltherefore failed with “sphinx-build: command not found” on a machine with a working Sphinx sitting in the tree. That message reads as “Sphinx is not installed”, and was believed — so the manual went unbuilt across several releases that reported it as building.docs/Makefilenow looks for the virtualenv before falling back toPATH.3.49.0 — How you come by a mutation, and how you get rid of one (PLR-14, PLR-34, DEC-24). Every documented acquisition path that does not go through a magic realm now works: a Lord of the Courts replaces one favour in six with a mutation (the DEC-38 carry-over, filled in), the chaos gift makes a Lord take an interest in a character who never swore to one, unresisted chaos changes you one time in three, and Polymorph Self rerolls what chaos has made of you. The two remaining routes are Chaos and Death realm spells and wait for PLR-08.
Three new effects carry all of it —
GAIN_MUTATION,LOSE_MUTATIONandPOLY_SELF— so a patron’s ladder, a potion, a breath and a building can each reach the same machinery from data.The Chaos Tower is built (DEC-24). Zangband names it in
t_info.txt, routes it through a Lua hook and ships no script to fill the hook, so it had a building, a door and no behaviour;spoilers/mutation.txtlists it as one of only six ways to lose a mutation, which is why DEC-24 kept it when it cut the Casino and the Weaponmaster. Great cities only, 2500 gold, and it is the only route that lets the player choose which mutation goes.The potion of New Life came back. CNT-11 deferred it on the single ground that “mutations are unbuilt”, and M8 built them. Half of Zangband’s potion survives —
cure_all_mutations()does,do_cmd_rerate()does not, because 4.2 fixes hit points at birth — and its description was overridden to stop promising the half that no longer happens.A buffer overflow that mutations made reachable.
menu_dynamic_add_label()writeslabel_list[m->count]into a copy oflower_case, which is twenty-six characters and a terminator. Nothing in 4.2 ever built a menu longer than that. A character can carry eighty-nine mutations at once — ninety-six less the seven the cancelling pairs make unreachable together — so the power list and the Chaos Tower could both run off the end of the allocation. Both now stop at twenty-six and say how many they did not show, rather than truncating in silence.3.48.0 — The mutations that act on their own (PLR-14, PLR-35). Twenty-one of the twenty-seven random mutations now fire on their own timer, and all five melee mutations land an extra blow in the attack round. That completes what a mutation does; the ways you come by one are next.
Each random mutation is rolled separately every turn, so three of them are three chances rather than one, which is Zangband’s arrangement. Every one is suppressed by anti-magic except cowardice — Zangband tests the flag inside sixteen of the seventeen blocks that could have it and pointedly not in that one, because being too frightened to act is not sorcery. That reading was kept rather than tidied into a blanket rule.
The melee dice were written the wrong way round in Zangband, all five of them.
natural_attack()fills indssanddddand then callsdamroll(ddd, dss), whose parameters are(num, sides)— so a scorpion tail is described as “3d7” and rolls 7d3, and an elephantine trunk is described as “1d4” and rolls a flat 4d1. Every one of the five hits harder in the code than in its own text: 14 against 12, 9 against 7, 6 against 5, 4 against 2.5, 7.5 against 6. DEC-20 puts the source ahead of the documentation on algorithm, so the dice are the code’s and the descriptions have been corrected to match — they are the only place a player ever sees these numbers.Six random mutations are deferred with reasons: wraith form (4.2 has no incorporeal player state), the warning (a piece of user interface rather than an effect), the two hit-point/spell-point exchanges (no effect reads both pools), losing a mutation at random (Phase 5’s removal paths), and the chaos gift (Phase 5’s patron).
3.47.0 — Twenty-three mutation powers, and nine honest gaps (PLR-16). The activatable mutations join the same list as racial and class powers, after them, and use the same machinery — the same failure roll against a stat, the same fall back to hit points when the mana is short.
There is no mechanical route from Zangband’s
mutation_power_aux()to a 4.2 effect chain: it is nine hundred lines of hand-written C in which every power calls whatever it happens to need. So the translations live intools/zconv/mutmap.toml, one judgement per mutation, each carrying the source line it was read off and what the translation costs — and the converter emits them, somutation.txtstays generated.Nine cannot be expressed and say so. Telekinesis, swapping places, sensing curses, the Midas touch, growing molds, weighing magic, sterility and the launcher all need machinery 4.2 has not got, and building it to carry one mutation would be the wrong way round in most cases; Polymorph Self is deferred to the acquisition paths it belongs to rather than away. All nine appear in the power list marked not yet rather than being hidden, because the character sheet describes them and a player would come looking.
Level-scaled radii became
power-whenbands, which is what DEC-37 added them for: Zangband’s spit-acid radius is1 + level / 30, so it is one band to level 29 and another after, not an average.Twelve descriptions were advertising charisma, which 4.2 removed in 4.2.0. A silly squeak read “(-4 CHR)” and is in fact inert; warts read “(-2 CHR, +5 AC)” when only the armour was real. The converter now takes out the charisma term and the punctuation holding it in place, and nothing else.
3.46.0 — The thirty-two standing mutations act (PLR-14, PLR-15, PLR-17). A continuous mutation is now simply true of your character: stats, armour, speed, searching, stealth, saving throw, resistances, vulnerabilities and object flags all land through
calc_bonuses()the way a shapechange does, and the character dump gains a [Mutations] section.The modifiers come out of
mutation_effect()rather than out ofspoilers/mutation.txt, and the difference is not small. The spoiler gives the headline of each mutation and stops, and the headline is generally the good half: superhuman strength is “+4 STR” there and +4 STR, -1 INT, -1 WIS in the code; being puny is “-4 STR” and also +2 DEX; a moronic mind is “-4 INT/WIS” and never mentioned as conferring immunity to fear and confusion; iron skin is +25 AC and -3 DEX rather than -1. Built from the documentation, every good mutation would have been better than Zangband’s and every bad one kinder.Two things needed measuring rather than assuming. Zangband’s stealth and searching skills are on the same scale as 4.2’s — of the nineteen races both games share, stealth is identical for seventeen and searching for twelve — but
calc_bonuses()multipliesOBJ_MOD_SEARCHby five on its way into the skill and leaves stealth alone, so extra eyes areSEARCH[3]and notSEARCH[15]. And saving throws are a skill in both games but not an object modifier in 4.2, so they get their own field — with a scale as well as an amount, because magic resistance is15 + level/5.Two mutations are inert and are meant to be. A silly voice and an illusory normal appearance moved charisma and nothing else, and 4.2 removed charisma in 4.2.0. Both are still gained, described and saved; neither does anything. Left in rather than dropped, because a mutation that disappears from a savefile is worse than one that does nothing.
3.45.0 — Ninety-six mutations, and the model that carries them (PLR-13, PLR-36, PLR-37, PLR-38). Zangband built mutations with no data file at all: the table is a C array in
tables.cand the selection weighting exists only as the widths of the case runs in a switch.zconv mutationsreads both and writesmutation.txt, so this is generated like everything else rather than hand-copied.Reading the source rather than the spoiler turned up three things the documentation never mentions. Three mutations have prerequisites — the Midas touch wants a thousand gold per level in hand, and a silly voice and elemental vulnerability want three mutations already, which is chaos compounding on itself. The race affinities are not uniform: a Vampire takes hypnotic gaze six times in ten and a Beastman polymorph self only one in ten. And the regeneration penalty the spoiler warns about is not in 2.7.5 — it was a 2.2.2d mechanic its authors removed, and reinstating it would not be finishing an unfinished feature (DEC-45).
Four kinds rather than the plan’s three: the five melee mutations sit in the same machine word as the random ones and behave nothing like them (DEC-44).
Charisma bites again. Three activatable mutations cast off
CHR, which 4.2 removed in 4.2.0; all three work on a mind rather than the world — the spoiler compares hypnotic gaze to the Mindcrafter’s Domination by name — so they cast off wisdom here, recorded as a substitution in the conversion report.The savefile block is written and read by name and by count, and the player block goes to version 4 with three older loaders kept.
M8: the virtues, finished — 30 August 2026#
3.44.3 — The last gap between the local gate and CI is closed, and it found nothing. Homebrew’s GCC now sits beside the system clang as
gcc-16— not ascc, which stays clang for the macOS build the waypython3.13sits beside the system Python.scripts/check-buildruns a third pass with it, at-Werror -Wvla -Wlogical-op.The whole tree is clean under it: 214 sources, zero warnings, zero errors.
-Wlogical-opturned up no real defects, which is worth saying plainly rather than dressing up.The pass earns its place all the same, and this was checked rather than assumed.
depth > 0 || depth > 0is an error to GCC and silent to clang with-Werror— introduced deliberately, the clang builds returned 0 and the GCC build returned 1. One nuance: the two CI jobs that ask for-Wlogical-opdo not pass-Werror, so CI prints such a warning and builds on. Locally it stops the gate. That is one step stricter than CI on purpose, and affordable only because the tree is clean of it.Also fixed: the script sent build output to
/dev/null, so a failure gave an exit code and no reason — the compiler talking to nobody again, by a third route. It is quiet while it works and prints everything the moment it does not.3.44.2 — The gate covers the build it was missing, and 3.44.1’s account of the gap was wrong. That entry framed the residual risk as clang against CI’s GCC. It is not: the macOS CI job builds with
env OPT="-Werror" make -f Makefile.osxand the same clang on this machine rejects the bad pointer exactly as GCC did. There was never a parity problem to reason about — the local build was the identical compiler run without the flag that makes a warning fatal.So
scripts/check-buildnow runs both build systems, because CI does and they are not interchangeable.Makefile.osxasks for-Wshadow,-Wwrite-strings,-Wmissing-prototypes,-Wnested-externsand-Wunused-macros; CMake asks for-pedantic; and they compile to different C standards, c99 against gnu99. A defect can pass either alone. Both are clean over the whole tree — 214 sources each, and the same 214.This also collapses the two failures of the last day into one. The grep that did not match
error:and the build that was not asked for-Werrorare the same mistake twice: the compiler was reporting a problem and the local check was not reading the answer.3.44.1 — A pointer of the wrong type, and the gate that should have caught it.
virtue_note_kill()passed amonster_racetomonster_is_living(), which takes amonster. Clang built it and the undefined behaviour ran; GCC with-Werroron a Linux runner refused it. The predicate reads only race flags, so the fix is arace_is_living()for callers that have a race and no instance —monster_is_living()now delegates to it, and the substitution keeps the semantics it was given: the living check, on the unique branches only.The local build was never running CI’s flags.
cmake -S . -B buildtakes the defaults; every Linux job configures with-Werror. That gap let two defects through in a day — this one, and two invented identifiers a few hours earlier that were reported as a clean build because the output was being filtered with a pattern that does not match what clang prints.scripts/check-buildnow configures with CI’s flags and lets its exit code be the answer. Run over the whole tree it finds nothing else of the kind.It deliberately does not go beyond CI:
-Wshadowalone objects to a dozen places in the existing suite that every Linux job builds happily, and a gate that fails on things CI accepts is one people learn to ignore.The kill hook also gained the test it did not have. Every one of the suite’s other tests passed with the bad pointer in place, because nothing called the function.
3.44.0 — Zangband’s unfinished feature, finished (PLR-18 to PLR-21). Topi Ylinen wrote the virtue system in 1998 and it never worked: by 2.7.5-pre1 there were 168 places that wrote a virtue and none that read one, and the knowledge screen that would have shown them was commented out with the note “Display virtues option is always left out”. Eight counters per character, moving with everything they did, and nothing looked at them for seven years.
The selection is kept because it is the part that did real work: a character is measured against eight of eighteen, chosen at birth by class, then race, then the realm they cast from, deduplicated, and padded from a weighted table. Two Warriors of different blood are asked different questions. The tables live in
class.txt,p_race.txtandrealm.txtrather than in a switch statement, so what a Necromancer is measured against is legible without reading C.Two things read them, which is the whole point. A Lord of the Courts of Chaos reaches for the bottom of its reward ladder less often for a character strong in Chance and Individualism, and more often for one strong in Harmony and Temperance — a Lord of Chaos is not impressed by a well-run life. And the dream at the inn runs truer for Enlightenment and Knowledge and darker for Unlife and Chance. Both sit on mechanisms that already existed.
The writers are sized by coverage rather than by census (DEC-43). Porting all 169 of Zangband’s nudges was one option and porting only what the two consumers read was the other; this is neither. The rule is that every virtue a character can be given must be one their play can move, because a Warrior carrying eight counters that nothing moves and nothing reads is the condition Zangband shipped in, reached by a different road. Nine hook sites cover all eighteen: what you kill, striking something asleep, a chaotic weapon discharging, digging and clearing trees, hasting, making yourself invulnerable, enchanting, mapping a level you have not walked, and collecting a bounty.
One substitution, marked as one. Zangband’s death block asks three times whether the dead thing was
GOOD, and 4.2 has no such flag — nor is it recoverable, Zangband having put it on 230 monsters running from Farmer Maggot to Mughash the Kobold Lord.monster_is_living()stands in on the unique branches only; the rest is dropped rather than approximated. Without it Unlife would have had no writer that raises it, and a Necromancer would carry a counter that only ever fell.The savefile block is written and read by count, and the player block goes to version 3 with version 2 kept as a loader — so the thirty-five characters in
tests/savesstill open. A reader looping to its own compile-time constant is exactly what broke them a fortnight ago.This closes M8’s gate. DEC-39 kept virtues on condition that something read them; two things do.
3.43.1 — The Incandescent Globe of Sawall is dropped (DEC-42). Zangband has two artifacts of that name — a light and a helm, both real, and neither a draft of the other. It could carry both because it saved an artifact by index and put the base object in front of the name; Angband saves the name alone and takes the first match, so a character could come back holding the wrong one. Renaming was available — three of the nine Chaos patrons are of House Sawall — and was declined as not worth it for one artifact. The helm stays, because it is the one already shipped and renaming a shipped artifact makes every savefile holding it refuse to load. The count stays at 51.
The converter now finds any duplicate name among the candidates rather than only the one it was told about, and reports an unruled pair as a problem instead of quietly importing both.
One artifact also gained a property it should have had a version ago: the Staff of the Wandering Wizard carries Zangband’s
SP, which became a mana modifier in 3.42.0 — the artifact file was not regenerated at the time, so it had been sitting on a mapping that existed and was not applied.3.43.0 — A character who has cheated says so.
Cheat, in red, at the left of the status line, for as long as that character exists.The game has always known.
noscoreis set the moment wizard mode is entered or a debug command is used, the cheat options each have a hidden twin set the moment they are turned on, and neither is ever cleared — that is what bars the character from the high score list. None of it showed on screen, so a character being walked about invulnerable looked exactly like one that had earned its way.Which is where this came from. A screenshot of a level 6 character with 1012 hit points, sent to report a stalking deer, read as a possible bug in hit point calculation; it was wizard mode, and the game could have said so.
Both records are now asked through one function,
player_has_cheated(), which the score list uses too — so the indicator and the scoring cannot drift into disagreeing about whether a character was played straight.3.42.3 — The white deer bounds out of earshot. Reported from play a second time, as still stalking me. The blessing itself was fixed in 3.33.2 and was working: the deer gave its gift once and shied away from every touch after. What it would not do was leave. It bounded ten grids, and a deer moves at speed 130 and hears at 40, so ten grids is a distance it undoes in a few of your turns — arriving back beside you to be touched and refused, over and over.
The bound is now
world:blessing-bound(100 grids by default, and tunable without a rebuild), chosen to clear the beast’s ownhearingrather than its hooves. Past that it no longer knows where you are and goes back to being a deer somewhere else. It is still neither killed nor removed; you can meet it again, and it will still have nothing more to give you.3.42.2 — Every character saved before today stopped loading, and the app terminated outright on some of them. Thirty-five real savefiles now live in
tests/saves, and a unit test loads every one of them, saves it again and loads it back — because loading alone passes over a fault that only shows up on the next save.Four faults, all the same shape: the savefile names something this build no longer has.
rd_miscsized three of its arrays from the counts compiled into the running build rather than the counts written in the file, which the rest of the loader has always read back correctly — so adding an object flag, an element or a modifier invalidated every existing save.rd_monster_memoryspan forever on a monster that had been removed.rd_monstertreated a vanished race as fatal to the whole load, which is what actually killed most of the corpus.rd_trapandrd_ignorewill do the same at the first renamed trap or object; both are fixed before they fire.Setting
ZTK_SAVE_TRACEnow names each block as it is read and echoes the loader’s notes. Every one of these was invisible without it.
M2: the rest of Zangband’s objects — 30 August 2026#
3.42.1 — The Scroll of Rumour is dropped rather than deferred (DEC-41). It sat on the deferred list, which reads as a queue, and it was never going to reach the front of one: the mechanism is a morning’s work and the content is 647 individual judgements about which of Zangband’s rumours are still true of this game. It now sits with the artifact bases under rejected, where nothing is waiting for anything.
3.42.0 — Three of the objects CNT-11 held back, and two that stay held. A Ring of Wizardry needed a mana modifier, which Angband’s object model did not have;
MANAis appended to the eleven it did. The pval is mana per casting level rather than a flat bonus (msp += sp_bonus * levels), which is the part ofSPits name does not tell you and the reason the ring is worth a finger at all.A Rod of Havoc turned out to need no code.
call_chaos()picks one of thirty damage types and throws it, andeffect:RANDOMchooses one of the effects that follow it — so the whole table is a data entry. 25 of the thirty elements exist here; the shapes that could not be expressed, and the five elements that do not exist, are named in the conversion report rather than approximated.A Scroll of Mundanity strips an item back to the plain thing it was made as. Zangband reset the fields one at a time; 4.2 keeps an item’s power in runes and curses as well as numbers, and
object_prep()rebuilds all of it from the kind — the same operation, in one call.A Scroll of Artifact Creation is not built, and the reason is the savefile. Angband stores an artifact by name and builds
a_infofromartifact.txtat startup, so one invented during play would not exist when the character was loaded again. A Wand of Rockets stays held for the reason it was held before: its damage is its own element,el_info[]is written to the savefile by position, and the change reaches every object and monster in the game for the sake of one wand.A Scroll of Rumour is waiting on a decision, not on code. Zangband’s 647 rumours are in the archive and
hints.txtis already the same shape, so the mechanism is a morning. A good many of the lines name things this game does not have or contradict what it does, and a scroll whose purpose is to tell the player something true is the wrong place to ship text that is false — which makes it curation under DEC-19 rather than conversion.3.41.0 — Three flags that were waiting on nobody, and two that were waiting on more than they said.
CAN_FLYandCAN_SWIMreach 117 and 91 imported monsters. Angband gates damaging terrain on one resistance flag named by the terrain — deep water asks forIM_WATER— which answers whether a creature can survive a grid and not whether it can keep out of one. A raven does not resist drowning; it never touches the surface. Until now every bird and every swimmer in the import was stopped by the first river it met, which is invisible from the outside, because a monster that will not cross simply goes somewhere else.PATRONgives a Lord of the Courts an interest in whoever carries it, which the(Chaotic)ego does. Building it turned up something missing from work already shipped: Zangband’s condition for a level reward isTR_PATRON || (one_in_(7) && TR_STRANGE_LUCK), and the Ring of Fate’s fifth effect — the only one that is not about critical hits — had been left out whenSTRANGE_LUCKwas built two versions ago. It is in now.The cruelty runs the opposite way to the fiction. Zangband writes
nasty_chance *= 2for a character with no patron of their own, and that is the denominator of a one-in-N roll, so a Lord that was never yours is half as likely to reach the bottom of the ladder. This shipped backwards first, and the test written alongside it asserted the backwards version — caught because the wrong direction collapsed two of the four rates in an older test onto each other and made it fail about one run in eight.``WILD_SHOT`` and ``WILD_WALK`` are not built, and the reason they were recorded as unblocked was wrong. Both cancel a terrain penalty, and this game imposes neither: there is no terrain term anywhere in Angband’s combat for
WILD_SHOTto waive, andWILD_WALKremoves the cost of climbing a mountainside that is a wall here rather than slow ground, and the poison of a swamp that is not a terrain type at all. The blocker was never the wilderness arriving. One ego goes with them:of the WildhadWILD_WALKand nothing else, so it was reaching the game as boots that granted nothing. The converter now drops any ego left with no property at all, by rule rather than by name, so it returns on its own the day the flag is built.3.40.3 — Bookkeeping that had drifted far enough to mislead. The requirement coverage table accounted for 91 of the 109 requirements and said in a footnote that seven were unassigned. Rebuilt rather than patched, with the arithmetic written down so it can be checked: 104 scheduled, one standing rule, two built and never scheduled, two closed by decision. Three requirements — WLD-04a, WLD-04b and WLD-08a — appeared nowhere in the plan at all despite having shipped in M4.
BAL-07is recorded as met, with DEC-40 for the method. The mapping from Angband 2.8.1’ssleepto 4.2’ssleepinesswas derived from the 434 monsters both versions carry rather than assumed, and the interesting part is what it found: for 13 of the 21 distinct source values the recovered figure is the source figure, and 275 of the 434 monsters carry the same number in both. 4.2 never rescaled sleepiness — it edited individual monsters. The scale was identity with noise on it, and the median is what takes the noise off.The
# Source:line in the five generated data files is now written relative to the top of the tree instead of absolutely, so regenerating in a differently-named checkout produces the same bytes.check-build-listsfails if an absolute path comes back. It cannot check the stronger property — that the shipped file is what the generator produces today — because the Zangband sources are not in the repository, so that comparison is only available to someone who has them.3.40.2 — A potion that no game could ever have produced. Angband builds one row of its object allocation table per dungeon level, from zero to
obj-make:max-depth, and clamps every request into that range. Zangband’s dungeon went deeper than the table does, so its Potion of Invulnerability arrived atalloc:11:105 to 105— five rows past the end of a hundred-row table, present in the data and unreachable in play. Two more objects had ranges that ran past the end without being stranded by it. The converter now clamps both ends of every allocation band to the table’s last row, reading the ceiling from the game’s ownconstants.txtrather than carrying a copy of the number.The deeper half of this was investigated and not done. The dungeons run to 127 while the object table stops at 100, so the bottom 27 levels generate their loot from the depth-100 row — which reads like a bug and is not one. Widening the table means every object whose range ends at 100 stops being generated below it, and that is 343 of the 434 kinds that have an allocation line: nothing at all would be generated at depth 127. “To 100” is Angband’s way of writing “to the bottom” throughout its own data, and the clamp is what implements it. The flat deep end is not the table being too short; it is that no object in either game is authored below depth 80. Changing it means authoring deeper objects, which is content rather than a fix.
3.40.1 — Three imports that were Angband’s own content under new names. CNT-11 found that comparing object kinds by name let Zangband’s renames through as if they were new; the same scrutiny applied to the monster, artifact and ego imports found three more, and each file needed a different notion of identity to find them.
Ego indices are 98.9% stable between Angband 2.8.1 and Zangband, so the index is identity there, and the single divergence was a rename:
of Levitationis Angband’sof Slow Descent— same slot, same loneFEATHERflag — which Angband still ships. Artifact indices are 69% stable, and every one of the 35 divergences turned out to be Zangband putting one of its own artifacts in an inherited numbered slot: Frakir for the Ring of Barahir, the Crown of Chaos for the Crown of Morgoth. Those are replacements, not renames, and were correctly imported. The artifact row is clean.Monster indices are 0.5% stable — Zangband renumbered the whole bestiary — so identity had to come from somewhere else, and what survives a rename is the description, which is hand-written prose. 33 imported monsters keep a 2.8.1 monster’s paragraph word for word. Thirty-one of those are Zangband writing a much deeper monster on inherited text, or boilerplate that several Angband monsters already shared. Two were renames:
hobois the boil-covered wretch andraving lunaticis the village idiot, identical in glyph, colour, speed, armour, depth and experience, and both still in Angband under the older name.The bestiary is 387 imported monsters rather than 389, and there are 17 ego types rather than 18. The total of 1013 has not moved, because it was already two out of date.
One artifact was also being lost silently rather than skipped. Zangband has two called “of Sawall”, and the converter read artifacts into a dictionary keyed by name, so one overwrote the other. It now reads them as a list and defers the second with the reason recorded: Angband writes an artifact’s name into the savefile and reads it back with the first exact match, so it cannot carry two of one name. Giving that one a name of its own is a content decision rather than a conversion one.
Warning
A savefile holding a hobo, a raving lunatic, or an item of Levitation will refuse to load. Angband stores monsters and egos in savefiles by name and reports “Monster race … no longer exists!” when the name has gone. The failure is loud rather than silent, but the character is not recoverable without putting the entries back.
3.40.0 — Eighty-two object kinds, and the three properties they needed. Zangband’s own objects, which until now were the one part of its content that had not come across: thirteen swords, nine polearms, eight hafted weapons, nineteen pieces of armour, twenty-eight rings and amulets, and the consumables among them that Angband has an effect for. A new
objectssubcommand inzconvproduces them, and the data file it writes is generated rather than edited, like the other three.The count in the plan was measured by name, and names lie. Zangband renamed twenty-six of the objects it inherited from Angband 2.8.1 without changing what they are — its Ring of Skill is Angband’s Ring of Accuracy, its Scroll of Enchant Weapon Deadliness is Enchant Weapon To-Dam, its Ring of Levitation is Feather Falling — and ten of those are still in Angband today under the older name. Importing by name puts both in the game. The converter compares the slot an object occupies instead, which is its actual identity, and that also settles the reverse case: four slots where Zangband kept the number and put a different object in it, which are new content.
Sixteen more were refused for being artifact bases rather than objects. Seven of them are called “Ring”, exist only so an artifact has something to hang on, and importing one puts a plain ring in front of the dummy the Ring of Barahir needs — after which
artifact.txtwill not parse.A Ring of Fate is not a lucky charm. Its
STRANGE_LUCKsharpens every melee critical by half again, including the ones monsters land on you, and gives one monster in thirteen a chance of arriving from up to forty levels deeper than the floor should allow. Zangband’s own comment on the third of its four effects reads “Luck isn’t always good for you…”. Also new:PSI_CRIT, which spends one to three mana per critical and does nothing on an empty pool, andNO_MAGIC, which stops you casting outright.Nineteen objects were deferred with the reason recorded, not dropped: ten statues and a figurine whose names interpolate a monster Angband objects cannot refer to, two more that need pets, and the rest waiting on mutations, a chaos-effect table, a mana modifier, a rumour list, or an element that cannot be added without changing the savefile layout.
Three older faults surfaced on the way, all in shared conversion tables and all now fixed in the generator. Zangband’s
LITEis a boolean+1light radius, not one of the properties its single pval scales — read as one, it had been giving the Crown of Chaos a light radius of 125, and thirteen other artifacts too much.NO_MAGICon an artifact had been mapped to Angband’s impair mana recovery curse, which is a different thing and is not even legal on an amulet. And an item carrying both an immunity and a resistance to the same element emitted both, in that order, so Angband’s parser assigned the weaker one last and quietly downgraded the immunity.
M2 and M3: the flags that were left — 29 to 30 August 2026#
3.39.0 — The last of Zangband’s flags, and two that were never flags at all.
QUANTUM: one monster that half your blows pass through, that takes half its turns, and that now and then simply stops existing.GHOUL_TOUCH: what you strike bare-handed falls asleep.RETURN: a thrown weapon comes back nineteen times in twenty.LUCK_10: a better saving throw.EASY_ENCHANT: twice the enchantment attempts, and no artifact resistance.Two of the eight turned out not to be mechanisms.
SENSEis Zangband’s data-file name for the searching skill — its own C calls the same flagTR_SEARCH— andTHROWis an item meant to be thrown, which Angband already has asTHROWING. Both had been filed as things to build on the strength of their names, and both were being dropped from the conversion instead: the rename alone restores a search bonus to eight artifacts and two egos.PASS_WALLis refused rather than deferred. Angband has no form of player wall-movement to hang it on, so carrying it would mean building wall-phasing rather than reading a flag.That is the end of the flag work.
flagmap.tomlhas nothing left to implement at all, andobjflagmap.tomlhas onlyPASS_WALL, kept there with its reason rather than pretended away. Everything else in either is deferred: the pet allegiance model, which is M10, three ego flags whose blockers have since shipped, and the habitat tags, which this game answers another way.3.38.1 — Three build lists that had stopped matching the tree. The Visual Studio project, which is hand-listed, had never heard of two of the files below and failed to link on Windows; the DOS package had no 8.3 name for
monster_speech.txtand quit during init, printing nothing; and the autotools install list had been missing four data files since M2. All three are now compared against the tree byscripts/check-build-lists, which runs on every push rather than in the job that happens to break.3.38.0 — Five monsters that will not hold still.
SHAPECHANGER: the chaos shapechanger, the lord of chaos, the unmaker, Nyarlathotep and Dworkin are drawn as some other creature each time the display refreshes, and about one time in twenty-five as an object. Nothing about them changes but the glyph. Zangband nested the check inside the multi-hued draw, so the flag does nothing at all withoutATTR_MULTI— all five carry it, which is why that never showed.3.37.0 — Things that burn you back, and things that bounce.
REFLECTINGand the three auras, with the ego flagsREFLECT,SH_FIREandSH_ELECthat are the same mechanisms from the wearer’s side. Strike something wrapped in fire and it burns you; a bolt aimed at a reflector comes back, nine times in ten, to a grid beside whoever fired it. And a monster wrapped in fire will not enter water — a movement rule hidden inside a damage flag, and one Zangband never got to use for want of a sea.3.36.0 — A hundred and two monsters decline to be teleported.
RES_TELE, which is not the flat immunity it sounds like: a unique that has it is unaffected outright, and anything else gets a saving throw on its size. Teleport Other is no longer a reliable way to remove an inconvenience.3.35.0 — Eighty-nine monsters have something to say.
CAN_SPEAK: they speak in a fight, differently once they have turned to run, and once more when they die. And a unique that could talk turns out, one death in ten, to have been wanted for something — the bounty is real gold on a curve reaching 32,000, which the flag report had filed as pure flavour.3.34.0 — The blood of Amber is a family.
AMBERITEgives the twelve a shared kind, theS_AMBERITESsummon, and a blood curse laid on whoever kills one. Until now they were twelve separate uniques with nothing joining them, which is what CNT-02 asked them not to be.
M7: races and classes — 25 to 29 August 2026#
3.33.2 — The white deer is no longer a healing fountain. Reported from play: a deer followed the character about, healing them to full every few paces, forever. The beast did remember — but the memory was kept on the animal, and an ordinary deer is not a unique, so the wilderness destroys and re-rolls it every time the map around the player is rebuilt. A fresh deer has forgotten, and at depth 1 with rarity 4 there is always another one.
The memory is now kept on the character as well, for
world:blessing-turns(500 by default, and tunable without a rebuild). While it lasts every beast shies away. The Unicorn was never affected — she is unique, so the wilderness brings her back with her memory intact.3.33.1 — The open sea, and the manual catching up with it. The first pass at populating the water only filled the shallows: deep water is damaging terrain and the generator refused it outright, so the ocean proper stayed empty. Deep water now names
IM_WATERas the flag a creature needs to be in it — exactly as lava namesIM_FIRE— so fish live there and everything else still keeps out. The player is unaffected; their drowning was always decided by what they carry.The Wilderness said “nothing else can swim yet”, which had stopped being true, and Monsters said nothing about the aquatic monsters at all. Both now describe what lives in the sea, that it is never tame however orderly the coast, and that the waterline is a line in both directions — a shark cannot follow you onto the beach.
3.33.0 — There are fish in the sea. Reported from play: walking the coast turned up nothing. It turned out the sea held nothing whatever — not just no fish, no monsters at all.
wild_populate()askedsquare_isempty()of each grid, that asks for a floor, and neither depth of water carries the FLOOR flag, so every square of ocean was refused. The wilderness now populates in two passes, land and water, each drawing from its own kind.Three things had to be true for it to feel right. The sea has a density of its own, because the population parameter measures what the land supports and left open water the emptiest place in the world. It has a danger floor, because danger derives from how well the country is policed and nobody polices the sea — a calm bay off a lawful city had nothing in it, the shallowest fish being a swordfish at depth eight. And fish cannot leave the water, whether by swimming ashore or by being scattered there as a shoal, so a character who stays on the sand is in no danger at all.
3.32.1 — Sharks are no longer trees. Zangband drew its aquatic monsters with
land Angband draws trees withl, so the import gave every fish, shark, whale, squid and kraken in the game the tree’s base — and a base’s flags are inherited. All twenty-four were immune to fear, immune to confusion, and regenerating, and because dungeon dwellers are matched by base, Arden and the Grove of the Unicorn were spawning great white sharks while the Forest monster pit could fill with krakens. The bestiary filed whales under Trees/Ents.There is now a Fish/Aquatic base — glyph
N, the last free letter in the game — and the twenty-four have moved to it. The ents stayed where they were. Faiella-Bionin, the stairway beneath the sea, and Rebma, the drowned city, have fish in them for the first time.They lose their tiles in the process: they had been borrowing tree pictures, and there is no aquatic monster in Angband to borrow from, so they render as text until somebody draws them.
3.32.0 — The imported monsters have tiles. All 391 monsters ZangbandTK adds were missing from every one of the five tilesets, so in graphics mode a chaos hound, a serpent man and a death sword all rendered as bare letters among the pictures. 235 to 256 of them per tileset now borrow the tile of an Angband monster sharing their kind, colour and depth — generated by
scripts/tiles/borrow-tiles.pyand verified by its--checkmode against the two ways this fails silently.The 122 imported uniques are deliberately left as text. Matched by kind and depth, Oberon of Amber, Dworkin and Mandor all come out wearing the Mouth of Sauron’s face. For a game about getting away from that, three Amber princes in Tolkien’s clothing is worse than no picture: a letter says not drawn yet, a wrong portrait says something false.
3.31.4 — Four crashes left behind by the tileset filtering. Dropping uninstalled tilesets from the list made
get_graphics_mode()return nothing for an id that used to resolve — which is the point — but four places had never had to consider that.reset_visuals(), shared by every front end, asserted on the result and then dereferenced it, and an assertion is no check at all onceNDEBUGis set. In the SDL front end,sdl_BuildTileset()dereferenced it outright; the saved configuration was read straight intouse_graphicswith nothing checking the tileset existed; and the loop that binds the current mode matched a tileset by its position in the list rather than its id, so a missing low-numbered tileset silently selected the wrong one.The Cocoa and Windows front ends were already safe, and were left alone.
3.31.3 — A review of M7, and eleven fixes. Five of the findings were things the game did silently and wrongly: a Yeek’s scream and a Sprite’s sleeping dust did nothing at all — projected with no dice, so a power of zero, paid for in full; a patron’s destruction levelled one grid instead of fifteen; a Draconian’s breath had its arc written into the radius slot, making it a needle-thin cone at a fifth of its range; and two of the Mindcrafter’s level bands were identical, because a ball with radius zero is given a radius of two.
The worst was structural: a patron’s reward was handed out from inside the level-up loop, and two rungs on every ladder grant or drain experience — both of which call back into that loop. A single kill worth several levels could recurse. It now waits until the level has settled, which is what Zangband did and why.
Also: a patron with no reward list crashed during loading rather than reporting a parse error; patron messages are now validated as format strings the way martial-arts messages already were; Learn all monsters skipped the first monster in the bestiary; the two new class abilities had no descriptions, so a Monk was never told it fights unarmed and a Chaos-Warrior was never told it was owned; and the version-1 savefile reader was a verbatim copy of the version-2 one, which is now one body with a flag.
3.31.2 — A picture of the coast, and the machinery to take it. The capture harness can now stand a character on a waterline before the first level is built, which is where it has to happen: asking for a new level mid-turn is honoured by the full game loop and ignored by the headless front end the pictures are captured through. Adds a Learn all monsters debug command (
B), the companion to Learn object kinds that 4.2 has and this did not.3.31.1 — A Mindcrafter has mana again. It never had any: the armour allowance before mana starts draining is read out of a class’s magic block, a Mindcrafter has no magic block, and its own starting armour then cancelled more mana than the class had. It began the game unable to pay for anything and stayed that way for fifty levels, paying for every power in blood. The sidebar compounded it by refusing to show spell points to a class without spellbooks, so there was nothing on screen to say what was wrong. Both fixed, and has mana is now one question asked in one place rather than three sites guessing from spellbooks. Found from a player’s screenshot.
The manual also gains a third screenshot: a Sprite Mindcrafter’s power list, showing what a character can do that came out of neither a book nor a shop.
3.31.0 — The Chaos-Warrior, and nine Lords of Chaos (PLR-03/PLR-05). The twelfth class, and the only one that belongs to somebody. Sworn at birth to a Lord it did not choose and cannot leave, and every level it gains the Lord decides how it feels: a full healing, a raised stat, experience, an object out of nothing — or a cursed weapon, a ruined stat, monsters dropped on top of you, or the level destroyed around you. Your Lord is on the character sheet in red, because it is not a credential.
Level 13 is three times as dangerous as any other, and nothing in the game says so. Every other thirteenth level is twice as dangerous; every fourteenth is the safest a Chaos-Warrior ever gets.
The nine Lords are Zelazny’s. Zangband’s sixteen were Moorcock’s Elric gods plus Khorne, Slaanesh, Nurgle and Tzeentch from Warhammer, and not one had any connection to Amber — the clearest case of drift in the game. They are replaced by Swayvill, Suhuy, Mandor, Dara, Gramble, Jurt, Despil, Borel and Gilva, four of whom already walk around in the bestiary. See DEC-38.
Polymorph uses the game’s own shapechanges rather than Zangband’s version, which permanently mangled the character; and one reward in six was a mutation, which is left as a deliberate gap until M8.
3.30.0 — The Mindcrafter, and psionics (PLR-03/PLR-06). The eleventh class. Twelve powers that arrive by level rather than out of a book, on
Nalongside anything the character’s race can do — from Neural Blast at level 1 to a telekinetic wave at 28. Deliberately not a realm of magic: nothing to study, nothing to choose at birth, nothing that can be taken away.Most of the twelve grow into something else as the character does, which needed a mechanism Angband has no equivalent of. See what is coming spots monsters at level 2, finds traps and doors at 5, sees invisibility at 15, maps the level at 20, grants telepathy from 25, detects everything at 30 and lights the whole level at 45 — one power, one key, for an entire career. Racial powers gained the same ability, which is where a Draconian’s breath changing with level can now go.
And psionic force, a kind of damage the game did not have. Every damaging type Angband has is an element resisted by a flag or by armour; this one asks whether there is a mind there to hurt. A golem, a mould or an animated weapon is completely unaffected however hard it is hit; a strange or feeble mind takes a third. Against anything else nothing resists it at all.
The
Ncommand is now Use a power rather than Use a racial power, since it lists both.3.29.1 — “What’s Different” has a Races and Classes page. The largest visible difference between this game and Angband — twenty races instead of eleven, powers a race can use, a class that fights better with nothing in its hands — was described in the reference manual and nowhere in the section whose whole job is to say what is different. It has a page now. The features page lists races, powers and the Monk under what is in the game rather than under what is coming, and the four stat and ability tables in Creating a Character are generated from the data files instead of typed, having silently stopped at Angband’s eleven races and nine classes.
3.29.0 — The Monk, and martial arts (PLR-03/PLR-04). The tenth class, and the first from Zangband. A Monk fighting bare-handed climbs a ladder of seventeen techniques — a punch at level 1, a crushing blow at 48 — throwing two strikes a turn at the start and eight by the end, which is more than a Warrior gets from any weapon. Some techniques stun; a knee doubles a male opponent up; a kick to the ankle leaves anything that walks limping. Against 4.2’s own handling of an empty weapon slot — one point of damage, no criticals — a Monk measures about eight times the damage over a hundred blows.
Armour is the whole cost. Over ten pounds plus four tenths per level across the six armour slots and the strikes halve, the to-hit and damage bonuses go, and the armour class paid for empty slots stops. Under it, an empty body slot alone is worth three points of armour for every two levels. Zangband’s numbers throughout; the class mapping from Zangband’s table to 4.2’s was measured against the six classes both games share rather than guessed, and the Monk keeps Zangband’s experience factor, which makes it the only class in the game that costs extra to level.
3.28.1 — Racial powers work without mana. Powers took their cost from spell points only, and the two classes with no spellbooks — Warrior and Monk — have a maximum of zero, so a Draconian Warrior could never once breathe. Any character who had spent their pool was in the same position. Zangband’s rule applies: short of mana, the price comes out of hit points instead, and the menu says which before you commit. The failure chance was also clamped in the wrong order and could be reported above 100 per cent for a poor governing stat.
3.28.0 — Racial powers (PLR-02). Eight of the nine new races can now do something no class teaches, on
N: an Amberite shifts into shadow and walks the Pattern, a Vampire drinks blood, a Draconian breathes, a Golem turns to stone, a Yeek screams, a Sprite throws sleeping dust, a Mindflayer blasts a mind, a Half-Titan sizes up what it is looking at. Level, mana, governing stat and failure chance are Zangband’s own numbers rather than new ones, and a test holds them there. Powers draw on spell points, so a Warrior of one of these races has a small pool of mana that exists only to feed them; failing spends it, being refused does not. The Beastman has no power on purpose — its character was mutation, which is not a thing you choose to do.Two commands also gained the roguelike keyset bindings they should have had: the racial power is
&there and the quest log%, becauseNandJare both running keys in that keyset and the commands were reachable only from theEntermenu. There is no letter left to give them — between the two keysets every letter and every usable control key in the game is already taken.3.27.1 — The nine new races are described, and the four turned down are described too. Creating a Character had still been saying there were eleven races; there are twenty. Each of the new ones now says what it is and what it costs, with a note that the experience figure is not a formality here the way it is in Angband. And there is a short section on Half-Ogre, Half-Giant, Cyclops and Dark-Elf, which are not coming: the first three are Half-Troll with one resistance swapped, and the last is a dearer Elf. “It exists in the original” is not on its own a reason to carry something, and somebody noticing they are missing deserves the reasoning rather than a gap.
3.27.0 — M7 begins: nine new races. Amberite, Beastman, Yeek, Draconian, Mindflayer, Vampire, Golem, Sprite and Half-Titan, bringing the total to twenty. Curated rather than imported wholesale — of Zangband’s twenty-one, four are undead wanting a mechanism the game has not got and several are close variants of races already here. The mapping between the two games was measured, not guessed: ten races appear in both, and against those the stats, hit dice, disarm, device, save and stealth transfer verbatim, while 4.2 turns out to have widened melee and archery by roughly double and flattened nearly every race’s experience cost to 120. Zangband’s costs are kept, which is what makes an Amberite dear at 225 and a Half-Titan the most expensive thing you can be born as at 255.
M6: quests — 24 to 25 August 2026#
3.26.1 — The features page said the platform was macOS on Apple Silicon, full stop, which has not been true for some time: the game is built and packaged for Windows, Linux, DOS, the Nintendo DS and the 3DS as well, and the page is the first thing a visitor reads. It now says so, with the honest caveat that only macOS is actually played through. And reporting a job at any building that hires is now written down as the design rather than apologised for as a shortcut: whoever wanted the work done happens to be drinking in this inn too, which beats walking back across the world to be paid.
3.26.0 — All six kinds of work, and the features page catches up. The three that were missing are written: a killing at a named depth of a named dungeon (only dungeons you have found, only depths that dungeon reaches), a killing that counts only above ground, and fetching a particular thing — finished by having it, checked against your pack, so buying it or taking it from a chest counts as fetching it. Which kind you are offered depends on what the world can supply, and every one falls back to a bounty, which needs nothing but a bestiary. The features page had been claiming towns and quests were still to come since before either landed; it now describes the game as it is.
3.25.0 — A job can be given up. Walk into a building that hires while you owe it something and it lists what you are carrying and offers to let you off. This is the fourth thing the lifecycle had to be able to say and could not: M6 asks that a quest can be taken, tracked, completed and failed, and until now a bounty on something twenty levels too deep held one of your slots until the character died — taken, impossible to finish, impossible to be rid of. Every carried job is listed, not just the oldest. No penalty beyond the walk back.
3.24.2 — Screenshots, at last, and they cannot go stale. The Screenshots page shows the whole world of one game — 129 blocks of sea, coast, forest and mountain with the towns and dungeon mouths on it — and the surface with a town standing in it. Both are captured from the running game rather than drawn: the test front end reports every character, position and colour the game puts on screen, and a script replays that into SVG using the palette read out of
z-color.c. Regenerating them runs the current code, so they cannot drift the way a folder of old PNGs does. Two debug commands set the scene, and neither does anything a player could not do from^A.3.24.1 — A quest remembers how many it asked for. It said “6 small kobold — 0 of 0 done”, and both halves of that were wrong. Angband never stored the target count, because its two quests come from
quest.txtand only progress had to be remembered; work taken from a building has nowhere else to keep it, so it came back as nought of nought — and could never be finished either, since the count starts at one and never reaches zero. Saved now, in quests block version 4. A job taken under 3.24.0 cannot be reconstructed and is dropped on load with a note rather than left owing something impossible. And the plural:monster.txtcarries one only for the ninety-six names that need it, so everything else now takes a plain “s”.3.24.0 — Quests you can see, and errands that are not killing. Two new kinds of work: carry word to a named town, and go and look at a place nobody here has been. Both are finished by arriving, which needed a trigger of its own — the existing check only ever sees a monster die, so a delivery that named a monster would have been completed by killing one. And
Jnow shows what you have taken on, how far along each job is, and where the travelling ones point; until now the only way to find out was to walk back into an inn and ask. The quests the game is won by are deliberately not in that list: you are on those from birth, and “kill the Serpent of Chaos” at the top of a first-level character’s list gives away the ending and tells you nothing you can act on. The manual gains its Quests chapter.3.23.3 — Roads are two grids wide, not three. Three was the fix for a one-grid road being invisible where it turned, and it worked — but it read as a motorway running across the country. Two keeps what the widening was for: a corner is a two-by-two block of paving, which is enough to see a turn as a turn. Routing is untouched, so the same three blocks in a hundred carry a road; there is simply a third less paving on them.
3.23.2 — The magesmith and the recharger did nothing for your money. Both were handed a dice string of
"0"and no subtype. Zero is not a small amount, it is nothing: the enchant effect tests its subtype as a set of bit flags and zero matches none of them, so the magesmith took the fee and did not so much as ask which item — and the recharger asked for the item and then worked at strength zero, the worst odds in the game. The magesmith now asks whether you want a weapon or a suit of armour before naming its price, and a weapon gets both to-hit and to-damage. The recharger works at the strength of a scroll of Recharging.3.23.1 — All forty-four debug commands are written down, in the manual and in the in-game help, by submenu and by key. They were never documented anywhere — the manual named the nine submenus and gave five examples, and the in-game page did the same, so the only way to find out what
^Acould do was to open every menu and look. Generated from the source rather than transcribed, so it is right. The three ZangbandTK added are marked.3.23.0 — WLD-16d: somebody will give you work. Walk into an inn that is hiring and you are offered a job before you are offered a bed: kill so many of a creature, come back, be paid. Quest-giving is a property a building carries rather than a building of its own — nothing in that path knows what an inn is, and moving it to the magetower is one line. The inn has it because that is where people who have been somewhere else are sitting, and because a town that has fallen keeps no services, so the work dries up exactly where you would expect without a rule saying so. About half the towns in a world are hiring. The quest list gained eight slots for work taken this way, kept strictly apart from the two quests the game is won by.
3.22.1 — The place line actually appears now. Third attempt, and the first two failed for two different reasons. It was given the lowest priority in the sidebar, so it was dropped on any screen shorter than 25 lines (fixed at 3.21.2) — and then it still did not show, because the sidebar was already full: a 24-line terminal spends the top row on messages and the bottom on the status line, leaving twenty-two, and there were twenty-three entries. The priority filter passed it and the row counter ran it off the bottom, where it was drawn underneath the status line and instantly covered. A blank spacer row has been given up to make room.
3.22.0 — Two more cheats, for testing a world you have to walk across.
=thenx: Take hit points asks a number and sets them there — written into the character’s hit dice rather than the total, so it survives levelling and the next step taken — and Know every place puts every town and dungeon mouth on the map at once, which turns the magetower into a way of getting anywhere. Both are also on the debug menu (^A), asiandk. Also: the manual stops calling this game Angband. Forty of the forty-seven references audited became ZangbandTK. The FAQ’s development section is cut — it described the Angband project’s process and pointed anyone who wanted to help at the wrong project. And the quick demo no longer tells new players they cannot leave the town, which has been the opposite of true since M5.3.21.3 — No more stranded paving. A road is routed to a town, so its last stretch runs across ground the town is then drawn on top of — and the wall and gate could leave a grid of paving stranded just outside them, with nothing but wall, door and street around it. Nought to three grids across forty-eight towns, which is rare and is still a road that goes nowhere. They are taken up now rather than joined up, because the town is standing where the rest of that road used to be.
3.21.2 — The sidebar says what kind of place you are in, and whether it has fallen. Under the name:
village,town,city,great city— and in red,city, takenfor one held by monsters,town, wildfor one the animals have taken back,village, emptyfor one that simply stands there. A fallen town keeps its shops but no services at all, so it reads as a large well-supplied city with no inn, no healer and no magetower and nothing said why; the only clue was its name coming from the lawless list, which no player should have to know. Reported twice from play, once as “there is no mage tower in this town” and once as “not many stores in this great city, no inn, no healer” — both correct behaviour, both invisible. The line was added at 3.21.0 and never actually appeared: the sidebar drops rows whose priority exceeds the terminal height less two, and it had been given the lowest priority of all.3.21.1 — Your world is your world again. Adding the magic fractal at 3.14.0 put one more draw into the middle of the shared world-seed stream, which shifted every draw after it — so every existing character’s world was quietly replaced: different rivers, different towns in different places with different names. The world is never saved (it regenerates from the seed) and what you know of it is stored by name, so a character who had walked to a great city called Helgram was loaded into a world with no such place, and the visit was discarded. That is why the magetower took your money and carried you nowhere: the destination it offered had ceased to exist. Magic now draws from a stream of its own and the shared one is left exactly as it was, so worlds made before 3.14.0 come back the way they were. Characters made between 3.14.0 and here will shift once more, back to their original world.
3.21.0 — A road out of a gate goes somewhere. Reported from play: the road leaves the town and just stops. It did. The approach paved three grids out of every gate and then went looking for the road network to join, so every gate the network did not reach kept a stub pointing into open country — measured, 147 of 508 gates over six worlds, better than one in four. A town has four gates and the roads commonly reach one or two, so this was never rare, and it is worse than having no road at all: a road is a promise that it goes somewhere. Nothing is paved now until the join is found. Also, the sidebar says what kind of place you are in — village, town, city or great city, under the name. How big a place is says what is in it: a village keeps three or four trades and no services at all unless it is the one you started in. The name alone said none of that, so standing in a village wondering where its magetower was is a question the screen can now answer.
3.20.0 — The game ends in Amber now. Angband is won by killing Sauron and then Morgoth; ZangbandTK is won by killing Oberon and then the Serpent of Chaos, at the bottom of the Courts of Chaos. That is Zangband’s own replacement, and it was already the right one — the Serpent is Zelazny’s, and the bestiary’s description of it has been sitting there all along: “The Unicorn of Order fought with Serpent and stole one of its eyes, known as the Jewel of Judgement. With the Jewel, Dworkin drew the Pattern.” The Unicorn went in yesterday; this is the other half of her. Quests now also name the dungeon they are in, because a depth no longer names a place: the Courts run 75 to 110 and the Abyss 90 to 127, so depth 100 is two different places and the ending would otherwise have been completable in the wrong one.
3.19.0 — M6 begins: quests have a lifecycle. Untaken, taken, complete, finished — four states, saved with the character. Angband has two quests, both alive from birth, and records one as done by zeroing the depth it lives at; that says everything about a quest which can only be finished and nothing about one you accept, carry around, and hand back. Nothing is player-visible yet: this is the floor the rest of M6 stands on. One existing rule did change, though — winning now counts only the quests the game ends on, because “you have no outstanding quests” stops meaning “you have finished the game” the moment a townsman can give you an errand.
M5: towns, roads and dungeons — 19 to 24 August 2026#
3.18.0 — The Unicorn of Amber. Silver-white, watching you with an interest you have not earned, and always about to leave — there is one of her in the world. Walk into her and she makes you whole: not your wounds only, but everything the town healer sells and charges for, at once and for nothing. Once, and then she is gone about her business. She carries the same flag the white deer does; being unique is what makes the blessing the greater one, which is the difference itself rather than a second mechanic. The first content here written from the Amber novels rather than filtered out of Zangband.
3.17.0 — CNT-20: not everything is there to be fought. A white deer stands too still and too unafraid of you; walk into it and it heals you — everything, whatever you were down to — then bounds away ten grids and carries on grazing. Once per deer, and the deer remembers, savefile included, so following it about does not work. It will not fight back, carries nothing and is worth no experience. They are uncommon and live in quiet country, so finding one is luck rather than a plan. Neither Angband’s nor Zangband’s — it goes in a third bestiary file so its provenance stays obvious. Also: the in-game manual gains a Towns and services page, which is where the inn’s dreams, the shop quality ladder and the one-house rule now live properly rather than crammed into the symbol reference.
3.16.0 — PLR-41: the inn’s dream, which closes the last open item from DEC-32. Zangband’s inn carried a nightmare that reached the sanity blast we dropped; the dream survives it. Sleep and you may see a true dream, which puts the nearest place you have not found onto the world map — it tells you where to walk and does not carry you there, because the magetower goes only to places you have actually stood in. Or you may have a dark dream, and be hunted through your sleep by something you have met: hold your nerve and you merely remember it, fail and you wake frightened or confused. Which you get depends on the law of the town, so a well-ordered city is a good place to sleep and a frontier town is not. The constraint from DEC-32 held: no insanity, no amnesia, no mutation trigger. It is also the exact mirror of the lotus — that takes places off your map, this puts one on.
3.15.0 — PLR-40, and the first thing in this project that is not a port of anything: the lotus. A mushroom that looks like any other until you have eaten one, and five turns after you do, you forget everything — the map under your feet, the world map and every place on it, every monster you had learned, what all your potions are, and every spell you knew. It takes nothing you cannot earn back and nothing comes back quickly, so the cost is hours rather than a character. You always still know where home is, because the magetower only travels to places you have found and a character who had forgotten everywhere would have nowhere to go. The idea is Zelazny’s: the first Amber novel opens on a man with no memory who knows only that there is a place called Amber and that he belongs to it. Also fixes a test that failed about one run in five for a correct reason — it walked the window and asserted the other axis held still, which is only true if the character did not start near the window’s edge, and block alignment decides that.
3.14.1 — The manual now says that there is only one house. Every town keeps a home and they are all the same one, so what you leave in your village is on the shelf when you open the door in a city four days’ walk away. Asked from play, which is fair warning that the machinery reads the other way: a shop in another town is another shop and is stocked afresh. A house per town would strand your spare gear wherever you happened to be standing when you outgrew it, so the house travels with you and the only cost of leaving is the walk back.
3.14.0 — WLD-15 and WLD-16a: shops come at a standard as well as a trade. A plain Weapon Smiths, or an Advanced, Expert or Arcane one, written above the door. A better shop draws its goods from the deeper end of what that trade sells, puts better magic on them — about three times the plusses at the top rung — and keeps a fuller shelf. Which rung a town’s shop gets is scored from the country around it, and mostly from magic, which is a new fourth parameter on the world map: population and law were already there, but towns are sited on law, so law had nothing left to say. Seven shops in ten are plain and one in fifty is arcane. Zangband hand-authored 113 building types to get this, about 73 of them rungs on a ladder; here it is three records in
quality.txtand a score. One thing you may notice as a side effect: a trade’s shelves are restocked when you carry your custom to a different town, so towns no longer all show the same stock. Walking out and back in does not re-roll it.3.13.2 — Roads are three grids wide. One grid was invisible where a road turned a corner in the block you were standing in, and a road that turns read as a road that ends — reported from play as a road that “appears to end at the beach” after a long walk. Corners are squared off so a turn looks like a turn. Also: a service building could be promised and then demolished. The generator places services on lots off the streets and the ruin pass that follows skips lots that already have a building — but it only recognised shops, so a ruin could be built on top of a freshly built magetower. That is what made the magetower go missing from towns that said they had one: about one service in ten, and one in six in a village. Services also now have a street cut up to them, which previously only shops did.
3.13.1 — The healer’s “restore your memory” is now “restore your lost levels”, which is what it actually does — Angband calls it life force, and memory was another game’s idiom. The manual explains what drains it and what losing it costs you.
3.13.0 — WLD-16c finished: the inn, the healer, the magesmith and the recharger join the magetower. A town has an inn, a healer and a magetower; a city adds the magesmith and the recharger; a town that has fallen has none of them. All of them charge, with the prices in
constants.txt. The inn is the one that earns its keep here more than it would in Angband — daylight is what reveals the overworld, so a bed until morning beats waiting a hundred turns at a time.3.12.2 — The magetower’s menu had to be dismissed once for every shop you had visited that game. Its handler was registered where the game re-registers it on leaving a store, and nothing removes handlers before adding them, so they stacked.
3.12.1 — Every town above a village keeps a magetower, unless it has fallen. It was scored on population and law before, and the scoring turned out to put every town just under the threshold — so only cities had one, and nothing said so. Eleven of twelve towns on a measured world now have one.
3.12.0 — The status line says
DayorNightabove ground. Angband never said, because its surface is one town-sized level; here daylight is what reveals the country, and at night you see as far as your lamp — which is indistinguishable from a broken map if nothing tells you the hour. Both cheats and the debug commands are now documented, in the manual and in the in-game help as page (e).3.11.2 — Your home village always has a magetower now: travel was unusable from the one place every journey starts. The status line also refreshes the place name as you walk out of a town, where before it kept showing the town you had left. And the gold cheat is in the cheat options screen (
=thenx) as well as the debug menu.3.11.1 — Three things found by playing. Roads reached towns but often stopped at a blank wall with the gate up to twenty-six grids along it, which is a dead end from where you are standing; the gate is now cut where the road arrives, and the last stretch of road is drawn up to it. The magetower did not offer your own village, because the visited flag was only set by taking a step and your steps were taken before the flag existed. And
^A$puts gold in your pocket, for testing things that cost money.3.11.0 — The magetower (WLD-16c), and with it the mechanism services are built on (WLD-18): a building is a door with behaviour behind it. Walk into a tower and the mages will carry you to a town you have stood in or a dungeon mouth you have seen, for a fare that rises with the distance. About a third of towns keep one — never a village, never a town that has fallen. The savefile’s wilderness block goes to version 5; all twenty existing saves still load.
3.10.0 — Towns have names, and the status line shows the one you are standing in. Where Angband said
Townat depth zero — which here meant the whole world, market square and empty moorland alike — it now names the place, or sayswilderness. A place in governed country and a place that has fallen draw their names from different lists, so what somewhere is called tells you something before you arrive.3.9.0 — Towns differ in who lives in them (WLD-11): about three in five have people, one in four has been emptied and taken back by animals, one in ten is held by monsters, and one in twenty stands empty with its shops intact. Law decides it, and a large town holds out longer than a hamlet. Your home village always has people in it. The download page also caught up with the fact that there are Windows, Linux and Nintendo builds now, not just a macOS image.
3.8.0 — Every dungeon mouth now has a road to its door. Six of the thirteen happened to sit on one; the rest were between eleven and sixty-two blocks away, which is up to a thousand grids of open country to search with nothing to follow. Better siting could not fix it — the deep dungeons belong a long way from any town — so the road goes to them instead.
3.7.1 — A code review of the whole milestone, and it found that eleven of the thirteen dungeons could not be entered: the depth clamp treated “above the top of this dungeon” as one case when it is two, so stepping onto the mouth of the Abyss asked for depth 1, found it above the top of 90, and returned the surface. Also: a dungeon’s inhabitants followed the player home into the open country, the object theme was defeating Angband’s rejection of unreadable spellbooks, and the fix for walls behind trees only worked when touching them. Fifteen findings in all.
3.7.0 — Each dungeon is now home to its own kinds of creature (CNT-05). The Caverns of Kolvir run to trolls and giants, Tir-na Nog’th to wraiths and vortices, Arden to animals and trees. A stranger from elsewhere turns up now and then, at rather less than a fifth of what is met.
3.6.0 — Each dungeon yields its own kind of treasure (CNT-12). Rebma is rich, Garnath is not, the Grove of the Unicorn runs to magic and the Caverns of Kolvir to tools. What a dungeon yields changes; how much does not.
3.5.1 — The in-game manual learned the dungeons: a page of its own listing all thirteen with their depths, the rule that a dungeon ends at its bottom, and how recall behaves per dungeon. The symbol list gained the
>of a dungeon mouth out in the open.3.5.0 — Thirteen dungeons (WLD-14), Amber’s own places, each covering a range of depths and ending at the bottom of it. Each remembers how far down it has been explored, so recall returns you to the dungeon you were last in. Each has its own floor and its own preferred layout. The savefile’s wilderness block goes to version 4; older saves still load.
3.4.1 — A town wall you were standing next to was invisible if you were standing in trees. Angband lights a wall only if the grid between it and you carries light onto its face, which assumes anything blocking sight is a wall nobody can stand in; ZangbandTK’s trees are passable and block sight, so the grid you occupied was judged to block the light. Only the stretches of wall with grass in front of them lit up, which is what made it puzzling.
3.4.0 — What you know of the surface now survives a save made while you are underground, not just a dungeon trip within one session. The savefile’s wilderness block goes to version 3; older saves still load.
3.3.1 — The in-game help caught up with the game: it points at zangbandtk.com rather than Angband’s manual,
Mis described as showing the world map out of doors, and the symbol list learned the overworld — grass, trees, water, mountainside, road — along with a note that not every town holds all eight shops.Also in 3.3.1: the overworld window scrolls rather than jumping the view across. A rebuild re-anchored both axes when only one had asked for it, and threw the panel away and chose it afresh, so a long walk west appeared to drop the character a dozen rows they had never walked.
3.3.0 — Roads (WLD-08). The towns are joined by routed roads that follow the valleys and go round the mountains; every town is reachable from the starting village along them. The overworld map is also remembered across a trip to the dungeon, which it was not before: coming back up put you in a town you had to learn again.
3.2.1 — Three faults that made a town away from home unusable. Larger towns came out as empty fields inside their walls, because every building lot was clamped to the starting village’s size; no town was drawn at all once the window no longer covered home; and about one town in fifty lost its down staircase to the gate-cutting.
3.2.0 — A dozen places in four sizes (WLD-10, WLD-11, WLD-11a, WLD-12), keeping different trades according to the country around them. The starting village became the smallest place in the world, deliberately: see DEC-29.
Releases gained a Windows zip beside the disk image: a 32-bit build, packaged with the manual and a
README.txtfor SmartScreen. Windows had been dropped from the release as untested; the mingw cross build, MSBuild, MSYS2 and Cygwin all pass now.
3.1.1 — the first release, 18 August 2026#
Tagged at the end of M5’s first day, so everything below it shipped in it: M0 to M4 complete, and M5 begun. CI builds the release from the tag.
M5 begins: the world map and the front door — 18 August 2026#
3.1.1 — Four faults: the Visual Studio project had never been told about the wilderness source, so the one build system that lists its sources by hand would not link; no savefile would open, because three additions to the wilderness block had all been left at version 1, so its shape could not be told from its version; and the world map and the menu each had one of their own.
3.1.0 — An overhead map of the world (WLD-25).
Mout of doors draws the whole world, one character to a block, from the same table the ground is drawn from, so what the map calls forest is what you walk into. It pans rather than scales, because squeezing 129 rows into 22 would lose the coastlines. In the dungeonMis still the level map.3.0.0 — The game’s own name and version on the title screen and in the About panel, in place of Angband’s and in place of the
git describestring the build systems had been handing it, and an icon of its own, cleaned up from Zangband’s.The town has gates. The boundary is sealed and four gateways are cut, one per side, each two tiles wide with doors that swing shut, in place of the ragged holes Angband’s starburst clearing happened to leave.
Releases are cut from a tag. CI builds a disk image whose bundle is signed ad-hoc and verifies that signature before publishing, with the source archive beside it. Ad-hoc signing is not notarization, so the first launch has to be allowed by hand once — Download says how.
Fixes: the 14 issues a code review turned up before M5 started, and a window scroll that moved the player with it.
The players’ guide became this game’s rather than Angband’s, and the unused Read the Docs configuration went.
Documentation and site — 17 August 2026#
The manual became ZangbandTK’s rather than Angband’s, and moved off sphinx-better-theme (unmaintained since 2013) onto pydata-sphinx-theme.
The site is published to zangbandtk.com from the Sphinx sources on every push, replacing a documentation build that was never deployed anywhere.
The manual’s chapters were grouped rather than left as 32 top-level entries.
Renaming was scoped to the site’s identity. Angband is still named throughout the manual body where the reference is correct, and the credits and licence pages are unchanged.
M4: the wilderness — 16 to 17 August 2026#
Zangband’s defining feature, and the largest piece of work so far.
The world map: fractal terrain over a parameter space, 2064 grids square, generated from a seed and never stored.
One continuous surface, not a set of levels, scrolling as the player walks.
The town stands in the world, with roads out of it, and you can walk out.
Rivers run to the sea and lakes sit in the hollows (WLD-08).
Deep water is crossed, not walled off — it can be waded and drowned in — and the world is not known in advance.
The world ends in sea, not in a wall.
The wilderness is inhabited, and the world is Zangband-sized (CNT-05).
The world remembers what you left, and forgets it in time (WLD-04). A unique you wounded and left is still out there (WLD-04b).
Zangband’s own wilderness glyphs, rather than Angband’s.
Fixes: the crash on quit, the missing townspeople, and 4.2’s town rock kept. Farmer Maggot turned up three faults, one of them serious.
WLD-05 and WLD-06 were withdrawn, along with the code WLD-05 had asked for.
M3: curses and vaults — 15 August 2026#
The Ancient and Foul Curse, with its cascade intact, and random object abilities.
Pit themes and vaults — including a finding about Zangband’s vaults.
M2: the bestiary and the loot — 15 August 2026#
389 monsters imported from Zangband, including the princes of Amber and the Mythos deities, bringing the total to 1013.
51 artifacts, all of them, including Grayswandir and Frakir.
18 ego types, all of them, including
(Vampiric),(Chaotic)and(Trump Weapon).Vampiric, vorpal and chaotic weapons — three mechanics Angband has no equivalent of.
Manual chapters for monsters and objects.
M1: lethality — 15 August 2026#
The lethality scalar: every monster at 73% of Angband’s hit points and 50% of its armour class, the measured difference between Zangband 2.7.5 and the Angband it forked from. This was the first build that played as something other than Angband.
M0: foundations — 15 August 2026#
Project identity, the
tools/zconvdata conversion tooling, and the manual scaffolding.The project was renamed to ZangbandTK on 16 August 2026.