---
title: "[bojemoi] feat(alert-agent): intégration Claude API (Haiku) comme backend LLM principal"
date: 2026-07-21T20:23:41+02:00
draft: false
tags: ["commit", "bojemoi", "main"]
categories: ["Git Activity"]
summary: "Commit 3217f0b par grafana-watcher dans bojemoi"
author: "grafana-watcher"
---

## Commit `3217f0b`

| | |
|---|---|
| **Repository** | bojemoi |
| **Branch** | `main` |
| **Author** | grafana-watcher |
| **Hash** | `3217f0b22abb1c42444282a068e544257b0dc725` |


### Description

Remplace Ollama (trop lent) par Claude API (Anthropic). Le backend est
sélectionnable via LLM_BACKEND: claude/kimi/ollama. Le LLM évalue
désormais la sévérité réelle de l'alerte (critical/high/medium/low/info)
indépendamment du label Prometheus.

- llm.py: ajout _call_claude + _extract_json pour parser la réponse
- config.py: claude_api_key (secret), claude_model, llm_backend=claude
- requirements.txt: anthropic>=0.40
- alerter.py: emoji selon sévérité LLM (🔴🟠🟡🔵⚪)
- webhook.py: sévérité Telegram = decision["severity"] (LLM) > label Prometheus
- stack: LLM_BACKEND=claude, secret claude_api_key monté

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

### Files Changed

```
M	alert-agent/alert_agent/__main__.py
M	alert-agent/alert_agent/config.py
M	alert-agent/alert_agent/llm.py
M	alert-agent/requirements.txt
M	stack/48-service-alert-agent.yml
```

### Diff Summary

```
 alert-agent/alert_agent/__main__.py |  2 +-
 alert-agent/alert_agent/config.py   | 11 +++++++--
 alert-agent/alert_agent/llm.py      | 48 +++++++++++++++++++++++++++++++++----
 alert-agent/requirements.txt        |  1 +
 stack/48-service-alert-agent.yml    |  6 ++++-
 5 files changed, 59 insertions(+), 9 deletions(-)
```