IncidentEnv — Autonomous Incident Remediation Environment

IncidentEnv is an OpenEnv-compatible RL environment for microservice incident response.
It includes a Rust simulation core, Python environment wrappers, reward/graders, a FastAPI runtime, Torch/PPO training, a React dashboard, and historical trace replay in reality mode.

What this repo contains

Meta_Hackathon/
├── incident-env/               # Main runtime + training package
│   ├── engine/                 # Rust simulation engine (PyO3)
│   ├── envs/                   # Gym/OpenEnv wrappers
│   ├── rewards/                # Reward components
│   ├── graders/                # Programmatic + LLM grading
│   ├── training/               # PPO training/eval/backtest
│   ├── api/                    # FastAPI server
│   ├── dashboard/              # React UI
│   ├── scenarios/              # Scenario configs + historical traces
│   └── tests/                  # Test suite
├── API_CONTRACT.md
├── ARCHITECTURE.md
├── DEMO_GUIDE.md
├── command.md                  # End-to-end commands
├── trackA.md
└── trackB.md

Core contract

Surface	Value
Observation	(72,) float32 in [0,1]
Structure	12 services × 6 metrics
Metric order	cpu, memory, error_rate, latency_p50, latency_p99, request_rate
Action	MultiDiscrete([12, 7])
Action layout	[service_id, action_type]
Reward	step reward in [-1.0, 1.0]

Execution modes

• benchmark: synthetic incident overlays for fast iteration.
• reality: historical trace replay + safety rails + operational scoring + audit trail.

Reality mode supports:

• timeline-based trace events
• action cooldowns and dependency checks
• high-risk action approvals (approval_token)
• action justification checks
• operational metrics (mttr_minutes, false_positive_rate, slo_recovery, customer_impact_minutes)

Quick start (Windows)

From C:\Users\ssang\Downloads\Meta_Hackathon\incident-env:

.\bootstrap.ps1
.\.venv\Scripts\python.exe -m pytest tests\

Run the stack

API

cd C:\Users\ssang\Downloads\Meta_Hackathon\incident-env
.\.venv\Scripts\python.exe -m uvicorn api.main:app --host 127.0.0.1 --port 8000

Dashboard (dev)

cd C:\Users\ssang\Downloads\Meta_Hackathon\incident-env\dashboard
npm install
npm run dev

Dashboard via API static route (/ui)

cd C:\Users\ssang\Downloads\Meta_Hackathon\incident-env\dashboard
npm run build
cd ..
.\.venv\Scripts\python.exe -m uvicorn api.main:app --host 127.0.0.1 --port 8000

Open: http://127.0.0.1:8000/ui

Runtime endpoints

Required:

• GET /
• GET /health
• POST /reset
• POST /step
• GET /state
• GET /metadata
• GET /schema
• POST /mcp

Supplemental:

• GET /episode/options
• POST /episode/start
• WS /episode/stream/{episode_id}
• POST /backtest/run

Reality mode examples

Reset with trace replay

Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:8000/reset" -ContentType "application/json" -Body '{"scenario":"bad_deploy","execution_mode":"reality","trace_id":"bad_deploy_trace_001"}'

Step with justification + approval

Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:8000/step" -ContentType "application/json" -Body '{"action":[3,2],"justification":"Rollback deploy for active deploy and error_rate symptoms while reducing escalation risk.","approval_token":"INC-APPROVED","operator_id":"oncall"}'

Training, eval, inference, backtest

cd C:\Users\ssang\Downloads\Meta_Hackathon\incident-env
.\.venv\Scripts\python.exe training\train.py --epochs 100
.\.venv\Scripts\python.exe training\eval.py --episodes 25
.\.venv\Scripts\python.exe inference.py --agent greedy
.\.venv\Scripts\python.exe -m training.backtest --agent-mode greedy --max-incidents 50 --output backtest_report.json

Validation command

cd C:\Users\ssang\Downloads\Meta_Hackathon\incident-env
.\.venv\Scripts\python.exe -m pytest tests\

Docker

cd C:\Users\ssang\Downloads\Meta_Hackathon\incident-env
docker build -t incidentenv .
docker run --rm -p 8000:7860 incidentenv

Troubleshooting

GET /episode/options returns 404

You are likely running the wrong backend copy. Start API from incident-env (hyphen), not a legacy incidentenv folder.

trained mode requested ... falling back to random

No checkpoint is available/selected. Use greedy, random, or four_stage, or train first to generate checkpoints\latest.pt.

Browser console content-all.js errors

Errors like Cannot find menu item with id translate-page are from browser extensions, not IncidentEnv.

Important docs

• incident-env\README.md — deep technical details
• command.md — command cookbook for setup/test/run/validate
• API_CONTRACT.md — integration contract
• ARCHITECTURE.md — system architecture
• DEMO_GUIDE.md — demo flow
• trackA.md, trackB.md — progress and handoff status