Can AI farm shrimp?

# Excerpt from the measured ReAct caretaker loop.
# Source: reef-shrimp-cohort/envs/reef_freshwater_env/examples/react_agent.py
# This is the agent decision surface only — not the full runner script.

ALLOWED_KINDS = {
    "feed",
    "water_change",
    "vacuum_substrate",
    "rinse_filter",
    "dose_minerals",
    "dose_ial",
    "test_kit",
    "buy_instrument",
    "sell_batch",
    "buy_catalog",
    "check_next",
}


def policy_decision(observation, history, budget, limits, reasoning_effort):
    """One ReAct turn: observe → reason → choose exactly one tool call."""
    public_history = []
    for transition in history[-6:]:
        prior_observation = json.loads(json.dumps(transition["observation"]))
        prior_scorecard = prior_observation.get("scorecard") or {}
        prior_scorecard.pop("catalog", None)
        prior_scorecard.pop("juvenile_resale_sources", None)
        prior_scorecard.pop("purchase_history", None)
        public_history.append({
            "turn": transition["turn"],
            "action": transition["action"],
            "observation": prior_observation,
            "reward": transition["reward"],
            "done": transition["done"],
            **({"action_error": transition["action_error"]} if transition.get("action_error") else {}),
        })

    prompt = f"""You are the caretaker policy for a long-horizon Crystal Red Shrimp breeding game.
Objective: maximize net USD reward. Juveniles become saleable only in complete same-grade batches of 20. Reward is saleable juvenile batch value by grade minus cumulative equipment purchases.
You may sell a complete batch, but the 20 sold juveniles are permanently removed and cannot mature into the next breeding generation. Weigh realized revenue against future colony growth.

CURRENT HARD LIMITS AND BUDGET LEFT:
{json.dumps(limits, separators=(",", ":"))}

The episode terminates at the first of: the simulation-day limit; no living adult shrimp; the action-turn limit; or refusal of the next model call because its conservative cost reserve would exceed the inference-dollar limit. Terminal accounting values realized sales and complete same-grade juvenile batches only. Adults, shrimplets, incomplete juvenile batches, unused supplies, and durable equipment have zero terminal value. As a termination boundary approaches, liquidate saleable juveniles unless keeping them has enough remaining time and budget to produce more counted value.

Crystal Red Shrimp targets: pH 5.5-6.5, GH 4-6 dGH, KH 0-2 dKH, ammonia 0-0.05 ppm, nitrite 0 ppm, and nitrate below 20 ppm. pH 5.6 and KH 1 are normal, not emergencies. In a stable filtered tank, chemistry kits generally need monthly checks, not checks after every time jump or dose. Do not chase normal acidic CRS parameters with repeated mineral dosing.

Use only this public observation; hidden chemistry and private grade counts are unavailable:
{json.dumps(observation, separators=(",", ":"))}

Recent trajectory:
{json.dumps(public_history, separators=(",", ":"))}

Choose exactly one action. Available actions and parameters:
- feed {{"discs": 1..4}}
- water_change {{"fraction": 0.1..0.35, "source": "ro_gh_plus", "powder_scoops": 1.0}}
- vacuum_substrate {{"intensity": 0.1..1.0}}
- rinse_filter {{"id": "filter"}}
- dose_minerals {{"scoops": 0.25..1.0}}
- dose_ial {{"leaves": 0.25..1.0}}
- test_kit {{"channel": "ph"|"nh3"|"nitrite"|"nitrate"|"gh"|"kh"}}
- buy_instrument {{"kind": "ph_meter"|"do_meter"|"thermometer", "tier": "cheap"|"standard"|"lab"}}
- buy_catalog {{"item": <any item in scorecard.catalog>, "quantity": 1..20}}
- sell_batch {{}}
- check_next {{"after_days": 1..90}}, {{"after_hours": 6}}, or {{"until": "next_day"}}

Prefer gathering missing evidence before risky interventions. Avoid buying equipment unless expected value exceeds its cost. When the tank is stable, use check_next with after_days to advance to biologically meaningful results; do not poll every six hours. Use 1-3 days when monitoring a recent concern and 30-90 days when waiting for breeding or juvenile growth. Do not jump beyond the episode horizon. Return a short reason, action kind, and params_json. params_json must be a string containing the JSON object of action parameters."""

    decision = model_decision(prompt, SCHEMA, "shrimp_farm_action", budget, reasoning_effort)
    raw_params = decision.pop("params_json")
    try:
        decision["params"] = json.loads(raw_params)
    except (json.JSONDecodeError, TypeError):
        decision["params"] = {}
        decision["format_error"] = f"invalid params_json: {raw_params!r}"
    if decision.get("kind") not in ALLOWED_KINDS:
        raise ValueError(f"policy selected unsupported action: {decision}")
    return decision


def react_loop(env, setup, budget, max_turns, max_days, reasoning_effort):
    """Measured episode: setup once, then up to max_turns observe → decide → act."""
    obs = env.reset()
    history = []
    for turn in range(1, max_turns + 1):
        limits = decision_limits(public_observation(obs), budget, turns_used=len(history), max_turns=max_turns, max_days=max_days)
        decision = policy_decision(public_observation(obs), history, budget, limits, reasoning_effort)
        if decision["kind"] == "check_next":
            obs = env.check_next(decision["params"])
        else:
            obs = env.step(ReefAction(kind=decision["kind"], params=decision["params"]))
        history.append({
            "turn": turn,
            "action": decision,
            "observation": public_observation(obs),
            "reward": env.last_reward,
            "done": env.last_done,
        })
        if env.last_done:
            break
    return history