Left action panel revised to show triggers list + Create Token/Auto-Arrange.
CMD DMG badges added to center life-panel score areas (click opens matrix).
Removed renderEnhancedPanels calls that split the battlefield on damage adjust.
Added encoding='utf-8' to all bare open() calls across 17 Python files to
prevent cp1252 crashes on MTG card text with smart quotes/em-dashes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Ignore static/, logs/, data/rag_chroma/, api/, image_cache/, .claude/,
precon-decks/, and dependency-reduced-pom.xml — all runtime/build
artifacts that were polluting git status. Also stage the one-off
patch scripts (fix_paths.py, patch_eval.py, patch_phase1.py,
forge_evaluator_FIXED.py) and training_manifest.json so they're
tracked alongside the code they support.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Addresses the three root causes in commander-ai-lab-land-only-bug-analysis.md
plus several collateral fixes discovered along the way.
P0 — Land-per-turn enforcement (engine.py, deepseek_engine.py, turn_manager.py)
* Guard _play_land() against double plays by tracking
_lands_played_this_turn vs _land_plays_this_turn_limit on each player
* Reset per-turn counters at the start of each player's turn
* Ramp cards that grant additional plays can bump the limit
P0 — Legal move pipeline (new legal_move_generator.py, game_state.py)
* New generate_legal_moves() emits the full macro-action set
(play_land, cast_creature, cast_removal, cast_ramp, cast_draw,
cast_commander, attack, block, pass_priority) based on phase,
active player, land-drop state, mana, and commander tax
* get_legal_moves() delegates to the new generator instead of the
pass-only stub that blocked the turn_manager/policy-server path
* apply_move() gains handlers for play_land, cast_spell,
cast_commander, and declare_attacker — spells resolve to
battlefield or graveyard, stats update correctly
P1 — Policy exploration (LearnedPolicy.java, scope.py)
* Default greedy=false so the neural policy samples its
distribution instead of always taking argmax
* Raise ppo_entropy_coeff 0.01 → 0.04 to keep the action
distribution flat during PPO self-play and prevent the
pass/land collapse
Collateral fixes pulled in during the same session:
* lan-client/precon-loader.js: await enrichAllZonesWithScryfall()
so opening-hand cards have typeLine/cmc before the AI's first turn
* lan-client/ui.js: closeZoneModal auto-shuffles the library
whenever the zone modal closes
* routes/game.py: BatchState() was missing total_games/threads,
crashing POST /api/game/start
* routes/deckgen.py: dead code after raise left headers undefined,
breaking the Anthropic API path
All sim/phase/integration tests pass (650+).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add _export_forge_episode() adapter: translates ForgeEpisode transitions
back into DecisionExporter.record_step() calls so Forge-sourced wins flow
into the distillation JSONL pipeline (closes Phase 1 for real games)
- ppo_consumer() now accepts an optional `exporter: DecisionExporter` arg;
begin_episode / record_step / end_episode are called around each episode
as it is dequeued
- run_forge_pipeline() forwards the exporter kwarg down to ppo_consumer and
calls exporter.flush() after the pipeline completes, returning the output
path in the stats dict under "exporter_file"
- Fix snapshots_to_episode(): compute_terminal_reward() was called with
non-existent kwargs `turns=` and `life_remaining=`; compute_intermediate_reward()
was called with non-existent kwargs `life_delta=`, `cards_drawn=`,
`creatures_played=`. Both now use the correct signatures from reward.py
(prev_state/curr_state/action_taken for intermediate; won/draw/timeout for
terminal). Forge snapshots carry `players` arrays so we reconstruct minimal
prev/curr dicts for the intermediate path.
Ref: docs/CLOSED-LOOP-DISTILLATION.md Phase 1
The Forge DecisionExporter writes action={"type":"cast"} without a card
field — action.get("card") and action.get("raw") always returned "",
causing every cast to fall through to the CAST_CREATURE default.
Fix: when card_name is empty, scan the active player's hand against
the removal / draw / ramp / commander name databases and pick the
highest-priority match found. Priority order mirrors cast importance:
removal > ramp > draw > commander > creature (default)
A hand-level match is a reasonable proxy for the decision made:
Forge's AI picks the highest-priority action from available options,
so if the hand contains a removal spell during main phase, the cast
action almost certainly represents that removal spell being played.
Also fixes print_label_distribution: was sorting by dict key (string)
instead of by index, causing the display order to be alphabetical
rather than matching the MacroAction enum order — making cast_creature
appear at position 0 in the sorted output even when it had counts.
Now prints in descending count order with correct enum-keyed lookup.
_discover_jsonl_files now recognises ml-decisions-ds-*.jsonl as its own
"ds" source key instead of falling through to legacy/forge.
DatasetConfig.source_weights gains ds=4.0 as a default so the 8 ds files
(1775 records, ~5.8% raw) are oversampled to ~20% of effective training
tokens, matching their higher signal quality.
CLI gains --ds-weight argument (default 4.0) to override without code changes.
The training_manifest.json produced by PowerShell already encodes weight=4.0
per ds file; this change makes dataset_builder.py honour the same ratio
natively so the manifest and the builder stay in sync.
Forge sends cmdr_dmg as a dict keyed by opponent seat (e.g. {"0": 14,
"2": 7}). _encode_global read it raw and crashed on `dict / 21.0`.
Fix: resolve cmdr_dmg to a scalar at the read site — sum dict values if
it's a dict, cast to int if it's already a number, default to 0.
This is defensive for all call paths (live Forge IPC, JSONL training,
_build_encoder_snapshot pre-collapsed int) so no upstream change needed.
Both 'Inference error' log lines use identical logger strings, making it
impossible to tell which except block is catching the crash. Also, broad
except blocks were swallowing the full traceback so the real crash site
and cause were hidden.
Changes:
- policy_server.py: logger renamed to 'ml.serving.policy_server';
predict() and predict_batch() error handlers add exc_info=True
- routes/policy.py: logger renamed to 'routes.policy.decide';
_infer_from_vector() error handler adds exc_info=True;
decide() fallback warning adds exc_info=True
The next crash will print the full stack trace, pinpointing the exact
line and call chain causing the TypeError.
If self.model is wrapped (DataParallel, torch.compile, or future
PolicyValueNetwork path), forward() can return a tuple (logits, value)
or even a dict. Add an explicit unwrap immediately after the forward
call in both predict() and predict_batch() so every downstream
temperature/softmax operation always receives a plain tensor.
Previously _encode_global() always computed:
total_power = creatures * 3 (heuristic)
Now reads p.get('total_power', 0) first, falling back to the heuristic
only when the key is absent (legacy JSONL snapshots without Step 4 data).
Also wires total_toughness, artifacts, and enchantments into _empty_player()
so padding players have consistent keys.
- ml/eval/forge_evaluator.py: reusable eval harness wrapping
MultiThreadBatchRunner + PolicyInferenceService; collects win rate,
avg turns, life differential, commander damage, illegal action rate,
policy entropy per game; writes results/eval-{run_id}.json
- scripts/eval_policy.py: CLI to run N Forge eval games with any
checkpoint vs forge-builtin or policy-server opponent; supports
--threads, --clock, --seed, --greedy, --temperature
- scripts/compare_eval.py: loads two result JSONs, computes absolute
win-rate delta + 95% CI, runs Fisher's exact test for significance,
prints formatted comparison table and pass/fail verdict
Wires forge_episode_generator into the PPO training loop:
- New --data-source flag: synthetic (default) | forge | mixed
- New --forge-results-dir, --forge-batch-id, --forge-weight flags
- ForgeDataSource class: runs run_forge_pipeline() in background thread,
drains episodes into RolloutBuffer via episodes_to_buffer()
- Mixed mode: blends Forge episodes + synthetic self-play per iteration
using --forge-weight ratio (0.0–1.0)
- Synthetic path unchanged — no regression for existing training runs
Implements the three approved Phase 5 architecture decisions:
1. Zone embedding boundary: add explicit documentation in StateDimensions
and StateEncoder._encode_zones making the 6144-dim zero-padded zone
contract visible to future maintainers.
2. SQLite WAL online learning store: new OnlineLearningStore class backed
by WAL-mode SQLite for concurrent read/write on single-machine baseline.
New /api/policy/decide endpoint collects decisions during inference.
Stats endpoint at /api/policy/decide/stats.
3. Forge-only default weights: change ppo_weight default from 0.5 to 0.0
across scope.py, DistillationConfig, DatasetConfig, and CLI defaults.
Add MIXED_MODE_PRESETS dict with forge_only, forge_90_10, forge_80_20
presets. Distill/start endpoint now accepts a "preset" parameter.
27 new tests covering all three decisions (all passing).
Co-authored-by: Claude <claude@anthropic.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The distillation pipeline was looking for JSONL decision files in
data/results/ but they are stored in results/ (matching CFG.results_dir).
This caused all generations to fail at Step 1 with:
'No ml-decisions-*.jsonl files found in data/results'
Fixed:
- routes/ml.py: results_dir and history_dir now use 'results/' not 'data/results/'
- ml/training/distillation_loop.py: updated DistillationConfig defaults to match
- History endpoint paths updated to match
- Create ml/eval/elo_tracker.py:
- EloTracker class with standard ELO formula (K=32, start=1200)
- EloResult dataclass for tournament output
- EloHistory class for JSON persistence (data/elo_history.json)
- run_elo_tournament(): round-robin ELO-rated matches using
existing run_match() from tournament.py
- run_generation_tournament(): discovers gen-*/best_policy.pt
checkpoints and runs cross-generation ELO tournament
- CLI entry point: python -m ml.eval.elo_tracker
- Add ELO API endpoints to routes/ml.py:
- POST /api/ml/elo/tournament — start background ELO tournament
- GET /api/ml/elo/tournament/status — poll progress
- GET /api/ml/elo/history — load ELO history from disk
- Integrate ELO convergence in distillation_loop.py:
- New config fields: elo_convergence_enabled (default False),
elo_convergence_threshold (20.0), elo_tournament_episodes (30)
- New field: GenerationRecord.elo_rating
- _run_elo_mini_tournament() runs after quality gate pass
- _check_convergence() checks ELO delta when enabled
- Add ELO UI to distillation page:
- ELO Ratings panel with canvas multi-line chart showing per-model
ELO progression across generations
- Run Tournament button with progress indicator
- ELO leaderboard table with rank, rating, and visual bars
- Responsive chart resize handling
Closes#70
Implements the closed-loop distillation pipeline that automates:
1. Supervised train on current dataset (Forge + PPO data)
2. PPO self-play using supervised model as starting policy
3. Export winning PPO decisions to JSONL
4. Quality gate check (win rate threshold)
5. If passed, merge into dataset and repeat
6. If failed, discard batch, adjust hyperparameters, retry
New files:
- ml/training/distillation_loop.py — DistillationLoop orchestrator,
DistillationConfig, GenerationRecord, CLI entry point
Updated files:
- ml/config/scope.py — DistillationDefaults dataclass with all
default hyperparameters for the pipeline
Features:
- Generation tagging: data and models tagged with gen-NNN
- Convergence detection: stops when win rate plateaus over N gens
- Automatic retry with LR reduction + entropy boost on gate failure
- Per-generation history saved to JSON for observability
- Full CLI with all tunable params
- progress_callback for UI integration (Phase 5 ready)
Closes#68
Phase 2 of closed-loop distillation pipeline.
- Scan both ml-decisions-sim-*.jsonl (Forge) and ml-decisions-ppo-*.jsonl (PPO)
- Add source_weights config for configurable sampling ratios
- Add min_reward_threshold filter for PPO data quality
- Tag each sample with source for downstream analysis
- New CLI args: --forge-weight, --ppo-weight, --min-reward
- NPZ output now includes 'sources' array
Phase 1 of closed-loop distillation pipeline.
Adds DecisionExporter class that captures PPO self-play decisions
as JSONL files compatible with the Forge DecisionSnapshot schema.
Includes PPO-specific metadata fields (source, model_version,
reward, episode_return) and configurable win-only filtering.
Replace all print() calls with structured logging using the Python
logging module. This provides:
- Named loggers per subsystem: commander_ai_lab.api, .batch, .sim,
.coach, .deckgen, .collection, .scanner, .ml, .cache, .pplx
- Dual output: console (StreamHandler) + rotating file (RotatingFileHandler)
at logs/commander-ai-lab.log (25 MB x 5 backups = 125 MB max)
- Configurable via env vars: CAL_LOG_DIR, CAL_LOG_MAX_MB, CAL_LOG_BACKUP_COUNT
- --verbose / -v flag for DEBUG-level output
- overnight-run.py: separate rotating log at logs/overnight-run.log (10 MB x 5)
- Sub-module loggers in ml/, scanner/, deck_sources/, src/ all propagate
to the root commander_ai_lab logger
- Stripped redundant [TAG] prefixes from messages (logger name carries context)
- Format: '2026-03-13 14:00:00 INFO [commander_ai_lab.batch] message'
Files changed: lab_api.py, overnight-run.py, and 11 sub-module .py files
176 total print() calls converted (101 in lab_api.py, 76 in sub-modules)
