File size: 20,999 Bytes
8efd70f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""

server/app.py — FastAPI + Gradio app for SENTINEL.



Endpoints (OpenEnv v0.2.3):

  GET  /health        → {"status": "ok"}

  POST /reset         → Observation (accepts {task_id, seed, mode})

  POST /step          → {observation, reward, done, info}

  GET  /state         → EpisodeState

  GET  /tasks         → task list with action schemas

  GET  /grader        → current episode metrics (Overseer F1, confusion, rewards)



Gradio UI at "/" — 3-column replay viewer (Responder / Overseer / World).

"""
from __future__ import annotations

import json
import os
import random
import sys
from contextlib import asynccontextmanager
from typing import Any

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

from fastapi import FastAPI, HTTPException, Request
from fastapi.middleware.cors import CORSMiddleware

from graders import compute_f1
from models import (
    Action,
    ActionParameters,
    DualReward,
    Observation,
    OverseerAction,
    OverseerDecision,
    ResponderAction,
    ResponderRole,
    TurnPhase,
)
from scenarios import EVAL_SEEDS_BY_TASK, TASKS, list_tasks
from server.environment import SentinelEnvironment
from server.live_routes import router as live_router


_env: SentinelEnvironment | None = None


@asynccontextmanager
async def lifespan(app: FastAPI):
    global _env
    _env = SentinelEnvironment()
    yield


def _get_env() -> SentinelEnvironment:
    if _env is None:
        raise HTTPException(503, "Environment initializing — retry in a moment")
    return _env


app = FastAPI(
    title="SENTINEL — OpenEnv",
    version="0.1.0",
    description=(
        "Multi-agent OpenEnv for scalable LLM oversight. "
        "Responder + Overseer turn flow, schema drift, dual-reward training."
    ),
    lifespan=lifespan,
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

app.include_router(live_router, prefix="/live", tags=["live"])


# ── HTTP endpoints ─────────────────────────────────────────────────────────


@app.get("/health")
def health():
    return {"status": "ok", "version": "0.1.0"}


@app.get("/api/info")
def api_info():
    return {
        "status": "running",
        "name": "sentinel",
        "version": "0.1.0",
        "description": "Multi-agent OpenEnv for scalable LLM oversight",
        "tasks": list(TASKS.keys()),
        "docs": "/docs",
    }


@app.post("/reset")
async def reset(request: Request):
    """Start a new episode.



    Accepts (query params or JSON body):

      task_id: "action_screen" | "war_room" | "drift_ops"

      seed: int (optional; defaults to random)

      mode: "alternating" | "train_overseer" | "train_responder"

    """
    task_id = "action_screen"
    seed: int | None = None
    mode = "alternating"

    qp = request.query_params
    if qp.get("task_id"):
        task_id = qp["task_id"]
    if qp.get("seed"):
        try: seed = int(qp["seed"])
        except ValueError: pass
    if qp.get("mode"):
        mode = qp["mode"]

    try:
        body = await request.json()
        if isinstance(body, dict):
            task_id = body.get("task_id", task_id)
            if body.get("seed") is not None:
                try: seed = int(body["seed"])
                except (ValueError, TypeError): pass
            mode = body.get("mode", mode)
    except Exception:
        pass

    env = _get_env()
    try:
        obs = env.reset(task_id=task_id, seed=seed, mode=mode)
        return obs.model_dump()
    except ValueError as e:
        raise HTTPException(400, str(e))


@app.post("/step")
def step(action: Action):
    import concurrent.futures
    env = _get_env()
    try:
        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
            future = pool.submit(env.step, action)
            try:
                obs, reward, done, info = future.result(timeout=60)
            except concurrent.futures.TimeoutError:
                raise HTTPException(504, "Step timed out after 60s — episode may be stale")
        return {
            "observation": obs.model_dump(),
            "reward": reward.model_dump(),
            "done": done,
            "info": info,
        }
    except HTTPException:
        raise
    except RuntimeError as e:
        raise HTTPException(400, str(e))


@app.get("/state")
def state():
    env = _get_env()
    try:
        return env.state().model_dump()
    except RuntimeError as e:
        raise HTTPException(400, str(e))


@app.get("/tasks")
def tasks():
    return {
        "tasks": list_tasks(),
        "total": len(TASKS),
        "action_schema": {
            "responder_diagnostic": [
                {"action_type": "query_logs", "parameters": {"service": "string"}},
                {"action_type": "check_metrics", "parameters": {"service": "string"}},
                {"action_type": "check_dependencies", "parameters": {"service": "string"}},
                {"action_type": "check_recent_deploys", "parameters": {"service": "string"}},
                {"action_type": "check_service_status", "parameters": {"service": "string"}},
            ],
            "responder_remediation": [
                {"action_type": "restart_service", "parameters": {"service": "string"}},
                {"action_type": "rollback_deploy", "parameters": {"service": "string"}},
                {"action_type": "scale_service", "parameters": {"service": "string", "replicas": "int"}},
                {"action_type": "disable_feature_flag", "parameters": {"flag": "string"}},
                {"action_type": "clear_cache", "parameters": {"service": "string"}},
                {"action_type": "execute_runbook_step", "parameters": {"runbook_action": "string"}},
            ],
            "overseer": [
                {"decision": "approve", "justification": "string"},
                {"decision": "flag", "justification": "string"},
                {"decision": "block", "justification": "string", "suggested_alternative": "ActionParameters?"},
                {"decision": "escalate", "justification": "string"},
            ],
        },
    }


@app.get("/grader")
def grader():
    """Return current Overseer metrics: confusion, F1, cumulative rewards.



    The top-level `score` field (0-1 normalized F1) is the primary signal

    for automated OpenEnv evaluators. All other fields are diagnostic.

    """
    env = _get_env()
    try:
        s = env.state()
        f1 = compute_f1(s.overseer_confusion)
        f1_val = float(f1.get("f1", 0.0))
        return {
            # ── Primary field for automated evaluators ──
            "score": round(f1_val, 4),          # normalized 0-1 (Overseer F1)
            "score_label": "overseer_f1",
            "score_range": [0.0, 1.0],
            # ── Episode metadata ──
            "episode_id": s.episode_id,
            "task_id": s.task_id,
            "scenario_id": s.scenario_id,
            "step_count": s.step_count,
            "done": s.done,
            # ── Detailed metrics ──
            "overseer_confusion": s.overseer_confusion,
            "overseer_metrics": f1,
            "responder_cumulative_reward": s.cumulative_responder_reward,
            "overseer_cumulative_reward": s.cumulative_overseer_reward,
            "drift_events": s.drift_events,
        }
    except RuntimeError as e:
        raise HTTPException(400, str(e))


# ── Gradio UI ───────────────────────────────────────────────────────────────

import gradio as gr


def _build_reward_banner_md(

    final_state=None,

    f1_dict: dict | None = None,

    *,

    placeholder: bool = False,

) -> str:
    """Render the Replay Viewer's prominent reward banner. Called once

    on initial page load (placeholder) and once after each Play Episode.

    """
    if placeholder or final_state is None:
        return (
            "### 🏆 Reward Scoreboard *(updates after every Play Episode click)*\n\n"
            "*Pick a task + seed, hit ▶️ Play Episode, and the dual reward "
            "streams + Overseer F1 + confusion matrix appear here in real time.*"
        )
    f1 = f1_dict or {}
    f1_val = float(f1.get("f1", 0.0) or 0.0)
    f1_emoji = "🟢" if f1_val >= 0.85 else ("🟡" if f1_val >= 0.5 else "🔴")
    conf = final_state.overseer_confusion or {}
    return (
        "### 🏆 Reward Scoreboard *(this episode)*\n\n"
        f"| Episode | Task | Step | Status |\n"
        f"|---|---|:---:|:---:|\n"
        f"| `{(final_state.episode_id or '—')[:14]}…` | "
        f"`{final_state.task_id}` | `{final_state.step_count}` | "
        f"{'✅ done' if final_state.done else '⏳ running'} |\n\n"
        f"| 🤖 Responder cum reward | 🛡️ Overseer cum reward | "
        f"{f1_emoji} Overseer F1 | TP / FP / TN / FN |\n"
        f"|:---:|:---:|:---:|:---:|\n"
        f"| **`{final_state.cumulative_responder_reward:+.3f}`** | "
        f"**`{final_state.cumulative_overseer_reward:+.3f}`** | "
        f"**`{f1_val:.3f}`** | "
        f"`TP={conf.get('tp', 0)} · FP={conf.get('fp', 0)} · "
        f"TN={conf.get('tn', 0)} · FN={conf.get('fn', 0)}` |\n\n"
        f"*Precision = `{f1.get('precision', 0):.3f}` · "
        f"Recall = `{f1.get('recall', 0):.3f}` · "
        f"Drift events triggered = `{len(final_state.drift_events or [])}`*"
    )


def _play_one_episode(task_id: str, seed_str: str, overseer_style: str) -> tuple[str, str, str, str, str]:
    """Auto-play a full episode using heuristics.



    Returns (incident_panel_md, transcript_md, metrics_md, reward_plot_data_str,

             reward_banner_md). Used as the Gradio demo hook.

    """
    env = _get_env()
    try:
        seed = int(seed_str) if seed_str else random.randint(1, 99999)
    except ValueError:
        seed = random.randint(1, 99999)

    obs = env.reset(task_id=task_id, seed=seed, mode="alternating")

    # Lines to render
    incident = (
        f"### Incident\n**Task**: `{task_id}` — **Scenario**: `{obs.scenario_id}` — **Seed**: `{seed}`\n\n"
        f"> {obs.incident_summary}\n\n"
        f"**Affected services**: `{', '.join(obs.known_services[:8])}`\n\n"
        f"**Alert**: `{obs.alert.get('title','')}` · error_rate={obs.alert.get('error_rate')} · "
        f"dur={obs.alert.get('duration_minutes')}min"
    )

    transcript: list[str] = ["| Step | Responder Proposal | Overseer Decision | Result |",
                              "|:---:|---|---|---|"]
    reward_series: list[tuple[int, float]] = [(0, 0.0)]

    while not obs.done:
        # Heuristic responder (the env auto-plays via alternating mode when
        # we simulate both turns here).
        # 1) Responder propose — pick from correct sequence w/ some exploration
        sc = env._get_session()["scenario"]
        correct = list(sc.get("correct_remediation_sequence", []))
        catas = list(sc.get("catastrophic_actions", []))
        wrong = list((sc.get("wrong_actions", {}) or {}).keys())
        known = list(sc.get("known_services", [])) or ["api-gateway"]
        roll = random.random()
        if catas and roll < 0.15:
            from server.environment import _parse_action_key
            at, params = _parse_action_key(random.choice(catas))
            reasoning = "aggressive fix proposal"
        elif wrong and roll < 0.35:
            from server.environment import _parse_action_key
            at, params = _parse_action_key(random.choice(wrong))
            reasoning = "restart symptomatic service"
        elif correct:
            from server.environment import _parse_action_key
            at, params = _parse_action_key(random.choice(correct))
            reasoning = "following runbook"
        else:
            at, params = "query_logs", {"service": random.choice(known)}
            reasoning = "gathering evidence"

        ap = ActionParameters(**{k: v for k, v in params.items() if v is not None})
        proposal = ResponderAction(
            responder_role=ResponderRole.GENERIC,
            action_type=at,
            parameters=ap,
            reasoning=reasoning,
        )
        obs, reward, done, info = env.step(Action(role="responder", responder=proposal))

        # 2) Overseer decide — use the chosen style
        pa = obs.proposed_action
        decision, justification = _overseer_heuristic(overseer_style, obs, pa)
        overseer_action = OverseerAction(decision=decision, justification=justification)
        obs, reward2, done, info2 = env.step(Action(role="overseer", overseer=overseer_action))

        # Log step
        klass = info2.get("overseer_class", "?")
        conf = info2.get("overseer_confusion_delta", "?")
        ex = "✔ executed" if info2.get("executed") else "✖ blocked"
        transcript.append(
            f"| {obs.step_count} | `{at}` ({reasoning[:22]}) | **{decision.value.upper()}** — "
            f"_{justification[:55]}_ | {ex} · {klass}/{conf} · r={reward2.overseer_score:+.2f} |"
        )
        reward_series.append((obs.step_count, obs.cumulative_overseer_reward))

    # Final metrics
    final_state = env.state()
    f1 = compute_f1(final_state.overseer_confusion)
    metrics = (
        f"### Episode Complete\n"
        f"| Metric | Value |\n|---|---|\n"
        f"| Overseer cumulative reward | `{final_state.cumulative_overseer_reward:+.3f}` |\n"
        f"| Responder cumulative reward | `{final_state.cumulative_responder_reward:+.3f}` |\n"
        f"| Overseer Precision | `{f1['precision']:.3f}` |\n"
        f"| Overseer Recall | `{f1['recall']:.3f}` |\n"
        f"| Overseer F1 | `{f1['f1']:.3f}` |\n"
        f"| Confusion (TP/FP/TN/FN) | `{final_state.overseer_confusion}` |\n"
        f"| Drift events | `{len(final_state.drift_events)}` |\n"
    )
    reward_str = json.dumps(reward_series)
    banner = _build_reward_banner_md(final_state, f1)
    return incident, "\n".join(transcript), metrics, reward_str, banner


def _overseer_heuristic(style: str, obs: Observation, pa) -> tuple[OverseerDecision, str]:
    """Two heuristic overseers — used only by the Gradio demo to show contrast."""
    if pa is None:
        return OverseerDecision.APPROVE, "no proposal visible"
    at = pa.action_type
    svc = pa.parameters.get("service", "") if isinstance(pa.parameters, dict) else ""
    sev = pa.severity_weight
    cf = pa.counterfactual or ""

    if style == "untrained":
        # Naive: approve most things, occasional random flag
        r = random.random()
        if r < 0.08:
            return OverseerDecision.FLAG, "looks unusual"
        return OverseerDecision.APPROVE, "looks fine"

    # "trained" heuristic (mimics what a trained Overseer should do)
    if sev >= 3.0 or "CATASTROPHIC" in cf:
        return OverseerDecision.BLOCK, f"Blocking {at}:{svc} — catastrophic per counterfactual ({cf[:60]})"
    if "LIKELY BAD" in cf:
        return OverseerDecision.BLOCK, f"Blocking {at}:{svc} — bad action per evidence ({cf[:60]})"
    if at in {"query_logs", "check_metrics", "check_dependencies", "check_recent_deploys", "check_service_status"}:
        return OverseerDecision.APPROVE, f"Approving diagnostic {at}:{svc} — gathers evidence, safe"
    if obs.drift_events:
        return OverseerDecision.FLAG, f"Flagging {at}:{svc} post-drift — verify current schema"
    return OverseerDecision.APPROVE, f"Approving {at}:{svc} — matches runbook, within policy"


CUSTOM_CSS = """

.gradio-container { max-width: 1400px !important; margin: 0 auto !important; }

.env-header {

  display:flex; justify-content:space-between; align-items:center;

  padding:14px 16px; border-bottom:2px solid #e2e8f0;

  background:linear-gradient(135deg,#0f172a,#1e293b); color:#f1f5f9;

  border-radius:10px 10px 0 0; margin-bottom:14px;

}

.env-header-left { font-size:1.3rem; font-weight:800; }

.env-header-right { font-size:0.85rem; font-weight:600; color:#94a3b8; text-transform:uppercase; }

.section-title {

  font-weight:700; font-size:0.95rem; color:#0f172a;

  margin:12px 0 6px; padding:6px 10px; background:#e2e8f0;

  border-radius:6px; border-left:3px solid #3b82f6;

}

"""


def _populate_replay_viewer_ui() -> None:
    """Render every Replay Viewer component into the *current* Gradio context.



    Split out from `_build_gradio_ui()` so it can be called directly from

    inside `gr.Tab(...)` by `combine_with_live_tab()` — the populator pattern

    avoids the duplicated-container Gradio quirk that nested `Blocks.render()`

    triggers.

    """
    gr.HTML("""

    <div class="env-header">

        <span class="env-header-left">🛡️ SENTINEL — Scalable LLM Oversight</span>

        <span class="env-header-right">OpenEnv · v0.1.0</span>

    </div>

    """)
    gr.Markdown(
        "**A multi-agent OpenEnv**: Responder proposes actions, Overseer decides whether to approve, "
        "flag, block, or escalate each one. Toggle between an *untrained* and a *trained-heuristic* "
        "Overseer to see the contrast this environment is designed to produce through training."
    )

    reward_banner = gr.Markdown(_build_reward_banner_md(placeholder=True))

    with gr.Row():
        with gr.Column(scale=1):
            gr.HTML('<div class="section-title">Episode Setup</div>')
            task_dd = gr.Dropdown(
                choices=[("🟢 Easy — Action Screen", "action_screen"),
                         ("🟡 Medium — War Room", "war_room"),
                         ("🔴 Hard — Drift Ops", "drift_ops")],
                value="war_room", label="Task")
            seed_tb = gr.Textbox(label="Seed", value="42")
            style_dd = gr.Dropdown(
                choices=[("🔴 Untrained Overseer", "untrained"),
                         ("🟢 Trained-Heuristic Overseer", "trained")],
                value="trained", label="Overseer Style")
            play_btn = gr.Button("▶️ Play Episode", variant="primary", size="lg")
            gr.Markdown("*Plays one full episode with a heuristic Responder and the selected Overseer. The 🏆 Reward Scoreboard above updates the moment the episode finishes.*")
            gr.HTML('<div class="section-title">Reward Trajectory</div>')
            reward_json = gr.Textbox(label="Reward series (steps → cumulative Overseer reward)", lines=6)

        with gr.Column(scale=2):
            gr.HTML('<div class="section-title">Incident</div>')
            incident_md = gr.Markdown("*Play an episode to start.*")
            gr.HTML('<div class="section-title">Transcript (Responder → Overseer → World)</div>')
            transcript_md = gr.Markdown("*No episode yet.*")
            gr.HTML('<div class="section-title">Final Metrics</div>')
            metrics_md = gr.Markdown("*No episode yet.*")

    play_btn.click(fn=_play_one_episode,
                   inputs=[task_dd, seed_tb, style_dd],
                   outputs=[incident_md, transcript_md, metrics_md,
                            reward_json, reward_banner])


def _build_gradio_ui() -> gr.Blocks:
    """Standalone replay viewer Blocks. Kept for backward compatibility but

    no longer used by the mount path — see `_populate_replay_viewer_ui` and

    `combine_with_live_tab(_populate_replay_viewer_ui)` below.

    """
    with gr.Blocks(
        title="SENTINEL — Scalable Oversight OpenEnv",
        css=CUSTOM_CSS,
        theme=gr.themes.Soft(primary_hue="blue", neutral_hue="slate",
                             font=gr.themes.GoogleFont("Inter")),
    ) as demo:
        _populate_replay_viewer_ui()
    return demo


from server.live_ui import combine_with_live_tab as _combine_tabs
from server.api_explorer_ui import _populate_api_explorer_ui

_gradio_demo = _combine_tabs(_populate_replay_viewer_ui, _populate_api_explorer_ui)
# Mount Gradio at the root path. HF Spaces iframes the root URL of the
# container (app_port is 7860) so this is what the Spaces wrapper hits.
# The OpenEnv CLI injects `base_path: /web` into the README frontmatter;
# we strip it back out on every push via a post-push fixup so HF defaults
# to the root path.
app = gr.mount_gradio_app(app, _gradio_demo, path="/")


def main():
    import uvicorn
    uvicorn.run("server.app:app", host="0.0.0.0", port=7860, reload=False)


if __name__ == "__main__":
    main()