File size: 23,029 Bytes
489366d
bb69d16
 
61bbf36
5931364
 
 
 
 
 
 
 
 
489366d
 
a2a2b07
 
5931364
 
 
 
 
 
 
52a9933
489366d
a2a2b07
5931364
bb69d16
489366d
5931364
52a9933
a2a2b07
 
5931364
bb69d16
489366d
5931364
 
 
 
52a9933
bb69d16
 
 
489366d
bb69d16
 
 
61bbf36
5fd4180
52a9933
1e7d39f
5931364
52a9933
bb69d16
 
 
489366d
bb69d16
 
a2a2b07
489366d
bb69d16
 
 
 
52a9933
bb69d16
 
52a9933
bb69d16
5931364
 
 
 
52a9933
 
5931364
 
bb69d16
 
52a9933
 
 
 
5931364
 
 
bb69d16
 
 
 
5931364
bb69d16
 
52a9933
 
 
bb69d16
 
52a9933
 
 
 
 
 
 
 
 
1e7d39f
52a9933
 
 
 
 
 
1e7d39f
52a9933
1e7d39f
bb69d16
 
 
 
52a9933
 
 
 
 
 
 
 
 
1e7d39f
52a9933
1e7d39f
bb69d16
 
 
5931364
 
bb69d16
 
5931364
61bbf36
5931364
 
 
 
 
 
 
 
 
 
 
 
 
 
 
61bbf36
5931364
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb69d16
 
 
 
 
 
 
 
1e7d39f
bb69d16
 
52a9933
bb69d16
 
 
 
5931364
 
a2a2b07
 
bb69d16
 
 
a2a2b07
bb69d16
5931364
 
 
 
 
 
 
 
 
 
 
bb69d16
5931364
5fd4180
5931364
 
 
 
52a9933
bb69d16
 
5fd4180
bb69d16
5931364
 
1e7d39f
5931364
 
 
 
 
 
 
 
52a9933
 
bb69d16
5931364
5fd4180
bb69d16
5931364
 
5fd4180
52a9933
bb69d16
 
 
 
a2a2b07
5931364
 
 
 
 
 
 
 
 
 
a2a2b07
bb69d16
 
5931364
bb69d16
 
 
52a9933
bb69d16
5fd4180
5931364
 
 
 
 
 
 
 
 
 
bb69d16
 
 
 
5fd4180
5931364
 
 
 
 
 
 
 
 
a2a2b07
5931364
 
a2a2b07
52a9933
5931364
 
 
 
 
52a9933
5fd4180
489366d
bb69d16
 
 
5fd4180
a2a2b07
52a9933
 
 
 
 
 
 
 
 
 
5931364
52a9933
 
 
 
5931364
489366d
bb69d16
 
5931364
bb69d16
5931364
 
bb69d16
 
52a9933
 
bb69d16
 
1e7d39f
bb69d16
5931364
 
 
 
bb69d16
52a9933
5931364
52a9933
bb69d16
5931364
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb69d16
5931364
 
 
 
 
bb69d16
1e7d39f
52a9933
 
 
5931364
 
 
 
 
 
 
 
 
 
 
 
52a9933
5931364
 
 
 
 
 
 
 
 
 
 
 
52a9933
5931364
52a9933
 
 
5931364
52a9933
 
bb69d16
489366d
 
61bbf36
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
"""
inference.py – PhishGuard-Env Baseline Inference Script
========================================================

Structured stdout logs (required by OpenEnv validator):
  [START] task=<level>
  [STEP]  task=<level> step=N reward=R is_correct=true|false
  [END]   task=<level> score=S steps=N

The episode score in [END] comes directly from info["score"] returned by
/step when done=True β€” which is GRADERS[level](metrics) from grader.py.
This guarantees the validator sees the same grader-based score that env.py
computes internally.
"""

from __future__ import annotations

import sys
import io
# Force UTF-8 output so emoji in env.py feedback strings don't crash on Windows cp1252
if sys.stdout.encoding and sys.stdout.encoding.lower() != "utf-8":
    sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8", errors="replace")
    sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding="utf-8", errors="replace")

import argparse
import json
import os

import textwrap
import time
from datetime import datetime, timezone
from typing import Any, Dict, List, Optional

import requests
from dotenv import load_dotenv
from openai import OpenAI

# Load .env file first so HF_TOKEN / OPENAI_API_KEY are available via os.getenv
load_dotenv()

from grader import PASS_THRESHOLD, GRADERS, grade_performance

# ─────────────────────────────────────────────────────────────────────────────
# Configuration
# ─────────────────────────────────────────────────────────────────────────────

API_KEY      = os.getenv("HF_TOKEN") or os.getenv("OPENAI_API_KEY")
API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
MODEL_NAME   = os.getenv("MODEL_NAME",   "Qwen/Qwen2.5-72B-Instruct")
ENV_BASE_URL = os.getenv("ENV_BASE_URL", "http://localhost:7860").rstrip("/")

MAX_STEPS_PER_LEVEL = 15
HTTP_MAX_RETRIES    = 3
HTTP_BACKOFF_BASE   = 1.5

# ─────────────────────────────────────────────────────────────────────────────
# LLM client
# ─────────────────────────────────────────────────────────────────────────────

if not API_KEY:
    print("[ERROR] No API key found. Set HF_TOKEN or OPENAI_API_KEY.", flush=True)
    sys.exit(1)

client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)

SYSTEM_PROMPT = textwrap.dedent("""
    You are a SOC (Security Operations Centre) Analyst triaging incoming emails.

    Analyse the email data provided and respond ONLY with valid JSON in this exact format:
    {"action": "<ACTION>", "reasoning": "<one sentence technical justification>"}

    Valid actions:
    - MARK_SAFE       : Deliver to inbox (confirmed legitimate email)
    - MOVE_TO_SPAM    : Bulk/unsolicited mail with no active threat
    - QUARANTINE      : Hold for analyst review (suspicious but unconfirmed)
    - BLOCK_DOMAIN    : Block sender domain at perimeter (confirmed phishing/malware source)

    Signal interpretation:
    - SPF fail + DMARC fail + urgency + links β†’ likely PHISH or MALWARE β†’ BLOCK_DOMAIN or QUARANTINE
    - Known sender, SPF pass, DMARC pass, no suspicious links β†’ likely SAFE β†’ MARK_SAFE
    - Bulk unsolicited with no malicious payload β†’ SPAM β†’ MOVE_TO_SPAM
    - Wire transfer / CEO fraud / financial urgency from unknown domain β†’ BEC β†’ QUARANTINE
    - Malware attachment confirmed by AV β†’ QUARANTINE (isolate, do not deliver)
    - Confirmed phishing domain β†’ BLOCK_DOMAIN (sever attack vector)

    confidence_hint field:
    - This is a contextual signal from your SIEM, mail gateway, or threat-intel feed.
    - It is intentionally noisy β€” treat it as one data-point, not ground truth.
    - If it directly contradicts other signals (SPF, DMARC, links), weigh all evidence.
""").strip()


# ─────────────────────────────────────────────────────────────────────────────
# HTTP helpers
# ─────────────────────────────────────────────────────────────────────────────

_session = requests.Session()


def _post(endpoint: str, payload: dict) -> dict:
    url = f"{ENV_BASE_URL}{endpoint}"
    last_exc: Optional[Exception] = None
    for attempt in range(HTTP_MAX_RETRIES):
        try:
            resp = _session.post(url, json=payload, timeout=30)
            resp.raise_for_status()
            return resp.json()
        except (requests.ConnectionError, requests.Timeout) as exc:
            last_exc = exc
            wait = HTTP_BACKOFF_BASE ** attempt
            print(f"  [WARN] POST {endpoint} failed (attempt {attempt+1}): {exc} β€” retrying in {wait:.1f}s", flush=True)
            time.sleep(wait)
        except requests.HTTPError as exc:
            if exc.response is not None and exc.response.status_code < 500:
                raise
            last_exc = exc
            wait = HTTP_BACKOFF_BASE ** attempt
            print(f"  [WARN] POST {endpoint} server error (attempt {attempt+1}): {exc} β€” retrying in {wait:.1f}s", flush=True)
            time.sleep(wait)
    raise RuntimeError(f"POST {endpoint} failed after {HTTP_MAX_RETRIES} attempts: {last_exc}")


def _get(endpoint: str) -> dict:
    url = f"{ENV_BASE_URL}{endpoint}"
    last_exc: Optional[Exception] = None
    for attempt in range(HTTP_MAX_RETRIES):
        try:
            resp = _session.get(url, timeout=10)
            resp.raise_for_status()
            return resp.json()
        except (requests.ConnectionError, requests.Timeout, requests.HTTPError) as exc:
            last_exc = exc
            wait = HTTP_BACKOFF_BASE ** attempt
            print(f"  [WARN] GET {endpoint} failed (attempt {attempt+1}): {exc} β€” retrying in {wait:.1f}s", flush=True)
            time.sleep(wait)
    raise RuntimeError(f"GET {endpoint} failed after {HTTP_MAX_RETRIES} attempts: {last_exc}")


# ─────────────────────────────────────────────────────────────────────────────
# Rule-based fallback triage  (used when LLM is unavailable / errors out)
# Covers all 10 PhishGuard scenarios deterministically.
# ─────────────────────────────────────────────────────────────────────────────

def _rule_based_triage(obs: Dict[str, Any]) -> tuple[str, str]:
    """
    Deterministic SOC triage using email header signals and confidence hints.
    Returns (action, reasoning) β€” same signature as the LLM path.

    Decision priority
    -----------------
    1. Threat-intel IOC hit on domain/link   β†’ BLOCK_DOMAIN
    2. QR-code / credential-harvesting hint  β†’ QUARANTINE
    3. AV-flagged or macro attachment        β†’ QUARANTINE
    4. Suspicious attachment + auth failure  β†’ QUARANTINE
    5. Spam content keywords                 β†’ MOVE_TO_SPAM
    6. BEC / financial urgency keywords      β†’ QUARANTINE
    7. URL redirect chain + auth failure     β†’ BLOCK_DOMAIN
    8. Auth failure + links                  β†’ BLOCK_DOMAIN
    9. Auth-OK, no threats                   β†’ MARK_SAFE
    10. Default (uncertain)                  β†’ QUARANTINE
    """
    spf  = obs.get("spf_record",  "").lower()
    dmarc = obs.get("dmarc_record", "").lower()
    urgency = obs.get("urgency_level", "").lower()
    links = obs.get("links", [])
    has_attach = obs.get("has_attachments", False)
    subject = obs.get("subject", "").lower()
    body    = obs.get("body",    "").lower()
    hint    = obs.get("confidence_hint", "").lower()

    auth_ok   = (spf == "pass" and dmarc == "pass")
    auth_fail = spf in ("fail", "softfail") or dmarc in ("fail", "none")

    # 1. Threat-intel IOC hit β†’ block the domain
    if "ioc feed" in hint or "ioc" in hint:
        if links:
            return "BLOCK_DOMAIN", "Domain appears on threat-intel IOC feed β€” block at perimeter"
        return "QUARANTINE", "IOC hit with no links β€” quarantine for analyst review"

    # 2. QR-code / credential harvesting phishing
    if "credential-harvest" in hint or "credential harvest" in hint:
        return "QUARANTINE", "QR-code credential-harvesting page detected β€” quarantine attachment"

    # 3. AV-flagged attachment (PE binary, macros, unsigned)
    if has_attach and any(kw in hint for kw in ("av:", "macro", "pe binary", "unsigned")):
        return "QUARANTINE", "AV/macro-flagged attachment β€” isolate from delivery"

    # 4. Attachment with authentication failure
    if has_attach and auth_fail:
        return "QUARANTINE", "Suspicious attachment combined with SPF/DMARC failure"

    # 5. Spam: prize / lottery / mass-marketing content
    spam_kw = ("prize", "congratulations", "claim", "won", "lottery", "$1m", "million")
    if any(kw in subject + " " + body for kw in spam_kw) and urgency != "critical":
        return "MOVE_TO_SPAM", "Bulk prize/lottery spam β€” no active threat payload"

    # 6. BEC: financial urgency keywords in body
    bec_kw = ("wire", "transfer", "account below", "fund", "bank details")
    if any(kw in body for kw in bec_kw) and urgency in ("critical", "high"):
        return "QUARANTINE", "BEC wire-transfer / financial-fraud pattern detected"

    # 7. URL redirect chain with auth failure β†’ confirmed phishing source
    if ("redirect" in hint or "url shortener" in hint) and auth_fail:
        return "BLOCK_DOMAIN", "Multi-hop URL redirect chain with auth failure β€” block domain"

    # 8. Auth failure + suspicious links β†’ block
    if auth_fail and links:
        return "BLOCK_DOMAIN", "Domain authentication failure with outbound links β€” block"

    # 9. Clean authentication, no threat signals β†’ safe
    if auth_ok and not has_attach:
        safe_negative = ("ioc" not in hint and "malware" not in hint
                         and "phish" not in hint and "credential" not in hint)
        if safe_negative:
            return "MARK_SAFE", "SPF/DMARC pass, no threat indicators β€” deliver to inbox"

    # 10. Default: hold for analyst review
    return "QUARANTINE", "Uncertain signals β€” quarantine as precaution"


# ─────────────────────────────────────────────────────────────────────────────
# LLM action selection  (rule-based fallback when LLM errors)
# ─────────────────────────────────────────────────────────────────────────────

def _choose_action(observation: Dict[str, Any]) -> tuple[str, str]:
    try:
        completion = client.chat.completions.create(
            model=MODEL_NAME,
            messages=[
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user",   "content": json.dumps(observation, indent=2)},
            ],
            response_format={"type": "json_object"},
            temperature=0,
            max_tokens=256,
        )
        parsed    = json.loads(completion.choices[0].message.content)
        action    = parsed.get("action", "QUARANTINE").strip().upper()
        reasoning = parsed.get("reasoning", "")
        return action, reasoning
    except Exception as exc:
        print(f"  [WARN] LLM unavailable ({type(exc).__name__}) β€” using rule-based fallback", flush=True)
        return _rule_based_triage(observation)


# ─────────────────────────────────────────────────────────────────────────────
# Run one level
# ─────────────────────────────────────────────────────────────────────────────

def run_level(level: str) -> Dict[str, Any]:
    """
    Run a complete episode for the given difficulty level.

    The episode score is taken from info["score"] on the terminal step
    (done=True) β€” this is GRADERS[level](metrics) computed by env.py,
    the same value the OpenEnv validator uses.

    Falls back to /state's overall_score only if no terminal step score
    was captured (e.g. MAX_STEPS_PER_LEVEL reached without done=True).
    """
    print(f"\n{'='*60}", flush=True)
    print(f"  LEVEL: {level.upper()}", flush=True)
    print(f"{'='*60}", flush=True)

    # ── Emit [START] ─────────────────────────────────────────────────────────
    print(f"[START] task={level}", flush=True)

    reset_resp  = _post("/reset", {"difficulty": level})
    obs         = reset_resp["observation"]
    total_tasks = reset_resp["total_tasks"]
    print(f"  Tasks in this level: {total_tasks}", flush=True)

    steps: List[dict] = []
    step_num            = 0
    done                = False
    step_resp: Dict[str, Any] = {}
    # episode_score is populated from info["score"] when done=True.
    # It comes from GRADERS[level](metrics) inside env.py.
    episode_score: Optional[float] = None
    # episode_metrics is populated from info["metrics"] when done=True.
    episode_metrics: Optional[dict] = None

    # Track current scenario ID: starts from reset, then updated after each step.
    current_scenario_id = reset_resp.get("task_id", "?")

    while not done and step_num < MAX_STEPS_PER_LEVEL:
        step_num += 1
        print(f"\n  Step {step_num} | scenario={current_scenario_id}", flush=True)

        action, reasoning = _choose_action(obs)
        print(f"    -> Action   : {action}", flush=True)
        print(f"    -> Reasoning: {reasoning[:80]}", flush=True)

        step_resp  = _post("/step", {"action": action, "reasoning": reasoning})
        reward     = step_resp["reward"]
        done       = step_resp["done"]
        is_correct = step_resp["is_correct"]
        info       = step_resp.get("info", {})

        # task_id in response is the scenario JUST processed β€” update for display
        graded_scenario_id = step_resp.get("task_id") or current_scenario_id

        print(f"    <- Graded   : scenario={graded_scenario_id}  correct={is_correct}  reward={reward:.4f}  done={done}", flush=True)
        feedback_raw  = info.get('feedback', '')
        feedback_safe = feedback_raw.encode('ascii', errors='replace').decode('ascii')[:120]
        print(f"    <- Feedback : {feedback_safe}", flush=True)

        # ── Emit [STEP] ───────────────────────────────────────────────────────
        print(f"[STEP] task={level} step={step_num} reward={reward:.4f} is_correct={is_correct}", flush=True)

        steps.append({
            "step":       step_num,
            "task_id":    graded_scenario_id,
            "action":     action,
            "reward":     reward,
            "is_correct": is_correct,
            "reasoning":  reasoning,
        })

        # Advance scenario ID tracker: next obs comes from the following scenario
        # (we don't know its ID until after the next step, so use graded+1 label)
        current_scenario_id = step_resp.get("task_id", "?")  # refreshed next iteration

        # Capture episode score and metrics from the terminal step.
        # info["score"] is non-None only when done=True (set by env.py via GRADERS).
        if done:
            episode_score   = info.get("score")
            episode_metrics = info.get("metrics")

        obs = step_resp.get("observation")
        if obs is None and not done:
            print("  [WARN] obs is None but done=False β€” breaking.", flush=True)
            break

    # ── Fallback: fetch from /state if episode ended without done=True ────────
    # (happens when MAX_STEPS_PER_LEVEL is reached before all tasks complete)
    if episode_score is None:
        state_resp    = _get("/state")
        episode_score = state_resp.get("overall_score", 0.01)
        episode_metrics = state_resp.get("metrics")

    print(f"\n  {'-'*50}", flush=True)
    print(f"  Level {level.upper()} complete | steps={step_num} | score={episode_score:.4f}", flush=True)

    # ── Emit [END] ────────────────────────────────────────────────────────────
    print(f"[END] task={level} score={episode_score:.4f} steps={step_num}", flush=True)

    return {
        "level":           level,
        "total_tasks":     total_tasks,
        "steps":           steps,
        "overall_score":   episode_score,
        "episode_metrics": episode_metrics or {},
    }


# ─────────────────────────────────────────────────────────────────────────────
# Main
# ─────────────────────────────────────────────────────────────────────────────

def main() -> None:
    parser = argparse.ArgumentParser(description="PhishGuard-Env Baseline Inference")
    parser.add_argument(
        "--level",
        choices=["easy", "medium", "hard"],
        default=None,
        help="Run a single difficulty level instead of all three.",
    )
    parser.add_argument(
        "--output",
        default=None,
        help="Path to write JSON results.",
    )
    args = parser.parse_args()

    levels_to_run = [args.level] if args.level else ["easy", "medium", "hard"]
    output_path   = args.output or f"results_{datetime.now(timezone.utc).strftime('%Y%m%dT%H%M%S')}.json"

    try:
        health = _get("/health")
        print(f"  server status: {health.get('status', 'unknown')}", flush=True)
    except Exception as exc:
        print(f"[ERROR] Cannot reach environment server at {ENV_BASE_URL}: {exc}", flush=True)
        print("        Make sure `python env.py` is running in another terminal.", flush=True)
        sys.exit(1)

    results: List[Dict[str, Any]] = []
    for level in levels_to_run:
        result = run_level(level)
        results.append(result)
        time.sleep(1)

    # ── Aggregate scoring ─────────────────────────────────────────────────────
    # Weighted by task count so all 10 tasks contribute equally
    # (easy=3, medium=4, hard=3).
    total_steps   = sum(len(r["steps"]) for r in results)
    total_correct = sum(s["is_correct"] for r in results for s in r["steps"])
    weighted_sum  = sum(r["overall_score"] * r["total_tasks"] for r in results)
    total_tasks   = sum(r["total_tasks"] for r in results)
    avg_score     = weighted_sum / total_tasks if total_tasks else 0.0

    # Cross-level grade_performance over combined metrics (mirrors FocusAI)
    if len(results) > 1:
        combined_metrics: Dict[str, Any] = {
            "total_tasks":      sum(r["episode_metrics"].get("total_tasks",      0) for r in results),
            "completed_tasks":  sum(r["episode_metrics"].get("completed_tasks",  0) for r in results),
            "perfect_tasks":    sum(r["episode_metrics"].get("perfect_tasks",    0) for r in results),
            "on_time":          sum(r["episode_metrics"].get("on_time",          0) for r in results),
            "breach_count":     sum(r["episode_metrics"].get("breach_count",     0) for r in results),
            "disruption_count": sum(r["episode_metrics"].get("disruption_count", 0) for r in results),
            "total_steps":      sum(r["episode_metrics"].get("total_steps",      0) for r in results),
        }
        performance_score = float(grade_performance(combined_metrics))
    else:
        performance_score = avg_score

    print(f"\n{'='*60}", flush=True)
    print(f"  BASELINE SUMMARY", flush=True)
    print(f"{'='*60}", flush=True)
    print(f"  Total steps      : {total_steps}", flush=True)
    print(f"  Correct steps    : {total_correct}", flush=True)
    print(f"  Weighted score   : {avg_score:.4f}  (pass threshold: {PASS_THRESHOLD})", flush=True)
    print(f"  Performance score: {performance_score:.4f}  (grade_performance)", flush=True)
    for r in results:
        print(f"  {r['level']:8s} score: {r['overall_score']:.4f}  ({r['total_tasks']} tasks)", flush=True)

    success = avg_score >= PASS_THRESHOLD

    # Build per-task summary for the results file
    all_tasks: List[Dict[str, Any]] = []
    for r in results:
        level_correct = sum(1 for s in r["steps"] if s["is_correct"])
        all_tasks.append({
            "task_id":    r["level"],
            "is_correct": level_correct > 0,
            "reward":     r["overall_score"],
            "level":      r["level"],
            "steps":      r["steps"],
        })

    run_summary = {
        "timestamp":         datetime.now(timezone.utc).isoformat(),
        "model":             MODEL_NAME,
        "env":               ENV_BASE_URL,
        "levels":            levels_to_run,
        "total_steps":       total_steps,
        "total_correct":     total_correct,
        "avg_score":         round(avg_score, 4),
        "performance_score": round(performance_score, 4),
        "pass_threshold":    PASS_THRESHOLD,
        "success":           success,
        "tasks":             all_tasks,
        "level_results":     results,
    }

    try:
        with open(output_path, "w", encoding="utf-8") as fh:
            json.dump(run_summary, fh, indent=2)
        print(f"\n  Results saved -> {output_path}", flush=True)
    except OSError as exc:
        print(f"\n  [WARN] Could not save results: {exc}", flush=True)


if __name__ == "__main__":
    main()