Mistral Sysadmin Script (Sysdawg)

A read-only SRE copilot powered by Mistral 7B for system administration troubleshooting. Provides advisory diagnostics and safe command suggestions without autonomous execution capability.

Features

• Interactive Chat Mode: Full conversation with context history (up to 10 turns)
• One-Shot Query Mode: Headless mode for scripting and automation
• System Diagnostics: Automatic collection of memory, disk, processes, and system status
• SSH-Based: Communicates with remote Mistral instance on dedicated machine
• Read-Only by Design: Never executes commands; strictly advisory analysis only
• Safe Defaults: Explicit command whitelists and blocklists prevent dangerous operations
• Conversation Context: Maintains history for multi-turn diagnostics

Quick Start

Prerequisites

• bash
• ssh (with key-based authentication to apollo)
• Mistral 7B instance running on a remote machine (apollo)

Initial Setup

chmod +x setup-ssh.sh sysdawg
./setup-ssh.sh

This creates and configures .env with SSH details to your Mistral host.

Usage Examples

Interactive Chat Mode (Default)

./sysdawg

Start a diagnostic conversation. Type questions and receive AI analysis with diagnostic context. Maintains conversation history for follow-up questions.

Example conversation:

You: Why is my load average so high?
[System diagnostics collected and analyzed]
Sysdawg: Based on your current metrics...
You: What processes are consuming the most CPU?
[Refined diagnostics sent with conversation history]
Sysdawg: Looking at your top processes...

One-Shot Query Mode

./sysdawg -p "Why is my disk full?"

Ask a single question and get an immediate response. Perfect for automation and scripting.

Show Help

./sysdawg --help
./sysdawg -h

Configuration

The .env file stores connection details to your Mistral instance:

APOLLO_HOST=apollo.local          # Hostname or IP of Mistral machine
APOLLO_USER=jay                   # SSH username
APOLLO_PORT=22                    # SSH port
OLLAMA_MODEL=mistral              # Model name in Ollama
OLLAMA_API_URL=http://localhost:11434/api/chat  # Ollama API endpoint

Run ./setup-ssh.sh to create this file interactively. Never commit .env to version control.

Custom Models

To make the persona more robust, you can bake the system prompt directly into a custom Ollama model. This is the recommended approach for better adherence to safety rules.

1. Create the Model

You can create the custom model either locally (if you run Ollama locally) or on your remote host (apollo).

Option A: Create on Remote Host (Recommended)

1. Copy the Modelfile to your server:

   scp ollama_model/Modelfile user@apollo:~/sysdawg.Modelfile

2. SSH into the server and build the model:

   ssh user@apollo "ollama create sysdawg -f ~/sysdawg.Modelfile"

Option B: Create Locally

If you are running Ollama on your local machine:

ollama create sysdawg -f ollama_model/Modelfile

2. Configure sysdawg to Use It

Update your .env file to use the new model name:

# In .env
OLLAMA_MODEL=sysdawg

Now sysdawg will use your custom-tuned model which has the persona and safety rules permanently embedded.

Architecture

Multi-Machine Setup

Local Machine (orion)          Remote Machine (apollo)
    sysdawg script        ──SSH──>  Mistral 7B
  - Collects diagnostics          - Analyzes diagnostics
  - Builds prompts                - Returns advice
  - Manages conversation          - No execution

How It Works

1. Initialization: Loads configuration from .env and reads SYSDAWG.md persona
2. Diagnostic Collection: Gathers system state (memory, disk, processes, logs, etc.)
3. Prompt Building: Combines diagnostics + user question + AI persona
4. SSH Communication: Sends prompt to Mistral instance via SSH
5. Response Analysis: Receives advisory response and displays to user
6. Context Maintenance: Tracks conversation history for follow-up questions

Command Safety Rules

Sysdawg is fundamentally designed to be read-only. It never executes commands.

Allowed Commands (Advisory Only)

• System inspection: cat, grep, ls, find, ps, top, free, df, uptime
• Logging: journalctl --no-pager, systemctl status, dmesg
• Utilities: awk, sed (read-only), netstat, ss

Forbidden Commands (Never Executed)

• Privileged execution: sudo
• Destructive operations: rm, mkfs, dd, iptables -F
• Service control: systemctl restart/stop
• Configuration changes: sed -i (in-place editing)
• Package management: apt, yum, pip

This is not a security feature—it's a fundamental design principle. All suggestions are advisory. You decide what to run.

Project Files

• sysdawg: Main executable script (interactive + one-shot modes)
• setup-ssh.sh: Configuration helper for initial setup
• .env: Configuration file (generated by setup-ssh.sh, never committed)
• .env.example: Example configuration template
• SYSDAWG.md: System prompt and personality definition for the AI
• commands.conf: Safety rules and command configurations
• INSTRUCTIONS.md: Development roadmap and phase descriptions
• IDEA.md: Original design philosophy and safety principles
• CLAUDE.md: Detailed documentation for AI code assistants
• CHANGELOG.md: Version history
• README.md: This file

Development Roadmap

Phase 1 (Current)

• Basic shell wrapper with system diagnostics
• Interactive chat and one-shot modes
• SSH communication to Mistral
• Read-only advisory responses

Phase 2 (Planned)

• Structured output format (Summary, Likely Causes, What to Check, Suggestions, Confidence)
• JSON/YAML formatted responses
• Reduced hallucination through constraints

Phase 3 (Planned)

• Command risk classifier (SAFE / REVIEW / DANGEROUS)
• Allowlist and blocklist validation
• Enhanced safety rails

Phase 4 (Planned)

• Python migration
• Modular collectors (power.py, network.py, storage.py, services.py)
• YAML-based configuration
• Plugin architecture

Phase 5 (Planned)

• Multi-device SSH tunneling
• Fleet-wide diagnostics
• Device registry management

Phase 6 (Planned)

• Web UI and dashboard
• Multi-device health visualization
• Container and process management

Phase 7+ (Future Exploration)

• Automation of low-risk operations
• Safe command execution with safeguards
• Dry-run mode patterns

Dependencies

Required:

• bash (core functionality)
• ssh (communication with Mistral host)
• Standard utilities: grep, sed, awk, cat, cut, tr

On Mistral Host (apollo):

• Mistral 7B LLM (via Ollama, llama.cpp, or compatible)
• bash (for receiving and processing prompts)

Usage Tips

For Recurring Issues

Use interactive mode to ask follow-up questions. The AI maintains context across turns and can refine analysis based on additional checks.

For Automation

Use one-shot mode in scripts:

./sysdawg -p "Check if postgres is running"

For Learning

Ask the AI to explain what commands to run:

./sysdawg -p "How would I check if my network is properly configured?"

The response includes suggested read-only commands and explanations.

For Complex Diagnostics

Interactive mode is ideal:

1. Ask the initial question
2. Get diagnostic suggestions
3. Run the suggested commands
4. Ask follow-up questions with new data

Important Notes

• The AI is an advisor, not an executor. You remain in control of all operations.
• The Mistral 7B model is capable and clever but can hallucinate. Always verify suggestions against your system state.
• Conversation history is stored in memory only; it's not persisted across sessions.
• SSH connectivity to apollo is required. Test with ssh apollo before using sysdawg.
• The tool is designed to fail safely—if the AI suggests something wrong, nothing breaks unless you run it.

Troubleshooting

SSH Connection Failed

# Verify SSH connectivity
ssh -p 22 user@apollo.local echo "Success"

# Reconfigure with setup-ssh.sh
./setup-ssh.sh

Mistral Not Responding

• Verify Mistral instance is running on apollo: ollama list
• Check Ollama service status: systemctl status ollama
• Verify network connectivity to apollo

Diagnostic Collection Issues

• Ensure standard utilities are installed (grep, sed, awk, etc.)
• Check .env configuration is correct
• Review logs with ./sysdawg -h for debug options

License

This project is available under several open-source license options:

• MIT License: Simple and permissive
• Apache 2.0: Provides explicit patent rights
• GPL v3: Copyleft license
• BSD 3-Clause: Similar to MIT with additional restrictions

Choose the license that best fits your project goals.

Contributing

This is an advisory-only AI tool designed to be safe by default. Contributions should maintain this safety-first principle. Never add autonomous execution capabilities.