Files
tidewall/docs/ARCHITECTURE.md
T
claudeandClaude Opus 5.5 05c7568ec1 N8: 24-client network soak, and the server fixes it found
make soak runs a real server and 24 headless clients in three processes for
two whole matches in real time. Each client reaches the server through its own
UdpRelay, which adds 60 ms, +/-20 ms jitter and 2% loss each way, so ENet runs
unchanged over a bad link. SoakPlayer drives each client from what a client
knows: it drops, glides, heads for high ground, harvests and tries to build.
tools/soak_report.gd checks world hashes against the server's, errors,
unreadable packets, every client playing every match with its connection
intact, memory between matches, and the p99 tick with the network's share.
make soak-selfcheck proves a drifted client world is caught.

Server changes: --matches=N, --world-hash-every=N, tick timing that includes
after_tick, and memory in match and progress lines.

What the soak found and fixed:
- Snapshot building for 24 clients took the p99 tick to 13.8 ms, over the
  12 ms budget, against a 2.8 ms simulation. Encoded records are now shared
  between clients for the tick, unchanged records skip the field diff, and
  clients are spread over the snapshot interval by handle. p99 is 7.3 ms.
- ENet peers had a timeout limit of 0, which gives up on a peer as soon as any
  reliable message has waited the minimum timeout. Now ENet's default, 32.

Not fixed: one or two clients still lose their connection about 15 s into the
second match's drop in each run (see the issue), so the soak still fails and
N8 stays open.

Refs #39

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-12 21:56:01 -05:00

32 KiB
Raw Blame History

Tidewall architecture

Tidewall is server-authoritative from day one. A match is always a server plus clients. Offline play starts a local server with bots and connects to it through the same interfaces, so there is one code path, and multiplayer is never a retrofit.

Rules that do not bend

  1. The simulation decides everything that affects the match. Clients ask; the server rules.
  2. Simulation code has no dependency on rendering, audio or UI. src/sim/ doesn't use any nodes or servers from those engine areas, and a test scans it to enforce this.
  3. The simulation advances only in fixed ticks of 30 Hz, and takes no input other than commands and the seed.
  4. Gameplay numbers live in src/sim/tuning.gd.

Simulation and presentation

          commands                     snapshots, events
Humans ─┐                                                 ┌─ Renderer
Bots  ──┼──> Server simulation (30 Hz, authoritative) ──> ├─ Audio
Net   ──┘                                                 └─ UI / HUD
  • Simulation (src/sim/): plain RefCounted objects and data arrays. Testable without a scene tree. Runs identically on the server and in prediction on clients.
  • Presentation (src/client/, src/ui/, src/audio/): reads snapshots and events and draws them, interpolating between ticks. It never changes simulation state.
  • The simulation emits events (shot fired, piece built, player downed) in a per-tick list. Presentation turns them into effects and sound.

Tick rate

The simulation runs at Tuning.TICK_RATE_HZ (30 Hz, 33.33 ms). The server runs a fixed accumulator and may run several ticks to catch up, with no more than 5 per frame. Snapshots go out every second tick, 15 a second. The client renders at any frame rate, predicts its own player and draws everyone else about 130 ms in the past.

Player command stream

A PlayerCommand is one tick of one player's intent. Humans, bots and the network all produce the same struct, so the simulation can't tell them apart.

Field Type Meaning
tick u32 Tick the command applies to
move_x, move_y i8 Movement axes, scaled to -1..1
yaw u16 Look yaw, full turn over 65536 steps
pitch i16 Look pitch, ±90° over ±32767 steps
buttons u16 Bit flags: fire, aim, jump, crouch, sprint, reload, interact, build
selection u8 Active inventory slot or build piece
material u8 Active build material
view_lag u8 How far in the past the player was seeing everyone else, back from tick, in quarter ticks. 0 for bots and local play

Clients stamp commands for the tick they predict, and each packet repeats recent commands to survive loss. The server applies the command for each tick, or a neutral command with the player's last look direction if none has arrived.

Look convention. At yaw 0 the player faces +z; forward is (sin yaw, 0, cos yaw). Right is forward × up, (-cos yaw, 0, sin yaw), matching Godot's right-handed, y-up frame, so turning right decreases yaw. Positive pitch looks up. Movement and the client's CameraRig both follow this, and a test checks that strafing right moves towards the camera's screen-right.

The binary encoding is a fixed 15 bytes per command, little-endian, in the order of the table above. See src/sim/player_command.gd. Decoding a truncated buffer returns nothing, and out-of-range values are clamped, so bad input can't reach the simulation. The player a command belongs to comes from its connection, never from the packet.

Entity IDs

Every entity has a u32 ID, allocated from a counter starting at 1 and never reused within a match. ID 0 means "none". Snapshots refer only to IDs, which removes a whole class of stale-reference bugs. Iteration is always in ascending ID order, so it's deterministic.

Seeded RNG rules

  • A match has one u64 seed, always logged.
  • Gameplay randomness comes only from named streams, created from the match seed and a stable name: rng.stream("loot"), "weapons", "drop", "worldgen", and so on, on one SimRng. One system's extra rolls can't disturb another's.
  • Bots don't draw from those streams. Each bot has its own generator, seeded from the match seed and its player ID, so nothing a bot thinks changes the simulation's state.
  • Simulation code never calls randf, randi, Time, or anything else that varies between runs. Presentation can use whatever randomness it likes, for effects.
  • The same seed and the same command log produce the same state hash, on the same engine build. Worldgen runs on both ends from the same seed, so the terrain itself is never sent.

Simulation.state_hash() is a SHA-256 over a canonical encoding of the tick, the next entity ID, every RNG stream's state and every entity in ID order, with positions rounded to millimetres.

Determinism is for tests, replays and worldgen. Clients don't rely on bit-exact matches with the server. Reconciliation corrects any difference.

Command log and replay

CommandLog (src/server/) holds a match's seed, its rules and every command the server was given: the simulation tick it arrived on, the handle it came from and the command's 15 bytes. ServerLoop.submit() is the one way commands get in, for people, scripts and bots alike, so a recording server logs them all, accepted or not. On disk it is a JSON header and the records compressed with Zstandard, about 0.6 MB for a 24-player match.

MatchReplay plays a log again: the same seed and rules, the same commands on the same ticks, and nobody thinking. It ends on the same state hash as the recording, which make match checks after every run. A replay can stop at any tick, which is how the match captures draw a moment of a real match. Replays are exact only on the same engine build, like everything else that relies on determinism.

ScriptedPlayer (src/bots/) is a stand-in for a person at the controls. It plays a human seat as a bot would, and also cuts driftwood and builds a breakwater and a wall, so a scripted match covers the drop, loot, combat, harvesting, building and the tide from a player's seat. ServerLoop.play_humans_by_script() turns it on; no shipped mode uses it.

Transport and handshake

The transport is ENet over UDP, used directly through ENetConnection (not Godot's high-level multiplayer API, so the server controls every byte it accepts). Channel 0 is reliable control traffic; channel 1 is for commands and snapshots.

NetServer and NetClient (src/net/) own the two ends, and NetProtocol the message layouts and the protocol version. A new connection is a guest until its HELLO is accepted:

  1. The client sends HELLO: a magic number, its protocol version and the name it wants.
  2. The server answers WELCOME (the player's handle, the match seed, tick rate, player limit and the name as accepted) or REJECT with a reason, then disconnects.

A guest gets one try. A HELLO that is malformed, on the wrong channel or too long is refused as such; a different protocol version, a full server and silence past the handshake timeout each have their own reason, and the client turns the reason into a sentence for the player (NetProtocol.reason_text). The reason also rides on the disconnect itself, in case the REJECT is lost. Names are cleaned on both ends: control and invisible characters removed, 16 characters at most, never empty, and a number added if someone already has it. Handles are positive, count up and are never reused while the server runs.

Leaving cleanly tells the server at once. A client that vanishes is dropped by ENet's timeout (5 to 15 seconds without an acknowledgement), and the owner sees the same "left" event with a different reason. Decoders return nothing for bad input and never raise an engine error; anything a joined player sends that isn't understood is counted and dropped.

Finding and joining games

People join a server three ways, none needing a third-party service:

  • By address. --connect=HOST[:PORT]; the game port defaults to 24650.
  • On the LAN. A listening server answers discovery QUERY packets on its game port plus one (ServerBeacon), unless started with --no-lan. A scan (ServerBrowser.scan_lan()) broadcasts a QUERY to the discovery ports of eight game ports, 24650 and the next seven even ports, so several servers on one machine are found.
  • Through a list server. Anyone can run one: --list-server [--port=N] (ListServer, default UDP 24660). A game server started with --list=HOST[:PORT] sends a REGISTER every 15 s. The list server answers an unknown or wrong token with a CHALLENGE carrying a token derived from a per-start secret and the sender's address and port, and lists the server only when a REGISTER echoes it. Servers are listed under the address their packets come from, never a claimed one. Entries expire after 45 s, with at most 8 per address and 256 in all. Clients ask for the list four servers per page, one page after another.

The messages are in NetDiscovery. Every QUERY is padded to 512 bytes and no answer is larger, so neither the beacon nor the list server can amplify a spoofed request. Both rate-limit each source address (SourceLimiter: 20 packets a second, at most 4,096 addresses tracked per second).

Lobby, backfill and late arrivals. People who connect during the lobby or the countdown take places (Match.join()); bots are only added when the match starts, into whatever places are still empty, so people joining and leaving in the lobby never displace or strand a bot. Someone who connects after the barge leaves, or during results, becomes a watcher (Match.watch()): they spectate a living player (the winner during results) through the same snapshots an eliminated player gets, and take a place ahead of any bots when the next lobby opens. The server still turns people away when NetServer.max_players people are connected.

Trusting nothing from clients

Tidewall is open source, so anyone can build a client that sends whatever bytes it likes. The server's defences, and the tests that attack them (test/server/test_hostile_client.gd, test/net/test_net_handshake.gd, test/net/test_parser_fuzz.gd, make fuzz):

  • Intent only. A command carries movement axes, look, buttons, a slot, a material and a view lag. There is no position, velocity, hit, target or item in it, so speed, teleport, aimbot-by-claim and duplication can't even be expressed. Every field is clamped on decode, and the simulation applies its own limits: sprint speed, the weapon's fire interval, build reach (Tuning.BUILD_RANGE) and material cost, pickup reach, one pickup or drop per press, lag compensation at most Tuning.LAG_COMPENSATION_MAX_TICKS.
  • One command per tick. Each tick's command is taken once, the first time it arrives. A tick already past or more than Tuning.MAX_COMMAND_LEAD_TICKS ahead is ignored, and a far-future tick can't move the dedupe marker and lock out the real ones.
  • Bounded messages. An INPUT has at most NetSnapshot.MAX_INPUT_COMMANDS commands and an exact length, or it is dropped whole. At most 32 game packets per player are kept per poll. Anything on the control channel after the handshake is dropped.
  • Bounded connections. At most 32 new connections per address per 10 s and 8 handshakes in progress per address (REASON_BUSY), a 5 s handshake timeout, and ENet's own peer limit.
  • Parsers. Every decoder refuses malformed input by returning an empty result, never by raising an error. make fuzz feeds 1,000,000 mutated and random messages to all of them, the snapshot receiver and the list server, and checks for engine errors and memory growth.

Snapshot format

NetSnapshot (src/net/) owns the layouts. An entity travels as a record: a list of whole numbers, already quantized, laid out per archetype. The server compares records to see what changed, the wire carries only the changed field groups, and the client keeps records as its picture of the world.

A snapshot is a binary, little-endian packet on channel 1, unreliable:

Part Contents
Header Message type (u8), server tick (u32), baseline tick (u32, 0 = builds on nothing), simulation tick (u32), the newest command tick received from this client plus one (u32, 0 = none yet), the number of the newest world message made (u32), the entity the client is watching (varint)
Match and tide Match phase (u8), tide phase (u8), sea level in centimetres (s16)
Own block A flag (u8), then, while the client's own player is alive and is who it watches, 37 bytes nobody else is sent: position and velocity as f32, the water level around the player as f64, which of the nine flood cells around the player that water reaches (u16), swimming stamina (u16) and state flags (u8)
Removed Count (u8), then IDs (varint)
Entities Count (u16), then records of ID (varint), archetype and changed-group mask (u8), then only the groups that changed

The server tick counts every tick the server has run and never goes back, so it orders snapshots across matches; the simulation tick is the clock commands name. Positions are quantized to 1/64 m (u16 per axis: 1,024 m across, heights from -128 m), yaw to u16, pitch to s8, health and armour to u8. Field groups: players have position, look, vitals, state flags and held item; pickups and shots have position and item.

Deltas and loss. SnapshotSender remembers what it told each client in its last 40 snapshots. Each snapshot is a delta against the newest one that client has acknowledged (its input packets carry that tick), or builds on nothing if there is none or it is too old. A delta only ever builds on something the client confirmed holding, so a lost snapshot is never resent: the next one covers it. SnapshotReceiver ignores snapshots older than its newest, and if it is ever sent a delta on something it doesn't hold, acknowledges 0 to ask for a fresh start.

Interest. A client is told only about entities near the player it is watching (its own, or the one it spectates): players and shots within Tuning.NET_INTEREST_PLAYERS_METRES (380 m, a little past the longest weapon) and items on the ground within NET_INTEREST_PICKUPS_METRES (90 m), measured on the map. An entity that leaves range is removed from the client's picture and nothing more is said about it. Outside a match there is no viewer and a snapshot is only its header.

Budget. A snapshot is at most Tuning.NET_SNAPSHOT_MAX_BYTES (1,200), so 15 a second is at most 18 KB/s before packet headers and never fragments. When more has changed than fits, the client's own player goes first, then the nearest, and each snapshot an update has waited counts as 6 m closer, so nothing waits for ever. What didn't fit is still owed and goes in a later snapshot.

Cost. Building 24 snapshots is the largest part of a busy server's tick, more than the simulation. Records are made once a tick and shared, each (entity, mask) encoding is made once a tick and shared, an unchanged record is skipped with one native comparison, and clients are spread over the ticks of the snapshot interval by handle (NetGame.snapshot_due_for()), so each tick builds half of them.

Input. The client answers on channel 1 with the snapshot tick it holds and its last few commands (Tuning.NET_INPUT_REDUNDANCY), oldest first. NetGame (src/server/) offers each tick's command to ServerLoop.submit() once, and echoes the newest tick it has received in every snapshot, taken or not.

Match notice. On joining and at every change of phase the server sends, reliably on channel 0, the match phase, index and seed, so the client can generate the island.

World messages

Built pieces, props, chests and the tide's schedule aren't entities and don't go in snapshots. A wall the client never heard about is a lie it would keep walking into, so they travel on channel 0, reliably and in order, as WORLD messages (NetWorld, src/net/).

Part Contents
Header Message type (u8), flags (u8: whole state, loot placed, tide follows), sequence number (u32), simulation tick (u32)
Tide Only when it starts: the tick it started (u32) and each phase's level (f32)
New pieces ID, kind, material and facing in one byte, slot (s16, s8, s16, u8), owner, tick placed (u32), ticks of building done, damage taken
Damaged pieces ID and damage taken so far
Removed pieces IDs
Props Prop and health left, for each one damaged
Chests Indices of chests opened

Counts and IDs are varints. WorldSender makes one message after each tick in which something changed and everyone is sent the same one; a quiet tick makes none. The simulation keeps the lists it reads (StructureSet.changed and .removed, PropSet.changed), so nothing is scanned. Someone joining is sent the whole state in one message after the MATCH notice, and so is everyone when a new match's simulation appears.

What isn't sent: props, chests' places and their contents, which the client rolls itself from the seed, as it does the island; and a piece's health, which follows from how far its building had got, the ticks since, and the damage it has taken. So a piece being built costs nothing on the wire. A scripted match of 12,198 ticks with 24 players made 121 messages, 2.3 KB in all. Bots don't build yet, so a match of people building will make far more: a new piece is about 16 bytes to each client.

ClientWorld applies them to the client's world. NetWorld.world_hash() covers everything replicated this way, and a client that has applied every message up to a tick agrees with the server's simulation at that tick, which a test checks through a whole match.

Each player is also told what they carry, as an OWN message on the same channel whenever it changes: held selection, materials, the five slots and ammunition.

Placing ahead of the server. When the player's command places a piece and the client's own run of Building.check passes, ClientWorld puts a provisional piece in its world at once, under an ID the server never uses, and holds its cost back from the materials shown. The server's piece arriving in that slot replaces it. Every snapshot carries the number of the newest WORLD message made, so the client knows when it has heard everything that snapshot's world had in it: once a snapshot from after the placing tick has come, and the world messages up to its number have too, and no piece came, the server turned it down (or never got the command) and the provisional piece is taken away. Losing a race for a slot ends the same way, with the winner's piece standing.

Lag compensation

A client draws other players about 130 ms in the past, and its commands take the trip to the server and wait their turn, so when the server fires a shot the target is no longer where the shooter saw it: 8 ticks on at a 50 ms round trip, 12 at 200 ms. Without help, nobody could hit a walking target.

So every command says how far back its player was seeing (view_lag), and the simulation keeps where every player's body has been for the last Tuning.LAG_COMPENSATION_MAX_TICKS (15 ticks, half a second) in HitHistory (src/sim/). A hitscan shot is traced against other players where they were at the moment the shooter saw, between ticks if need be, and against terrain, props and built pieces as they are now. Projectiles fly in the server's present and aren't rewound.

The client claims only when it was looking and where it aimed. Whether that is a hit is worked out by the server, from its own record. Three limits bound the claim:

  • Never more than the cap back. A round trip much over 300 ms needs more than the cap, and that shooter has to lead the target.
  • Never the future: the lag can't be negative.
  • Never earlier than a moment the same player has already claimed. A view only moves forward, so nobody can fire at the present and the past by turns and keep whichever hits.

Because the lag travels in the command, it is in the command log, and a replay judges every shot exactly as the match did. HitHistory itself isn't in the state hash: a replay rebuilds it from the same ticks.

Measured (test/net/test_lag_compensation_link.gd, simulated link, headless): a client firing a signal revolver once a second at the middle of a target walking across at 4 m/s, 25 m away, as drawn on its own screen. At 50, 100 and 200 ms round trips, 41 of 41 shots hit; with the lag ignored, 0 of 41. At 500 ms, past the cap, 0 of 41.

The client: prediction, reconciliation and interpolation

ClientGame (src/net/) is a client's side of a match. It works on packets, so it runs the same over a NetClient or a test's simulated link. From the seed in the MATCH notice it generates the island itself (a Simulation it never steps) and plans the same barge route, so it can run the server's movement code on the server's ground.

The command clock. A command must reach the server before the tick it names, so the client's tick runs ahead of the server's by the trip there plus a margin. Nobody measures the trip. Each snapshot says which tick the server was about to run and the newest command tick it had received; the difference is how many commands were waiting. ClientClock holds that at Tuning.NET_COMMAND_MARGIN_TICKS (3): it jumps ahead if commands arrive late two snapshots running, waits if they are far too early, and otherwise runs up to 2% fast or slow. With three commands waiting and each packet repeating the last three, jitter and a lost packet or two cost nothing.

Prediction. ClientPrediction keeps a copy of the client's own player and runs every command on it as it is made, through Movement.step, Water.update and Drop.ride, the same code the server runs. The player moves at once, not a round trip later.

Reconciliation. The own block in each snapshot is the player's exact state before the snapshot's simulation tick: 32-bit floats, as the server holds them, not the quantized record. If that equals what was predicted before that tick, nothing happens, which is the normal case. If not, the copy is put where the server says, every command since is run again, and the distance the player's present position moved is the correction. A correction is never shown at once: the drawn position keeps an offset that fades at NET_CORRECTION_SMOOTH_RATE, unless it is over 3 m, which is shown as the teleport it is. Built pieces and broken props are solid on the client as soon as their WORLD message arrives, a moment after they are on the server, so a client running into a wall that new is corrected. What the client can't predict: a basin still filling after a breach, and anything another player does to it.

Interpolation. SnapshotInterpolator draws every other entity Tuning.NET_INTERP_DELAY_TICKS (4 ticks, two snapshots) in the past, on a straight line between the snapshots either side of that moment, with yaw turning the short way. The moment drawn is a clock of its own, steered gently towards where arrivals say it should be, so jitter in the network never becomes jitter on screen. When a snapshot is lost the entity carries on along its last heading for up to 3 ticks; if that guess was wrong, the difference is blended away when the news arrives.

Measured (test/net/test_prediction_link.gd, simulated link, headless): two players through the drop and 75 s of walking, sprinting, strafing, jumping and crouching on a real island, each with 100 ms of latency each way, 20 ms of jitter and 2% loss. Own movement: about 1,400 snapshots compared per client and none needed a correction. The other player was drawn on average 0.4 mm from where a perfect link would draw them, and 9 mm from where the server had them at that moment; a lost snapshot during a sudden stop or start showed as at most 13 cm of drift, eased out. At 150 ms, 40 ms and 10% loss a client took 3 to 7 corrections of at most 22 cm, all blended. These are measurements of the netcode's arithmetic, not of how it feels in the hand, which needs a real playtest.

Module map

Path Holds May depend on
src/sim/ Tick loop, entities, RNG, commands, tide, combat and its lag compensation, building src/sim/ only
src/bots/ Bot brains that produce PlayerCommands src/sim/
src/net/ Transport, snapshot encoding, the client's side of a match (prediction, interpolation) src/sim/
src/server/ Headless and listening servers, joining the network to the match src/sim/, src/bots/, src/net/
src/client/ World rendering, camera, input, effects src/sim/, src/net/
src/ui/ HUD, menus, map src/sim/ (read only), src/client/
src/audio/ Sound playback, music src/sim/ (read only)
src/app/ Startup and mode selection Everything

Bots

Bots live in src/bots/ and play through the same door as humans: each tick a bot reads the simulation and returns one PlayerCommand, which the server submits like any other. They never write to the simulation.

NavGrid is the shared map they plan on: one point per 2 m terrain sample, solid where the ground is a cliff or a solid stands between step and head height, and eight times as costly where the water is deeper than wading. Path search is Godot's AStarGrid2D. The server builds the grid while the lobby waits (about 0.4 s, next to the island's own generation) and syncs it every tick before bots think: new solids block points at once, removed solids are found by a rolling scan within a few ticks, and points turn wet as the tide and flood reach them. Every change is stamped with its tick, and each bot checks its remaining path against the stamps twice a second, so a new wall or rising water across a path makes it plan again. Only one path is planned per server tick across all bots; others wait their turn.

NavBot follows a path: it steers at the next waypoint, sprints on long dry stretches, jumps when it stops making progress, and after a few tries marks the point blocked and plans around it. A goal the grid's regions show to be cut off (inside a ring of rock, behind a hut's walls) is turned down without a search, which would otherwise cover the whole island before giving up.

BotBrain decides where a bot goes and what it presses, and gives NavBot its goals. It knows what a player would: the map (BotWorld: loot markers at the points of interest and the ground each tide level will leave dry), the tide forecast, items within a few metres, opponents in view with a clear line to them, and who just hit it. Each tick, in order of urgency, it gets to ground the coming tide won't reach, fights a visible opponent if it is armed, bandages itself when nobody is in view, picks up useful items and opens chests, walks to the next loot marker, and otherwise wanders the safe ground. Difficulty (Tuning.BOT_DIFFICULTIES) sets its reaction time, how far it sees and how far its aim wanders.

Two things the brain handles itself, because the grid is flat terrain only:

  • Huts, lofts and piers. Their doors and steps are finer than the grid, so a bot walks to the foot of the point of interest's own walking route and follows its points in, then back out the same way, before anything else, even if it must flee.
  • Safe ground. For each tide phase it is the dry land joined to the island's highest walkable ground, so it only shrinks, each phase's inside the last. A bot off it leaves when the time before the next rise only just covers the walk.

tools/bot_matches.gd (make bot-matches) plays whole bot-only matches on a run of seeds and reports winners, causes of death and how each difficulty placed.

Terrain

The island's heightfield (Island, in src/sim/) is the only terrain collision, on the server and in client prediction. Nothing in the renderer is ever used for collision.

The client draws it with TerrainView (src/client/terrain/): 64 m chunks at four levels of detail chosen by distance to the camera. TerrainChunkBuilder builds chunk geometry on WorkerThreadPool threads, and the main thread only uploads finished meshes, within a per-frame budget. Each chunk hangs skirts from its edges deeper than any two levels can disagree, so a level boundary never opens a crack. Normals come from the full-resolution heightfield at every level, so lighting doesn't change across a boundary. The terrain shader blends procedural sand, marsh, grass and rock layers using splat weights from the island's elevation bands, with rock on steep ground.

The match view

MatchView (src/client/) is the client's picture of one match: sky, sun, terrain, sea, points of interest, props, built pieces and EntityView, which draws players, the barge, items and chests. All of it reads one Simulation and never writes to it, and a new match gets a new view. The camera sits on a player's shoulder through CameraRig. The figures, barge and items are stand-in shapes for the art pass to replace.

In play, the simulation it reads is the client's own copy of the world: the Simulation that ClientGame generates from the match's seed and never steps. ClientWorld keeps its pieces, props, chests and tide in line with the server, and ClientMirror (src/net/) writes the rest of the picture into it every frame: each entity the server has told this client about, where it is being drawn now (everyone else from the recent past, the client's own player where it was predicted), the barge, the clock and the sea level. So the same views draw a networked match, a local server's simulation in a test, or a replay in a capture, and there is one set of drawing code. Captures and tests still point a MatchView straight at a server's simulation.

The client program

ClientSession (src/app/) is the game while it is in a match: a ClientGame, a ClientInput turning the keyboard and mouse into one command a tick, a ClientMirror and a MatchView. Its pump() does a frame's work and needs no scene tree, so tests drive it.

The server is either across the network (--connect=HOST:PORT, a NetClient) or in the same process (LocalServer, src/server/): the same ServerLoop and NetGame a dedicated server runs, with the packets handed across in memory, reliable ones first, in the order a socket would deliver them. The client can't tell the difference, so offline play is the network path with no network, as "Rules that do not bend" asks.

src/app/main.gd picks the mode: --server runs the dedicated server; otherwise the game starts an offline match, or joins a server with --connect. Without a window (tools/godot.sh is always headless) a mode has to be asked for. --autoplay puts a small script at the controls and --client-seconds=N quits after N seconds with one JSON line saying how far the client got, which is what make client-check reads. There are no menus yet (N6, A10) and no HUD (A9): one line of text says when the client is connecting, in the lobby or watching.

Performance budgets

Budget Target
Server tick Under 12 ms with 24 players, including the network's share (make soak checks the p99)
Client 60 fps on a GTX 1060-class GPU at medium settings
Terrain streaming Under 8 ms of main-thread work in any frame (measured headless; GPU upload unmeasured)

Software rendering in CI can't verify the client budget. It needs a real GPU playtest.

Soak testing

make soak (tools/soak.sh) is the multiplayer gate: a real server process and 24 headless clients in three processes, each client behind its own UdpRelay, a UDP relay that delays, jitters and loses datagrams both ways, so ENet runs unchanged over a bad link. SoakPlayer plays each client from what a client knows: it drops, glides, heads for high ground, harvests and tries to build.

Desync is checked by hash: the server, started with --world-hash-every=N, logs NetWorld.world_hash() on the simulation ticks from each multiple of N to the snapshot interval, and each client hashes its own world on the snapshots at those ticks when its world messages have caught up exactly (world_seq) and it has no provisional pieces. Tick 0 is never compared: it lasts the whole lobby while the world is set up. tools/soak_report.gd compares them and checks errors, unreadable packets, that every client played every match, the p99 tick, and memory from one match to the next. make soak-selfcheck tampers with one client's world to prove a drift is caught.

A client process holds back a new match's notice until it has made that client's island, at most one a frame: a dozen clients making islands in one frame starved their connections of service for long enough that ENet dropped them.