Sentinel-AIOps: Instructions file for Codex

🛡️ Sentinel-AIOps Mission Briefing

🎯 Project Mission

Sentinel-AIOps is an autonomous system designed for CI/CD log anomaly detection and remediation within the Antigravity ecosystem.

🧠 Technical Core

Our intelligence relies on Python-based Machine Learning models:

• LightGBM Multiclass Classifier: Supervised model classifying CI/CD logs into 10 failure categories using numerical telemetry features (CPU usage, build duration, memory consumption, retry count). Macro F1 ≈ 0.89 on a balanced 10-class test set. Ablation-verified: removing TF-IDF features reduces F1 by only 0.004 points, confirming the model learns from operational telemetry — not log text.
• Isolation Forest: Unsupervised anomaly detector for out-of-distribution logs that don't match any known failure pattern, flagging novel infrastructure incidents.
• PSI Drift Monitor: Population Stability Index tracking that compares live inference distributions against the training baseline across all numerical and categorical features. Triggers a retrain recommendation when any feature PSI exceeds 0.20.

🏗️ Infrastructure

The system follows a robust, local-first architecture:

• Inference Engine: Powered by a FastMCP server for low-latency, localized model inference.
• Observability Interface: A FastAPI dashboard providing real-time health badges, PSI drift heatmaps, inference history, and GitHub webhook ingestion.

📜 Workflow Rules

All autonomous agents interacting with this project MUST adhere strictly to the following operational protocols:

1. Always log 'Reasoning' before execution: Every action must be preceded by a clear, documented rationale.
2. Save all ML metrics as 'Artifacts': Performance metrics (specifically F1-Score and PR AUC) must be rigorously tracked and saved as permanent project artifacts.

📁 Directory Map

The project is structurally divided into the following key domains:

• /data: For storing raw logs and processed datasets used in model training and inference.
• /models: For housing the trained weights and configurations of our Isolation Forest and LightGBM models.
• /mcp-server: For containing the FastMCP-based local inference logic and API endpoints.
• /dashboard: For the Next.js frontend code providing the observability and monitoring interface.