Apppublic

samnitmehandiratta/my-env

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

🚨 SRE Incident Triage

Can your AI survive an on-call shift?

An OpenEnv reinforcement learning environment where agents triage real production incidents — reading noisy alerts, tracing cascading failures, and recommending the right fix under pressure.

-orange?style=for-the-badge)

</div>

🧠 What Is This?

Every company has an on-call rotation. At 3am, an engineer gets paged — alerts firing, logs streaming, services cascading. They have minutes to find the root cause and act.

This environment turns that into an RL benchmark.

An AI agent receives the same raw signals a human SRE would see: production alerts, service logs, and real-time metrics. It must investigate, diagnose, and resolve — just like a real engineer would.

Unlike toy environments, this one models a task where reasoning quality directly determines the score. LLMs that have absorbed engineering knowledge genuinely outperform those that haven't.

🎯 Three Incident Scenarios

EASY ──────────────────────────────────────────── HARD
  │                    │                            │
  ▼                    ▼                            ▼
cpu_spike       cascading_failure            memory_leak

🏷️ Task	🎚️ Difficulty	💥 Incident	🔍 What Agent Must Find
`cpu_spike`	🟢 Easy	Traffic flood hitting api-server	No rate limiting, IP making 8K req/min
`cascading_failure`	🟡 Medium	4 services down in a cascade	payment-service v2.3.1 set DB pool to 500 (was 10)
`memory_leak`	🔴 Hard	Multi-service OOM kills over 2h	Shared logging-middleware v2.1.0 never flushes its buffer

⚡ Agent Action Space

The agent speaks JSON — clean, structured, LLM-native:

json

// 🔍 Drill into a specific service
{"action_type": "investigate", "service": "payment-service"}

// 🩺 Submit your diagnosis
{
  "action_type": "diagnose",
  "severity": "critical",
  "root_cause": "DB connection pool exhausted after v2.3.1 deploy",
  "affected_services": ["payment-service", "checkout-service"]
}

// 🛠️ Recommend the fix (highest scoring action)
{
  "action_type": "resolve",
  "severity": "critical",
  "root_cause": "payment-service v2.3.1 set DB_POOL_SIZE=500, exceeding DB max_connections=100",
  "affected_services": ["payment-service", "checkout-service", "order-service"],
  "recommended_action": "rollback payment-service to v2.3.0 and restart connection pool"
}

// ✅ End episode
{"action_type": "done"}

👁️ Observation Space

Every step, the agent sees:

json

{
  "alerts": [
    "CRITICAL: payment-service health check failing (0% healthy pods)",
    "HIGH: checkout-service error rate at 78%"
  ],
  "logs": [
    "16:44:30 INFO  deployment: payment-service v2.3.1 rolled out",
    "16:44:31 ERROR payment-service: Failed to acquire DB connection (pool exhausted)",
    "16:45:02 ERROR checkout-service: payment-service call timed out after 3000ms"
  ],
  "metrics": {
    "payment-service": {"cpu_percent": 12, "error_rate": 100, "healthy_pods": 0},
    "payment-db":      {"cpu_percent": 98, "active_connections": 500, "max_connections": 100}
  },
  "step": 2,
  "additional_info": null  // populated after investigate actions
}

📊 Reward Function

Scores are partial — every step provides signal, not just the final answer:

Component	Max Weight	How It's Earned
🎚️ Severity classification	+0.30	Exact match: `critical / high / medium / low`
🔎 Root cause accuracy	+0.45	Keyword matching — partial credit per relevant term
🗂️ Affected services	+0.05	Any services identified
🛠️ Recommended action	+0.25	Keyword match on correct remediation

All scores clamped to (0.01, 0.99) — never binary, always informative.

Strategy that works best: investigate → investigate → resolve (2-step exploration before committing)

📈 Baseline Results

Tested with Qwen/Qwen2.5-72B-Instruct via HuggingFace Router:

cpu_spike          ████████████████████░  0.92  ✅ Success
cascading_failure  ████████████████████░  0.96  ✅ Success
memory_leak        █████████████░░░░░░░░  0.66  ✅ Success

The hard task (memory_leak) genuinely challenges frontier models — the shared component spans 4 services with no obvious single alert.

🚀 Quick Start

bash

# Clone and build
git clone https://github.com/samnitmehandiratta/my-openenv
cd my-openenv
docker build -t sre-triage .

# Run
docker run -p 7860:7860 \
  -e HF_TOKEN=hf_... \
  -e API_BASE_URL=https://router.huggingface.co/v1 \
  -e MODEL_NAME=Qwen/Qwen2.5-72B-Instruct \
  sre-triage

🔌 API Endpoints

Method	Endpoint	Description
`POST`	`/reset`	Start new episode — body: `{"task": "cpu_spike"}`
`POST`	`/step`	Take action — body: `{"action": {...}}`
`GET`	`/state`	Current episode state

🧪 Run Inference Locally

bash

HF_TOKEN=hf_... \
API_BASE_URL=https://router.huggingface.co/v1 \
MODEL_NAME=Qwen/Qwen2.5-72B-Instruct \
python inference.py

Built for the Scaler × HuggingFace OpenEnv Hackathon 2025

Because the best RL environments are the ones where intelligence actually matters.

</div>