CogsGuard is the New Default Game
cogames missionsnow shows CogsGuard (cogsguard_arena) as the default and only visible game- Legacy missions (training_facility, hello_world, machina_1) are hidden from the CLI but still accessible by name
- Default season changed to
beta-cogsguard
Season-Aware Policy Validation
cogames uploadandcogames validate-policynow accept--seasonto validate against the correct game for that season- Validation uses the game appropriate for the target season (e.g.,
betaseason validates againsttraining_facility.harvest,beta-cogsguardvalidates againstcogsguard_machina_1.basicusing the Machina1 map) - This ensures policies are validated against the same game they'll compete in
CLI Flag Updates
Breaking changes to command-line arguments for improved consistency:
cogames submissions- replacedpolicy_namewith an optional--policy/-pflag instead of a positional argumentcogames diagnose- removed-mshort form for--experimentscogames diagnose- replaced--repeatswith--episodes/-ecogames make-mission- removed-hand-wshort form for--heightand--widthcogames login- replaced--server/-swith--login-servercogames pickup- added--mission/-mand--variant/-vflags (previously hardcoded tomachina_1.open_world)cogames validate-policy- replacedpolicypositional argument with--policy/-pflag
Tournament CLI Updates
cogames upload- Upload a policy without submitting to a tournamentcogames submit- Submit an uploaded policy to a tournament seasoncogames submissions- View your uploaded policies and tournament submissionscogames leaderboard --season <name>- View the leaderboard for a seasoncogames seasons- List available tournament seasonscogames pickup- Evaluate a candidate policy against a fixed pool and compute VOR
CLI Command Restructuring
-
cogames evalhas been renamed tocogames run -
Commands intended for demonstration purposes to get you up and running are housed under
cogames tutorial:cogames train->cogames tutorial train. This command offers a thin wrapper around pufferlib traincogames make-policy->cogames tutorial make-policy. This command creates an example python implementation of a scripted policy for you to modify- Introduced
--scriptedand--trainablemodes forcogames tutorial make-policy. Each creates an example policy implementation from a template in different style: scripted or neural-net based. cogames tutorial->cogames tutorial play. This is an interactive tutorial that walks you through the basic mechanics of the game.
What's new
Breaking change: The format for specifying a policy in cogames command is now updated to support passing keyword
arguments.
-
cogames
-por--policyarguments should now take the formclass=...[,data=...][,proportion=...] [,kw.name=val] -
Example: uv run cogames play -m hello_world -p class=path.to.MyPolicyClass,data=my_data/policy.pt,kw.policy_mode=fast
Update: cogames submit now accepts much larger checkpoint files. It makes use of presigned s3 URLs for uploading your
checkpoint data instead of going through web servers that enforced lower file size caps.
Update: cogames submit does a light round of validation on your proposed submission before uploading it. It ensures
that your policy is loadable (using the latest copy of cogames in pypi, not necessarily the one you have installed),
and can perform a step on a simple environment.