mp_test_driver.gd was 1458 lines doing six unrelated jobs. It is now
orchestration only — the scripted sequence, the RPC plumbing between peers, and
the manual keyboard controls — with the work in modules that each have one:
MpWorldView finding things in the world and describing what they are doing
MpSteps the simulated player actions (reach, grab, carry, drop)
MpAsserts the per-step checks
MpSnapshot the cross-peer sync audit
MpReport the ledger, the log, the overlay, the screenshots
The scenario list and the audit logic carry over unchanged. That audit compares
what is actually RENDERED on both peers, not just the replicated values behind
it, which is the only thing that catches a plate whose contents arrived but
whose visuals were never rebuilt — so it was worth moving verbatim.
Assertions that tested the old model now test the new one: "who is holding
this" is the authority of the object's NetXform, not of the object itself.
Suite: 146/146 passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adding an object to the game no longer requires any networking code. The
MultiplayerSpawner runs in its default mode — no spawn_function, no payload
dictionary — and NetReplication builds each object's SceneReplicationConfig
from a convention, so scenes carry no hand-authored replication at all.
Every replicated node gets two generated synchronizers: NetSync for script
state, always server-owned, and NetXform for position, handed to whoever is
holding the object. That split is what makes grab prediction work — a
synchronizer never applies inbound state on the peer that owns it, so a
player's own hand drives an object with no round trip while is_dirty and
friends keep flowing one way from the server.
Interaction is now two RPCs for the whole game (NetGrab), and client gating is
one rule applied to every object (NetWorld). Deleted: net_pickable.gd, the
replicated net_held_by field and its held-state juggling, the
grant/reject/force-release negotiation, the static-item despawn RPC, and the
per-scene replication configs. Authority is the single source of truth for who
simulates an object.
Two things the convention had to learn, both found by the test suite:
* Addon scripts are excluded. godot-xr-tools' snap zones and pickables expose
a public `enabled`, which is exactly the flag each peer must set for itself
— so replicating it meant the server sent `enabled = true` back over every
client's gate, and stations went on grabbing objects out of the local
player's hands.
* Arrays of nodes are excluded. Array[Node3D] and Array[FoodItem] would
otherwise try to serialise live node references.
table.gd's replicated state loses its underscore prefix, which now marks a
variable as private and unreplicated; two in-place array mutations there were
skipping their setters, and the progress bar could divide by zero on a client.
Suite: 146/146 passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Rebuild the multiplayer layer on stock MultiplayerSpawner/Synchronizer so
adding an object to the game needs no networking code.
MultiplayerSpawner only replicates node creation and deletion, so "sync
everything regardless of what it is" has to come from a SceneReplicationConfig
built by convention in code. NetReplication does that, giving every node two
generated synchronizers: NetSync for script state (always server-owned) and
NetXform for position (handed to whoever is holding the object).
test/spike/ establishes the four engine behaviours the design rests on. Two
are worth flagging: per-peer visibility CANNOT be used to stop the server
fighting a client's held object, because MultiplayerSpawner despawns and
respawns the node on every visibility flip; and set_visibility_for is only an
override on top of public_visibility, so calling it alone does nothing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds test/multiPlayerTest.tscn plus a driver that runs the kitchen flow
across two game instances: grab/drop, dirt station, sink washing, hob
cooking, counter combining and plating. Runs headless (run_mp_test.ps1),
in two visible windows (run_mp_test_windowed.ps1), or by hand with
keyboard controls (play_mp_test.ps1). 108 checks, exits non-zero on
failure.
After every step both peers snapshot every item's position and rendered
state and the server diffs them. Targeted assertions only look at the
thing a step touched, which misses desyncs elsewhere - that audit is
what caught the last bug below.
Bugs found and fixed:
- net_pickable: apply_held_state() only wrote `enabled` in its
non-authority branch, so once a client grabbed an item every other
peer set enabled=false and regaining authority never restored it. The
server could then never pick that item up again, and a station would
"snap" it (emitting has_picked_up, so a plate still got marked dirty)
while pick_up() bailed out on the disabled item - leaving the zone
holding an item with no grab driver.
- network_manager: station gating only happened in the spawn path, so
stations baked into a scene file kept running their snap zones on
clients and grabbed items straight out of the local hand. Added
gate_existing_stations().
- network_manager: despawn_item() only freed the server's copy. Items
baked into a scene aren't tracked by the MultiplayerSpawner, so
consuming one left a ghost on every client, which then blocked the
station it sat in and got grabbed instead of its replacement.
- network_manager: the snap-into-station decision read the server's own
copy of the item position, but the reliable release RPC routinely
overtakes the synchronizer's unordered position updates - so it acted
on a stale position and teleported items back into the station they
had just been carried away from. The releasing peer now sends its
final transform and the server adopts it first.
- container: contained_ids.append()/erase() mutate the array in place,
which never fires the setter that rebuilds the plate's visuals. The
peer that put food on a plate was the only peer that never redrew it;
remote peers looked right because the synchronizer assigns there.
Also null-guards XRServer.get_tracker() in the vendored xr-tools hand
grab point, which threw on every successful grab without an XR runtime,
and adds multiplayer_world.populate_from_layout so debug scenes can bake
their own content instead of spawning the whole kitchen.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pairs with the debug-log-heavy investigation workflow: Godot writes its
own engine log to logs/godot.log alongside the existing NetworkManager
temp-file log, and the directory is excluded from version control since
it's runtime output, not source.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The snap-zone crash fixed just before this was invisible in the logs
until traced through addon source by hand — the log showed grabs,
drops, and authority handoffs, but nothing about which peer/hand/zone
actually held an item at each step, or when a guard clause silently
changed behavior.
Adds a _holder_desc() helper (reports "loose", "hand(<path>)",
"zone(<station>)", or "other(...)") and logs every net_held_by
transition, every reclaim that has to restore freeze_mode/collision_mask
from a stuck state, every time the grab-race guard protects an actively-
held item from a stale sync, every force-drop triggered by a losing
authority race, and every pickup/drop event including the ones that
don't get forwarded (picked up by something other than a hand, or
dropped while not authority — exactly the silent case behind the
snap-zone crash). NetworkManager's snap/release-from-zone calls get the
same treatment: which station an item is pulled from, and whether a
snap attempt succeeded, was skipped by the race guard, or found nothing
in range.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
algodoogle