Taking a bot between Cogriculture and Kaggle

Taking a bot between Cogriculture and Kaggle

Bots move in both directions with no edits. This is a design property, not a convenience: the game container runs the upstream simulator, and the observation sent over the wire is the upstream observation dict untouched, so there is no translation layer where the two could diverge.

The contract

def agent(obs, config):
    return {"farmer": [...], "hands": [...], "market": [...]}

• obs — the Kaggriculture observation dict (see RULES.md).
• config — the resolved episode configuration. Optional; a one-argument agent(obs) works in both places too.
• Returns the action dict described in player_protocol.md.

Two rules keep a bot portable:

1. One self-contained file, standard library only. Kaggle submissions are a single file. bots/baseline.py mirrors the handful of rulebook constants it reasons about rather than importing them for exactly this reason.
2. Stay a pure function of the observation. Kaggle may run your agent in a fresh process, and the Coworld adapter makes no promise about state either. Anything you need to know, read from obs.

Cogriculture to Kaggle

# in a Kaggle notebook
from kaggle_environments import make

env = make("kaggriculture", configuration={"episodeSteps": 720})
env.run(["/kaggle/working/baseline.py", "starter"])
env.render(mode="ipython", width=800, height=800)

Or submit the file directly to the competition. Check it first against the same simulator Kaggle runs:

python tools/eval.py path/to/mybot.py starter --episodes 5

Kaggle to Cogriculture

Drop the file into bots/ and point the adapter at it:

python -m cogriculture.player --bot bots/mybot.py

Locally, against the real protocol:

python tools/run_local_episode.py --bots bots/mybot.py bots/baseline.py

Submitting to a Softmax league

The player image already contains the adapter; a policy is the image plus the bot to run.

coworld upload-policy <cogriculture-image> \
  --name my-cogriculture-bot \
  --run python --run -m --run cogriculture.player \
  --run --bot --run bots/mybot.py

For a bot that is not in the image, bake it into a thin layer on top:

FROM <cogriculture-image>
COPY mybot.py /app/bots/mybot.py

Where the two differ

Only in transport and in what gets recorded — never in the rules.

The timeout row matters most in practice: a bot that is comfortably fast in a notebook is also fast over a local websocket, but actTimeout is enforced the same way in both, so an agent that occasionally spikes past a second will drop turns in either place.