SamandClaude Fable 5.1
fd1bb10261
- a refused character login left the player on a loading screen with no world behind it and no way off — reported as sitting at character select with no message, and the symptom was worse than the report: `SMSG_CHARACTER_LOGIN_FAILED` (0x41) had a name and no parse arm, so the client never heard it, and it was not at character select when the refusal landed, because the IO thread sends `CMSG_PLAYER_LOGIN` and announces the connection in the same breath (the entry's head start, 0777), so `ClientState::InWorld` was already set and the loading cover was up waiting for a snap the server had just declined to send; the wire half parses the opcode into `ServerPacket::CharacterLoginFailed` / `SessionEvent::CharacterLoginFailed` and carries the byte RAW, since vmangos sends a bare `1` and mangos-classic its own `CharLoginFailReasons`, and a wire layer that interpreted it would be picking a server; the refusal is an entry being REVOKED, not a failure to enter — `Cycle::LoginRefused` runs the same `SessionEnd::LoggedOut` teardown a `/logout` does, the cover cuts to black, `back_on_login_refused` takes the select screen back with the relist's roster underneath, and clearing `pending_pick` is load-bearing: it is 0065's reconnect memory and the relist is auto-answered with it, so keeping it would re-enter the refused character forever from behind a black screen; the byte is a 1-BASED REASON INDEX, not a status code — wow-re read `OnCharLoginResult` (`0x5aad70`): `dec eax; cmp eax,5; ja` into a six-entry jump table mapping 1..=6 onto `CHAR_LOGIN_*` 0x3e/0x3f/0x40/0x42/0x43/0x44 and SKIPPING 0x41 `CHAR_LOGIN_FAILED`, reachable only as the default, so the tempting `0x3d + byte` is right for four rows and wrong for the rest; `char_login_refusal_text` is that table, resolved against the player's own `GlueStrings.lua` and asserted against it (vmangos's `1` therefore reads "World server is down" — the server's dialect, and what the reference would say); the select screen runs the shared `login::drive_dialog` rather than a second copy of the reference's one `GlueDialog`, which found two holes — `to_select_on_roster` closed ANY dialog on an arriving roster while the refusal's own relist produces one a second later, so it would have wiped the message (the original bug one layer up; now scoped to status/queued), and `preflight`'s descriptor latch outlived the session that armed it, announcing "the avatar never streamed in" from character select, now disarmed on any disconnect; a live probe against vmangos pins the three facts this rests on — the byte arrives as 0x01, the roster still serves after it, a retry enters the world on the same socket — and `WOW_REFUSE_LOGIN=1` refuses the first pick of a run by sending a non-player guid, the one guard that needs no server state, so a path no real account can stage by hand is walkable on demand (2071)
- the ocean's depth ramp was never a placeholder, and its deep row is darker than we drew it: four files called the ADT ocean's `V = byte/255` a stand-in "pending its own RE" — the director's report had the sea fading to deep colour about 6x too slowly — on the claim that ocean takes a non-LUT UV path via `0x68d890`, which is the MAGMA vert-fill, as `benilla_adt::LiquidVertex::texcoords` had said by that same address all along; the ocean fill is `0x68d690` and it reads a depth LUT exactly as the river fill `0x68d790` does, `0x68c4c0` building both side by side (`0xc81768 = min(d/42, 1)` river, `0xc7fcd8 = min(d/255, 1)` ocean), the two fills differing in exactly 3 of 246 bytes — one `fld` displacement — so there is no second ocean-only depth term, and wow-re's cross-check found `/255` bit-exact for all 256 entries; the divisors differ because the authored bytes do — a census over every MCLQ block on Azeroth + Kalimdor, each wet vertex's depth byte against its own geometric depth, reads a river at 8.96 byte/yd (reproducing the independently-measured ≈8.5, the control) and the sea at 1.72, 5.2x gentler, so `/255` saturates at ~148 yd and `/42` at ~4.7 yd, the same ramp in yards and not a 6x slower one, while 83.4% of ocean vertices carry byte 255 outright, so the open sea is authored fully deep and the ramp is only a shore band; a test pins both divisors against the raw MCLQ bytes (swapping the ocean onto `/42` fails it), `examples/liquid_depth_census` is the measurement with named shore-band pins, `WOW_OCEAN_DEPTH_DIV` is the A/B, and the alpha-ramp deferral closes with it — `0xc7fbc0`'s `1.6·(i/63)^8` is the sky glare falloff, bound at `0x685257` inside a branch proven dead, so the linear `127 + 2·row` alpha shipped was right; and the finding nobody asked for: `0x68a830` fills the 8×64 swatch as an exact integer accumulator in BYTE space, `row(i) = c0 + floor(i·(c1−c0)/64)`, so the ramp never reaches the deep endpoint (row 63 is ~98.4% of the way), and an ocean-only tail `[0x68a9c0, 0x68aa36)` — gated on last row AND selector 0, river being selector 1 — runs RGB→HSV, `V ×= 0.9` (`0x3f666666` at `0x8102ec`), HSV→RGB and forces that row's alpha to 255, which reduces to `floor(0.9·byte)` per channel (99.03% bit-exact over all 2^24 triples, max 1/255); byte 255 ⇒ `V = 1.0` ⇒ row 63 exactly, so the darkened row IS the open sea, and benilla had rendered deep ocean ~10% too bright over most of the water in the world — an error that survived every regression diff because no capture scenario holds any ocean; `liquid.wgsl`'s `swatch_row` / `swatch_at` port the accumulator, the LINEAR/LINEAR CLAMP sampling (so the darkening ramps in across the final 1/64 of V rather than stepping) and the ocean-only tail, on the ADT lane only — the WMO arms keep the plain lerp, their opacity being a different 256-entry ramp — pinned by a test against wow-re's published byte vectors (row 62 = 12,36,41; row 63 = 11,35,40, green one ABOVE the deep endpoint; after the tail 9,31,36) that also asserts the WGSL still carries the formula; rivers move at most 7/255 from the row quantization (2069, 2074)
- the weapon swing trail — `SpellVisualKit` CharProc type 8, which benilla had dropped on the floor with the rest of the dispatcher's unmodelled cases: 34 of 1772 kits carry it, 27 are reached from a live `Spell.dbc` chain, and 16 of those name NO emitter slot at all, so the ribbon is their entire visual — kit 324 alone reaches 218 spells (Heroic Strike, Overpower, Mortal Strike, Bloodthirst, Rend) and rendered nothing beyond the swing animation, with 506 (Revenge, Riposte, Mongoose Bite), 557 (Sunder Armor), 646 (Hamstring, Wing Clip), 399 (Sinister Strike), 822 (Garrote) and 823 (Feint, Disengage) in the same shape; the proc draws nothing — `0x60d80a` latches a packed colour and a duration on the unit, and the unit's next `PlayAnimation` (`0x5fe48e`, the fields' only reader) hands both to its two weapon-hand trail objects and clears the duration, so an arm fires exactly once and a second proc on a live trail RESETS its ring rather than stacking; each object rings 128 samples of its weapon model's `$WTB`/`$WTT` markers (455 of the 571 weapon models author both) and draws them as a strip that fades from opaque at the OLDEST end of the arc to ≈0 at the blade, the alpha being a clock and not just an opacity; `weapon_trail.rs` transcribes the arithmetic rather than approximating it — the truncating `_ftol` decode (not the small-int idiom every other integer param takes, which would yield colour 0 and no trail at all), `fadeStep`'s low-byte floor (`cmp al,1`, so a ≈768 ms hitch takes a floor of 1 instead of its own huge value), the alpha-derived termination, and the ring's writer/reader modulus mismatch (`& 127` stored, `% 127` read — shipped 5875 behaviour, reachable only by Whirlwind's 10 000 ms spin); the draw state was dispatched rather than guessed and the guess was wrong in three of seven places: the callback's six EGxRs writes say `0x07 = 2` (ordinary SRC_ALPHA / INV_SRC_ALPHA — additive is mode 3, the difference between a ribbon that whites out crossing itself and a coloured smear) and `0x10 = 0` (depth test OFF: the arc draws over the world and over the swinging character, never eaten by his own shoulder), so `EffectDrawSpec::no_depth_test` is a new pipeline-key bit on the shared effect lane with every other producer explicitly false, where the state had only existed as the global `WOW_PARTICLE_NODEPTH` debug lever; the third — lighting ON over a vertex format with no normal — is a verified state that is not yet a verified appearance and ships unlit as a named deviation with the question out; the consume edge is `AnimDriver::started_anim`, not the one-shot alone, because `0x5fe2f0` is the image's SINGLE animation entry point, locomotion included — a unit that simply starts running consumes the arm, which is how Charge (kit 44, 38 spells, anim id −1) trails at all; kit 324's 600 ms duration is on screen for 1250 ms at 60 Hz, since `fadeStep` is `trunc(dt·alpha/300)` and shrinks with the alpha it decays, twice wow-re's own estimate; the warm pass gets the depth-test-off pipeline as the one combination that ships (alpha, unlit, no rasterizer settle) rather than as another factor of the cross product — 36 more warm draws per camera for pipelines nothing asks for — so the first Heroic Strike does not compile one live; and `benilla-extract charprocs` grows the trail census that found the 34/27/16 split (2076)
- a zero injury column plays nothing: benilla walked down the wound-grunt family when the selected column was 0 — crushing to critical to ordinary — on a comment with no citation ("crushing rows are often 0 in data"), a true observation about the data and a false inference about the client; `0x623490` tests the selected id once (`0x6234e6 test ebx,ebx ; je 0x62350e`, the shared epilogue) and no fallback is even reachable — the row pointer dies at `0x6234e4`, so neither the bus pick `0x623b10` nor the play `0x458890` ever sees the row to re-read a column from, and the CGPlayer twin `0x62f880` is identical; it mattered because the class-2 chance roll had just landed beside it (2073): a crushing blow skips the roll (class 9 carries threshold 100) AND fell through to the ordinary kit, so it was louder than an ordinary hit on the same creature by two routes; `sound/combat.rs` now plays the selected column or nothing — crushing → column 5, critical → 4, ordinary → 3 — and a crushing blow on a creature with no crushing kit, which is most of the 406 rows (`Skeleton`, `Silithid`, `MountainGiant` among them), is silent, the reference's behaviour rather than an omission (2075)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Important
Issues and pull requests are closed here. benilla is a solo project developed in a private tree; this repo is its export, published as squashed snapshots, so a PR here has nothing to land on. The best way to contribute is to join the Discord and report the bugs you find. Questions and ideas are welcome in the same place.
benilla speaks the original 1.12.1 protocol, so it connects to any server the real client could, and reads its game data at runtime from your own 1.12.1 install. Every file format and the network protocol are implemented from scratch, with no original client code, no third-party WoW crates, and no bundled game assets.
What works
- Formats: readers for the full asset stack (MPQ patch chain, BLP, DBC, ADT/WDT/WDL, M2, WMO), wired into Bevy as an asset source.
- World: streamed terrain out to the horizon, portal-culled WMOs with interior lighting, doodads and ground clutter, swimmable liquids, sky and weather, and the client's own day/night lighting, fog and gamma passes.
- Models: GPU-skinned M2s with the full animation controller, a near feature-complete particle system, ribbons, and animated gameobjects from doors to lifts.
- Characters: customization end to end, the armor texture composite, weapons with sheathing and enchant glows, shapeshift forms, stealth and mounts.
- Movement: a WoW-feel controller, networked movement in both directions, the server-granted modes from slow fall to roots, a follow camera with collision, boats, zeppelins and taxi flights.
- Networking: SRP6 auth through world-session crypto, the object mirror into the ECS, and live wire coverage from movement and chat through spells, party, quests, mail, trade, vendors, bank, loot, the auction house and PvP honor.
- UI: a from-scratch FrameXML + Lua engine driving the built-in interface, from the login and
character screens through the full HUD, the classic windows (guild, macros and key bindings
included), chat, nameplates, floating combat text and tooltips; third-party addons load from
a
benilla-config/AddOns/folder beside the executable (partial: AtlasLoot and Bagnon run, see below). - Combat: melee on the faithful swing law, ranged and Auto Shot, casting with GCD and cooldowns, combo points, crowd control that really holds you, and the spell visual pipeline.
- Audio: music, ambience and SFX under the client's own selection and crossfade rules, with interior and underwater transitions and zone reverb.
Where it's going
benilla is done when a 1.12.1 player can do everything here that they could in the original client, it looks and feels the same, and it runs from a download on Windows, Linux and macOS. No dates; the order is what is likely, not a promise.
- Battlegrounds and meeting stones, then the long tail of small features that separates a working client from a finished one.
- Addons, options and performance, ongoing.
- The no-brainer fixes from VanillaFixes, SuperWoW and the like.
- Playable downloads for Windows, Linux and macOS. Linux first.
Not planned: other expansions or client versions, Warden (anticheat).
Running it
You need a 1.12.1 (build 5875) client install for game data, a vanilla server to connect to, and stable Rust. Any 1.12.1 core works; vmangos is what development runs against, and cMaNGOS and the rest speak the same protocol.
WOW_DATA=/path/to/WoW/Data cargo run --release -p benilla
The server defaults to localhost:3724, the stock realmd auth port. Point WOW_HOST
at any IP or hostname, appending the auth port if yours is remapped
(WOW_HOST=play.example.com:5000). Credentials go in at the login screen, or set WOW_USER /
WOW_PASS to skip it.
Early inspiration and file format guidance came from the wowemulation-dev community, and warcraft-rs in particular.
benilla is an independent fan project, not affiliated with or endorsed by Blizzard Entertainment.
It ships no Blizzard content — no art, models, sounds, maps, MPQ contents or FrameXML; you
provide your own legally obtained 1.12.1 client. The interface code under
crates/benilla-app/assets/ui/ is ours, written to the client's own layout and API names so that
the windows look right and 1.12.1 addons find the names they expect.
World of Warcraft is a trademark of Blizzard Entertainment, Inc. Our own code is licensed under MIT or Apache 2.0, at your option.