cmu-agents/assignment-1/select-move

Use available tools to select a good move.

The two tools

Everything below is built out of exactly two calls, and the distinction between them matters:

• simulate_move(fen, move=None) — the search tool. It returns a dict for

the position, with fen, legal_moves, game_over, winner and result. Passing move=None just inspects a FEN; passing a UCI move plays that one ply and returns the position after it. It never touches the live board and never invokes the opponent, so chain its returned fen values to roll out a line as deep as you like, for either side.

• play_move(move) — the commit tool. It plays one move on the real board

and the opponent answers. Call it exactly once, at the very end, with the move the search chose. Never call it to explore.

Both are available inside run_python as ordinary synchronous functions — assume they are already imported when your code is executed, and write only the relevant snippet, with no import lines and no tool-call boilerplate.

Step 1: Opening preferences

If it is the first full move of the game, skip the search entirely. Take the first of these that appears in legal_moves, and play it:

e2e4, d2d4, c2c4, g1f3

If none is legal, fall through to Step 2. Beyond this step, you should use Python code to choose the move. Write code that executes the described strategy so you reliably reason about board state.

Step 2: Order the candidates

Call simulate_move(fen) with move=None on the current position and take legal_moves from the result. Sort by UCI string, ascending — a2a3 before b1c3. This order is what settles ties later, so keep it.

Step 3: Score each candidate two plies deep

For each candidate in order, call simulate_move(fen, candidate). That result is the position after your move; take its legal_moves as Black's replies, and for each reply call simulate_move(after_white["fen"], reply) and score the position that comes back with the evaluation in Step 4. Black is the minimizer, so the candidate's score is the lowest of those replies. There is nothing to undo — every simulate_move starts from the FEN you hand it, so the live board is untouched throughout.

That is the whole search: your move, Black's answer, evaluate. Nothing deeper is ever examined. Stop early at any position whose game_over is true and score it directly from winner and result rather than simulating replies from it.

Step 4: Evaluate a position

Terminal positions first, claiming draws where they are available:

Otherwise sum, in centipawns, from White's point of view:

• Material for every piece on the board — pawn 100, knight 320, bishop 330,

rook 500, queen 900, king 0. Add for White, subtract for Black.

• Centralization: ±12 for each piece standing on d4, e4, d5, or e5. Add for

White, subtract for Black. Every piece counts the same — a pawn on e4 is worth the same 12 as a queen there.

• Check: if the side to move is in check, ±25 — plus if that side is Black,

minus if it is White.

There is nothing else. No king safety, no pawn structure, no mobility, no development or castling term.

Step 5: Play the best score

Take the candidate with the highest score. A later candidate must strictly beat the current best to replace it, so when scores are equal the earliest move in the Step 2 ordering wins.

Then call play_move(best) — once, with that move, as the last statement of the snippet. This is the only call in the whole procedure that changes the actual game board. Do not leave this out -- the computation will have no effect otherwise. Do not just print out the move, call play_move at the end to actually play the move, this will work like a standard play_move tool use.

That call is your move for this turn. The observation you get back shows the board after it, with the opponent's answer already made. Read that new position and start the next turn from it — do not call the play_move tool again for the same move, which would play a second move from a board that has already moved on.

Notes on the shape of this strategy

• The 25-point check term is smaller than a pawn, so it only ever breaks ties.

It never justifies giving up material for a check.

• A 2-ply search sees Black's immediate recapture, so material is not hung for

free, and it finds and avoids mate in one. Anything that pays off on Black's second move is outside the horizon.

• Draws score 0 rather than as a loss, so a repetition is accepted readily from

a worse position.