Improve quiescence search

[luccabb]

Summary

• Add draw detection (fifty-move rule, insufficient material, repetition)
• Properly handle check positions (search all evasions, not just captures)
• Add is_tactical_move() helper for cleaner tactical move detection
• Use best_score tracking instead of just alpha

Details

Draw Detection

The quiescence search now detects draws:

• Fifty-move rule: board.is_fifty_moves()
• Insufficient material: board.is_insufficient_material()
• Repetition: Check board.is_repetition(2) after each move

Check Handling

When in check, the position is unstable and we can't use stand-pat for pruning:

• Search ALL legal moves (evasions), not just tactical moves
• Don't use stand-pat beta cutoff
• Initialize best_score to -inf instead of stand_pat

Tactical Move Detection

New is_tactical_move() function that identifies:

• Captures (material change)
• Promotions (significant material gain)
• Checks (forcing moves)

Previously used is_zeroing() which included quiet pawn pushes.

Test plan

• All 64 unit tests pass
• Verified proper handling of check positions

---

🤖 Generated with Claude Code

[github-actions]

🔬 Stockfish Benchmark Results

vs Stockfish Skill Level 3

Metric	Wins	Losses	Draws	Total	Win %
Overall	32	61	7	100	32.0%
As White	17	30	3	50	34.0%
As Black	15	31	4	50	30.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 5

vs Stockfish Skill Level 4

Metric	Wins	Losses	Draws	Total	Win %
Overall	20	75	5	100	20.0%
As White	12	36	2	50	24.0%
As Black	8	39	3	50	16.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 5

vs Stockfish Skill Level 5

Metric	Wins	Losses	Draws	Total	Win %
Overall	6	86	8	100	6.0%
As White	3	42	5	50	6.0%
As Black	3	44	3	50	6.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 6
• Draw by fifty moves rule: 1

Configuration

• 5 chunks × 20 rounds × 3 skill levels = 300 total games
• Each opening played with colors reversed (-repeat) for fairness
• Moonfish: 60s per move
• Stockfish: 60+5 time control

[greptile-apps]

Greptile Overview

Greptile Summary

This PR significantly improves the quiescence search implementation by adding draw detection (fifty-move rule, insufficient material, threefold repetition), properly handling check positions by searching all evasions instead of just captures, and refactoring tactical move detection into a cleaner is_tactical_move() helper function.

Key improvements:

• Draw detection prevents the engine from searching drawn positions indefinitely
• Check handling ensures all legal evasions are searched when the king is in check, preventing tactical oversights
• The new is_tactical_move() function correctly identifies captures, promotions, and checks (excluding quiet pawn pushes that were incorrectly included via is_zeroing())
• Best score tracking replaces returning alpha, which properly handles positions where no moves improve the position

The changes align well with standard quiescence search algorithms and should improve tactical accuracy.

Confidence Score: 4/5

• Safe to merge with high confidence - improves search correctness with standard techniques
• The implementation follows well-established quiescence search patterns and all unit tests pass. The logic is sound for draw detection, check handling, and tactical move filtering. Minor style issue with unnecessary type annotation.
• No files require special attention

Important Files Changed

Filename	Overview
moonfish/engines/alpha_beta.py	Improved quiescence search with draw detection, proper check handling, and best_score tracking
moonfish/move_ordering.py	Added is_tactical_move() helper for cleaner tactical move detection (captures, promotions, checks)

Sequence Diagram

sequenceDiagram
    participant QS as quiescence_search
    participant Board
    participant MO as organize_moves_quiescence
    
    QS->>Board: is_check()
    alt In Check
        Board-->>QS: true
        QS->>QS: Set best_score = -inf
        QS->>Board: Get all legal moves (evasions)
    else Not In Check
        Board-->>QS: false
        QS->>Board: eval_board() → stand_pat
        alt stand_pat >= beta
            QS-->>QS: Return beta (cutoff)
        end
        QS->>QS: Set best_score = stand_pat
        QS->>QS: Update alpha = max(alpha, stand_pat)
        QS->>MO: organize_moves_quiescence()
        MO->>MO: Filter tactical moves (captures, promotions, checks)
        MO-->>QS: tactical_moves
    end
    
    loop For each move
        QS->>Board: push(move)
        QS->>Board: is_repetition(2)?
        alt Repetition Draw
            Board-->>QS: true
            QS->>QS: score = 0
        else Not Repetition
            Board-->>QS: false
            QS->>QS: Recursive quiescence_search()
            QS-->>QS: score = -recursive_result
        end
        QS->>Board: pop()
        
        alt score > best_score
            QS->>QS: Update best_score
        end
        
        alt score >= beta
            QS-->>QS: Return beta (cutoff)
        end
        
        QS->>QS: Update alpha = max(alpha, score)
    end
    
    QS-->>QS: Return best_score

Loading

[greptile-apps]

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

unnecessary type annotation on variable assignment

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

Prompt To Fix With AI

This is a comment left during a code review.
Path: moonfish/engines/alpha_beta.py
Line: 143:143

Comment:
unnecessary type annotation on variable assignment

<sub>Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!</sub>

How can I resolve this? If you propose a fix, please make it concise.

[github-actions]

🔬 Stockfish Benchmark Results

vs Stockfish Skill Level 3

Metric	Wins	Losses	Draws	Total	Win %
Overall	28	64	8	100	28.0%
As White	15	31	4	50	30.0%
As Black	13	33	4	50	26.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 7

vs Stockfish Skill Level 4

Metric	Wins	Losses	Draws	Total	Win %
Overall	27	72	1	100	27.0%
As White	19	30	1	50	38.0%
As Black	8	42	0	50	16.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 1

vs Stockfish Skill Level 5

Metric	Wins	Losses	Draws	Total	Win %
Overall	13	82	5	100	13.0%
As White	11	37	2	50	22.0%
As Black	2	45	3	50	4.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 4

Configuration

• 5 chunks × 20 rounds × 3 skill levels = 300 total games
• Each opening played with colors reversed (-repeat) for fairness
• Moonfish: 60s per move
• Stockfish: 60+5 time control