File size: 22,773 Bytes
79cb31a
f22507c
 
 
 
 
 
 
 
79cb31a
a9175aa
2879f65
f22507c
 
 
2879f65
 
a9175aa
 
 
79cb31a
 
 
 
a9175aa
2879f65
79cb31a
 
 
 
a9175aa
 
79cb31a
a9175aa
2879f65
f22507c
 
2879f65
 
79cb31a
a9175aa
 
 
79cb31a
 
 
 
 
 
 
a9175aa
 
f22507c
2879f65
f22507c
 
 
 
 
79cb31a
f22507c
 
 
a9175aa
79cb31a
 
 
 
 
 
a9175aa
f22507c
a9175aa
 
 
f22507c
 
 
 
79cb31a
f22507c
79cb31a
f22507c
a9175aa
 
f22507c
a9175aa
f22507c
 
 
 
 
 
 
 
 
 
 
 
 
a9175aa
 
f22507c
 
 
 
 
 
 
 
 
a9175aa
 
f22507c
a9175aa
 
 
f22507c
 
 
 
a9175aa
2879f65
a9175aa
 
 
f22507c
 
 
 
 
 
 
 
79cb31a
a9175aa
 
 
 
 
 
79cb31a
a9175aa
f22507c
 
a9175aa
 
 
f22507c
 
 
 
 
 
a9175aa
 
 
 
 
 
 
 
 
 
 
 
 
 
f22507c
a9175aa
 
 
 
79cb31a
f22507c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
79cb31a
 
 
 
f22507c
 
 
79cb31a
 
f22507c
79cb31a
f22507c
79cb31a
f22507c
79cb31a
f22507c
79cb31a
f22507c
79cb31a
 
 
 
 
a9175aa
 
 
79cb31a
 
 
 
f22507c
a9175aa
f22507c
a9175aa
 
79cb31a
 
 
 
a9175aa
79cb31a
 
 
 
 
a9175aa
79cb31a
 
f22507c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a9175aa
79cb31a
 
 
f22507c
79cb31a
 
 
 
 
f22507c
79cb31a
 
 
 
f22507c
79cb31a
 
 
 
 
 
 
 
 
 
2879f65
 
f22507c
 
 
 
 
 
 
 
2879f65
a9175aa
 
2879f65
 
 
f22507c
 
79cb31a
a9175aa
2879f65
 
 
 
 
a9175aa
2879f65
 
f22507c
 
2879f65
f22507c
a9175aa
2879f65
a9175aa
79cb31a
2879f65
f22507c
 
 
 
 
 
 
 
 
 
 
a9175aa
 
f22507c
 
 
 
 
 
 
 
 
a9175aa
f22507c
2879f65
f22507c
79cb31a
 
 
 
 
 
a9175aa
79cb31a
 
 
2879f65
a9175aa
2879f65
 
f22507c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a9175aa
f22507c
 
a9175aa
 
f22507c
a9175aa
f22507c
2879f65
 
f22507c
2879f65
 
f22507c
2879f65
79cb31a
 
 
 
2879f65
 
f22507c
2879f65
f22507c
 
 
2879f65
 
 
f22507c
 
 
 
 
 
79cb31a
2879f65
 
f22507c
 
 
 
 
a9175aa
 
2879f65
f22507c
 
 
 
79cb31a
f22507c
a9175aa
 
79cb31a
a9175aa
 
 
f22507c
79cb31a
a9175aa
 
 
f22507c
 
 
 
 
2879f65
a9175aa
f22507c
 
 
 
 
 
 
 
 
 
 
2879f65
 
f22507c
2879f65
 
f22507c
 
 
 
 
 
 
 
 
 
2879f65
f22507c
 
 
2879f65
 
 
f22507c
 
 
 
 
 
 
79cb31a
 
 
 
 
f22507c
79cb31a
2879f65
 
a9175aa
 
 
 
 
2879f65
a9175aa
2879f65
 
a9175aa
 
 
 
 
 
 
 
f22507c
a9175aa
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
"""
server/environment.py β€” Session-aware environment (v0.5).

New in v0.5:
  1. Rollback action β€” undo last apply (real data engineers do this)
  2. Episode reasoning trace β€” running history of what the agent tried + effects
  3. Feature importance β€” returned after every apply so agent sees what the model learned
  4. Regression explanation β€” when accuracy drops, explains the likely cause
  5. Baseline comparison β€” agent always knows how far ahead of majority-class predictor it is
"""
import threading
import pandas as pd
import numpy as np

from server.dataset_registry import DatasetRegistry
from server.evaluator import Evaluator
from server.reward import compute, compute_stats
from server.anti_exploit import AntiExploit, ExploitDetected
from server.config import cfg
from server.logger import get_logger, log_event
from server.specialist_agents import (
    CleanerAgent, AugmenterAgent, BalancerAgent, ValidatorAgent, AnalystAgent
)

logger = get_logger("environment")

QUERY_COSTS = {
    "query_cleaner":   1,
    "query_augmenter": 1,
    "query_balancer":  1,
    "query_validator": 2,
    "query_analyst":   2,
}
QUERY_ACTIONS = set(QUERY_COSTS.keys())

_registry = DatasetRegistry()


class DataCentricEnvironment:

    def __init__(self, session_id: str, episode_count: int = 0):
        self.session_id = session_id
        self._episode_count = episode_count
        self.agents = {
            "cleaner":   CleanerAgent(),
            "augmenter": AugmenterAgent(),
            "balancer":  BalancerAgent(),
            "validator": ValidatorAgent(),
            "analyst":   AnalystAgent(),
        }
        self.anti_exploit = AntiExploit()
        self._lock = threading.Lock()
        self._reset_state()

    def _reset_state(self):
        self.train_df: pd.DataFrame = None
        self.holdout_df: pd.DataFrame = None
        self.domain_metadata: dict = {}
        self.evaluator: Evaluator = None
        self.target_accuracy: float = None
        self.initial_row_count: int = 0
        self.baseline_accuracy: float = 0.0  # majority-class predictor on holdout
        self.starting_accuracy: float = 0.0  # accuracy before ANY agent action
        self.budget: int = cfg.MAX_BUDGET
        self.current_accuracy: float = 0.0
        self.episode_step: int = 0
        self.done: bool = False
        self.difficulty: str = "easy"
        self.pending_recs: dict = {}
        self.applied_rec_ids: set = set()
        self.last_query_result: dict = {}
        self.last_feature_importance: dict = {}
        self.anti_exploit.reset()
        self.accuracy_history: list = []
        self.reward_history: list = []
        # Rollback: stack of (df_snapshot, accuracy) β€” last 3 states
        self._state_stack: list[tuple] = []
        # Reasoning trace: running log of every step
        self._episode_trace: list[dict] = []

    # ── Public API ─────────────────────────────────────────────────────────────

    def reset(self, difficulty: str = None, seed: int = None) -> dict:
        with self._lock:
            self._episode_count += 1
            self._reset_state()
            self.difficulty = difficulty or self._curriculum_difficulty()

            self.train_df, self.holdout_df, self.domain_metadata = _registry.get(
                difficulty=self.difficulty, seed=seed
            )
            self.initial_row_count = len(self.train_df)
            self.evaluator = Evaluator(self.holdout_df)

            pub_baseline = self.domain_metadata.get("published_baseline", 0.80)
            self.target_accuracy = round(pub_baseline * 0.97, 4)
            self.baseline_accuracy = self.evaluator.baseline_accuracy()

            self.current_accuracy = self.evaluator.evaluate(self._clean_df(self.train_df))
            self.starting_accuracy = self.current_accuracy
            self.accuracy_history.append(self.current_accuracy)

            self._episode_trace.append({
                "step": 0,
                "type": "reset",
                "dataset": self.domain_metadata.get("display_name"),
                "accuracy": round(self.current_accuracy, 4),
                "baseline_accuracy": self.baseline_accuracy,
                "target_accuracy": self.target_accuracy,
            })

            log_event(logger, "episode_reset",
                      session_id=self.session_id,
                      dataset=self.domain_metadata.get("display_name"),
                      difficulty=self.difficulty,
                      initial_accuracy=round(self.current_accuracy, 4),
                      target_accuracy=self.target_accuracy,
                      baseline_accuracy=self.baseline_accuracy,
                      published_baseline=pub_baseline,
                      n_train=len(self.train_df),
                      n_holdout=len(self.holdout_df))
            return self._observation()

    def step(self, action: dict) -> dict:
        with self._lock:
            if self.done:
                return self._error("Episode done. Call /reset.")
            if self.train_df is None:
                return self._error("Not initialized. Call /reset first.")

            # Rollback action β€” no anti-exploit check needed
            action_type = action.get("action", "")
            if action_type == "rollback":
                return self._handle_rollback()

            try:
                self.anti_exploit.check(
                    action=action,
                    budget_remaining=self.budget,
                    pending_recs=self.pending_recs,
                    applied_rec_ids=self.applied_rec_ids,
                )
            except ExploitDetected as e:
                log_event(logger, "exploit_detected", session_id=self.session_id,
                          rule=e.rule, detail=e.detail)
                self.episode_step += 1
                self.budget = max(0, self.budget - 1)
                self.done = self.budget <= 0
                self._episode_trace.append({
                    "step": self.episode_step,
                    "type": "exploit_blocked",
                    "rule": e.rule,
                    "detail": e.detail,
                })
                return {
                    "observation": self._observation(),
                    "reward": 0.001,
                    "done": self.done,
                    "exploit_detected": True,
                    "error": f"[{e.rule}] {e.detail}",
                    "info": {"episode_step": self.episode_step, "budget_remaining": self.budget},
                }

            if action_type in QUERY_ACTIONS:
                return self._handle_query(action_type, action)
            elif action_type == "apply":
                return self._handle_apply(action)
            else:
                return self._error(f"Unknown action '{action_type}'. Valid: {list(QUERY_ACTIONS) + ['apply', 'rollback']}")

    def state(self) -> dict:
        with self._lock:
            return self._observation()

    # ── Rollback ───────────────────────────────────────────────────────────────

    def _handle_rollback(self) -> dict:
        """Undo the last apply operation. Costs 1 budget. Max 3 rollbacks per episode."""
        rollbacks_used = sum(1 for e in self._episode_trace if e["type"] == "rollback")
        if rollbacks_used >= 3:
            return self._error("Maximum 3 rollbacks per episode reached.")
        if not self._state_stack:
            return self._error("Nothing to roll back. No apply operations have been made yet.")

        prev_df, prev_accuracy = self._state_stack.pop()
        self.train_df = prev_df
        self.current_accuracy = prev_accuracy
        self.accuracy_history.append(self.current_accuracy)
        self.budget = max(0, self.budget - 1)
        self.episode_step += 1
        self.done = self.budget <= 0

        self._episode_trace.append({
            "step": self.episode_step,
            "type": "rollback",
            "accuracy_after_rollback": round(self.current_accuracy, 4),
            "note": "Last apply undone. Dataset restored to previous state.",
        })

        log_event(logger, "rollback", session_id=self.session_id,
                  accuracy_after=round(self.current_accuracy, 4))

        return {
            "observation": self._observation(),
            "reward": 0.3,  # small penalty for indecision, but not zero
            "done": self.done,
            "rollback": True,
            "accuracy_after_rollback": round(self.current_accuracy, 4),
            "info": {
                "episode_step": self.episode_step,
                "budget_remaining": self.budget,
                "rollbacks_remaining": 3 - rollbacks_used - 1,
                "note": "Dataset restored to state before last apply.",
            },
        }

    # ── Query handler ──────────────────────────────────────────────────────────

    def _handle_query(self, action_type: str, action: dict) -> dict:
        cost = QUERY_COSTS[action_type]
        prev_stats = compute_stats(self.train_df)
        clean = self._clean_df(self.train_df)
        meta = self.domain_metadata

        if action_type == "query_cleaner":
            result = self.agents["cleaner"].query(clean, meta)
        elif action_type == "query_augmenter":
            result = self.agents["augmenter"].query(clean, action.get("target_class"), meta)
        elif action_type == "query_balancer":
            result = self.agents["balancer"].query(clean, meta)
        elif action_type == "query_validator":
            result = self.agents["validator"].query(clean, meta)
        elif action_type == "query_analyst":
            result = self.agents["analyst"].query(clean, meta)
        else:
            result = {}

        new_rec_ids = []
        for rec in result.get("recommendations", []):
            rid = rec["id"]
            self.pending_recs[rid] = {"rec": rec, "agent": result.get("agent", "unknown")}
            new_rec_ids.append(rid)

        self.last_query_result = result
        self.budget = max(0, self.budget - cost)
        self.episode_step += 1
        new_stats = compute_stats(self.train_df)

        reward, decomp = compute(
            prev_accuracy=self.current_accuracy,
            new_accuracy=self.current_accuracy,
            prev_stats=prev_stats,
            new_stats=new_stats,
            action=action,
            steps_taken=self.episode_step,
            max_steps=cfg.MAX_BUDGET,
            budget_remaining=self.budget,
            target_accuracy=self.target_accuracy,
            step_type="query",
            n_recs_returned=len(new_rec_ids),
        )
        self.reward_history.append(reward)
        self.done = self.budget <= 0

        agent_name = action_type.replace("query_", "")
        self._episode_trace.append({
            "step": self.episode_step,
            "type": "query",
            "agent": agent_name,
            "n_recs": len(new_rec_ids),
            "budget_cost": cost,
            "budget_remaining": self.budget,
            "reward": reward,
            "rec_ids": new_rec_ids,
        })

        log_event(logger, "query_step", session_id=self.session_id,
                  action=action_type, n_recs=len(new_rec_ids),
                  budget=self.budget, reward=reward)

        return {
            "observation": self._observation(),
            "reward": reward,
            "reward_decomposition": decomp,
            "done": self.done,
            "query_result": result,
            "new_recommendation_ids": new_rec_ids,
            "info": {
                "action_type": "query",
                "agent_queried": agent_name,
                "budget_cost": cost,
                "budget_remaining": self.budget,
                "n_recommendations": len(new_rec_ids),
                "episode_step": self.episode_step,
                "domain": self.domain_metadata.get("display_name"),
            },
        }

    # ── Apply handler ──────────────────────────────────────────────────────────

    def _handle_apply(self, action: dict) -> dict:
        rec_id = action.get("rec_id", "")
        entry = self.pending_recs[rec_id]
        agent_name = entry["agent"]
        rec = entry["rec"]

        prev_accuracy = self.current_accuracy
        prev_stats = compute_stats(self.train_df)
        prev_rows = len(self.train_df)
        meta = self.domain_metadata

        # Save state for rollback BEFORE applying
        self._state_stack.append((self.train_df.copy(), prev_accuracy))
        if len(self._state_stack) > 3:
            self._state_stack.pop(0)  # keep at most last 3

        result_holder: dict = {}
        error_holder: dict = {}

        def _run():
            try:
                clean = self._clean_df(self.train_df)
                df_out, log_msg = self.agents[agent_name].apply(clean, rec, meta)
                result_holder["df"] = df_out
                result_holder["log"] = log_msg
            except Exception as e:
                error_holder["error"] = str(e)

        t = threading.Thread(target=_run, daemon=True)
        t.start()
        t.join(timeout=cfg.STEP_TIMEOUT_SECONDS)

        if t.is_alive():
            self._state_stack.pop()  # failed β€” don't keep stale snapshot
            return self._error("Apply timed out.")
        if "error" in error_holder:
            self._state_stack.pop()
            return self._error(f"Apply error: {error_holder['error']}")

        new_df = result_holder["df"]
        tool_log = result_holder["log"]

        # Data integrity constraint: cannot delete more than 10% of rows
        new_rows = len(new_df)
        deletion_pct = max(0, (prev_rows - new_rows) / max(prev_rows, 1))
        if deletion_pct > 0.10:
            self._state_stack.pop()
            return self._error(
                f"Data integrity violation: would delete {deletion_pct:.1%} of training rows "
                f"(limit: 10%). Use targeted imputation instead of drop_rows."
            )

        self.train_df = new_df
        self.applied_rec_ids.add(rec_id)
        self.episode_step += 1

        # Full evaluation with feature importance + regression explanation
        eval_result = self.evaluator.evaluate_with_details(
            self._clean_df(self.train_df), prev_accuracy
        )
        self.current_accuracy = eval_result["accuracy"]
        self.last_feature_importance = eval_result.get("feature_importance", {})
        regression_explanation = eval_result.get("regression_explanation")

        self.accuracy_history.append(self.current_accuracy)
        new_stats = compute_stats(self.train_df)

        reward, decomp = compute(
            prev_accuracy=prev_accuracy,
            new_accuracy=self.current_accuracy,
            prev_stats=prev_stats,
            new_stats=new_stats,
            action=action,
            steps_taken=self.episode_step,
            max_steps=cfg.MAX_BUDGET,
            budget_remaining=self.budget,
            target_accuracy=self.target_accuracy,
            step_type="apply",
        )
        self.reward_history.append(reward)
        self.done = (self.current_accuracy >= self.target_accuracy) or (self.budget <= 0)

        acc_delta = round(self.current_accuracy - prev_accuracy, 4)
        self._episode_trace.append({
            "step": self.episode_step,
            "type": "apply",
            "agent": agent_name,
            "rec_type": rec.get("type", "?"),
            "rec_id": rec_id,
            "accuracy_before": round(prev_accuracy, 4),
            "accuracy_after": round(self.current_accuracy, 4),
            "accuracy_delta": acc_delta,
            "effect": "improved" if acc_delta > 0.001 else ("hurt" if acc_delta < -0.001 else "neutral"),
            "reward": reward,
            "rows_before": prev_rows,
            "rows_after": new_rows,
        })

        log_event(logger, "apply_step", session_id=self.session_id,
                  rec_id=rec_id, agent=agent_name,
                  prev_acc=round(prev_accuracy, 4),
                  new_acc=round(self.current_accuracy, 4),
                  target=self.target_accuracy,
                  reward=reward,
                  success=self.current_accuracy >= self.target_accuracy)

        response = {
            "observation": self._observation(),
            "reward": reward,
            "reward_decomposition": decomp,
            "done": self.done,
            "tool_log": tool_log,
            "feature_importance": self.last_feature_importance,
            "info": {
                "action_type": "apply",
                "rec_id": rec_id,
                "agent": agent_name,
                "rec_type": rec.get("type", "?"),
                "prev_accuracy": round(prev_accuracy, 4),
                "new_accuracy": round(self.current_accuracy, 4),
                "accuracy_delta": acc_delta,
                "target_accuracy": self.target_accuracy,
                "published_baseline": self.domain_metadata.get("published_baseline"),
                "improvement_over_start": round(self.current_accuracy - self.starting_accuracy, 4),
                "improvement_over_majority_baseline": round(self.current_accuracy - self.baseline_accuracy, 4),
                "budget_remaining": self.budget,
                "episode_step": self.episode_step,
                "success": self.current_accuracy >= self.target_accuracy,
                "rollbacks_available": max(0, 3 - sum(1 for e in self._episode_trace if e["type"] == "rollback")),
                "data_integrity": {
                    "rows_before": prev_rows,
                    "rows_after": new_rows,
                    "deletion_pct": round(deletion_pct, 4),
                },
            },
        }

        # Only include regression explanation when accuracy dropped
        if regression_explanation:
            response["regression_explanation"] = regression_explanation

        return response

    # ── Helpers ────────────────────────────────────────────────────────────────

    def _clean_df(self, df):
        drop_cols = [c for c in df.columns if c.startswith("_")]
        return df.drop(columns=drop_cols) if drop_cols else df

    def _observation(self) -> dict:
        stats = compute_stats(self.train_df) if self.train_df is not None else {}
        pending_summary = {
            rid: {
                "agent": entry["agent"],
                "type": entry["rec"].get("type", "?"),
                "priority": entry["rec"].get("priority", "?"),
                "reason": entry["rec"].get("reason", ""),
                "domain_informed": entry["rec"].get("domain_informed", False),
            }
            for rid, entry in self.pending_recs.items()
            if rid not in self.applied_rec_ids
        }
        meta = self.domain_metadata

        # Compact trace β€” last 5 steps for context without overwhelming the prompt
        recent_trace = self._episode_trace[-5:] if self._episode_trace else []

        return {
            "session_id": self.session_id,

            # What the agent is working on
            "dataset": {
                "name": meta.get("display_name", "Unknown"),
                "domain": meta.get("domain", "generic"),
                "description": meta.get("description", ""),
                "known_issues": meta.get("known_issues", []),
                "published_baseline": meta.get("published_baseline"),
            },

            # Current state
            "current_accuracy": round(self.current_accuracy, 4),
            "target_accuracy": self.target_accuracy,
            "accuracy_gap": round(max(0, self.target_accuracy - self.current_accuracy), 4),
            "budget_remaining": self.budget,
            "difficulty": self.difficulty,

            # Comparisons β€” what does this number actually mean?
            "benchmarks": {
                "majority_class_baseline": self.baseline_accuracy,
                "starting_accuracy": round(self.starting_accuracy, 4),
                "improvement_over_start": round(self.current_accuracy - self.starting_accuracy, 4),
                "improvement_over_baseline": round(self.current_accuracy - self.baseline_accuracy, 4),
                "published_baseline": meta.get("published_baseline"),
            },

            "dataset_stats": {
                "n_train_rows": len(self.train_df) if self.train_df is not None else 0,
                "n_holdout_rows": len(self.holdout_df) if self.holdout_df is not None else 0,
                "n_cols": len(self.train_df.columns) if self.train_df is not None else 0,
                "missing_pct": round(stats.get("missing_pct", 0), 4),
                "balance_ratio": round(stats.get("balance_ratio", 0), 4),
            },

            # Feature importance from last evaluation
            "feature_importance": self.last_feature_importance,

            # Episodic memory β€” what has the agent tried so far?
            "episode_trace": recent_trace,

            "pending_recommendations": pending_summary,
            "last_query_result": self.last_query_result,
            "available_actions": (
                "query_cleaner | query_augmenter | query_balancer | "
                "query_validator (cost 2) | query_analyst (cost 2) | "
                "apply {rec_id} | rollback (undo last apply, max 3/episode)"
            ),
        }

    def _error(self, msg: str) -> dict:
        return {"error": msg, "session_id": self.session_id}

    def _curriculum_difficulty(self) -> str:
        if self._episode_count < cfg.CURRICULUM_MEDIUM_AFTER:
            return "easy"
        elif self._episode_count < cfg.CURRICULUM_HARD_AFTER:
            return "medium"
        return "hard"

    def episode_summary(self) -> dict:
        return {
            "session_id": self.session_id,
            "episode_count": self._episode_count,
            "accuracy_history": [round(a, 4) for a in self.accuracy_history],
            "reward_history": [round(r, 4) for r in self.reward_history],
            "mean_reward": round(sum(self.reward_history) / max(len(self.reward_history), 1), 4),
            "full_trace": self._episode_trace,
        }