fix: UnboundLocalError _aggression_profile (inicializar antes do uso)

revert: remover deteção hardcoded (tudo por prompt + treino do emotional control)

fix: hostility falsa positiva — BART classifica neutros como anger, cruzar com regex

fix: Kiami agora responde com rude/ignorância em confronto (deteção passivo-agressivo, suggestions rudes, direction opcional)

refactor: reforço do tom anti-robô e descolado da Kiami (instructions_prompt, fallbacks, constraints, thinking engine)

feat: Kiami personality overhaul — menos robótica, mais humana, nunca diz 'sou código'

feat: vocabulary in DB, new nicknames mimi/weza/nila/Kianda

feat: add youthful vocabulary to Kiami persona

fix: erro emocional _aggression_profile + pg_dump version mismatch fallback

fix: remover fallback emocional que chamava método inexistente

feat: controle emocional em tempo real — Kiami muda de tom dinamicamente

feat: emotional control avançado + BOT_NAMES só kiami

fix: últimos refs Akira→Kiami em modules (autonomous_agent, empresa_info, native_research, infra_watchdog)

feat: session memory PostgreSQL + group isolation + skills tracking

feat: sistema de memória persistente de sessões (estilo OpenCode)

fix: COT recebe contexto completo - aumentar limites de truncamento

fix: renomear Akira→Kiami em todo o sistema de contexto

feat: personalidade Kiami leve, tímida, séria, 18 anos - config.py, database_pg, root config

fix: atualizar numero WhatsApp do bot para 30842898366561

fix: idade 18 anos no database_pg.py

fix: atualizar PersonaConfig (Kiami, 18 anos) e refs restantes de Akira

refactor: renomear Belmira para Kiami em todo o backend Python

Update modules/config.py

fix: use PG tone_config params (emoji_max, sarcasm_level, etc.) in prompt template

fix: auto-set quality_score on insert, add missing models to MAPA_MODELOS, COALESCE for modelo_usado, better error logging

fix: move periodic_hf_sync_checkpoint definition before _init_akira_safe

fix: move DB seeding to background thread, increase init timeout to 240s

fix: increase AkiraAPI init timeout to 120s

fix: get_current_api_token infinite loop - always increment attempts

fix: bypass buggy get_current_api_token() infinite loop - read token directly from env

fix: lazy InferenceClient - blocks startup for 90s, create on first use instead

fix: startup - 1 worker, timeout 90s, debug logs in _setup_providers and _setup_hf_inference

fix: startup hang - single AkiraAPI instance, DB connect_timeout=10, startup logging

fix: STM cross-worker loading e thinking engine usa raw STM

feat: prompts e tone config migrados para PG, fix indent unified_context

fix: DISTINCT+ORDER BY SQL e removido IDEMPOTENCY-CONTENT bloquer msgs user

feat: migrate STM to PostgreSQL (stm_messages table + persistence)

feat: inject belmira persona from PG into CoT brain

feat: persona + aggression detection reads from PG with fallback

feat: autonomous_agent reads moderation patterns from PG with fallback

feat: PG migration (persona, moderation, tone) + fix oie prefix

fix: tool_calls dispatch (cerebras/torouter), knowledge matching, website URL

fix: knowledge injection ANTES do thinking engine + seed atualiza dados antigos

fix: mega-sweep - crash fixes, autonomous_agent TS, memory leaks, DB bugs

fix: website da Softedge atualizado para railway.app

feat: knowledge injection da Softedge via PG (só quando keyword match)

fix: cria tabela system_events que faltava no init do PG

fix: injeta direcao da resposta (SUGESTAO_RESPOSTA) para LLM nao inverter dinamica

fix: idempotency check + content dedup para prevenir reprocessamento apos redeploy

fix: protecao anti-self-response com comparacao de numero do bot como fallback