Files
transcience/CLAUDE.md
T
claude b351bc2d55
ci / verify (push) Successful in 48s
Stage 4: upgrades, and a quartermaster to spend them at
Every level banks one choice. Choices queue, and are spent at an NPC in the
hub: walk to it, press E, take one of three weighted options. Seven upgrades,
all data — split shot, glass cannon, spread, sniper, doubleshot, poison,
eraser — and SimWorld gained no per-upgrade branch to run any of them.

The four ambiguities in the brief were settled with the user first, since
each changes what gets written:

  damage      base x (1 + sum additive) x product multiplicative. The flat
              +5% every upgrade carries, spread's -10%, doubleshot's -50%
              and glass cannon's +100% pool; sniper multiplies on top, so
              two snipers is 4x and not +200%.
  glass       half the LEVELLED maximum, multiplying if taken twice, so the
              price does not fade to a rounding error by level 15.
  poison      independent stacks, not a refresh.
  split       +/-45 degrees from the original heading.

Independent poison stacks sound expensive and are not: every dose lasts the
same number of ticks, so doses expire in the order they were added, the
pending expiries are a plain FIFO, and PoisonTrack only ever looks at its
front. O(1) per actor per tick however many are live.

Stats are derived from the upgrade list and never stored, the way level is
derived from experience -- a saved stat cannot disagree with the upgrades
that produced it. Upgrade riders (split charges, poison, erase chance) travel
on the bullet instead, because a shot in flight has to keep what it was fired
with rather than gaining Poison because the shooter just took it.

Two invariants this collided with, both now pinned:

  - bullet speed gained a ceiling. Wall collision samples once per tick, so
    anything over a tile per tick tunnels; two snipers asked for 2480 u/s
    against a 1920 threshold, and a tunnelling bullet looks like a bullet.
  - BULLET_INTEREST_RADIUS rose to 2900, because an upgraded player shot is
    now the longest-travelling bullet in the game. test_interest measured
    the worst case from static content, which upgrades quietly invalidated.

Choosing is intent checked three ways: a choice must be owed, the index must
name one of the three options the SERVER put on the table, and the player
must be standing at the NPC. The offer is rolled once and persisted, so
closing the screen is not a reroll and neither is a crash.

tools/diag_upgrades.tscn covers level -> banked choice -> refused in a
dungeon and refused across the room -> taken at the NPC -> new stats ->
on disk. Bots never walk to the quartermaster, so the smoke test cannot.

Known gap recorded in the roadmap: at PLAYER_BULLET_DAMAGE = 6, the +5% the
first upgrade carries rounds back to 6 and visibly does nothing. It comes out
right in aggregate, but the fix is a balance edit across content.gd and so is
the user's call.

check.sh clean, 357 tests, SMOKE PASS (18 assertions), all four diagnostics
green.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-06 15:34:58 +02:00

13 KiB

Transcience

Top-down twin-stick bullet-hell with a dedicated, server-authoritative backend. Godot 4.7, GDScript only. One executable is both server and client.

Starting a new session? docs/ROADMAP.md says what is built, what is next, and which file implements each feature. docs/DECISIONS.md records what the user has already decided and why — read it before asking a design question, several answers there are not the obvious default.

Commands

tools/check.sh      # parse-check every script (~5s) -- run after every edit
tools/test.sh       # GUT suite, headless (~3s)
tools/smoke.sh      # real server + 4 bot clients over ENet (~40s)
tools/server.sh     # dedicated server
tools/client.sh --listen            # host and play, no menu

Diagnostics. Each runs as a scene (they need the Net autoload) and exits non-zero on failure, so they gate like tests:

godot --headless --path . res://tools/diag_prediction.tscn    # prediction gap; injects clock drift
godot --headless --path . res://tools/diag_progression.tscn   # kill -> xp -> level -> death -> roster
godot --headless --path . res://tools/diag_loot.tscn          # drop -> pick up -> persist -> use -> drop
godot --headless --path . res://tools/diag_upgrades.tscn      # level -> choice -> taken at the NPC -> new stats
godot --headless --path . --script tools/bench.gd             # sim cost per tick
python3 tools/build_local_assets.py                           # rebuild the local-only bullet atlas

diag_progression, diag_loot and diag_upgrades exist because the bot smoke test cannot cover any of them: bots are poor shots, so they rarely kill anything, which means they neither earn levels, produce drops, nor ever reach the quartermaster.

Everything after -- goes to GameOpts.parse():

Flag Effect
--server Dedicated server, no view.
--join Skip the menu and connect.
--listen Skip the menu and host a listen server with a local player.
--bot Scripted input instead of the keyboard. Implies --join.
--host, --port, --name Connection details.
--autoquit N Quit after N physics ticks.
--boss-rush Server-side: dungeons spawn the boss and no trash.
--depth N Server-side: depth of new dungeons, which drives map size.
--account N Client-side: override the local account, so several clients can coexist on one machine.
--store PATH Server-side: character store location. Use a scratch path in tests.
--verbose / --quiet Log level.

A change is done when check.sh, test.sh and — if it touched networking, instances or the simulation — smoke.sh all pass. Say so explicitly; do not report a networking change as working on the strength of unit tests alone.

Hooks

tools/install-hooks.sh once per clone (points core.hooksPath at .githooks/, committed in the repo — plain bash, no pre-commit framework, so cloning costs nothing extra to run tools/*.sh). Own .gd files (not vendored addons/) trigger check.sh + test.sh on commit (~7s); smoke.sh runs on push (~35s, skip deliberately with SKIP_SMOKE_HOOK=1 git push).

The one rule

The server decides everything; the client only sends intent.

A client can send exactly two things: an [InputFrame] (move vector, aim angle, five button bits, an inventory slot) and a handshake — plus the two low-rate character-roster requests, which are also pure intent. There is no message for "I moved here", "I hit that", "I took damage", "my escape finished" or "I now own this item". Adding one would collapse the whole security model, so don't — validate-after-the-fact is strictly weaker than having no code path at all.

When a new player action needs a message, look at whether it fits in the input frame first. Item use and drop did, and got the redundancy, the replay guard and the per-tick rate limit for free. The cost was one rule: anything in the input frame that must not repeat has to be edge-triggered (see prev_buttons), because frames are resent and a starved server coasts on the last one it holds.

SimWorld.authoritative is true on the server and false on the client. In replica mode the world runs no AI, fires no emitters and resolves no hits; it only integrates bullets it was told about. tests/unit/test_server_authority.gd and tests/integration/test_replica_parity.gd pin this down.

Layout

Path What lives there
src/sim/ The whole game as plain RefCounted objects. No nodes, no physics server, no rendering.
src/actors/ Data-only Resource definitions: EnemyDef, BossDef, ItemDef, LootDrop, DungeonDef, UpgradeDef. Shapes, not instances.
src/sim/patterns/ Bullet emitters — the authoring surface for every enemy and boss.
src/sim/map_grid.gd Tile grid: collision, line of sight, chunk streaming.
src/sim/map_gen.gd Dungeon generation; build() is the only entry point.
src/content/rooms.gd Hand-authored room stamps (hub, boss arenas) as text.
src/meta/ Accounts, characters, persistence, XP curve. Server-owned.
src/content/content.gd All enemies and bosses, defined in code. Source of truth.
src/content/items.gd All items, same idea. Items.ORDER is the wire format — append only.
src/content/dungeons.gd The kinds of run. Dungeons.ORDER is both a wire format and the hub's portal order.
src/content/upgrades.gd The seven upgrades and their draw weights. Upgrades.ORDER is a wire format.
src/net/ Codec, ServerRuntime, ClientRuntime.
src/instances/ Lobby hub and dungeon runs.
src/view/, src/ui/ Read-only rendering. Never decides anything.
src/autoload/net.gd The only autoload. RPC surface.
tools/ Headless tooling.

The simulation must not import anything from src/net/, src/view/ or src/ui/, and must not touch Net. That is what lets tests drive a thousand ticks in milliseconds with no SceneTree.

Godot gotchas that will waste your time

  1. New class_name needs a cache refresh. .godot/global_script_class_cache.cfg is only rebuilt by the editor or godot --headless --path . --import. Until then every use of the new class reports Identifier not declared, which looks like a real error. check.sh does the refresh for you.
  2. Autoload names do not resolve under --script. A --script run has no main loop, so Net is Identifier not found. Tools that need autoloads must run as a scene (see tools/check.tscn); tools that don't can use --script. This is why GameLog and GameOpts are static classes rather than autoloads.
  3. ResourceLoader.load() returns non-null for a broken script. Never test the return value to detect a parse error; read the engine's stderr instead. And never call Script.reload() on the script you are running — it hangs.
  4. Input events default to device = 16, which matches nothing. Bindings must use device = -1. tools/setup_input_map.gd generates the input map correctly; edit that file, not the [input] block in project.godot.
  5. set_anchors_preset(preset) does not zero the offsets. keep_offsets defaults to false, which despite the name means "recompute offsets to keep the control's current rect on screen" — for a freshly created Control that rect is (0,0)-sized, so it comes out pinned to the top-left corner regardless of the anchors. Use set_anchors_and_offsets_preset() for any Control built in code. Separately: .position assigns an absolute coordinate even on an anchored control; offset_left/offset_top are the anchor-relative ones. This combination silently broke the main menu and three pieces of the HUD.
  6. ProjectSettings.save() drops settings equal to the engine default and strips comments. Anything load-bearing (the 60 Hz tick) is asserted in code in src/main.gd instead of trusted to project.godot.

Non-obvious invariants

  • ClientRuntime.process_physics_priority = -10. The client must sample and send input before ServerRuntime ticks, or a listen server's drawn ship sits a permanent tick ahead of the authoritative one and bullets trail it. Measure with godot --headless --path . res://tools/diag_prediction.tscn (~0.1px is healthy, 4px means the ordering broke).
  • PLAYER_RADIUS (hitbox) < PLAYER_VISUAL_RADIUS (sprite), and PLAYER_MUZZLE_OFFSET derives from the visual one. Prefer a visible near-miss over an invisible hit; keep the muzzle clear of the sprite.
  • INPUT_MAX_LEAD must stay well above INPUT_LEAD_MAX. The server's input acceptance window has to be wider than the band in which the client re-syncs its own numbering. Violate it and drifting clocks land in a silent dead zone where the server rejects everything and the client never notices — the ship and the authoritative position separate permanently. Pinned by tests/unit/test_input_lead.gd.
  • Never send the map, or its seed. Geometry is streamed per peer in chunks around that peer's player (ServerRuntime._stream_map). The seed would let any client regenerate the whole dungeon. MAP_STREAM_RADIUS must stay wider than FOG_VIEW_RADIUS, or prediction runs on terrain the client lacks.
  • Bullet speed must stay under one tile per tick. Wall collision samples position once per tick, so anything faster tunnels. Upgrades multiply bullet speed, so SimConfig.MAX_BULLET_SPEED clamps the result — without it two Snipers put shots through walls. Pinned by test_bullet_speeds_stay_below_the_tunnelling_threshold.
  • A player's combat numbers are derived, never stored. PlayerStats.build() recomputes them from the character's upgrade ids every time, so a saved stat cannot disagree with the upgrades that produced it. Upgrade riders (split charges, poison, erase chance) travel on the bullet instead, because a shot in flight must keep what it was fired with.
  • Only ServerRuntime writes progression and persistence. The simulation reads a player's level and max health, and moves items between the ground and a bag; it never grants experience, retires a character, or touches the store. It announces what happened and ServerRuntime banks it. One writer means a level can never disagree with the experience that earned it, and an inventory on disk can never disagree with the one in the world.
  • Items.ORDER is a wire format. An item's index in it is the byte that rides the snapshot and every item event. Append, never reorder — reordering makes every existing client decode a potion as a ration, so it needs a Protocol.VERSION bump. Dungeons.ORDER is the same, and additionally decides which hub portal opens which dungeon (Nth P marker in the lobby stamp → Nth entry).
  • Which dungeon you enter comes from where you are standing, never from the client. SimWorld.portal_at() resolves the player's server-side position to a portal, and the PORTAL_USED event carries the answer. There is no message that names a dungeon, and adding one would let any client pick the easy variant's loot rate.
  • Loot has two visibilities, and the instanced one is enforced in the codec. NetCodec.encode_snapshot filters items owned by another peer, exactly like the actor interest radius. Never move that check into the client: hiding an entity the client was handed defends nothing.
  • LocalAuthProvider is insecure on purpose. Any client can claim any account. It exists to have the same shape as Steamworks (opaque ticket in, 64-bit account id out) so swapping is one class. Do not ship it.
  • No contact damage. Every enemy threatens through bullets only; touching one is harmless. tests/unit/test_content.gd enforces that every hostile has an emitter.
  • No i-frames. Every bullet that touches a player lands; spawn_grace is the only invulnerable state. Do not reintroduce post-hit immunity — it makes dense patterns safer than sparse ones.
  • A disconnect is not an exit. Dropping in a dungeon keeps the player in the world as linkdead, channelling out over the same second the escape costs. Damage must never cancel the escape channel, or quitting beats the button. See docs/NETCODE.md.

Adding content

A new enemy or boss is data, never code. Add a builder to src/content/content.gd returning an EnemyDef / BossDef made of the emitters in src/sim/patterns/, register its id in enemy() / boss(), and add a test. tests/unit/test_boss.gd::test_a_brand_new_boss_needs_no_engine_changes builds a boss from scratch and asserts the simulation needs no changes to run it — if you find yourself adding a per-boss branch to SimWorld, stop and add an emitter type instead.

tools/export_content.gd writes .tres copies into resources/ for tuning in the editor inspector. Those are an export, not the source; port changes back.

Style

Typed GDScript everywhere (untyped_declaration is a warning). Tabs, snake_case files, PascalCase class names. Comments explain why a thing is the way it is — the netcode and anti-cheat decisions especially. Keep the existing density.