statico's NetHack Tools

Player tools for NetHack 3.6, 3.7 and 5.0.

• Site: nethack.statico.io
• Source: github.com/statico/nethack-tools
• Contact: @statico on Discord — bug reports and corrections welcome, especially if you can point at the line of C that proves me wrong.

Also by me: the NetHack Tiles Client, a desktop app for macOS, Windows and Linux that logs into nethack.alt.org or hardfought.org over SSH and renders the game with tiles instead of ASCII. It's linked from the home page — as its own block rather than a seventh card in the tools grid, since it isn't part of this site and the grid is pinned to three across.

Identify unknown items from their shop price. The reason this exists: no calculator on the wiki has been updated for 5.0.

• Exact integer math. A direct port of get_cost() and set_cost() from src/shk.c, including the round-half-up tweak (((tmp * 10) / divisor) + 5) / 10, and the artifact ×4 and angry-shopkeeper tmp += (tmp + 2) / 3 surcharges that are applied after rounding. That ordering is what most calculators get wrong.
• Reverse lookup by brute force. Every base cost from 1 to 30,000 is run forward through the algorithm across both branches of the unknowable 1-in-4 ID surcharge, and exact matches are collected. Inverting the arithmetic analytically gives wrong answers because of the rounding tweak.
• Two-observation intersection. Enter a buy price and a sell price — or prices from two shops at different Charisma — and intersect the candidate sets. This usually collapses the answer to one item.
• Probability ranking. Candidates ranked by how often they actually generate, using the prob field from the object tables.
• Quick-pick prices for the scenario you're in. The twelve base costs backing the most objects in your enabled classes — 4, 5, 7, 10, 20, 50, 60, 100, 150, 200, 300, 500, the same ladder in both versions — run forward through the same algorithm under your current Charisma, direction and modifiers, so the chips show what you'd actually be quoted. At Cha 8 they read 5/7/9/13/27/67/80/133/200/267/400/667; at Cha 18, 3/4/5/7/13/33/40/67/100/133/200/333. The list is derived from the object table, so it tracks the version switch, the enchantment field and the class filter rather than being a hardcoded list that drifts.
• Charisma defaults to 11, the bottom of the 11–15 band where the ladder applies no adjustment at all, so an untouched form prices at face value. The hint under the field names the band you're in.

It reports honestly: when six items share a base cost, it says six.

Build a wish string and find out what the game will actually give you.

• Canonical output. Adjectives emitted in the order readobjnam() in src/objnam.c parses them, so it pastes straight into the game.
• The annotation is the point. For each part of the wish it says whether the game honors it, silently drops it, clamps it, or rolls for it — with the file:line that proves it. The rnd(5) enchantment test zeroes your enchantment rather than reducing it, so +7 is never granted; charge counts are capped by whatever mksobj() already rolled; negative Luck turns "blessed" into cursed and quietly discards erodeproof and poison.
• Artifact odds. Failure is rn2(nartifact_exist()) > 1, so the first two artifacts of a game are free and after that it's (n−2)/n.
• Enchantment survival probabilities are computed against the real mksobj_init() distributions, not assumed.

Every monster from the source tables, with a threat filter.

• What can spawn on me here? Enter experience level and depth and it runs the game's own rndmonst() logic from src/makemon.c — the level_difficulty/6 too-weak floor, the (depth + XL)/2 ceiling, the G_NOGEN / G_UNIQ / Gehennom exclusions, uncommon(), align_shift() and temperature_shift(). The percentages shown are exact selection probabilities, not estimates.
• Full attack lists, resistances held and conveyed, and the M1/M2/M3 flags decoded.
• Dangerous attacks are color-coded off the AD_* damage type — a table fact, not an opinion about which monsters are scary.

Every branch on one vertical map, parsed from dat/dungeon.def (3.6) and dat/dungeon.lua (3.7 / 5.0) — two unrelated file formats.

• Branches attach at the exact level ranges their entrances can generate on, with a bracket showing each range.
• Click any dungeon, branch or special level for its level range, entry rules, flags and map-variant count.
• Depth is random, so the diagram is drawn at each dungeon's minimum depth and bottom-anchored levels are labelled symbolically (levels N−4 to N−1) rather than pinned to a number they don't have.

The dungeon files say nothing about shops, temples or altars, so neither does this page.

All eight vanilla levels, with maps rendered from dat/soko*.lua rather than copied from a screenshot.

• Clickable level previews with the solution video from each level's wiki page, loaded via a click-to-load facade so eight iframes don't load at once.
• Horizontal and vertical flip toggles, because 3.7 and 5.0 mirror Sokoban levels at generation — so the reference map can be made to match what's actually on screen. 3.6 never flips. The toggles are repeated inside the level panel, so you can re-orient without scrolling back to the top of the page; both sets stay in sync.
• The walkthrough video mirrors with the map. Maps are flipped as grid data — the text stays selectable — but a video can only be flipped as pixels, so the embed is the one place a CSS transform is the right answer. Only the media is transformed, not the surrounding box, so the play button doesn't become a backwards triangle. A mirrored YouTube player has a mirrored control bar too, so there's an Unmirror video escape hatch. Flipping while a video is playing patches the DOM in place rather than re-rendering, so the video keeps playing instead of restarting.

Search over every “You feel …” and “You hear …” message. Nothing is shown until you start typing.

• 302 messages scraped from the wiki's You feel and You hear articles, each one then handed to a subagent that grepped NetHack-3.7 and NetHack-5.0 for the literal C string producing it, recording the exact file:line on a match.
• 190 more messages found the other way: a source sweep read every You_feel()/You_hear() call site in 5.0 and kept the ones with no matching wiki article.
• 3.7 backfilled deterministically. The reverse sweep only reads 5.0, but a 5.0 citation hands over the exact C string literal, and finding that literal in the 3.7 tree is a plain search — backfill_3_7() in gen-messages.py does it, resolving 188 of those 190 against 3.7. Same file and nearest line wins when a literal appears at several call sites; ambiguous across files stays unverified rather than guessed.
• Honest verification coverage, not a claim of completeness. Of the 492 total: 3.7 is 484 confirmed present, 3 confirmed absent (the wiki's wording doesn't match anything in the source — shown with the closest text that was found), and 5 unverified. 5.0 is 486 present, 3 absent, 3 unverified. Unverified means the call site is pure format strings with no literal to search for, not that the message is missing. Every row shows its status per version rather than a single pass/fail.

Goals, ascension kit, and a death log, saved to localStorage. Nothing is uploaded.

• Every item name and mechanic links to its wiki article. All link targets across the site — 331 of them — are validated against the wiki dump by scripts/check-links.py; the build fails on a dead link.
• Prayer timer. Tracks turns since your last prayer against the real mechanic — the timeout starts at 300 and resets to rnz(350), which is heavy-tailed. So it shows a gradient (risky / marginal / safe) rather than claiming a guarantee it can't make.
• Export and import as JSON.

Things worth recording, verified against NetHack-5.0 @ a8a13bed8:

• 3.7 and 5.0 have identical object tables. include/objects.h differs between the two branches only in header comments. They share one dataset here, labelled as both.
• 5.0 vs 3.6 adds amulet of guarding, amulet of flying, gold dragon scales and scale mail, helm of caution, shields of drain and shock resistance, silver mace, spellbook of chain lightning, and wand of stasis; renames huge chunk of meat to enormous meatball and the venoms to splash of ....
• In 3.6, every ring has prob: 0 in the C table. That's genuinely what the source says — 3.7 changed it to 1. Ring probabilities are shown as uniform for 3.6.
• 5.0 weapon probabilities sum to 1002, not 1000 — silver mace was added without rebalancing. Denominators come from the data, never a hardcoded 1000.
• Unidentified gems sell for a flat 3–8 zorkmids, derived from the shopkeeper's id. A gem's sell price carries no information at all.
• Sokoban floor 4 has no up staircase. It's the top of the branch, not a missing glyph.
• The wiki's level letters are inverted relative to the source files on floors 1–3: wiki "Level 1a" is soko4-2.lua. Floor 4 is not inverted.
• Rolling-boulder traps render as floor, not ^ — they stay hidden on a premapped level. Some wiki maps draw them as ^.
• 3.7 and 5.0 are identical in more places than expected. Beyond the object tables, the monster tables and dat/dungeon.lua are also byte-identical apart from comments and version stamps. Each page says so rather than implying its version switch reveals a difference that isn't there.
• The Rogue level is defined in all three versions, not just 3.6. It has no .des/.lua file because mklev.c generates it procedurally.
• 3.6 has no mstrength(). makedefs -m was deprecated there — do_monstr() prints a notice and emits a stub — so 3.6 difficulty comes only from the mons[] table. 3.7/5.0 compute it in src/mondata.c, and the formula disagrees with the shipped table for exactly two monsters (cleric and wizard), both G_NOGEN player-monsters where hand-set difficulty is explicitly allowed. The table value wins; the discrepancy is asserted in a test.
• Horned devil, erinys and barbed devil can never be randomly generated in any version. G_HELL excludes them outside Gehennom, and inside it uncommon() rejects maligntyp > A_NEUTRAL. They only arrive via demon summoning.
• Cross-alignment and crowning do not restrict artifact wishes. readobjnam() checks only is_quest_artifact() and the rn2() roll. The alignment field is used by mk_artifact() for sacrifice gifts, which is where the folklore comes from.
• The Palantir of Westernesse is freely wishable — it sits in the quest-artifact index range but is nobody's questarti.

Static site. No build step, no dependencies, no npm install.

python3 -m http.server 8000    # serve from the repo root

Absolute paths (/assets/smui.css) require the server to be rooted at the repo root.

The generators read a local NetHack git checkout and a NetHackWiki database dump, neither of which is vendored here.

python3 scripts/gen-objects.py     # -> data/objects.json
python3 scripts/gen-sokoban.py     # -> data/sokoban.json
python3 scripts/gen-wish.py        # -> data/wish.json
python3 scripts/gen-monsters.py    # -> data/monsters.json
python3 scripts/gen-dungeon.py     # -> data/dungeon.json
python3 scripts/gen-messages.py    # -> data/messages.json (needs messages-verification.json)
python3 scripts/gen-og-images.py   # -> assets/og/*.png (social cards; needs ImageMagick)
python3 scripts/check-links.py     # validates every wiki link target
python3 scripts/gen-og-images.py --check   # every page has a social card

node scripts/test-prices.mjs       # 44 pricing assertions
node scripts/test-wish.mjs         # 135 wish-rule assertions
node scripts/test-monsters.mjs     # 106 monster-table assertions
node scripts/test-messages.mjs     # message data + search/filter engine assertions

Every generator refuses to write output if its verification assertions fail. They locate rules by pattern in the C source and die if the pattern is gone, so an upstream change breaks the build loudly instead of silently producing stale numbers.

data/checklist.json is authored by hand, not generated — but its link targets are validated like everything else.

scripts/gen-og-images.py reads each page's own <title> and og:description to draw its 1200×630 card, so the card can't drift from the page it represents — edit the page's metadata and re-run it. Cards are drawn with ImageMagick's drawing commands rather than SVG, since the usual rsvg delegate isn't assumed to be installed.

scripts/gen-messages.py is the odd one out: locating each message's C source line is unbounded reading, not a pattern match, so that half is done by fanned-out agents and checked in as scripts/messages-verification.json rather than reproduced by the script itself. gen-messages.py --extract-only regenerates the deterministic half (messages-raw.json, not checked in); the full run merges that sidecar in.

Add the page, then add one entry to TOOLS in assets/nav.js. Every page picks it up.

Styled after SMUI — Nord palette, JetBrains Mono, square corners. The whole system is assets/smui.css; pages use its classes and its CSS custom properties, and don't define colors of their own.

Light and dark are both first-class. The switcher in the nav cycles system → light → dark; system follows your OS live, and an explicit choice persists in localStorage. assets/theme.js is loaded synchronously from <head> so a stored choice is applied before first paint rather than flashing the wrong theme. color-scheme is set per theme, so native selects, number spinners and scrollbars follow too.

Nothing renders below 12px, including the JS-scaled Sokoban maps.

Every numeric field keeps its native up/down spinners on screen. WebKit hides them until hover or focus, which makes them undiscoverable on values people nudge one at a time — Charisma, dungeon level, enchantment — so smui.css forces them visible in both themes.

.stack spaces its children with margin-top, which has two consequences worth knowing before you copy a pattern: a gap set on a .stack does nothing, and a child carrying its own margin: 0 silently cancels the spacing. Both were happening. Use .stack-tight where the spacing has to survive that — it is a real flex column, so gap applies and no child margin can override it. Where an element only means "no trailing space", write margin-bottom: 0, never margin: 0.

Almost nothing here is original. The work was reading other people's work carefully and wiring it together.

The NetHack DevTeam and everyone who contributed to it since 1987. All game data on this site — every base cost, generation probability, Sokoban map, and the entire pricing algorithm — comes from the NetHack source, specifically:

NetHack is licensed under the NetHack General Public License.

NetHackWiki and its contributors, licensed CC BY-SA 3.0. Used for:

• Sokoban level notes and per-level strategy, condensed from the Sokoban page and the eight level pages
• The video links on each level page
• Every reference link across the site — 331 targets, validated against the wiki database dump
• Human-readable dungeon and level names, and the article titles they link to
• Cross-checking the price calculator against Price identification

Note that this site's numbers are derived from the C source, not transcribed from the wiki's tables. The wiki was the check, not the input.

All eight are from the Concise Nethack series by Larry Fluckiger, linked from the NetHackWiki level pages. They are embedded here with credit and all rights remain with the author. Nothing is loaded from YouTube until you press play.

If you're the author and would rather not be embedded, open an issue or find me on Discord and I'll switch to plain links.

• SMUI — the design system this is styled after
• Nord by Arctic Ice Studio — color palette, MIT
• JetBrains Mono by JetBrains — typeface, SIL Open Font License

MIT for the code written here (assets/, scripts/, the HTML). Data in data/ is derived from the sources above and carries their licenses.

Built by statico, mostly by pointing Claude at the source. Not affiliated with the NetHack DevTeam.