Play and fuzz games on gorge¶
Three tools drive gorge without a server. All read the corpus in .cards.
| Tool | Use it to |
|---|---|
forgec |
fetch the Forge corpus, compile it to the IR, report which cards are playable |
mtgsim |
play seeded games between bots and replay each one to check its event chain |
cardfuzz |
play many random and exploring games, record failures, and repeat one with -repro |
forgec¶
go run -trimpath ./cmd/forgec fetch # pinned corpus into .cards
go run -trimpath ./cmd/forgec compile # write the IR cache
go run -trimpath ./cmd/forgec report # playable cards and the primitives still missing
forgec coverage writes coverage.md and summary.md into -out. The card tables are on
Card support.
mtgsim¶
go run -trimpath ./cmd/mtgsim -games 20 -verify -seed 1
go run -trimpath ./cmd/mtgsim -seats 2 -a bot -b lethal-pressure
go run -trimpath ./cmd/mtgsim -seats 4 -format commander -policies az-redeal,lethal-pressure,-,forge-heuristic
Game i uses seed + i. A game that doesn't finish, or doesn't replay under -verify, is an error and a non-zero exit.
-policies names a policy for every seat (- keeps the plain heuristic); without it -a and -b name seats 0 and 1.
Each game line ends with policies=, and a seat that searched gets a search seat line with its decisions, simulations
and tree nodes. Seats are asked the way gorged asks them: a search policy such as az-redeal answers from an honest
root dealt from its own observation feed, and a heuristic seat from the same reused board.
-sb-agent name=/path adds a SpellBench v2 stdio agent as a policy.
cardfuzz¶
go run -trimpath ./cmd/cardfuzz -games 1000 -workers 4
go run -trimpath ./cmd/cardfuzz -report -missing # which cards and abilities were never used
go run -trimpath ./cmd/cardfuzz -repro cardfuzz-failures.jsonl -line 1
It keeps its coverage in -state and appends failures to -failures. Each failure replays with -repro.
The flags for -spellbench and -manabrew run the same games through those wires. See every flag.