Integrate arpahls/rooms as the multi-agent room runtime (replace in-monolith room.py)

[rosspeili]

Problem

OPSIIE ships an in-process multi-agent room implementation (room.py + /room handling in OPSIIE_0_3_80_XP.py). It works for basic “temporal nexus” sessions but diverges from the official arpahls/rooms framework that ARPA maintains for multi-agent orchestration.

What we have today (bundled)

Area	Current behavior
Entry	/room nyx, g1: <theme> — regex ^/room\s+([^:]+):\s*(.+)$
Agents	opsie (always) + nyx, g1, kronos validated against AGENT_DISPLAY_NAMES
Backends	OPSIE → Ollama (get_opsie_response / stream_room_response, hardcoded llama3); others → agentic_network.ask_model (OpenAI/Gemini HTTP)
Turn selection	Query all agents, score with MiniLM + length/uniqueness heuristics (_select_best_response)
Memory	In-process Chroma collection room_<slug> + conversation_history
Export	Optional CSV under outputs/rooms/ on /close
UX	Agent-colored terminal output, speak_agent_response for voice; direct address via agent <prompt> prefix

Docs already point to arpahls/rooms as the standalone path (docs/COMMANDS.md, docs/ARCHITECTURE.md, docs/MODULES.md, README Related repos) but there is no code integration—two parallel implementations.

What arpahls/rooms provides

• Session orchestration: round_robin, argumentative, dynamic (expertise-weighted routing)
• LiteLLM routing (ollama/llama3, commercial APIs, local-first)
• @AgentName forced turns, PASS mechanic, optional global orchestrator, human-in-the-loop
• Pydantic agent/session config, custom Python inference per agent
• Timestamped RAM history → Markdown / CSV export (tested via pytest + CI)

Goal

Use arpahls/rooms as the room engine inside OPSIIE while preserving the familiar operator surface:

• Keep /room <agents>: <theme> and /close (or document a deliberate alias).
• Keep OPSIIE terminal colors and voice (speak_agent_response, agent voice IDs).
• Map authenticated user identity from kun.py (call_name, optional is_current_user) into the rooms user profile.

Avoid long-term duplication of orchestration logic in room.py.

Proposed integration (phased)

Phase 1 — Adapter behind existing /room UX (MVP)

1. Add rooms as an optional dependency (pip install from Git tag or PyPI if published) in requirements.txt (or requirements-rooms.txt extra).
2. Add opsie_rooms_adapter.py (name TBD) that:
   • Parses the existing /room command into a rooms session config (topic, agent list, turn mode default dynamic).
   • Registers each OPSIE agent as a custom_function (or LiteLLM model string where keys exist in .env):
     • opsie → wrapper around get_opsie_response / Ollama chat (same stack as main loop).
     • nyx / g1 / kronos → wrapper around ask_model(..., suppress_output=True) so existing API keys and personas stay in agentic_network.py.
   • Reuses get_agent_description() / get_agent_display_names() for personas and expertise lists.
3. Replace current_room = Room(...) with a thin OpsieRoomSession facade that delegates turns to rooms.session.Session and returns {agent, response} for the main loop to print/speak.
4. On /close, call rooms storage export (MD/CSV) and keep writing under outputs/rooms/ if we want path compatibility.

Phase 2 — Feature parity + cleanup

• Adopt @AgentName addressing in room mode (document alongside legacy agent <prompt> or migrate).
• Expose session type / turn limit via optional flags: /room dynamic nyx, g1: topic or rooms.yaml.
• Remove MiniLM load from room.py (saves boot cost) once dynamic routing lives in rooms.
• Fix duplicate logging: today get_best_response already calls add_conversation, but handle_user_query may add again—align with single source of truth in the adapter.
• Deprecate room.py with a migration note in docs/COMMANDS.md.

Phase 3 — Optional deeper coupling

• Shared .env / config for Ollama model names (see configurable Ollama issue) passed into LiteLLM as ollama/<model>.
• Optional bridge: append room transcript snippets to global mnemonic / Chroma (off by default; rooms is RAM-first by design).
• Document when to use OPSIIE /room vs standalone python cli.py from arpahls/rooms.

Non-goals (initially)

• Replacing agentic_network.py or /ask live agents.
• Forcing all OPSIE operators to install commercial API keys for room mode (OPSIE-only + local Ollama path should remain viable).

Acceptance criteria

• /room nyx, g1: <theme> starts a session backed by arpahls/rooms, not room.Room.
• At least opsie + one remote agent work in one session with correct personas.
• /close ends room mode and offers export (MD or CSV) without breaking OPSIIE main convo.
• Voice and terminal colors still work for room lines.
• docs/COMMANDS.md, docs/ARCHITECTURE.md, docs/MODULES.md, README updated; room.py marked deprecated or removed.
• Dependencies and setup documented (LiteLLM, local Ollama, API keys per agent).

Open questions

1. Dependency packaging: git submodule vs pip install git+https://github.com/arpahls/rooms@<tag>?
2. Kronos in rooms: currently limited in ask_model for non-live—skip in v1 or stub with clear message?
3. Chroma per room: drop in favor of rooms RAM history, or optional sync layer?
4. Config file: rooms.yaml at repo root vs env-only for v1?

References

• Bundled: room.py, OPSIIE_0_3_80_XP.py (/room, /close, room query block ~1996–2108)
• Official: https://github.com/arpahls/rooms — Architecture, Examples
• Related: arpahls/rooms linked from README Related repositories