4 subagentes analizaram os logs. Corrigido:
1. OPENROUTER_MODEL: nvidia/nemotron (404) → google/gemini-2.0-flash-exp:free
- nemotron é modelo de function-calling que requer registro NIM
- 404 era lookup de function ID no Nvidia, não modelo
2. Mistral timeout: 15s → 20s base + connect timeout separado (5, timeout)
- req.post(timeout=X) era connect+read combinados
- TLS handshake consumia budget de leitura
- size-bump floors também aumentados
3. Groq payload: user_prompt agora truncado a 2000, context a 600/mensagem
- Total máximo: 7000 chars (cabe no limite 7500)
- system_prompt já estava a 2000
4. Factual question gate: 'oq', 'oquê', 'o que', 'porque', 'como' adicionados
- Perguntas factuais com palavras-interrogativas nunca bloqueadas como trivial
- 3 locais corrigidos (autonomous search, light gate, tool gate)
5. Cerebras blacklist: revertido de 30d para 1h (302→1h)
6. Gemini fallback models: removido gemini-3.5-flash-lite (inexistente), adicionado 2.…
fix: provider chain + web search + gemini models
Isaac Quarenta
46e7b44
- Reordena provider chain: llama (HF Router) primeiro (3-8s), mistral por último
- Remove injeção automática de 'Angola' em queries curtas
- Adiciona timeout de 8s na pesquisa web autónoma
- Reduz scraping de páginas de 4-6 para 3 resultados
- Remove modelos Gemini deprecated (gemini-2.5-flash/-lite) → só 3.5-flash-lite
- Adiciona palavras ao whitelist de cumprimentos (valeu, blz, tmj)
- Remove keywords genéricas de _busca_direta (novo, hoje, pesquisa, etc)
- Adiciona timeout=10s no DDGS
fix: corrige falhas de treinamento em Cerebras/OpenRouter/Mistral + muda modelo Cerebras default para gpt-oss-120b
Isaac Quarenta
e7b8ee2
feat: MAC Drive System + proactive mode + agent_timeout fix
Isaac Quarenta
e830d82
- Add homeostatic drives (curiosity, alignment, memetic_efficiency) with exponential decay
- Add MAC-Thinking bridge for CoT enrichment
- Fix agent_timeout error: extract usable content before retry on internal markers
- Improve error fallback messages (less alarming)
- Add proactive skills: self_edit, self_delete, proactive_check_in, smart_react
- Add MAC API endpoints (/api/mac/state, /api/mac/satiate)
- Add listen engine proactive triggers
- Integrate MAC drives into api.py, thinking_engine, context_builder, emotional_control
- Add BotCore TS send_poll action handler
- Update .env.example and config.py with MAC settings
feat: training levels 1-3 now actually affect responses via API providers
Isaac Quarenta
5b47808
- New module: api_training_bridge.py - connects API providers (Cerebras, OpenRouter, Mistral, etc.) to training system
- Level 1 (Emotions): emotional context from DB injected into LLM prompts
- Level 2 (NLP): API-enriched responses generate better embeddings + finetuning data
- Level 3 (API Adapter): real knowledge distillation via best API provider, stores high-quality responses
- Level 4 (MoE): specialist-specific training data generated via APIs
- Level 5 (DoRA): API-generated examples when dataset is insufficient
- Training context injection in api.py: emotion history + distilled examples enrich every prompt
- Fixed: user variable scope bug in api.py, temperature parameter in _call_openrouter
