Live-test showed Grok 4.1 Fast inventing 'mcp_mychatarchive_search_brain'
when asked about Telegram chat history. Fix:
- Tool descriptions explicitly say 'TELEGRAM' and 'NICHT ChatGPT/Memory/RAG'
- SYSTEM_PROMPT_EXTRA lists all 4 exact tool names and forbids inventing
alternatives like 'mychatarchive' or 'search_brain'
- Forbids using rag_search/openmemory_search/web_search for Telegram-
related queries
- Demands honest 'no data' over hallucination
Grok 4.1 Fast was eskaliereing 'Stand zur G2010B-Vermietung?' to
perplexity/sonar instead of using chatlog_*. Strengthened
SYSTEM_PROMPT_EXTRA with explicit triggers (G2010B, D1603, ARAKAWA,
Concierge, Bnb, Kaution, Mieter) and a 'NO web search for this' rule.
Phase 2 of the chatlogger feature.
Multi-tenant:
- chatlogger-bot.service -> chatlogger-bot@.service (systemd template)
- /opt/chatlogger/<instance>.env per bot, all share /opt/chatlogger/chats.db
- existing bot migrated: chatlogger-bot@arakawa.service
- adding a second group bot is now: cp arakawa.env <name>.env, edit token+chats,
systemctl enable --now chatlogger-bot@<name>
Jarvis (hausmeister-bot) tool:
- homelab-ai-bot/tools/chatlog.py (autodiscovered by tool_loader.py)
- 4 tools: chatlog_list_chats, chatlog_recent, chatlog_search, chatlog_since
- Reads SQLite directly via sys.path += /opt/homelab-brain/chatlogger
- SYSTEM_PROMPT_EXTRA tells the LLM when to use these (Concierge, Airbnb,
Listing, 'aktueller Stand', 'wer hat zugesagt', 'muss ich reagieren').
Tested live: hausmeister-bot loads 48 tools (44 + 4 new), all chatlog
handlers callable, returns real messages from the @arakawa_concierge_bot
recording of -1003924901022.
