# Handoff: Linux Sysadmin Agent Implementation

## Context
Brainstorming session completed for building a Linux sysadmin AI agent.

## Plan Location
`/home/will/.claude/plans/shimmering-discovering-bonbon.md`

## What Was Decided

### Architecture
- Multi-agent system with **master-orchestrator** (Opus) overseeing all agents
- New **linux-sysadmin** agent (Sonnet) for Arch Linux workstation
- Existing K8s agents become managed by master-orchestrator
- Shared state files in `~/.claude/state/` for cross-agent awareness

### Linux Sysadmin Agent
- **Target**: Arch Linux with pacman/yay/homebrew
- **Scope**: Full sysadmin (maintenance, troubleshooting, config, security)
- **Autonomy**: Conservative by default (confirm all changes), session-configurable
- **Interaction**: CLI + scheduled, alerting deferred

### Key Design Decisions
- Claude Code compatible (Markdown + YAML frontmatter for agents)
- JSON files for state/data (not agent definitions)
- Model selection policy for cost efficiency (haiku → sonnet → opus)
- Agents can delegate to multiple subagents (parallel/sequential)
- `system-instructions.json` captures all processes
- `future-considerations.json` tracks deferred items (all agents aware)

### Settings.json Cleanup Required
- Remove non-standard `agents` field with `promptFile` entries
- Keep only standard Claude Code settings

## Implementation Order (20 steps, 7 phases)
1. Foundation (state files, CLAUDE.md)
2. Master orchestrator agent
3. Linux sysadmin agent
4. Update existing K8s agents
5. Clean settings.json
6. Skills & slash commands
7. Workflows

## Future Considerations to Track
- Prometheus/Alertmanager for workstation alerting
- Network admin agent
- Personal assistant agent
- External LLM integration
- Slash commands redesign

## Next Action
Read the full plan and begin Phase 1: Foundation