duffyduck
1ea614c26b
fix(brain): CPU-only torch — verhindert 5 GB CUDA-Bloat im Brain-Image
...
sentence-transformers zieht torch als Dependency, und der Default-Wheel
auf x86_64-linux ist die CUDA-Variante mit allen NVIDIA-Libs
(nvidia-cudnn, nvidia-cublas, cuda-toolkit, triton, ...). ~5 GB pro
Build-Layer, frisst die 22-GB-VM auf.
Fix: torch CPU-Wheel explizit zuerst installieren. Damit ist die
torch-Dependency erfuellt wenn sentence-transformers spaeter kommt,
und die CUDA-Libs werden nie gezogen.
Brain laeuft eh komplett auf CPU (MiniLM-Embeddings ~120 MB), GPU-Bloat
war reine Disk-Verschwendung.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com >
2026-05-23 15:45:51 +02:00
duffyduck
70d1500096
feat(brain): Phase B — Vector-DB-Memory, Conversation-Loop, Skills, Tool-Use
...
OpenClaw (aria-core) ist raus, ARIA laeuft jetzt mit eigenem Agent-Framework
im aria-brain Container. Vector-DB-basiertes Gedaechtnis statt Sessions,
eigener Conversation-Loop mit Hot+Cold-Memory + Rolling Window, Tool-Use
fuer Skills, Memory-Destillat-Pipeline.
aria-brain/ (neuer Container)
- main.py FastAPI auf 8080, alle Endpoints
- agent.py Conversation-Loop mit Tool-Use (skill_create + run_<skill>)
- conversation.py Rolling Window, JSONL-Persistenz, Distill-Marker
- proxy_client.py httpx-Wrapper zum Claude-Proxy, OpenAI-Format
- prompts.py System-Prompt aus Hot+Cold+Skills
- migration.py Markdown-Parser fuer brain-import/ → atomare Memories
- skills.py Filesystem-Layer fuer /data/skills/<name>/ (Python-only,
venv pro Skill, tar.gz Export/Import, Run-Logs)
- memory/ Embedder (sentence-transformers, multilingual MiniLM)
+ VectorStore (Qdrant-Wrapper)
docker-compose.yml
- aria-core (OpenClaw) raus, openclaw-config Volume raus
- aria-brain Service (FastAPI + Memory)
- aria-qdrant Service (Vector-DB) mit Bind-Mount aria-data/brain/qdrant/
- Diagnostic teilt jetzt Netzwerk mit Bridge (vorher: aria-core)
- Brain bekommt SSH-Mount fuer aria-wohnung + /import fuer brain-import/
bridge/aria_bridge.py
- send_to_core → HTTP-Call an aria-brain:8080/chat (statt OpenClaw-WS)
- aria-core-spezifische Handler raus: doctor_fix, aria_restart,
aria_session_reset, Auto-Compact-Logik, OpenClaw-Handshake
- Generischer container_restart-Handler (Whitelist Bridge/Brain/Qdrant)
- Side-Channel-Events aus /chat-Response (z.B. skill_created) werden
als RVS-Events forwarded
- file_list_request / file_delete_request → an Diagnostic forwarded
- Tote OpenClaw-Connection-Logik bleibt im Code als Referenz (nicht aktiv)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com >
2026-05-11 22:23:17 +02:00