jev-ops — Extensible AI-Powered Diagnostics for DevOps & SRE

Architecture

Why System One for SRE?

Traditional LLMs produce conversational prose that automation cannot safely rely on. TypeSafe Jev evaluates atomic questions directly against state, returning typed probabilities.

🎯

Atomic Questions

Deconstruct incidents into separate gut-check determinations: classification (Choice), severity level (Score), and intervention urgency (Noul).

⚡

Parallel Evaluation

All questions in a pack are evaluated concurrently against the same state in a single round-trip. Adding questions incurs zero context rot and minimal added latency.

🛡️

Pack = Data, Not Code

Packs are declarative YAML definitions. They never execute arbitrary shell scripts or spawn subprocesses, maintaining strict security boundaries across multi-tenant clusters.

Cloud & DevOps Ecosystem

Connect With Modern Cloud & Automation Tools

Drop jev-ops into AWS Lambda, Azure Monitor, GCP Cloud Run, Terraform pipelines, CheckMK, and Ansible as an autonomous decision layer.

# AWS Lambda / CloudWatch Logs Automated Remediation
import base64, gzip, json, subprocess, boto3

def lambda_handler(event, context):
    payload = gzip.decompress(base64.b64decode(event["awslogs"]["data"]))
    logs = "\n".join([e["message"] for e in json.loads(payload)["logEvents"][:20]])
    
    # Analyze via jev-ops aws-cloudwatch pack
    proc = subprocess.run(
        ["jev-ops", "analyze", "aws-cloudwatch", "--json"],
        input=logs.encode(), stdout=subprocess.PIPE
    )
    result = json.loads(proc.stdout)
    action = result["decisions"]["recommended_action"]["value"]
    conf = result["decisions"]["recommended_action"]["confidence"]

    # Automated remediation if confidence >= 90%
    if action == "ssm_restart" and conf >= 0.90:
        ecs = boto3.client("ecs")
        ecs.update_service(cluster="prod", service="payment-worker", forceNewDeployment=True)
        return {"remediated": True, "action": "ecs_cycled"}
    return {"remediated": False, "escalated": True}

Extensibility

Packs Own the Domain Vocabulary

Zero technology-specific logic exists in the Rust binary. Packs define their own failure modes, severity rubrics, and instructions.

api_version: "jev-ops/v1"
kind: "DiagnosticPack"

metadata:
  name: "linux"
  version: "0.1.0"
  description: "General Linux diagnostic classification"
  author: "jev-ops"

spec:
  input:
    type: "text"
    max_bytes: 1048576

  decisions:
    health:
      type: "choice"
      values: [healthy, degraded, unhealthy, unknown]
    category:
      type: "choice"
      values: [normal, cpu, memory, filesystem, network, application, security, hardware, unknown]
    severity:
      type: "score"
      min: 0
      max: 5
    needs_attention:
      type: "boolean"

Safety Engineering

Confidence-Gated Automation Policy

In SRE, a model's decision tells you what, but calibrated confidence tells you whether you are permitted to act.

Confidence Tier	Permitted Action	Automation Policy	Typical Scenarios
High (≥ 95%)	ACT AUTOMATICALLY	Safe autonomous remediation without human latency	Pod restart on CrashLoop, cache eviction, transient alert suppression
Medium (70% - 94%)	CONFIRM & VERIFY	Notify engineer with pre-filled recommendation and runbook	Replication lag investigation, node draining proposal, memory limit upgrade
Low (< 70%)	DO NOT ACT	Halt automation. Escalate immediately to human on-call	Ambiguous multi-service cascade, zero-day kernel panic, unknown failure mode

CLI Workflow

Built for Modern Linux, Cloud & Automation Pipelines

Pipes directly with aws logs, az monitor, gcloud logging, kubectl, journalctl, and jq with predictable process exit codes.

Command	Description	Example
`analyze <pack>`	Analyze input via stdin or file against a loaded pack	`journalctl -p err \| jev-ops analyze linux`
`CloudWatch ECS`	Triage AWS ECS Fargate container crash or OOM	`aws logs tail /aws/ecs/pay \| jev-ops analyze aws-cloudwatch`
`GCP Cloud Run`	Diagnose Cloud Run memory ceiling and termination	`gcloud logging read "resource.type=cloud_run" \| jev-ops analyze gcp-cloud-ops`
`analyze --json`	Emit clean machine-consumable JSON without debug pollution	`jev-ops analyze linux --json \| jq .decisions`
`packs list`	Discover packs in `./packs`, config, and data dirs	`jev-ops packs list`
`packs show <pack>`	Inspect decision schema, rules, and input limits	`jev-ops packs show aws-cloudwatch`
`packs new <name>`	Scaffold a new custom diagnostic pack template	`jev-ops packs new my-service --dir ./packs`
`packs validate <path>`	Validate YAML schema, types, and limits	`jev-ops packs validate ./packs/aws-cloudwatch/pack.yaml`
`completion <shell>`	Generate auto-completions for bash, zsh, fish, powershell	`source <(jev-ops completion bash)`

Production Verified

Tested on Live Linux Authentication Logs

Real production telemetry streamed from /var/log/secure to the TypeSafe Jev (jev-latest) flagship model.

# grep -E "Failed|Invalid|error" /var/log/secure | tail -n 100 | jev-ops analyze linux --json | jq .
{
  "schema_version": "1",
  "pack": {
    "name": "linux",
    "version": "0.1.0"
  },
  "provider": "typesafe-jev (jev-latest)",
  "input": {
    "bytes": 10414,
    "truncated": false
  },
  "decisions": {
    "category": {
      "type": "choice",
      "value": "security",
      "confidence": 1.0,
      "probabilities": {
        "application": 0.0,
        "cpu": 0.0,
        "filesystem": 0.0,
        "hardware": 0.0,
        "memory": 0.0,
        "network": 0.0,
        "normal": 0.0,
        "security": 1.0,
        "unknown": 0.0
      }
    },
    "health": {
      "type": "choice",
      "value": "degraded",
      "confidence": 0.4,
      "probabilities": {
        "degraded": 0.55,
        "healthy": 0.1,
        "unhealthy": 0.33,
        "unknown": 0.02
      }
    },
    "needs_attention": {
      "type": "boolean",
      "value": true,
      "confidence": 0.83
    },
    "severity": {
      "type": "score",
      "value": 2,
      "confidence": 0.6,
      "probabilities": {
        "0": 0.0,
        "1": 0.22,
        "2": 0.54,
        "3": 0.15,
        "4": 0.04,
        "5": 0.05
      }
    }
  }
}

Multi-Platform Releases

Install CalVer v2026.09.19

Automated GitHub Action release builds with verified SHA256 checksums.

🐧 Linux (x86_64 / musl)

Standalone pure static binary for Debian, Ubuntu, RHEL, Fedora, and Alpine.

sudo curl -fsSL https://github.com/nemke82/jev-ops/releases/download/v2026.09.19/jev-ops-linux-x86_64 -o /usr/local/bin/jev-ops && sudo chmod +x /usr/local/bin/jev-ops

🍏 macOS (Apple Silicon / Intel)

Standalone native Darwin executable for M1/M2/M3/M4 Apple Silicon Macs.

sudo curl -fsSL https://github.com/nemke82/jev-ops/releases/download/v2026.09.19/jev-ops-macos-arm64 -o /usr/local/bin/jev-ops && sudo chmod +x /usr/local/bin/jev-ops

🪟 Windows & Cargo

Standalone portable Windows executable or build directly from source.

curl.exe -fsSL https://github.com/nemke82/jev-ops/releases/download/v2026.09.19/jev-ops-windows-x86_64.exe -o jev-ops.exe