- app.py: decorator spaces.GPU so ativo se AKIRA_GPU_DURATION>0 (free=60s/chamada nao chega ao LLM local; PRO=120; L4 dedicada=0 e CUDA nativa funciona fora do decorator); Status mostra modo GPU; fallback de erro de GPU mais explicito
- api.py: guarda _in_graceful corta a recursao generate() -> circuit breaker -> graceful -> generate()
- database_pg.py: apos 3 falhas seguidas o PG desativa-se na sessao (fallback SQLite, sem spam de logs nem 1.5s de retry por mensagem)
- Dockerfile: base nvidia/cuda 12.1 + torch cu121 (imagem dedicada; o build do Space usa a base gradio e ignora este ficheiro)
fix(contexto): dedup de contexto, isolamento por conversa e placeholder de busca
Isaac Quarenta
04bb7b4
- api.py: dedup SECTION_4/5 do unified_context (historico 2x + mensagem atual 2x causavam repeticao), re-trunca final_prompt pos-injecoes, busca real quando a resposta e 'vou pesquisar', rotulos LSTM por relevancia (PV e grupo), guards para nunca substituir resposta por ''
- short_term_memory: isolamento por conversation_id (sem cid → [] em vez de memoria global), selecao ordenada por timestamp, clear_messages() real para /reset, retencao 400, fix datetime.fromtimestamp em get_messages_for_llm
- session_memory: turnos lidos por conversation_id real (antes JID/vazio cruzava PV↔grupo), filtragem por relevancia nao-destrutiva (troco recente sempre + ordem cronologica), regex de grupo corrigida
- database_pg: historico devolvido em ordem cronologica (reversed) para todos os consumidores
- lstm_extension: lock no read→modify→write, match de topico por palavra inteira, follow-up curto preserva topico
- loop_detector: decisao deterministica (confidence >= 0.75) em vez de random
- token_estima…
fix(memory): STM race idx_sm_user + pg_dump version mismatch suppress
Isaac Quarenta
23d9636
fix: Cerebras context_history, kota ban, gender detection, gíria rules
Isaac Quarenta
2f334ea
- _call_cerebras: inclui context_history (antes ignorava, causava falha silenciosa)
- Compressed path: user_prompt não é mais vazio
- mac_integration: remove 'Peço desculpas' hardcoded
- info_softedge: force update DB prompt em cada startup
- api.py tone: kota EXCLUSIVO Isaac, gírias integradas naturalmente (não sufixo)
- api.py _build_prompt: gender hint baseado no nome (mana/mano, parceira/parceiro)
- database_pg: kota removido do pool geral, mambo = coisa (não pessoa)
- database_pg: informal pronome = tu/você (não puto/kota)
fix: gírias angolanas no Postgres + diversity no prompt
Isaac Quarenta
76b8b99
- database_pg.py: nova tabela 'girias' (giria, intensidade 1-5,
tom_adequado[], exemplo_uso, ativo) com seed de 20 gírias angolanas.
Inserção idempotente via ON CONFLICT DO NOTHING.
- api.py: GIRO DE GÍRIAS section no _inject_tone_instruction (10 gírias
inline como fallback), ADAPTAÇÃO DE GÍRIA AO TOM DO UTILIZADOR (FORMAL
→ parceiro/mano; CASUAL → kota/cassules/cria; AGRESSIVO → puto/mambo/kenga).
TODO: substituir lista inline do api.py por query get_girias_for_tone()
contra a tabela girias — gírias devem crescer só via DB, sem rebuild.
- Added _safe_json_load() to database_pg.py to handle already-deserialized dict/list from psycopg2
- Replaced bare json.loads() in session_memory.py (get_memory, get_recent_skills) with _safe_json_load()
- Cast confianca to float at 3 emotion analysis points to prevent str > 0 TypeError
- Cast _last_ts to float at time isolation check to prevent str - float TypeError
- config.py: Remove duplicate AKIRA_TONE_CONFIG (was silently overwritten)
- config.py: Remove duplicate BOT_NUMERO (hardcoded overwrote env var)
- config.py: Fix is_privileged() stale SQLite import -> database_pg
- config.py: Fix SYSTEM_PROMPT in __all__ (undefined -> get_system_prompt)
- config.py: get_system_prompt() now loads from DB (info_softedge) first
- database_pg.py: Fix persona_usuario ALLOWED_PERSONA_COLS schema mismatch
- database_pg.py: Remove auto ON CONFLICT DO NOTHING injection in _convert_query
- api.py: Remove Flask compat layer (dead weight: _RequestCompat, _RequestProxy, jsonify)
- api.py: Deduplicate 8 MockToolCall classes into single top-level class
fix: pool decrementa _created ao retornar; info_softedge safe JSON
Isaac Quarenta
dc0543b
feat: session memory como agents - turnos reais, pool fix, lifecycle completo
Isaac Quarenta
24c55d8
fix: protege __init__ contra métodos ausentes no deploy
Isaac Quarenta
eb254d1
fix: use module-level _pool_return() to avoid method resolution issues
Isaac Quarenta
0c2342b
- Replaced all self.return_connection(conn) with _pool_return(conn)
- Module-level function bypasses any singleton/method resolution issues
- All 15 connection pool returns now use the direct function
- Fixes AttributeError: 'DatabasePG' object has no attribute 'return_connection'
fix: adiciona return_connection na DatabasePG para delegate ao pool
Isaac Quarenta
8b70b5f
fix: connection pool leak - return connections instead of closing them
Isaac Quarenta
a01da14
- Fixed duplicate else block in ConnectionPool.return_connection (syntax error)
- Fixed get_connection to wait with timeout instead of recursive call or unlimited creation
- Changed all conn.close() to return_connection() for pool-managed connections
- Affected: advisory locks, init_db, salvar_mensagem, cleanup, stats
- Added finally block to _init_context_isolation_tables for connection cleanup
fix(memory): conversation_id em saves, fallback por numero, pool de conexoes PG
Isaac Quarenta
6e1619d
fix: adiciona import threading faltante no database_pg.py
Isaac Quarenta
c850c45
feat: otimizacoes de contexto, prompts no DB e connection pooling
Isaac Quarenta
adcd7ae
- THINKING DECAY v4: Corrigida deteccao de queries isoladas para respostas curtas
- InfoSoftEdge: Sistema de armazenamento de prompts no PostgreSQL
- TokenEstimator: Estimativa precisa de tokens para portugues
- SmartCache: Cache LRU com invalidacao por TTL
- ConnectionPool: Pool de conexoes PostgreSQL
- Smart Truncation: Truncagem inteligente baseada em tokens
- Relevance Scoring: Scoring de relevancia para mensagens STM
- Memory Monitor: Monitoramento de memoria e contexto compacto
- Background Tasks: Executor para tarefas em background
- Enhanced LSTM: Contexto de problemas anteriores e resumo
fix: _execute_simple creates own connection without RealDictCursor to bypass psycopg2 bug
Isaac Quarenta
83cd2ac
fix: _execute_simple uses normal cursor to bypass psycopg2 RealDictCursor bug
Isaac Quarenta
e17c7f1
fix: brand_identity uses image_data from HF + full traceback for SQL error diagnosis
Isaac Quarenta
d3a2b37
fix: brand_identity returns media_response format for agent loop + SQL error shows query params
Isaac Quarenta
2ca844a
fix: recuperar_resposta_por_id adds modelo_usado and created_at to PG SELECT (matches SQLite schema)
Isaac Quarenta
832e882
feat: remove ALL dedup/anti-loop logic from Python
Isaac Quarenta
7020ecf
Removed:
- recently_processed dict + dedup_timeout in lstm_memory_system.py
- _detect_loop() + loop detection call in bot_registry.py
- is_duplicate_content_hash() now returns False (database.py + database_pg.py)
- check_idempotency() now returns False (database.py + database_pg.py)
- LSTM idempotency check in lstm_extension.py
Messages now flow through without silent drops.
fix: create akira_timers table in _init_db - was missing causing 500 errors
Antigravity Assistant
ea9d17f
fix: DORA training pipeline was broken - 3 critical bugs
Antigravity Assistant
5f3cff2
BUG 1: contar_exemplos_treino() defined twice with different param names
- First def used 'quality_min', second used 'min_quality'
- Python overwrites first with second, callers got TypeError silently
- Fixed: removed duplicate, all callers now use min_quality
BUG 2: Fallback SQL query incompatible with RealDictCursor
- SELECT COUNT(*) returns dict {'count': N}, not tuple (N,)
- Code accessed result[0][0] → KeyError → silently set total_ft=0
- Fixed: use result[0]['count'] with proper alias
BUG 3: except: pass blocks hiding all errors silently
- _purificar_e_segmentar_dataset() exceptions swallowed
- store_training_example() failures invisible (debug level)
- Fixed: added fallback to salvar_exemplo_treino() without embeddings
- Fixed: all bare except blocks now log at warning level
Additional fixes:
- Anti-explanation rule in system prompt (config.py)
- ThinkingEngine rule 11 prevents suggesting unprompted explanations
- Bot will use kota/puto/bué naturally without explaining meanin…
FIX: Listen Engine filtering, training pipeline, advisory lock, PersonaTracker
Antigravity Assistant
246a71f
- Listen Engine: classify group messages in /akira, reject context-only
(non-bot-directed) messages before processing. Store context in
ContextoGrupoManager from both /escutar and /akira endpoints.
- Training pipeline: remove advisory lock requirement from _run_loop()
(worker runs in separate process, lock always timed out)
- Advisory lock: change from wait_for_advisory_lock(15s+10s) to
non-blocking acquire_advisory_lock(). Per-conversation semaphore+queue
already handles concurrency.
- PersonaTracker: add defensive getattr/None checks for stm_manager
- unified_context.py: add missing factory functions (get_stm_manager,
get_unified_context_builder, build_unified_context)
- api.py: fix factory function indentation (were nested inside class)
fix: database_pg _init_db indentation + autocommit; treinamento_worker usa get_database()
Antigravity Assistant
c1b4fbc
Fix PG transaction abort: isolate each DDL with explicit rollback
Antigravity Assistant
8e092cf
- _init_db: wrap each DDL in exec_ddl() with explicit rollback on error
- _init_context_isolation_tables: same pattern
- _ensure_all_columns_and_indexes: same pattern
- Fixes 'current transaction is aborted, commands ignored' by ensuring
each DDL runs in isolated transaction with explicit rollback on error
fix: 8 bugs encontrados na auditoria final — DB, embeddings, NameError, SQL injection
Antigravity Assistant
427b7ed
database_pg.py:
- UNIQUE constraints em aprendizados, girias_aprendidas, dedup_messages
(upserts estavam falhando silenciosamente sem constraint)
- atualizar_persona: connection leak + SQL injection via whitelist
- recuperar_embeddings: isinstance dead code (psycopg2.extensions.adapt e funcao)
- Migracao CREATE UNIQUE INDEX IF NOT EXISTS para DBs existentes
treinamento.py:
- analise_emocao NameError quando emotion_trainer e None
api.py:
- _save_response_embedding_async usava SQLite em vez de PG
- vocabulario_autonomo.processar_mensagem duplicado removido
local_llm.py:
- FALLBACK_SYSTEM_PROMPT undefined removido do __all__
fix: inject mensagem_citada into ThinkingEngine historico_recente para contexto de reply
Antigravity Assistant
17f77b4
Quando o usuário respondia em reply (ex: 'como assim protocolo?'),
a ThinkingEngine recebia historico_recente vazio (devido a lock timeout
ou isolamento strict de 2 msgs), gerava CoT genérico sobre 'protocolo'
e a SUGESTAO_RESPOSTA imperativa fazia a Mistral ignorar a mensagem
citada que continha 'Hermes'.
Agora a mensagem citada é injetada como entry extra no historico_recente
antes da chamada _te.think(), garantindo que o bloco STM no CoT inclua
o contexto original.
ThinkingEngine: incluir SUGESTAO_RESPOSTA + coaching imperativo (para de ignorar contexto)
Antigravity Assistant
62bac87
fix: lock timeout 30s + training worker + PostgreSQL setup