Lumi_Nox v0.6.0

Lumi_Nox v0.6.0

This release makes the open-core repository easier to navigate and safer to
reuse without changing its central dual-AI coordination model.

Highlights

• Organized the five game integrations under games/, with one package per
  integration and documented module entry points.
• Completed the public memory layer with time decay and deterministic membership
  guard facts. These modules contain rules only; no viewer records or databases
  are included.
• Added a public voice registry adapter that accepts user-provided environment
  variables. Production cloned-voice IDs remain private.
• Added zero-dependency smoke tests and GitHub Actions checks for the runnable
  coordination demo, memory rules and Python compilation.
• Strengthened the curated release boundary to reject missing local imports,
  credentials, persona material, personal paths and production-only data before
  publication.

Path changes

Game modules that previously lived at the repository root now use package paths.
For example:

python -m games.buckshot.bot
python -m games.terraria.bot --watch
python -m games.kingdom_rush.ai --dry-run
python -m games.word_games.wordle.bot --no-llm

See games/README.md for the complete map.

v0.5.0 - Smarter Kingdom Rush AI + word-game solver in a worker process

Highlights

Kingdom Rush AI got a lot smarter. The tower-defense bot now pairs its algorithmic tower-placement skeleton with a per-wave LLM strategist that reads the in-game bestiary and adapts the build to each incoming wave — plus reworked tower valuation, summoner counters, and early pre-defense.

Word-game solvers no longer stall the show. The Wordle / Handle entropy solver now runs in a separate worker process (solver_worker.py / solver_client.py) with a timeout and an automatic in-process fallback, so heavy mid-game search never blocks the main loop.

Also in this release

• New support modules opened from the live system: run_architecture.py (text-vs-realtime pipeline source of truth), web_search.py (live web-search augmentation for the fast brain), viewer_name.py (viewer-name normalization + multi-join folding).
• Chat pipeline: dropped an obsolete emotion-tag gate; registered a role-play base model.
• General refresh of the open-core slice to match the daily-running system (buckshot prompt context, memory tweaks).

Open-core as always: the engine, games and architecture are here to read and build on; the characters' persona, IP and worldview stay closed.

v0.4.0 - Wordle and Handle word games

@
Adds two word games the AIs play on stream:

• Wordle - an English word game with a solver bot and a self-contained web frontend (wordle.html), driven over a bridge.
• Handle - the Chinese-idiom Wordle (汉兜), with its own engine, word database and frontend.

Both are self-contained (no external service, no browser-automation deps) - the .html pages are the game frontends and the bridge serves them and feeds state to the AIs. Five games are now open: Buckshot Roulette, Terraria, Kingdom Rush, Wordle and Handle.
@

v0.3.0 - Kingdom Rush game bot

Adds a third game the AIs play: Kingdom Rush (tower defense).

• A LuaJIT mod (kingdom_rush_bridge.lua) reverse-engineered into the game's LOVE (love2d) engine, exposing live game state and accepting commands over TCP.
• patch_kingdom_rush.py injects the bridge into a local game install (backs up the original exe first).
• kingdom_rush_ai.py / kr_strategy_llm.py - the tower-defense decision logic.
• docs/kingdom-rush-reverse-engineering.md - the full reverse-engineering notes (engine internals, injection architecture, game object structures).

As with the other games, the game itself is not included - bring your own legally-owned copy.

v0.2.0 - Terraria game bot

Adds a second game the AIs play: Terraria.

• terraria_bot.py - a full action/sandbox bot featuring A* pathfinding and a five-layer goal planner (decompose -> plan -> path -> act).
• terraria_bridge.py - TCP bridge to a tModLoader mod that exposes game state.
• docs/terraria-behavior-tree.md - the bot's atomic-behavior architecture.

The bot decides and acts while the two characters narrate it as a stream segment. As with Buckshot Roulette, the game itself and the tModLoader mod are not included - bring your own.

v0.1.0 - dual-AI co-hosting system

Two AI characters co-hosting one live show - the engine and orchestration that let them share a stage. Battle-tested by a real daily livestream.

This v0.1.0 opens a coherent slice of the production system:

• realtime dual-session voice engine (doubao SC2.0) + protocol codec
• turn-taking: dynamic speaker scheduling + one-voice-at-a-time arbitration
• cross-character history mirroring
• streaming TTS (voice-cloned) and ASR
• long-term per-viewer & self memory (SQLite + LLM extraction)
• a game the AIs play (Buckshot Roulette) over a TCP bridge
• main.py: a zero-setup runnable demo of the coordination core (no API keys)

Open-core: the skeleton, engine and capabilities are open; the characters' persona, IP and worldview are not. More game adapters (Terraria, Kingdom Rush) and pluggable model interfaces are on the roadmap.