README

ChessPas

ChessPas is a Free Pascal port of the C Gomoku project in ../ChessC/. The game is five-in-a-row on a 15 x 15 board, not standard chess. ChessC is preserved as a read-only reference.

Author and Contribution

This project was created and extended by OpenCode, an AI software engineering assistant, working in the shared repository with the user.

I translated the original ChessC Gomoku project from C to Free Pascal and implemented the following parts:

• Pascal units corresponding to the C source files, including the board, players, game manager, actions, stacks, queues, AI utilities, AC automaton, genetic algorithm, training code, and MT19937 random generator.
• A playable human-versus-AI command-line program in TRGoChessPas.pas.
• Coordinate input in forms such as h8 and 8h.
• Interactive commands for help, undo, serialization, and quitting.
• AI pattern evaluation using the original default pattern weights.
• A compiled Aho-Corasick matcher for efficient pattern detection.
• Alpha-beta search, negamax evaluation, tactical win detection, candidate ordering, and Zobrist caching.
• Multithreaded root-level AI search using Free Pascal threads and critical sections.
• Game save/load and game-record replay support.
• Debug and release build scripts.
• Regression tests covering board geometry, AI tactics, serialization, game state, coordinate parsing, pattern matching, random generators, and utilities.

Build and Run

Install Free Pascal (fpc). The native build has been verified with FPC 3.2.2 on Linux AArch64. On Debian/Ubuntu:

sudo apt update
sudo apt install fpc

Run from this directory (ChessPas/):

sh build.sh
./TRGoChessPas

The build script recreates .build/ for compiler unit output and writes the executable as TRGoChessPas. For Lazarus, open TRGoChessPas.lpi.

Alternatively, from the parent repository root:

sh ChessPas/build.sh
./ChessPas/TRGoChessPas

Playing

The human player is X and moves first. The AI player is O. Complete five or more contiguous stones horizontally, vertically, or diagonally to win. A full board without a winning line is a draw.

Enter coordinates such as h8 or 8h, using columns a through o and rows 1 through 15. Coordinates are case-insensitive; occupied squares and trailing characters are rejected.

Commands during a game:

?        Show help
-        Undo the last round
>        Print serialized game data
q        Quit

End-of-input also exits the program. Loading saved games is supported by the Pascal API, not by an interactive command.

Tests

From this directory, run:

sh test.sh

The script rebuilds and runs all five regression programs with range, overflow, I/O, assertion, and heap checks. Build output stays in .build/tests/. Set FPC=/path/to/fpc to select a different compiler for the test script.

All five programs passed on Linux AArch64 with runtime checks enabled and zero leaked blocks reported by heap tracing. Full training runs and cross-platform builds have not been verified.

Cleanup

From this directory, run:

sh clean.sh

This removes .build/, .release-build/, lib/, loose compiler output, and local executables. It preserves Pascal sources, Lazarus project/settings files, release artifacts, Git metadata, and everything under ../ChessC/. Training and supporting units remain part of the port even though the interactive executable does not invoke training.

Port Notes

Actions and integer fields retain the C eight-character, least-significant-nibble-first serialization. Raw board serialization converts Pascal's internal stone values to the C wire encoding. Game and record loaders reject malformed fields, illegal or duplicate moves, nonalternating turns, and inconsistent game state. Game loading preserves player types.

TGame does not own its players, matching FreeGame in C. Callers must free players separately, including players created by a successful LoadGame call.

Correctness fixes include anti-diagonal bounds, shared five-in-a-row detection, negamax terminal scores, exact-only search caching, undo turn restoration, full-board handling, complete pattern failure-chain output, RNG isolation, queue capacity handling, and genetic crossover parent preservation.

Known Limitations

The AI rebuilds its evaluation state from the board before each move. It uses a six-placement search with neighboring candidates, immediate-threat filtering, and multithreaded root search. It is not expected to produce identical moves to the C engine.

General opening searches at the default depth can take more than 20 seconds. The tactical tests are not a playing-strength or latency benchmark. Search depth and worker count are configured by AIDepth and CoreCount in Globals.pas; changing them requires a rebuild.

Release Builds

From this directory, run:

sh release.sh

The script is configured for a Linux AArch64 host: it requires a Linux AArch64 build and attempts Linux x86_64 and Windows x86_64 cross builds. Successful binaries are placed in the parent repository's release directory:

../release/linux-aarch64/
../release/linux-x86_64/
../release/windows-x86_64/

Each successful target also receives a BUILD_INFO.txt file. Cross builds require the corresponding Free Pascal compiler, RTL units, linker, target libraries, and, for Windows, MinGW-w64. Failed cross builds are skipped rather than replaced with placeholder binaries. Set FPC=/path/to/fpc if needed.