File size: 15,128 Bytes
4058302
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a574799
4058302
eb2d131
 
 
 
 
4058302
eb2d131
 
 
 
58af620
 
 
a574799
 
4058302
 
 
 
 
eb2d131
 
 
a574799
4058302
 
 
 
 
 
 
 
a574799
 
4058302
eb2d131
4058302
 
eb2d131
 
 
a574799
 
4058302
eb2d131
4058302
eb2d131
 
a574799
eb2d131
4058302
 
 
 
 
eb2d131
4058302
 
 
 
 
 
 
 
 
 
 
 
eb2d131
 
 
4058302
eb2d131
4058302
eb2d131
 
4058302
 
 
 
 
 
 
 
 
 
 
 
 
 
eb2d131
4058302
eb2d131
 
 
 
4058302
 
eb2d131
4058302
 
eb2d131
 
4058302
eb2d131
4058302
 
eb2d131
4058302
 
eb2d131
4058302
 
 
 
 
 
 
 
 
 
eb2d131
 
 
 
4058302
eb2d131
4058302
 
 
 
 
eb2d131
 
 
 
4058302
eb2d131
 
4058302
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb2d131
 
 
 
 
4058302
 
a574799
eb2d131
4058302
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb2d131
4058302
 
 
 
 
 
 
 
 
eb2d131
 
 
4058302
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb2d131
 
 
 
 
 
4058302
eb2d131
 
 
4058302
 
 
 
 
eb2d131
 
 
4058302
 
 
eb2d131
 
 
4058302
 
 
 
eb2d131
 
4058302
eb2d131
4058302
 
 
 
 
eb2d131
4058302
 
 
 
 
 
eb2d131
4058302
 
eb2d131
4058302
 
eb2d131
 
 
 
 
a574799
4058302
 
 
 
 
58af620
eb2d131
58af620
 
 
 
 
 
eb2d131
 
 
 
 
a574799
 
4058302
a574799
 
 
 
 
58af620
 
 
 
 
 
eb2d131
a574799
 
eb2d131
 
 
 
 
4058302
eb2d131
a574799
 
4058302
a574799
 
 
eb2d131
a574799
 
 
eb2d131
 
58af620
 
 
 
 
 
a574799
58af620
 
 
 
 
 
 
 
 
4058302
 
 
 
58af620
4058302
 
 
 
 
eb2d131
58af620
 
 
 
 
 
eb2d131
 
 
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
"""Baseline inference for the Incident Command Center environment.

Two policies are provided:

- `HeuristicCoordinator` — a deterministic state machine that exercises the
  full action space, picks role-appropriate actors, and consults the
  observation's `investigation_targets` and `playbook_hints` so the heuristic
  adapts to whatever the server is currently serving.
- `random_action` — a pure random baseline for comparison.

Running this script hits a deployed environment (local or Hugging Face Space)
and prints a structured trace the hackathon judges can follow.
"""

from __future__ import annotations

import asyncio
import json
import os
import random
from typing import Dict, List, Optional

from client import IncidentCommandEnvClient
from models import IncidentAction, IncidentObservation

ENV_URL = os.getenv("ENV_URL", "http://127.0.0.1:8000")
BENCHMARK = "incident_command_center_env"
RANDOM_BASELINE = os.getenv("RANDOM_BASELINE", "false").lower() == "true"
# When set, run an LLM-backed policy (base or fine-tuned checkpoint) instead
# of the heuristic / random ones. Point this at a HF hub id or a local dir.
POLICY_MODEL = os.getenv("POLICY_MODEL", "").strip()


# ---------------------------------------------------------------------------
# Logging helpers (structured line format, easy to grep)
# ---------------------------------------------------------------------------


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


def log_step(
    step: int,
    action: str,
    reward: float,
    done: bool,
    error: Optional[str] = None,
    components: Optional[Dict[str, float]] = None,
) -> None:
    error_val = error if error else "null"
    done_val = str(done).lower()
    comp_val = "-" if not components else ",".join(f"{k}={v:+.2f}" for k, v in components.items())
    print(
        f"[STEP] step={step} action={action} reward={reward:+.2f} "
        f"done={done_val} error={error_val} components={comp_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} score={score:+.3f} rewards={rewards_str}",
        flush=True,
    )


# ---------------------------------------------------------------------------
# Heuristic coordinator
# ---------------------------------------------------------------------------


class HeuristicCoordinator:
    """Deterministic multi-agent playbook agent.

    The state machine runs per incident and picks the correct specialist for
    each action so it never eats the wrong-actor penalty:

    1. Triage inspects logs + metrics using observation-provided targets.
    2. Investigator consults a KB article for the playbook.
    3. Ops Manager negotiates handoff to the owner the incident expects.
    4. Investigator applies a fix matched to inferred root cause.
    5. Ops Manager submits a postmortem when the incident marks it required.
    6. Ops Manager closes the incident with the inferred root cause.
    """

    def __init__(self) -> None:
        self._phase_by_incident: Dict[str, int] = {}
        self._root_cause_by_incident: Dict[str, str] = {}

    def select_action(self, observation: IncidentObservation) -> IncidentAction:
        incident_id = observation.incident_id
        phase = self._phase_by_incident.get(incident_id, 0)
        targets = observation.investigation_targets or {}
        log_targets = targets.get("logs", []) or []
        metric_targets = targets.get("metrics", []) or []
        kb_targets = targets.get("kb", []) or observation.playbook_hints

        # Haystack of all visible text we can mine for clues.
        haystack = " ".join(
            [
                observation.incident_title or "",
                observation.incident_description or "",
                observation.terminal_output or "",
                " ".join(observation.visible_signals or []),
            ]
        ).lower()

        if phase == 0 and log_targets:
            self._phase_by_incident[incident_id] = 1
            return IncidentAction(
                actor="triage_agent",
                action_type="inspect_logs",
                target=self._best_target(haystack, log_targets),
                reason="Initial triage: scan top logs for failure signature.",
            )

        if phase <= 1 and metric_targets:
            self._phase_by_incident[incident_id] = 2
            return IncidentAction(
                actor="triage_agent",
                action_type="inspect_metrics",
                target=self._best_target(haystack, metric_targets),
                reason="Correlate logs with dashboards.",
            )

        if phase <= 2 and kb_targets:
            self._phase_by_incident[incident_id] = 3
            return IncidentAction(
                actor="investigator_agent",
                action_type="consult_kb",
                target=self._best_target(haystack, list(kb_targets)),
                reason="Review runbook for candidate fix.",
            )

        if phase <= 3:
            self._phase_by_incident[incident_id] = 4
            owner = self._infer_owner(haystack, observation.customer_tier)
            return IncidentAction(
                actor="ops_manager_agent",
                action_type="negotiate_handoff",
                target=owner,
                reason="Route to accountable specialist.",
            )

        if phase <= 4:
            self._phase_by_incident[incident_id] = 5
            guess = self._infer_root_cause(haystack)
            self._root_cause_by_incident[incident_id] = guess
            return IncidentAction(
                actor="investigator_agent",
                action_type="apply_fix",
                resolution_summary=self._generate_fix_plan(guess),
                reason=f"Attempt mitigation for {guess}",
            )

        if phase <= 5 and observation.postmortem_required and not observation.postmortem_submitted:
            self._phase_by_incident[incident_id] = 6
            guess = self._root_cause_by_incident.get(
                incident_id, self._infer_root_cause(haystack)
            )
            return IncidentAction(
                actor="ops_manager_agent",
                action_type="submit_postmortem",
                postmortem_note=(
                    f"Incident {incident_id}: identified root cause {guess}. "
                    "Mitigation applied. Follow-up actions queued for "
                    "reliability review."
                ),
                reason="High-impact incident — postmortem required.",
            )

        guess = self._root_cause_by_incident.get(
            incident_id, self._infer_root_cause(haystack)
        )
        return IncidentAction(
            actor="ops_manager_agent",
            action_type="close_incident",
            root_cause=guess,
            resolution_summary=f"Closed with hypothesis {guess}.",
            confidence=0.75,
            reason="Enough evidence gathered to close incident.",
        )

    # -- helpers ------------------------------------------------------------

    def _best_target(self, haystack: str, candidates: List[str]) -> str:
        """Pick the candidate target whose tokens most overlap with the haystack."""
        best = candidates[0]
        best_score = -1
        for candidate in candidates:
            score = sum(1 for token in candidate.lower().split("-") if token in haystack)
            if score > best_score:
                best = candidate
                best_score = score
        return best

    def _infer_owner(self, haystack: str, tier: str) -> str:
        if tier == "enterprise":
            return "ops_manager_agent"
        if any(
            token in haystack
            for token in ["deploy", "rate", "sla", "rotation", "cert", "mtls"]
        ):
            return "ops_manager_agent"
        if any(
            token in haystack
            for token in ["schema", "export", "cache", "inventory", "search", "ranking"]
        ):
            return "investigator_agent"
        return "triage_agent"

    def _infer_root_cause(self, haystack: str) -> str:
        table = [
            (("redis", "pool"), "redis_connection_pool_exhausted"),
            (("jwt",), "jwt_clock_skew_mismatch"),
            (("token", "clock"), "jwt_clock_skew_mismatch"),
            (("spf",), "spf_record_misconfiguration"),
            (("cache", "invalidation"), "cache_invalidation_topic_lag"),
            (("timezone",), "timezone_normalization_bug"),
            (("offset",), "timezone_normalization_bug"),
            (("idempotency",), "idempotency_key_regression"),
            (("duplicate", "invoice"), "idempotency_key_regression"),
            (("mtls",), "mtls_cert_chain_mismatch"),
            (("certificate", "chain"), "mtls_cert_chain_mismatch"),
            (("feature", "flag"), "feature_flag_scope_misconfigured"),
            (("429",), "rate_limit_misconfigured_for_promo_segment"),
            (("promo",), "rate_limit_misconfigured_for_promo_segment"),
            (("schema", "drift"), "schema_version_drift"),
            (("schema", "mismatch"), "schema_version_drift"),
            (("dedupe",), "dedupe_rule_disabled"),
            (("alert", "storm"), "dedupe_rule_disabled"),
            (("out-of-order",), "event_ordering_race_condition"),
            (("oversell",), "event_ordering_race_condition"),
            (("deadlock",), "lock_escalation_on_reporting_view"),
            (("reporting", "lock"), "lock_escalation_on_reporting_view"),
        ]
        for tokens, guess in table:
            if all(tok in haystack for tok in tokens):
                return guess
        return "unknown"

    def _generate_fix_plan(self, root_cause: str) -> str:
        fixes = {
            "redis_connection_pool_exhausted": "increase redis pool and recycle stale connections",
            "jwt_clock_skew_mismatch": "sync clock tolerance and increase jwt leeway",
            "spf_record_misconfiguration": "fix spf record and align sending domain",
            "cache_invalidation_topic_lag": "scale invalidation consumer and replay partition 3",
            "timezone_normalization_bug": "patch timezone parser and use iana timezone map",
            "idempotency_key_regression": "restore idempotency guard and persist retry token first",
            "mtls_cert_chain_mismatch": "reissue certificate chain with full intermediate chain",
            "feature_flag_scope_misconfigured": "rollback feature flag and restrict experiment segment",
            "rate_limit_misconfigured_for_promo_segment": (
                "hotfix promo segment rate limits and enable exponential backoff"
            ),
            "schema_version_drift": "enforce schema negotiation and pin serializer to v11",
            "dedupe_rule_disabled": "restore dedupe rule and replay critical fingerprints",
            "event_ordering_race_condition": "enable sequence guards and quarantine out-of-order events",
            "lock_escalation_on_reporting_view": (
                "offload reporting to replica and schedule reporting off-peak"
            ),
        }
        return fixes.get(root_cause, "collect additional diagnostics and rollback last change")


# ---------------------------------------------------------------------------
# Random baseline
# ---------------------------------------------------------------------------


def random_action(observation: IncidentObservation) -> IncidentAction:
    action_type = random.choice(observation.available_actions or ["inspect_logs"])
    teams = observation.available_teams or [
        "triage_agent",
        "investigator_agent",
        "ops_manager_agent",
    ]
    actor = random.choice(teams)

    targets_pool: List[str] = []
    for _tool, values in (observation.investigation_targets or {}).items():
        targets_pool.extend(values)
    targets_pool.extend(
        ["payments-api", "auth-service", "dash-auth", "dash-redis", "kb-rate-limits"]
    )
    random_target = random.choice(targets_pool)

    return IncidentAction(
        actor=actor,  # type: ignore[arg-type]
        action_type=action_type,  # type: ignore[arg-type]
        target=random_target,
        root_cause="unknown",
        resolution_summary="random baseline action",
    )


# ---------------------------------------------------------------------------
# Episode driver
# ---------------------------------------------------------------------------


async def run_task(task_name: str, llm_policy=None) -> None:
    env = IncidentCommandEnvClient(base_url=ENV_URL).sync()
    if llm_policy is not None:
        policy_name = f"llm:{getattr(llm_policy, 'label', POLICY_MODEL)}"
    elif RANDOM_BASELINE:
        policy_name = "random_baseline"
    else:
        policy_name = "heuristic_coordinator"
    coordinator = HeuristicCoordinator()

    log_start(task=task_name, env=BENCHMARK, policy=policy_name)

    rewards: List[float] = []
    steps_taken = 0
    success = False
    score = 0.0

    try:
        res = env.reset(task_name=task_name)
        while not res.done:
            steps_taken += 1
            if llm_policy is not None:
                action = llm_policy.select_action(res.observation)
            elif RANDOM_BASELINE:
                action = random_action(res.observation)
            else:
                action = coordinator.select_action(res.observation)
            res = env.step(action)
            reward = float(res.reward or 0.0)
            rewards.append(reward)
            log_step(
                step=steps_taken,
                action=f"{action.actor}:{action.action_type}:{action.target or '-'}",
                reward=reward,
                done=res.done,
                components=getattr(res.observation, "reward_components", None),
            )

        score = sum(rewards) / len(rewards) if rewards else 0.0
        success = score > 0.1
    finally:
        try:
            env.close()
        except Exception:
            pass
        log_end(success=success, steps=steps_taken, score=score, rewards=rewards)


def main() -> None:
    llm_policy = None
    if POLICY_MODEL:
        from llm_policy import LLMPolicy

        llm_policy = LLMPolicy(POLICY_MODEL, label=POLICY_MODEL)

    for task in ["easy", "medium", "hard"]:
        asyncio.run(run_task(task, llm_policy=llm_policy))

    if llm_policy is not None:
        policy_label = f"llm:{POLICY_MODEL}"
    elif RANDOM_BASELINE:
        policy_label = "random_baseline"
    else:
        policy_label = "heuristic_coordinator"

    print(
        json.dumps(
            {
                "benchmark": BENCHMARK,
                "policy": policy_label,
                "env_url": ENV_URL,
            },
            indent=2,
        )
    )

    if llm_policy is not None:
        try:
            llm_policy.release()
        except Exception:
            pass


if __name__ == "__main__":
    main()