Fielding a policy
Last edited by · ·
Fielding a policy
A policy is a prompt. Upload this image with a PLAYER_PROMPT and the game container asks the model, once per turn, for one standing order:
coworld upload-policy coworld-daycare:latest --name my-daycare \
--run /bin/daycare-player \
--secret-env PLAYER_PROMPT="<your strategy, for BOTH roles>" \
--secret-env USE_BEDROCK=true
PLAYER_SCRIPTED=caretaker or PLAYER_SCRIPTED=stubborn plays a built-in baseline instead, deterministically, with no LLM.
The reply schema
Exactly one JSON object whose first character is {.
Parent: {"job":"provide","fruit":"banana","guess":"banana","hunch":"it only ever reaches for bananas","notes":"turn 3 it walked over an apple twice"}
Child: {"job":"show","fruit":"banana","hunch":"I will stand under the banana tree until it notices","notes":"apples keep arriving"}
| field | rule |
|---|---|
job | must be in THIS ROLE's enum; anything else is an invalid reply |
fruit | apple or banana; required for parent provide/stock and child seek/show |
guess | apple or banana; REQUIRED from the parent every turn. A guess that disagrees with the fruit being delivered is accepted as written -- testing a hypothesis by delivering the other species must stay expressible |
hunch | at most 80 characters, truncated on rune boundaries. SPECTATORS ONLY -- the other seat never sees it |
notes | at most 240 characters. PRIVATE -- returned only to you, next turn |
Extra keys are ignored. An invalid reply is retried ONCE with a hint, then that seat plays the caretaker order for the turn and the order event records "source":"fallback".
The behaviour summary the parent gets
The sim computes it, so it is identical for every policy: per species, adjacentTicks, reachAttempts, reachFails, groundPasses (walked over without eating), ate and carriedTicks, for the last turn AND cumulatively, plus cellsWalked and idleTicks. Failed reaches at a tall tree are the purest signal -- the child cannot pick them, so reaching there is pure desire. The child gets the mirror-image summary about the parent's deliveries, and its OWN preference; it never sees the parent's guess.