CosmicClash/docs/TECH_STACK.md

Tech stack

What this project is built with, and why each piece was chosen, sourced from the project's own docs and code comments. Where the reasoning for a choice isn't recorded anywhere, this doc says so rather than guessing.

Engine: Godot 4.7

The whole game — client and dedicated server alike — is one Godot 4.7 project, GDScript only. There is no C#, ONNX, or .NET code involved at runtime anywhere in the shipped product.

Why Godot, specifically: legal, not technical. Per README.md's "Legality" section, the concept of "vehicle soccer" cannot be copyrighted, but Rocket League's specific expression of it can be. Building on Unreal or Unity — the engines Psyonix and most Rocket-League-likes use — would invite comparison to that specific expression. Using a different engine (Godot) and different vehicles (space ships instead of cars) is a deliberate part of keeping the project's own expression original and legally distinct.

Physics: Jolt Physics

Set via Game/project.godot's 3d/physics_engine="Jolt Physics" — Godot 4's alternative physics backend, not the engine's own default (GodotPhysics3D). All ship and ball movement is force/torque-based (_integrate_forces), never kinematic.

Why Jolt over Godot's default physics: not written down anywhere in the project's own docs, but per the project owner, the goal was a physics engine whose behavior isn't tied to Godot's own release cycle — so upgrading to a future major Godot version doesn't silently change how the game feels, the way an engine-version upgrade has repeatedly worried Rocket League's own playerbase (players have specifically flagged that Unreal Engine upgrades risk changing timestep/continuous-collision behavior enough to break muscle memory built over thousands of hours).

This holds up under scrutiny. GodotPhysics3D (Godot's built-in default) is an internal engine subsystem, versioned and rewritten alongside Godot itself, and has a real history of behavior changing across Godot releases — for example a kinematic-body regression introduced between 4.3-dev4 and 4.3-dev5, and collision-detection differences reported across the 4.0 line. Jolt, by contrast, is developed as an independent upstream library (jrouwe/JoltPhysics) with its own semantic versioning and a user base beyond just Godot, so its collision behavior changes on its own release cadence rather than as a side effect of unrelated Godot core work. It's a real reduction in coupling, not a complete guarantee: Godot still pins (and can bump) a specific Jolt version per release, and Godot's own Jolt integration layer can itself introduce differences — e.g. a sleeping RigidBody3D wakes differently under Jolt than under GodotPhysics3D when another body approaches it.

Separately documented, and a real consequence either physics backend would share: Jolt is not bit-deterministic across platforms or even across differing contact orderings on the same platform, and Godot exposes no world snapshot/restore API. That fact is why the multiplayer architecture is server-authoritative with client-side prediction of only the local ship, rather than rollback/resimulation netcode — rollback would require deterministic replay, which no physics engine choice here provides (multiplayer-todo.md §1, decision 1).

Multiplayer transport: Godot's built-in MultiplayerAPI over ENet

The default and fully-supported transport is ENetMultiplayerPeer — Godot's built-in high-level multiplayer networking, direct-IP over UDP, port 7777 by default. A thin NetTransport abstraction (Game/scripts/net_transport.gd) exists specifically so a second transport (Steam) can be swapped in without touching the rest of the networking code.

Design choices layered on top of the built-in peer, and why:

• ENetMultiplayerPeer.server_relay is forced to false. It defaults to true, which lets any client rpc() any other client through the server — incompatible with a server-authoritative model. Called out in multiplayer-todo.md §2.1 as "the single highest-value one-line security change in the document."
• Manual multiplayer polling, not Godot's automatic idle-frame poll. NetworkManager calls set_multiplayer_poll_enabled(false) because the automatic poll runs on the idle frame, which would tax every RPC issued from _physics_process up to a full frame in each direction — unacceptable for a physics-tick-rate game.
• Server-authoritative simulation with client-side prediction of the local ship and ball only; no world rollback. Direct consequence of Jolt's non-determinism (see above).
• A custom binary wire format (net_codec.gd) rather than raw RPC argument marshalling, for compact, quantised input/snapshot packets sent at high frequency — no stated alternative was considered in the docs, but the packet-size/channel-intent design in multiplayer-todo.md §2 is extensive and deliberate.

Optional multiplayer transport: Steam (GodotSteam)

Game/scripts/steam_transport.gd implements the same NetTransport interface using SteamMultiplayerPeer over Steam's SDR (Steam Datagram Relay), from a custom GodotSteam-patched Godot build (not stock Godot — STEAM.md). It is entirely optional: the default build and every CI check use ENet only, and a build without the steam feature is fully functional without it.

Why it's optional and why raw ENet remains primary: multiplayer-todo.md states plainly that "Docker/VPS is the primary v1 deployment path. Raw ENet self-hosting needs port forwarding, and SDR is Phase 7 — so [the ENet phases] ship something that works on LAN or a VPS and nowhere else." Steam/SDR is being added later specifically to remove the port-forwarding requirement and to supply verified player identity — direct-IP ENet's slot-reclaim logic is keyed by display name today, which is insecure against a public server (see multiplayer-next.md).

Dedicated server hosting: Docker (primary) or native systemd

The dedicated server is not a separately-written service — it's the same Godot project, exported headless (res://scenes/server_boot.tscn) via Godot's own --export-release "Linux Dedicated Server" preset. Two deployment paths are documented (SERVER.md):

• Docker, the primary path: a multi-stage Dockerfile builds the export inside a pinned barichello/godot-ci:4.7.1 image and produces a slim ubuntu:24.04 runtime image. make verify-phase6 builds it, runs it, joins two independent client processes to it, and asserts on match/goal/ arena-rotation behaviour — this is also the entire Phase 6 GitHub Actions workflow.
• Native systemd, for a VPS: copy the exported binary to /opt/cosmic-clash, run it as a dedicated cosmicclash service user via deploy/cosmic-clash-server.service.

Per the project owner, Docker was chosen as the primary path for three reasons: it gives a pinned, reproducible build environment that behaves identically across local development, CI, and hosted production servers (rather than three separately-drifting setups); it's portable across hosting providers instead of assuming a specific Linux distro/init system the way the systemd unit does; and it's the tooling the team is already most familiar with. That matches what's independently visible in the repo — SERVER.md documents Docker as the one path CI actually exercises (make verify-phase6), while the systemd unit is native-deployment documentation only, with no automated verification of its own.

AI opponents: reinforcement learning, trained out-of-process, run in pure GDScript

Two entirely separate pieces, deliberately joined only at a JSON file:

• Training (Python, not shipped): Godot RL Agents (godot-rl==0.8.2, vendored bridge plugin at Game/addons/godot_rl_agents, MIT-licensed) drives self-play PPO via Stable-Baselines3 (==2.4.0) over PyTorch (==2.13.0) and Gymnasium (==1.0.0), against a headless instance of the actual game (training/train.py launches real parallel godot --headless processes from source — the training environment is the game, not a reimplementation of its physics). training/requirements.txt pins these versions strictly, because the training code (ship_action_codec.gd, train.py) depends on specific library-internal behaviour (godot_rl's discrete-action-space branch, SB3's logit layout, Gymnasium's dict-key sorting) that an unpinned upgrade could silently change mid-curriculum.
• In-game inference (Game/scripts/policy_network.gd): the trained checkpoint is exported to a small JSON file (training/export_policy.py) and evaluated at runtime by a hand-written, dependency-free GDScript MLP. Game/addons/godot_rl_agents/VENDORED.md notes the plugin's ONNX/C# files are present upstream but unused here — they require the .NET Godot build, which this project does not use.

Why this split, rather than shipping ONNX/.NET inference: stated directly in TRAINING.md — "the trained policy is exported to a small JSON file and runs inside the game in pure GDScript — shipped bots need no Python, no .NET, no network." Keeping the shipped game GDScript-only (no .NET Godot build) is consistent with the rest of the stack.

Tooling (not shipped with the game)

• mcp/godot-mcp (git submodule, Node/TypeScript) — drives a live Godot editor/runtime instance for AI-assisted development; not part of the game.
• mcp/blender-mcp (git submodule, Python/uv) — drives a live Blender instance for generating original 3D assets (ships, arenas), for the same originality reasons covered under "Why Godot" above.
• tools/blender/, tools/textures/ — standalone Python scripts (Blender's embedded Python, plus texture generators) used to produce the project's original meshes and textures.

What's deliberately absent

• No C# or .NET runtime anywhere in the shipped game or server, despite early project framing (see README.md's history) once describing a "C# backend" — that was never built and is not the current plan.
• No HTTP/WebSocket/gRPC layer for multiplayer — ENet/Steam SDR over UDP only, via Godot's own MultiplayerAPI.
• No ONNX or other ML runtime in the shipped game — see "AI opponents" above.