WarGames For Prime RL

This package publishes WarGames as a Prime Intellect Verifiers environment. Prime starts an episode, receives the current game frame, calls the normal computer-control tools, and receives the next frame plus the reward scored by WarGames.

The same game catalog, mission IDs, reward profiles, and trusted scoring probes used by the local runner are used here.

# from the WarGames repo root
uv pip install -e ./environments/prime

Run any shipped eval config:

prime eval run wargames \
  --config environments/prime/configs/doom/eval-map01.toml \
  -n 1 -r 1

Supported Games

Game	Missions	Eval config	RL config
Red Alert	297	`configs/redalert/eval-soviet-01.toml`	`configs/redalert/rl-soviet-01.toml`
FlightGear	14	`configs/flightgear/eval-c172p-takeoff.toml`	`configs/flightgear/rl-c172p-takeoff.toml`
SuperTuxKart	63	`configs/supertuxkart/eval-lighthouse.toml`	`configs/supertuxkart/rl-lighthouse.toml`
0 A.D.	390	`configs/zeroad/eval-arcadia.toml`	`configs/zeroad/rl-arcadia.toml`
Freeciv	12	`configs/freeciv/eval-earth-small.toml`	`configs/freeciv/rl-earth-small.toml`
Doom	204	`configs/doom/eval-map01.toml`	`configs/doom/rl-map01.toml`
SuperTux	321	`configs/supertux/eval-welcome-antarctica.toml`	`configs/supertux/rl-welcome-antarctica.toml`
Mindustry	27	`configs/mindustry/eval-veins.toml`	`configs/mindustry/rl-veins.toml`
Craftium	96	`configs/craftium/eval-chop-tree.toml`	`configs/craftium/rl-chop-tree.toml`
IKEMEN GO	9	`configs/ikemen/eval-kfm.toml`	`configs/ikemen/rl-kfm.toml`
Open Surge	33	`configs/opensurge/eval-sunshine-1.toml`	`configs/opensurge/rl-sunshine-1.toml`
Quaver	70	`configs/quaver/eval-crossover-beginner.toml`	`configs/quaver/rl-crossover-beginner.toml`
Naev	128	`configs/naev/eval-tutorial.toml`	`configs/naev/rl-tutorial.toml`

The full game docs live under ../../docs/games.

Config Format

Prime configs select the WarGames task:

game = "redalert"
mission = "redalert.soviet-01.normal"
reward_profile = "protective"
recorder_mode = "none"
max_steps = 500
rollouts_per_example = 8

Field	Meaning
`game`	WarGames game ID.
`mission`	Mission ID from `scenarios/<game>/missions/`.
`reward_profile`	Reward profile ID from `scenarios/<game>/profiles/`.
`recorder_mode`	`none`, `summary_only`, or `full`. Use `none` for high-throughput RL.
`max_steps`	Episode step budget. Budget stops are marked as truncated.
`rollouts_per_example`	Prime rollout count for training configs.

Every game ships standard. Red Alert also ships dense, protective, terminal, speedrun, and aggressive_stress_test.

Reward profile schema and per-game fields: ../../docs/reward_profiles.md.

Runtime Contract

For each turn:

Prime sends a model response.
WarGames reads the latest computer-control tool call.
WarGames applies it to the live game.
WarGames returns a JSON status block, the latest frame, and the updated reward.

The model sees rendered game output and the public tool/action history. Trusted game state stays inside WarGames and is only used by the reward profile.

Supported tool calls are the normal WarGames computer controls:

move_mouse
mouse_down
mouse_up
key_down
key_up
scroll
wait

Publishing

The Prime package bundles the WarGames runtime and mission catalog into the published wheel. Publish after main contains the runtime changes:

make -C environments/prime publish

The publish target runs the WarGames harness/evaluation tests, Prime conformance tests, a Prime source archive install check, and then pushes the public layerbrain/wargames environment.