TURN-BASED GAME DESIGN DOCUMENT

(Environment Name: “Stellar Orchard”)

1. Concept Paragraph

Stellar Orchard is a deterministic, turn-based strategy game where two rival horticulturists compete to cultivate the most thriving orchard on a distant exoplanet. The environment simulates a grid of orchard plots, each plot able to host an alien tree that produces lumen fruit—a glowing bioengineered product that yields energy points. Players must choose each turn to Plant, Nurture, or Harvest specific plots to maximize their total energy yield before the season ends. Weather and soil conditions are fixed by a deterministic seed at reset, ensuring reproducibility. The design, theme, and terminology are completely unrelated to any negotiation or trading example.

2. Roles and Win Condition

Player Roles:
- Player A (Solar Gardener)
- Player B (Lunar Gardener)
  Each controls their own half of the orchard (plots tagged A1–A5 for A; B1–B5 for B).
Objective:
Accumulate the highest Energy Points (EP) via strategic planting, nurturing, and harvesting of trees.
Win Conditions:
- The game ends after 10 turns or when both players have harvested all trees.
- The player with the highest cumulative EP wins.
- If total EPs are equal, the result is a draw.
- If a player performs two invalid moves consecutively, they forfeit the match and lose automatically.

3. Turn Structure and Determinism

Fixed alternating turns: Player A → Player B → Player A → ...
Each turn, a player chooses one valid action.
Game ends after Turn 10 (each player acts 5 times).
A random_seed parameter initializes soil fertility levels and initial weather pattern with reproducible deterministic effects.
No hidden randomness during play; outcomes are computed deterministically based on seed and previous actions.

4. Action Grammar (Machine-Parseable)

All player actions are deterministic commands describing their move this turn. Actions must match exactly one of the grammar patterns below.

Action Type	Format	Description and Rule	Example (Valid)	Example (Invalid) and Reason
Plant	`Plant:<plot>`	Plants a new tree on the specified empty plot. `<plot>` ∈ {A1–A5 (for Player A), B1–B5 (for Player B)}	`Plant:A3`	`Plant:C2` → Invalid (nonexistent plot)
Nurture	`Nurture:<plot>`	Boosts growth stage of one existing, unharvested tree on a valid occupied plot.	`Nurture:B4`	`Nurture:B6` → Invalid (out of range)
Harvest	`Harvest:<plot>`	Harvests a fully grown tree from the specified plot, removing it and collecting EP.	`Harvest:A1`	`Harvest:A1,A2` → Invalid (multiple plots not allowed)
Pass	`Pass`	Player skips the turn intentionally (strategic or necessary if no valid move).	`Pass`	`[Pass]` → Invalid (extra brackets not part of syntax)

Regular Expression Patterns

^Plant:(A[1-5]|B[1-5])$
^Nurture:(A[1-5]|B[1-5])$
^Harvest:(A[1-5]|B[1-5])$
^Pass$

All valid actions will be wrapped by players inside \boxed{{}} at runtime, e.g., \boxed{{Harvest:A1}}.

5. Game State Schema

Example game_state format:

{
  "turn_number": 3,
  "max_turns": 10,
  "active_player": "Solar Gardener",
  "plots": {
    "A1": {"owner": "A", "status": "grown", "growth_level": 3},
    "A2": {"owner": "A", "status": "empty", "growth_level": 0},
    "A3": {"owner": "A", "status": "seedling", "growth_level": 1},
    "B1": {"owner": "B", "status": "grown", "growth_level": 3},
    "B2": {"owner": "B", "status": "harvested", "growth_level": 0}
  },
  "energy_points": {
    "A": 15,
    "B": 12
  },
  "soil_fertility": {
    "A1": 0.9,
    "A2": 0.6,
    "B1": 0.8
  },
  "weather_pattern": "Radiant Skies",
  "transcript": [
    {"player": "A", "action": "Plant:A3"},
    {"player": "B", "action": "Nurture:B1"}
  ],
  "winner": null,
  "random_seed": 57
}

6. Initialization Rules

random_seed is set during environment reset; it governs deterministic generation of:
- soil_fertility for each plot (each between 0.5 and 1.0).
- weather_pattern (chosen deterministically from a fixed set: Radiant Skies / Lunar Mist / Crystal Winds).
energy_points[A] and [B] start at 0.
All plots start as "empty".
Observations include the common introduction, initial weather, soil fertility summary, and allowed actions.

7. Validation and Error Handling

On every step:

Extract literal content from \boxed{{...}} using _extract_answer_content.
Validate syntax against regex patterns.
Check logical validity:
- Plant → must target empty plot owned by that player.
- Nurture → must target already planted, not yet grown tree.
- Harvest → must target grown, unharvested tree.
- Pass → always valid.
Any violation triggers set_invalid_move(player, reason), e.g.:
- "Invalid format"
- "Plot not owned by player"
- "Plot already occupied"
- "Tree not ready to harvest"

8. Terminal Conditions and Scoring

Automatic End:
- When turn_number >= max_turns
- OR when all plots are harvested or empty (no active trees remain)
Scoring:
- For each harvested tree: EP += int(10 * soil_fertility[plot]).
- No fractional points.
- Final score is the total EP accumulated.
Result Determination:
- Higher total → Winner.
- Equal totals → Draw.
- Consecutive double invalid move by a player → automatic Loss.

9. Player Prompt Specification

Identity Context:
“You are a cosmic horticulturist tending bioluminescent trees on the exoplanet Selora. Your goal is to maximize your orchard’s energy yield before the season ends.”

Prompt Structure Includes:

Current turn number and remaining turns.
Your current EP and plots' status summary.
Weather pattern and soil fertility hints.
List of all valid actions:
- Plant:<plot>
- Nurture:<plot>
- Harvest:<plot>
- Pass
Explicit response requirement: “Put your final answer within \boxed{{}} at the end of your response.”

Few-shot Examples:

Example valid response:
I will start by planting my first tree in plot A2.
\boxed{{Plant:A2}}

Example invalid response:
Let's see how this goes!
\boxed{{Grow:A2}}    # Invalid because 'Grow' is not a recognized action type.

Dialogue Elements: None (pure command game). All player utterances (including justification text) get appended to transcript for transparency.

10. API Mapping Plan

reset(seed: Optional[int])
- Initializes game_state using the deterministic seed.
- Builds initial soil fertility and weather data.
- Clears all plots, transcript, and EPs.
- Returns the observation containing theme introduction and available actions.
step(action: str)
- Extracts action content via _extract_answer_content.
- Validates syntax + legality.
- Updates game state deterministically (adjust plot status, growth level, EPs).
- Adds the player’s message and resulting action to transcript.
- Advances to the next turn or marks the game as terminal if conditions met.
- Returns updated observation, reward, and termination flag.
_generate_player_prompt(player_id)
- Constructs text prompt including:
  - Game identity summary
  - Current turn, soil info, plot statuses
  - Weather condition and EP scores
  - List of valid actions and formatting requirement
  - Few-shot examples above
- Returns formatted prompt string.

11. Copy-Check Against the Example

All terms—plots, trees, lumen fruit, Energy Points, soil fertility, weather pattern, Solar/Lunar Gardeners, and orchard management theme—are original and entirely unrelated to any negotiation, trading, or economic dialogue example.
The game_state keys (plots, soil_fertility, energy_points, etc.) and all action types are unique to Stellar Orchard and do not replicate any element from other sample environments.

8.2 KiB Raw Blame History Unescape Escape