When Stockfish detects a forced mate, it sets the mate field and resets
the score to 0. The previous code was showing '0 cp' which was misleading.
Changes:
- Format evaluation strings to show 'Mate in X' when mate is detected
- Only show delta (centipawn change) when neither position has a mate
- Update eval instruction to specifically mention mate situations
- Add explanation of mate notation in the prompt
Example output now:
- Pre-Eval: -326 cp
- Post-Eval: Mate in 3
(instead of showing '0 cp' which was confusing)
The previous implementation tried to use undo() on a Chess object created from
a FEN string, which has no move history. This resulted in all three FEN positions
being identical.
Now we use the game object's PGN (which contains full move history) and load it
into temporary Chess objects before undoing moves. This correctly generates:
- FEN before user's move (undo both computer and user moves)
- FEN after user's move (undo only computer move)
- FEN after computer's reply (current position)
This ensures the LLM receives accurate position context for each stage of the
move exchange.
- Include FEN before user's move, after user's move, and current position
- Ensures LLM has complete position context to understand piece locations
- Fixes issue where LLM could 'forget where pieces are' during move analysis
- Add instruction to use FEN data for piece location understanding
- Change openingData from single to array in ChessGame.tsx
- Update all opening lookups to use lookupPossibleOpenings()
- Extract move sequence from PGN for opening lookup
- Update Tutor.tsx to accept array of openings
- Add smart opening instructions based on count:
- 1 opening: Confident identification with metadata
- Multiple: List possibilities, suggest general principles
- None: Focus on position without inventing names
- Pass up to 5 possible openings to LLM in all prompts
- Add FEN and opening data to all user question types (best move, hint, general)
- Prevents LLM from losing track of piece positions during conversation
- Ensures accurate responses based on actual board state
- Fixes issue where tutor would forget piece locations in general questions
- Fix TypeScript errors in analysis page with non-null assertions
- Fix Chessboard component API usage to use options prop
- Fix Analyze Game button placement to be always visible
- Add comprehensive test suite (20 tests) for tactic detection
- Fix pawn detection threshold bug (200 -> 100 centipawns)
- Add error handling to uciToSan for invalid moves
- Integrate tactical data into Tutor component for real-time feedback
- Pass missedTactics from ChessGame to Tutor via props
- Update LLM prompt to explain missed tactical opportunities
- Fix Jest configuration to handle react-markdown ESM issues
- All 50 tests passing
Bug Fixes:
- Fix turn validation: Players can now only move their own pieces
- Fix computer not moving when player moves too quickly (before evalP0 ready)
- Add safety check: Computer move triggered when chat messages sent
- Computer now always responds even if evalP0 is missing (only move history skipped)
New Personalities:
- Add 'Friendly Motivator' - encouraging, positive coach focused on building confidence
- Add 'Bloody Pirate' - theatrical trash-talker with Monkey Island style humor
Personality Reordering:
- Reordered from serious to playful:
1. Opening Professor (most serious)
2. Professional Coach
3. Friendly Motivator
4. Speedrun Super GM
5. Hype Streamer
6. Angry Prodigy
7. Drunk Russian GM
8. Bloody Pirate (most playful)
- Added section headers for clarity (Serious/Professional, Balanced/Entertaining, Spicy/Trash-talking)
Technical Details:
- ChessGame: Added checkAndMakeComputerMove() callback
- Tutor: Calls onCheckComputerMove after sending messages
- Both fixes ensure robust gameplay even with race conditions
- Remove auto-scroll to prevent page jumping when moves are added
- Add evaluation change column (Eval Δ) showing centipawn change per move
- Color-code evaluation changes (green=good, red=bad) based on player color
- Use moveHistory state instead of chess.js history for richer data display
- Keep chess board in view as primary focus
- Create chessFormatDetector utility for automatic format detection
- Update StartScreen with textarea supporting both FEN and PGN input
- Add real-time format detection with visual feedback indicators
- Update translations for all 4 languages (EN, DE, FR, IT)
- Add markdown rendering for Tutor chat messages
- Add comprehensive tests for format detection (20 tests)
- Update game initialization to handle both FEN and PGN formats
This completes the fix/stale-analysis-data branch with:
- Fixed stale evaluation data in hint/best move requests
- Clarified AI's dual role (opponent + tutor) to prevent hint rejection
- Stored complete evaluation history (P0, P1, P2) for all moves
- Improved end-game analysis with better mistake detection
- Fixed duplicate analysis runs with useRef flag
- Added markdown rendering for formatted analysis output
- Added unified FEN/PGN import with auto-detection
This commit fixes a bug where the AI tutor and end-game analysis would use stale data from previous evaluations.
The `Tutor` component now receives the live `game` instance and has a new function, `evaluateCurrentPosition`, which is called on-demand when a user requests a hint or the best move. This ensures the LLM receives up-to-date information.
The end-game analysis was also corrected to use the proper evaluation data when constructing the move history, preventing incorrect analysis of mistakes and blunders.
The test suite was improved by restoring deleted tests, adding a new test to verify the fix, and making existing tests more robust.
Stefan