Upload api.py

Require explicit web search requests

fix(persona): tom serio-coloquial-logico-racional + fim de respostas com aspas

fix(persona): tom serio-coloquial-racional + declaracoes de afeto + strip de aspas em loop

fix(chat+images): schnell primeiro no HF + interjeicoes/identidade sem web search

fix(chat): busca web com timeout 20s + aviso honesto quando nao ha resultados

fix(chat): gatilho de busca usa deve_pesquisar() (perguntas factuais/temporais tambem vao a web)

fix(chat): pesquisa web real no caminho Gradio - rewrite deixa de substituir a mensagem e injeta resultados da web

fix(chat): anti-loop de saida (_collapse_repetition) + busca factual alargada + PG retry apos cooldown

perf(ext-gpu): max 180 tokens + deteccao de tunel morto (ngrok 404)

feat(chain): external_gpu provider na chain (skills+websearch+persona) + fix historico Gradio6 sem [brackets]

fix(api): faltavam emotion_analyzer (property) e _warm_emotion_analyzer - __init__ rebentava ao inicio

fix(chat): singleton anti-veneno + _akira_reply defensivo (local_gpu_direct/graceful_sem_chaves) - fim do AttributeError providers

perf(gpu): warmup da API fora da janela GPU + resposta curta na 1a chamada (load consome ~30s dos 60s do FREE)

feat(chat): sem providers ativos responde ja com instrucao de Secrets (graceful_sem_chaves) em vez de recursar

feat(zerogpu): free tier por defeito + anti-recursao graceful + circuit breaker PG

fix(build): torch 2.8-2.13 (requisito ZeroGPU HF) + ativacao dinamica do provider local_gpu + mensagem de treino sem GPU dedicada

feat(gpu): migracao para Space Gradio GPU (Akiragpu/L4) - modelo local 4-bit em 1 prioridade da chain, QLoRA real, app.py Chat/Treino/Status

Fix: Finalize isolation and continuity boost adjustments

fix(contexto): dedup de contexto, isolamento por conversa e placeholder de busca

Fix: placeholder search fallback, tighten loop detection, improve LSTM topic extraction and conditional injection

fix(contexto): histórico cruza workers — grava conversation_id, fallback PG/usuario, repõe depth JEV, criatividade proactiva

fix(jev): boost de continuidade — follow-up encadeia tópico em curso (eleva depth +1.5 e recalcula consciousness)

fix(jev): continuidade de topico — historico no preclassify + continuity boost System 2 + injecao de encadeamento no prompt

fix(jev): gate consciência 1-10 define consciousness_10/depth_score antes do uso (fix NameError)

fix(jev): depth_score nao definido — NameError em /api/akira

feat(jev): consciência 1-10 → System-1 rápido / System-2 consciente total

feat(jev): System 1 vs System 2 control via depth score + intent criteria improvements

fix(jev): defensive getattr for jev_client — evita AttributeError em workers paralelos

feat(jev): multi-key rotation (JEV_API_KEY/_2) + /debug/jev-status live ping

fix(JEV): espelhar jev_client de LLMManager em AkiraAPI (corrigir AttributeError)

feat(jev): System One Model integration + loop_detector refactor for JEV hooks

feat(jev): add low-level noul/score/choice methods + jev_questions integration

feat(jev): System One Model integration — ultra-fast intent/emotion/routing/moderation

fix(mistral): remove cooldown/blacklist para 403 Forbidden

fix(api): _rewrite_query None-safe + _cerebras_tools NameError + query rewrite guard

fix(reply): REPLY Contexto peso máximo, IGNORE LSTM antigos, torna mais formal sobre citada

fix(api): NameError lstm_ctx em _build_context_isolation_layer (free variable)

fix: move explicit_mention_pattern and new_topic_signals to enclosing scope of _build_context_isolation_layer to fix NameError

fix(api): NameError explicit_mention_pattern em _build_context_isolation_layer

fix(api): NameError is_bot_self_response não definido em akira_endpoint

fix(vision+reply+tokens): HF router free Qwen2-VL-7B, LSTM reply lock, max token largo 8192

fix(anti-hallucination): CoT le contexto real, reconhece erro, nao defende alucinacao

fix(akira): henry reply visible + web_search ja entrega resumo na mesma hora

fix(api): implement missing _should_search and _rewrite_query methods

fix(api): remove obsolete Cerebras compact block and integrate query rewrite

feat(api): add global truncation, universal tool compact, query rewrite

chore(api): reorder provider chain (Mistral>Cerebras>Gemini>OpenRouter>Tokenra>Groq) and enable Cerebras

fix(mistral): TOP_P 1.5->0.9 clamp, remove penalties

fix(lid): sender_jid fallback robusto lid:/lid_ + ctx sender_jid para tag_everyone