File size: 15,305 Bytes
1fa95ff
8704391
 
 
1fa95ff
3ad8f7d
8704391
 
 
 
 
 
 
 
 
3ad8f7d
8704391
3ad8f7d
 
 
abba50b
 
 
 
 
 
1fa95ff
 
 
 
 
8704391
1fa95ff
 
 
 
 
 
8704391
1fa95ff
8704391
 
 
 
1fa95ff
8704391
 
 
 
 
1fa95ff
 
 
8704391
1fa95ff
 
82dd413
 
bf0878b
82dd413
bf0878b
 
82dd413
1fa95ff
8704391
 
 
 
 
 
 
 
 
1fa95ff
3ad8f7d
 
 
 
 
 
 
 
 
 
abba50b
 
8704391
3ad8f7d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
abba50b
8704391
1fa95ff
 
82dd413
1fa95ff
 
 
abba50b
1fa95ff
abba50b
1fa95ff
 
 
 
 
 
 
 
8704391
abba50b
 
1fa95ff
abba50b
1fa95ff
 
 
 
abba50b
1fa95ff
abba50b
1fa95ff
 
 
 
 
abba50b
 
 
 
 
 
1fa95ff
abba50b
 
 
1fa95ff
abba50b
1fa95ff
 
 
8704391
1fa95ff
 
 
 
 
 
 
 
 
 
abba50b
1fa95ff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8704391
abba50b
1fa95ff
8704391
 
 
1fa95ff
abba50b
 
 
1fa95ff
8704391
1fa95ff
 
abba50b
1fa95ff
abba50b
 
1fa95ff
8704391
 
 
 
 
 
 
1fa95ff
 
abba50b
8704391
1fa95ff
 
 
8704391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fa95ff
8704391
1fa95ff
 
8704391
 
 
 
3ad8f7d
8704391
 
3ad8f7d
abba50b
3ad8f7d
 
 
 
1fa95ff
abba50b
 
8704391
3ad8f7d
8704391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3ad8f7d
 
 
 
 
 
8704391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3ad8f7d
 
abba50b
 
3ad8f7d
 
abba50b
 
3ad8f7d
 
8704391
 
 
 
 
 
 
dbc7299
8704391
 
 
 
 
 
 
1fa95ff
 
3ad8f7d
1fa95ff
3ad8f7d
1fa95ff
3ad8f7d
1fa95ff
 
 
 
 
 
 
8704391
1fa95ff
8704391
1fa95ff
8704391
 
 
 
1fa95ff
 
abba50b
8704391
 
 
 
 
abba50b
1fa95ff
abba50b
3ad8f7d
8704391
 
 
 
3ad8f7d
1fa95ff
 
 
 
 
 
 
 
8704391
1fa95ff
abba50b
1fa95ff
 
abba50b
1fa95ff
 
 
8704391
1fa95ff
8704391
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
"""
inference.py — Baseline SRE agent for the OpenEnv SRE Incident Investigation environment.

Follows the exact pattern from the contest sample inference script.

Mandatory environment variables:
    API_BASE_URL        The API endpoint for the LLM
    MODEL_NAME          The model identifier to use for inference
    HF_TOKEN            Your Hugging Face / API key (used as LLM API key)
    LOCAL_IMAGE_NAME    Docker image name for the environment
                        e.g. registry.hf.space/arjun4707-sre-env:latest

Optional:
    ENV_BASE_URL        Direct URL to running env server (skips Docker)
                        e.g. http://localhost:8000 or https://arjun4707-sre-env.hf.space

STDOUT FORMAT (strictly required by contest evaluator):
    [START] task=<task_name> env=sre_env model=<model_name>
    [STEP]  step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>
    [END]   success=<true|false> steps=<n> score=<score> rewards=<r1,r2,...,rn>

Usage:
    python inference.py                   # runs all 3 tasks
    python inference.py --all-tasks
    python inference.py --task sre-easy-001
    python inference.py --difficulty hard
"""

from __future__ import annotations

import argparse
import asyncio
import json
import os
import sys
import time
from typing import Any, Dict, List, Optional

from openai import OpenAI

# ---------------------------------------------------------------------------
# Import our typed SRE environment client
# ---------------------------------------------------------------------------
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
try:
    from client import SREEnvClient
    from models import SREAction, SREObservation
except ImportError as e:
    print(f"[DEBUG] Import error: {e}", flush=True)
    print("[DEBUG] Make sure client.py and models.py are in the same directory.", flush=True)
    sys.exit(1)

# ---------------------------------------------------------------------------
# Config — all from environment variables (mandatory per contest rules)
# ---------------------------------------------------------------------------

# IMPORTANT: Must use API_BASE_URL and API_KEY exactly as injected by the
# contest LiteLLM proxy. Do NOT hardcode keys or bypass with other providers.
API_BASE_URL     = os.environ.get("API_BASE_URL")
MODEL_NAME       = os.environ.get("MODEL_NAME",   "gpt-4o-mini")
API_KEY = os.environ.get("API_KEY") or os.environ.get("HF_TOKEN")
LOCAL_IMAGE_NAME = os.environ.get("LOCAL_IMAGE_NAME") or os.environ.get("IMAGE_NAME", "")
ENV_BASE_URL     = os.environ.get("ENV_BASE_URL",     "")

BENCHMARK             = "sre_env"
MAX_STEPS             = 20
SUCCESS_SCORE_THRESHOLD = 0.1

ALL_TASKS = [
    {"task_id": "sre-easy-001",   "difficulty": "easy"},
    {"task_id": "sre-medium-002", "difficulty": "medium"},
    {"task_id": "sre-hard-003",   "difficulty": "hard"},
]

# ---------------------------------------------------------------------------
# Structured stdout logging — exact format required by contest evaluator
# ---------------------------------------------------------------------------

def log_start(task: str, env: str, model: str) -> None:
    print(f"[START] task={task} env={env} model={model}", flush=True)


def log_step(step: int, action: str, reward: float,
             done: bool, error: Optional[str]) -> None:
    error_val  = error.replace("\n", " ")[:120] if error else "null"
    done_val   = str(done).lower()
    action_str = str(action).replace("\n", " ").replace("\r", "")[:200]
    print(
        f"[STEP] step={step} action={action_str} "
        f"reward={reward:.2f} done={done_val} error={error_val}",
        flush=True,
    )


def log_end(success: bool, steps: int, score: float,
            rewards: List[float]) -> None:
    rewards_str = ",".join(f"{r:.2f}" for r in rewards)
    print(
        f"[END] success={str(success).lower()} steps={steps} "
        f"score={score:.3f} rewards={rewards_str}",
        flush=True,
    )


# ---------------------------------------------------------------------------
# OpenAI LLM client
# ---------------------------------------------------------------------------

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

SYSTEM_PROMPT = """You are an expert Site Reliability Engineer (SRE) investigating a production incident.

At each step output ONLY a single JSON action object. No markdown, no explanation.

AVAILABLE ACTIONS:

  {"action_type": "query_alerts"}

  {"action_type": "query_logs",
   "service": "<service_name>",
   "log_level": "ERROR",
   "time_window_minutes": 60}

  {"action_type": "query_metrics", "metric_name": "<n>"}
  Metrics: error_rate, latency_p99, latency_p50, cpu_usage,
           memory_usage, db_connections, request_rate, cache_hit_rate

  {"action_type": "annotate", "note": "<hypothesis>"}

  {"action_type": "submit",
   "root_cause_service": "<service>",
   "root_cause_type": "<type>",
   "affected_services": ["<svc1>", "<svc2>"],
   "severity": "<P1|P2|P3|P4>",
   "recommended_action": "<steps>",
   "confidence": 0.9}

Root cause types: resource_exhaustion, dependency_failure, configuration_error,
  code_bug, data_corruption, network_partition, cascading_failure, traffic_spike

STRATEGY:
1. query_alerts first.
2. query_logs for services in alerts and the topology.
3. query_metrics: error_rate, memory_usage, db_connections.
4. annotate hypothesis.
5. submit when confident. Find ROOT CAUSE, not the loudest symptom.

SEVERITY: P1=revenue loss/site down, P2=major feature broken, P3=degraded, P4=minor
AFFECTED SERVICES: list ALL services in the call chain, including indirect victims.
RECOMMENDED ACTION: for configuration_error include "rollback" and "revert".

Output ONLY valid JSON."""


def call_llm(messages: List[Dict]) -> str:
    """Call LLM via OpenAI client."""
    response = llm.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        temperature=0.1,
        max_tokens=512,
    )
    return response.choices[0].message.content.strip()


def parse_action(text: str) -> Optional[Dict]:
    """Parse JSON action from LLM output."""
    clean = text.strip()
    if clean.startswith("```"):
        lines = clean.split("\n")
        clean = "\n".join(lines[1:-1])
    try:
        return json.loads(clean)
    except json.JSONDecodeError:
        start, end = clean.find("{"), clean.rfind("}") + 1
        if start >= 0 and end > start:
            try:
                return json.loads(clean[start:end])
            except json.JSONDecodeError:
                pass
    return None


def format_obs(obs: SREObservation) -> str:
    """Format observation into LLM-readable text."""
    parts = []
    if obs.message:
        parts.append(f"[STATUS] {obs.message}")
    for a in (obs.alerts or [])[:10]:
        parts.append(
            f"[ALERT] [{a.get('severity','?').upper()}] "
            f"{a.get('alert_name')} @ {a.get('service')}: "
            f"{a.get('message')} [{a.get('status')}]"
        )
    logs = obs.logs or []
    if logs:
        parts.append(f"[LOGS] {len(logs)} entries:")
        for e in logs[-30:]:
            parts.append(
                f"  {e.get('timestamp','')} [{e.get('level','?'):5}] "
                f"{e.get('service','?')}: {e.get('message','')}"
            )
    metrics = obs.metrics or []
    if metrics:
        vals = ", ".join(str(p.get("value")) for p in metrics)
        parts.append(f"[METRIC: {obs.metric_name or '?'}] {vals}")
    if obs.grader_score is not None:
        parts.append(f"\n[FINAL SCORE] {obs.grader_score:.4f} / 1.0")
        bd = (obs.grader_breakdown or {}).get("breakdown", {})
        for k, v in bd.items():
            if k != "correct_answers":
                parts.append(f"  {k}: {v.get('score',0):.2f} (w={v.get('weight',0):.2f})")
    parts.append(f"\n[BUDGET] {obs.queries_remaining} queries remaining")
    return "\n".join(parts)


def action_to_repr(action_dict: Dict) -> str:
    """Short string repr of action for [STEP] log."""
    atype = action_dict.get("action_type", "unknown")
    if atype == "query_logs":
        return (f"query_logs(service={action_dict.get('service')},"
                f"level={action_dict.get('log_level')})")
    elif atype == "query_metrics":
        return f"query_metrics(metric={action_dict.get('metric_name')})"
    elif atype == "query_alerts":
        return "query_alerts()"
    elif atype == "annotate":
        return f"annotate(note={str(action_dict.get('note',''))[:40]})"
    elif atype == "submit":
        return (f"submit(root={action_dict.get('root_cause_service')},"
                f"type={action_dict.get('root_cause_type')})")
    return atype


# ---------------------------------------------------------------------------
# Single episode — async, matches contest sample pattern
# ---------------------------------------------------------------------------

async def run_episode(
    task_id: Optional[str] = None,
    difficulty: Optional[str] = None,
) -> Dict:
    """
    Run one full SRE episode using the OpenEnv async client.
    Always emits [START], [STEP]*n, [END].
    """
    task_label   = task_id or difficulty or "random"
    rewards:     List[float] = []
    steps_taken: int         = 0
    final_score: float       = 0.0
    success:     bool        = False

    log_start(task=task_label, env=BENCHMARK, model=MODEL_NAME)

    env = None
    try:
        # Connect to environment — Docker image takes priority (contest runner),
        # falls back to direct URL (local dev / HF Space URL)
        if LOCAL_IMAGE_NAME:
            print(f"[DEBUG] starting container from {LOCAL_IMAGE_NAME}", flush=True)
            env = await SREEnvClient.from_docker_image(LOCAL_IMAGE_NAME)
        elif ENV_BASE_URL:
            print(f"[DEBUG] connecting to {ENV_BASE_URL}", flush=True)
            env = SREEnvClient(base_url=ENV_BASE_URL)
            await env.connect()
        else:
            raise RuntimeError(
                "Set LOCAL_IMAGE_NAME (Docker) or ENV_BASE_URL (direct URL)"
            )

        # Reset
        reset_kwargs: Dict[str, Any] = {}
        if task_id:
            reset_kwargs["task_id"] = task_id
        if difficulty:
            reset_kwargs["difficulty"] = difficulty

        result = await env.reset(**reset_kwargs)
        obs    = result.observation

        messages: List[Dict] = [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user",   "content": format_obs(obs)},
        ]

        for step in range(1, MAX_STEPS + 1):
            if result.done:
                break

            steps_taken  = step
            error_msg: Optional[str] = None
            action_repr  = "error"

            try:
                action_text = call_llm(messages)
                action_dict = parse_action(action_text)
                if action_dict is None:
                    error_msg   = "parse_failed"
                    action_dict = {
                        "action_type": "submit",
                        "root_cause_service": "",
                        "root_cause_type": "",
                        "confidence": 0.0,
                    }
            except Exception as e:
                error_msg   = f"llm_error:{str(e)[:80]}"
                action_text = "{}"
                action_dict = {
                    "action_type": "submit",
                    "root_cause_service": "",
                    "root_cause_type": "",
                    "confidence": 0.0,
                }

            action_repr = action_to_repr(action_dict)

            # Step environment using typed SREAction
            sre_action = SREAction(**action_dict)
            result     = await env.step(sre_action)
            obs        = result.observation
            done       = result.done
            reward     = float(result.reward or 0.0)

            rewards.append(reward)

            # Emit [STEP] immediately after env.step() returns (contest rule)
            log_step(step=step, action=action_repr, reward=reward,
                     done=done, error=error_msg)

            messages.append({"role": "assistant", "content": action_text})
            messages.append({"role": "user",      "content": format_obs(obs)})

            if done:
                final_score = float(obs.grader_score or reward or 0.0)
                break

    except Exception as e:
        error_str = str(e)
        print(f"[DEBUG] episode error: {error_str}", flush=True)
        if not rewards:
            rewards = [0.0]
        log_step(step=max(steps_taken, 1), action="episode_error",
                 reward=0.0, done=True, error=error_str[:120])

    finally:
        # Always close env (contest rule: [END] emitted after env.close())
        if env is not None:
            try:
                await env.close()
            except Exception as e:
                print(f"[DEBUG] env.close() error: {e}", flush=True)

        final_score = min(max(final_score, 0.001), 0.999)
        success     = final_score >= SUCCESS_SCORE_THRESHOLD
        log_end(
            success=success,
            steps=max(steps_taken, 1),
            score=final_score,
            rewards=rewards if rewards else [0.0],
        )

    return {
        "task_id":     task_label,
        "difficulty":  difficulty or "?",
        "steps":       steps_taken,
        "final_score": final_score,
        "success":     success,
    }


# ---------------------------------------------------------------------------
# Multi-task runner
# ---------------------------------------------------------------------------

async def run_all_tasks() -> None:
    results = []
    for cfg in ALL_TASKS:
        try:
            r = await run_episode(
                task_id=cfg["task_id"],
                difficulty=cfg["difficulty"],
            )
            results.append(r)
        except Exception as e:
            print(f"[DEBUG] ERROR {cfg['task_id']}: {e}", flush=True)
            results.append({
                **cfg, "final_score": 0.0,
                "steps": 0, "success": False, "error": str(e),
            })
        await asyncio.sleep(2)

    avg = sum(r.get("final_score", 0) for r in results) / len(results) if results else 0
    print(f"\n[SUMMARY] average_score={avg:.4f} model={MODEL_NAME}", flush=True)
    print(json.dumps({
        "model": MODEL_NAME,
        "image": LOCAL_IMAGE_NAME or ENV_BASE_URL,
        "results": results,
        "average_score": round(avg, 4),
    }, indent=2), flush=True)


# ---------------------------------------------------------------------------
# Entry point
# ---------------------------------------------------------------------------

if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        description="SRE Incident Investigation — Baseline Inference"
    )
    parser.add_argument("--task",       type=str, default=None)
    parser.add_argument("--difficulty", type=str, default=None,
                        choices=["easy", "medium", "hard"])
    parser.add_argument("--all-tasks",  action="store_true")
    args = parser.parse_args()

    if args.task or args.difficulty:
        asyncio.run(run_episode(task_id=args.task, difficulty=args.difficulty))
    else:
        asyncio.run(run_all_tasks())