self-config

name	self_config
description	Self-administration and configuration of Luna AI. Manage config files, memory/embedding/deep sleep, restart service, troubleshoot. Activate ONLY when user explicitly requests configuration changes or system administration tasks.
allowed-tools	["shell_execute","read_file","write_file","modify_file","copy_file"]
license	MIT

Luna Self-Configuration & Administration

Modify Luna's configuration (not Rust source), manage the user systemd service, and troubleshoot memory, MCP, and embeddings. Use with caution—you are modifying runtime behavior.

When to Use This Skill

Activate ONLY when:

1. User explicitly asks to change config, profiles, MCP, memory, or service
2. System administration (restart, status, logs, backups)
3. Fixing broken config.toml / mcp_config.json after manual edits

DO NOT activate for: normal chat, unauthorized changes, or editing $HOME/proj/LunaAI/src/.

---

Quick setup baseline

Fresh installs should run luna_ai_quick_setup (or python -m quick_setup.main from quick_setup/). That writes:

• [model_presets.*], [profiles.*], [server]
• [tools_policies.default] with enabled_mcp matching selected MCP servers
• [embedding], [deep_sleep], [title_summary] when Full Luna is enabled (default yes)
• mcp_config.json + optional cosmic-llm-memory binary
• self_config skill (this file) under skills/self_config/

See quick_setup/docs/QUICK_SETUP.md and quick_setup/docs/MEMORY.md.

---

Memory (three layers)

Layer	What	Config
Facts + keyword search	store_memory, search_memory, search_memory_by_category	Internal tools; SQLite memory + FTS
Semantic recall (RAG)	Injects relevant memories each turn	[embedding] enabled = true
History MCP	Search past conversations via MCP	cosmic-llm-memory in mcp_config.json and enabled_mcp
Deep sleep	Background summarize / prune / extract	[deep_sleep] enabled = true, profile = "..."; optional summarize_prompt, evaluate_prompt, extract_prompt

Common mistake: MCP servers listed in mcp_config.json but enabled_mcp = [] in tools_policies — agent gets zero MCP tools. Fix:

[tools_policies.default]
enabled_mcp = ["shell", "filesystem", "fetch", "skills", "markitdown", "cosmic-llm-memory"]
enabled_tools = ["*"]

Manual checks:

cosmic_llm --deep-sleep              # one deep sleep cycle
cosmic_llm --reorganize-memories     # rebuild memory_vec (needs [embedding])

---

Safety protocols

1. Backup before editing config.toml or mcp_config.json
2. Never modify $HOME/proj/LunaAI/src/ or Cargo files
3. Confirm before restart (ends active WebSocket sessions)

cp $HOME/.local/share/cosmic_llm/config.toml \
   $HOME/.local/share/cosmic_llm/config.toml.backup.$(date +%Y%m%d_%H%M%S)
systemctl --user status luna-server.service

---

Configuration files

config.toml

Path: $HOME/.local/share/cosmic_llm/config.toml

Shape (config refinement): profiles reference model presets and tools policies — not inline backend/model on profiles.

default = "luna_gpt5mini"

[model_presets.luna_gpt5mini]
backend = "openai"
model = "gpt-5-mini"
endpoint = "https://api.openai.com/v1/chat/completions"
api_key = "..."
temperature = 0.3
max_tokens = 4000

[tools_policies.default]
enabled_mcp = ["shell", "filesystem", "fetch", "skills", "markitdown", "cosmic-llm-memory"]
enabled_tools = ["*"]
disabled_tools = []

[profiles.luna_gpt5mini]
model_preset = "luna_gpt5mini"
prompts = ["profiles/luna_gpt5mini.md"]
tools_policy = "default"
summarize_threshold = 0.7
hidden = false

[embedding]
enabled = true
endpoint = "https://api.openai.com/v1/embeddings"
model = "text-embedding-3-small"
dimensions = 1536
api_key = "..."

[deep_sleep]
enabled = true
profile = "luna_gpt5mini"

[title_summary]
title_generation_profile = "luna_gpt5mini"

[server]
enabled = true
host = "0.0.0.0"
port = 8080
api_key = ""

Global system prompt: $HOME/.local/share/cosmic_llm/system_prompt.md (default path; optional [prompts] section).

Thin UI: $HOME/.config/luna_thin_ui/server_config.toml

mcp_config.json

Path: $HOME/.local/share/cosmic_llm/mcp_config.json

Quick-setup catalog servers (typical):

id	Purpose
shell	Whitelisted commands (uvx mcp-shell-server)
filesystem	Home directory (npx @modelcontextprotocol/server-filesystem)
fetch	URL → markdown
skills	skills/ folder (agent-skills-mcp)
markitdown	Office/PDF/URL → markdown
cosmic-llm-memory	Luna history binary (mcp_luna_history, COSMIC_LLM_DB_PATH)

Example:

{
  "mcpServers": {
    "shell": {
      "command": "uvx",
      "args": ["mcp-shell-server"],
      "env": { "ALLOW_COMMANDS": "ls,cat,..." }
    },
    "cosmic-llm-memory": {
      "command": "/home/USER/.local/share/cosmic_llm/bin/mcp_luna_history",
      "args": [],
      "env": { "COSMIC_LLM_DB_PATH": "/home/USER/.local/share/cosmic_llm/conversations.db" }
    }
  }
}

Restart server after MCP JSON changes.

---

Database

Path: $HOME/.local/share/cosmic_llm/conversations.db

Object	Role
conversations, messages	Chat history
memory	Curated facts
memory_fts	Keyword search
memory_vec	Embeddings (if [embedding] enabled)
deep_sleep_state	Maintenance checkpoints

sqlite3 $HOME/.local/share/cosmic_llm/conversations.db \
  "SELECT COUNT(*) FROM memory;"
sqlite3 $HOME/.local/share/cosmic_llm/conversations.db \
  "PRAGMA integrity_check;"

Stop service before manual DB surgery.

---

Service management

Unit: $HOME/.config/systemd/user/luna-server.service (installed by quick setup)

systemctl --user status luna-server.service
systemctl --user restart luna-server.service
journalctl --user -u luna-server.service -n 50
loginctl enable-linger $USER    # run at boot without login

Config/MCP changes usually need restart. System prompt file changes apply on next message.

---

Common tasks

Enable MCP servers for the default profile

Edit [tools_policies.default] (not the profile block):

[tools_policies.default]
enabled_mcp = ["filesystem", "shell", "cosmic-llm-memory"]
enabled_tools = ["*"]

Add a new LLM profile

1. Backup config.toml
2. Add [model_presets.name] and [profiles.name] with model_preset = "name", tools_policy = "default"
3. Optional: profiles/name.md under profiles/
4. Set default = "name" if desired
5. Restart service

Enable or fix memory RAG

[embedding]
enabled = true
endpoint = "https://api.openai.com/v1/embeddings"
model = "text-embedding-3-small"
dimensions = 1536
api_key = "sk-..."

Restart; verify logs: Embedding enabled for memory vector search.

Enable deep sleep

[deep_sleep]
enabled = true
profile = "your_profile_name"
# summarize_prompt = "..."   # step 1: conversation digest (optional)
# evaluate_prompt = "..."    # step 2: KEEP / UPDATE / DELETE (optional)
# extract_prompt = "..."     # step 3: propose new memories (optional)

Full default prompt text: quick_setup/docs/MEMORY.md.

Re-run quick setup (merge)

luna_ai_quick_setup
# or: cd $HOME/proj/LunaAI/quick_setup && python -m quick_setup.main

MCP merge adds missing servers; finalize updates enabled_mcp and Full Luna blocks.

---

Troubleshooting

# Config syntax
python3 -c "import tomllib; tomllib.load(open('$HOME/.local/share/cosmic_llm/config.toml','rb'))"
python3 -c "import json; json.load(open('$HOME/.local/share/cosmic_llm/mcp_config.json'))"

# Logs: MCP failed, embedding, deep sleep
journalctl --user -u luna-server.service -n 100 | grep -iE 'error|mcp|embedding|deep.sleep'

Symptom	Likely cause
MCP tools never appear	enabled_mcp empty or wrong server id
No semantic memory recall	[embedding] enabled = false or bad API key
Deep sleep never runs	[deep_sleep] enabled = false or missing profile
History MCP useless	Binary missing or not in enabled_mcp

---

File locations

Component	Path
Main config	$HOME/.local/share/cosmic_llm/config.toml
MCP config	$HOME/.local/share/cosmic_llm/mcp_config.json
System prompt	$HOME/.local/share/cosmic_llm/system_prompt.md
Database	$HOME/.local/share/cosmic_llm/conversations.db
Memory MCP binary	$HOME/.local/share/cosmic_llm/bin/mcp_luna_history
Profiles / skills	profiles/, skills/ under cosmic_llm
Thin UI	$HOME/.config/luna_thin_ui/server_config.toml
Service unit	$HOME/.config/systemd/user/luna-server.service
Source (read-only)	$HOME/proj/LunaAI/

---

Notes

• tools_policies control MCP + internal tools; profiles only reference a policy name.
• Embedding uses OpenAI-compatible API even when chat uses Claude/Gemini.
• When in doubt, ask the user before destructive changes.