This checkout had no git history at all, so this one commit carries the whole
of the Everwood work as a single import on top of chairmanpow/benilla-octo's
`main` (706cd71). The two trees have genuinely diverged — each has files the
other does not — so this is a fork point, not a catch-up.
Today's work, which is the reason a baseline exists at all:
- Booth previews no longer cast world shadows (`NO_BOOTH_SHADOW`). Portrait
characters were entering the sun's shadow cascades, so every skinned-rig
shadow specialization compiled live mid-frame; the hitch detector logged
"main thread blocked 1316 ms" beside a `compiled LIVE` warning. This was
the Fashion/Tmog stutter.
- Per-pane dress-up: each `DressUpModel` pane assembles and stages its own
tried-on look instead of every pane mirroring the live player.
- Item template cache is append-only JSONL written on the IO pool, replacing
an 11.6 MB whole-map JSON rewrite on the main thread every 10 s.
- `WOW_SYSTEM_TIMING` rebuilt per-thread: it took 2 mutexes per system enter
and 4 per exit, costing 15 ms a frame (590 fps -> 60) and corrupting the
measurements it was taken to produce.
- The visibility-churn sweep moved to its own `WOW_FLASH_CHURN` switch, so a
frame-cost reading no longer pays 2.7 ms a frame for it.
Local state (`benilla-config/`, `logs/`, `target/`) is excluded by the
existing .gitignore; the account credential it holds is not in this commit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SgstFWtn1RxEJ5kVrpgYwn
6.5 KiB
METHOD — how benilla is built
benilla is a from-scratch World of Warcraft 1.12.1 client in Rust and Bevy. It reads the game's data from the player's own install, speaks the original protocol to any 1.12.1 server, and does not build a server or simulate the game. The reference client is the spec: benilla does what 1.12.1 does, as a modern, idiomatic client.
The rule for every change
A change is right when it makes benilla more like 1.12.1 or fixes a bug, with evidence from the
reference, in one small piece, with the gates green. A deviation from the reference is the
maintainer's call alone, and each one is written down where it lives: a CVar's Deviates row, a
comment naming the reference fact and why we differ. Anything else is a fork.
The loop
- The reference first. For anything fidelity-defining, get the mechanism from the real
client before building: its behaviour, a DBC field, a FrameXML line, a packet capture. The
names and shapes under
reference/are the surface benilla tracks. - Build it idiomatically. Modern Bevy and Rust, performance woven in from the start. Implement the mechanism well; do not ape a quirk that is a bug.
- Measure, never eyeball. Timing, delay, ordering and feel are settled with instruments: trace timestamps, byte-derived numbers, a numeric probe. A capture confirms an existence fact; it cannot measure time. Never search for a viewpoint: take the one simplest shot, and if it does not settle the question, hand over a precise script ("log in, look at X, does Y?").
- Close the loop on the symptom. A bug reported as seen is closed by reproducing it at the reported spot first, fixing, and showing the same instrument clean there plus one adjacent state. Green gates and an existing mechanism are never that evidence.
- Prove the run before reading the result. A live result is evidence once the run is valid from numbers: preflight banner clean, camera and body where intended, the subject in frame, the window long enough. A negative from an unproven run is not evidence.
- Commit small and often; land by pull request. Atomic commits, explicit paths, on a branch; main moves only by a squash-merged pull request, one commit per piece of work. No "done" and no commit without saying what was verified and how the result was judged.
Hard rules
- Never commit anything from the install. Art, models, sounds, maps, data: the install is
read at runtime from a gitignored path, and everyone provides their own. The one exception is
interface code: FrameXML and GlueXML run off the player's own patch chain, and our own
counterparts under
assets/uistay until they retire. - The install is read-only. benilla never writes into the WoW folder: no screenshot, log,
cache or scratch file.
scripts/smoke.shfails a run that leaves the install changed. - Local state lives in one folder,
benilla-config/, at the repo root in a dev build and beside the binary in the player build; every path to it resolves throughcrate::local_state. Player settings are CVars persisted as a diff inbenilla-config/config.toml. - UI is stock-first: a window is migrated, not authored. The end state is the stock 1.12
FrameXML executed off the player's own chain.
assets/uidoes not grow: a test names its files and fails on a new one. A window is built by pointingbenilla.tocat the stock file, deleting ours and building the engine verbs it calls, never stubbing one to make the file load. - A setting's default is the reference's default. Every option boots at the stock 1.12
value. Shipping another value costs an explicit
Deviatesrow with the reason, and a test fails a row that drifts either way. - Our code is MIT OR Apache-2.0, which is a claim about provenance: no original client code, no bundled assets.
Where things live
docs/METHOD.md, this file: rules and method. Never state, never a log.docs/MAP.md: what is built, generated from the code byscripts/genmap.shat every land. Read it to orient; never edit or commit it by hand.- git: the history. No changelogs, no status docs, no "what we did" narratives.
- The code: a comment says what the code does and the 1.12 fact behind it, in a line. No history, no stories.
Gates
- The land gates:
cargo fmt --all -- --check,cargo clippy --workspace --all-targets -- -D warnings,cargo test --workspace, the doc-link and render-pass lints, the player build without thedevfeature with its own tests, and the engine boot checks, asscripts/gates.sh. It runs on the tree that lands, before it lands; it is memoized on the tree. The tests that read the install or the addon corpus skip where the data is absent, and the gate prints the count; where the data is, a skip is a failure (BENILLA_REQUIRE_DATA=1). The workspace tests then run a second time withWOW_DATA=(no install), so a test that reads the install without opening withwow_data_or_skip!fails here and not on a clone. - Per round:
scripts/check.sh, scoped to the changed crates and their dependents. Run each gate once per round;teethe output to a file and grep the file. - A clean run is the fourth gate:
scripts/smoke.sh, a live login and logout against the local server. For wire work the gate is a live run with a trace (WOW_MOVE_TRACE=<path>, tags filtered), never a capture: capture mode has no network. - Platform seams: a
cfg(target_os), a[target.'cfg(…)']dependency, a#[link]or anextern "system"is invisible to the gates of the platform you are on. Say in the commit which platforms you built.
The local server
benilla talks to any 1.12.1 server: WOW_HOST names it (default localhost), WOW_USER and
WOW_PASS together log in without typing, WOW_CHAR picks the character. There is no default
account. These bind every session and every agent:
- A scripted run logs in as the account its checkout declares in
.probe-identity(the three variables, one per line, never committed), or as the three variables name; never as a player's, because a login kicks whoever holds the account, and the client refuses a scripted login from a checkout on any account but the declared one. - Every unattended run says so:
WOW_UNATTENDED=1. An agent's run is silent:WOW_NOSOUND=1. - No unattended combat probes.
- Anything about hostility, reaction colour, nameplates, threat, damage or breath runs with
WOW_GM=off. - Read the preflight banner before debugging anything else.
The switches for such a run are docs/CONTRIBUTING.md, "Running it unattended".