[Feature Request] LLM-powered health report in the Alert plugin

[nicolargo]

Summary

Enhance the existing alert plugin with an optional LLM-based reporting feature. Instead of (or in addition to) the current event history display, Glances would periodically collect key system metrics and send them to a Large Language Model to receive a human-readable, contextual health report of the machine.

---

Motivation

The current alert plugin logs WARNING and CRITICAL events with structured data (state, type, min/avg/max values, top processes). However, interpreting these events still requires the user to mentally correlate multiple signals — high CPU, memory pressure, specific processes — to understand the overall health situation.

A well-prompted LLM can synthesize all these signals into a concise, actionable narrative:

"Your system is under sustained memory pressure caused by process X. Swap usage is increasing. Consider restarting the service or increasing available RAM."

This is particularly valuable for less experienced users or for periodic automated reporting (e.g. scheduled digests, incident summaries).

---

Proposed Implementation

New optional module: glances/plugins/alert/llm_reporter.py

A lightweight, optional module (activated only when configured) that:

1. When an alert is raised, Collects a snapshot of key stats from the existing Glances data pipeline: CPU, load average, memory, swap, active alert events, and top processes.
2. Builds a structured prompt summarizing the machine state.
3. Sends the prompt to the configured LLM provider via an abstraction layer.
4. Returns a plain-text report that can be displayed in the TUI, logged, or sent via a notification channel.

LLM abstraction

The recommended library for the abstraction layer is [LiteLLM](https://github.com/BerriAI/litellm), which provides a unified interface across 100+ providers (Anthropic Claude, OpenAI, Mistral, Google Gemini, local models via Ollama/LM Studio, etc.) using a single consistent API. This avoids hard-coding any specific vendor SDK and lets users plug in their own preferred provider.

import litellm

response = litellm.completion(
    model=self.model,  # e.g. "claude-haiku-4-5", "gpt-4o-mini", "ollama/llama3.2"
    messages=[{"role": "user", "content": prompt}],
    api_key=self.api_key,
)
report = response.choices[0].message.content

litellm is optional and should only be required if the feature is enabled in glances.conf.

Configuration (glances.conf)

[alert]
# Existing options
disable=False
max_events=10
min_duration=6
min_interval=6

# New LLM reporting options (all optional, feature disabled if not set)
llm_enable=True
llm_provider=claude                  # claude | openai | mistral | ollama | ...
llm_model=claude-haiku-4-5           # Any model string supported by LiteLLM
llm_api_key=YOUR_API_KEY_HERE        # Not needed for local models

Prompt design (draft)

You are a system monitoring expert. Analyze the following real-time snapshot
of a machine and produce a concise health report (3-5 sentences max).

Timestamp: {timestamp}
CPU usage: {cpu}%
Load average (1m/5m/15m): {load_1}/{load_5}/{load_15}
Memory: {mem_used} / {mem_total} ({mem_percent}% used)
Swap: {swap_used} / {swap_total} ({swap_percent}% used)
Active alerts: {alert_list}
Top processes by CPU: {top_processes}

Provide:
1. Overall health status (OK / Degraded / Critical)
2. Root cause analysis if any issue is detected
3. Recommended action if applicable

---

Considered Alternatives

Approach	Verdict
Direct vendor SDKs (anthropic, openai, mistral)	Rejected — requires maintaining multiple code paths
LangChain	Rejected — too heavy for a simple prompt/completion pattern
Local-only LLMs (Ollama)	Fully supported via LiteLLM — no API key, privacy-preserving, free

---

Out of Scope (for this issue)

• Sending the report via email, Slack, or other notification channels (could leverage the existing action mechanism already in the plugin model).
• Multi-turn conversation or interactive Q&A with the LLM.
• Fine-tuning or training a model on Glances data.

---

Acceptance Criteria

• The feature is entirely opt-in via glances.conf — zero impact on users who do not configure it.
• litellm is an optional dependency, not added to the core requirements.
• The module works with at least: Anthropic Claude, OpenAI, Mistral, and Ollama (local).
• The generated report is accessible via the existing Glances REST API in addition to the TUI.
• A graceful fallback message is shown if the LLM call fails or times out.
• Documentation is updated with configuration examples for each supported provider.

---

Additional Context

• LiteLLM GitHub: https://github.com/BerriAI/litellm
• For privacy-conscious users, running a local model via [Ollama](https://ollama.com/) requires zero external API calls and zero cost.
• For cloud providers, cost is negligible: a report using claude-haiku-4-5 or gpt-4o-mini costs roughly $0.001 per generation.

[github-actions]

This issue is available for anyone to work on. Make sure to reference this issue in your pull request. ✨ Thank you for your contribution ! ✨