Commit History
-
fix(chat+hf): fluidez conversacional sem busca + HF tier com quota real por conta
Isaac Quarenta- web_search: e_mensagem_conversacional() bloqueia saudacao/interjeicao/piada na web; identidade do bot alargada (idade/nome/capacidades 2a pessoa); factuais de clima/previsao adicionados ao classifier - api: gate de busca com bloqueio previo (identidade/social), rewrite com contexto e sentinela NAO_PESQUISAR, query vazia nao pesquisa; bloco [FLUIDEZ CONVERSACIONAL] no system prompt (referentes, sem eco, sem parecer buscador) - cellcog: classificacao de erros (rate_limit/payment/gated/coverage/transient); 402 por modelo em cache 10min sem abortar a lista; conta marcada so com erro real de TODOS os modelos; Pollinations so apos esgotar tentativas - hf_inference_rotation: mark_account_limited(conta) — antes marcava a conta errada (indice da rotacao) e podia esvaziar o pool preventivamente5fb9e9a -
fix(chat+images): schnell primeiro no HF + interjeicoes/identidade sem web search
Isaac Quarenta02995ae -
Prioritize HF FLUX for image generation
Isaac Quarenta +1: CopilotCo-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>d841dbe -
fix(images): HF tier usa InferenceClient oficial (router+nscale dava 400 em tudo e caia sempre no Pollinations) - FLUX.1-dev primeiro
Isaac Quarentabdf1c54
-
fix(images): HF Inference com modelos sem gate + diagnostico de licenca gated-403
Isaac Quarenta6e7884a
-
fix(reply+lstm+image): LID reply ativa Akira, STM attribution, HF top antes Flux
Isaac Quarenta- listen_engine: is_quoted_from_bot com LID strip + is_reply_to_bot_hint, honra reply_to_bot do Node - api escutar: override reply_to_bot True -> requer_resposta True, observed_only=False - api reply isolation: 15 msgs p/ retry imagem, desativa is_isolated quando reply_to_bot - cellcog: HF FLUX.2-klein-9B TIER1 antes Pollinations - thinking: REGRA 17 vc=bot desambiguacao -> Vou gerar de novo8e00552
-
fix: HF Image Gen - Router API em vez de InferenceClient (deprecated)
Isaac Quarenta- HF Router: https://router.huggingface.co/{provider}/{model} - Provider: nscale (free tier para FLUX/SDXL/SD3) - Modelo mais potente adicionado: FLUX.2 Klein 9B - Ordem de prioridade: Klein 9B > FLUX.1 dev > schnell > SD3.5 > SDXL - InferenceClient (huggingface_hub) substituído por requests direto ao Router - Todos os tokens HF rotation testados com fallback153e1b1 -
fix: image gen chain - Pollinations primeiro (free, sem auth, sem moderation)
Isaac QuarentaReordenado de: HF → CellCog → Google → Cloudflare → Stability → Pollinations Para: Pollinations (TIER 1) → HF (TIER 2) → CellCog (TIER 3) → Cloudflare (TIER 4) → Stability (TIER 5) Rationale: - HF FLUX 402 em todas 3 contas (ann/hf_main/isaac) — sem créditos mensais - Google Imagen: nenhum modelo disponível no plano free - Cloudflare: SSL EOF constante - Stability: content_moderation bloqueia 'inferno', 'fogo' etc - Pollinations: free, no auth, modelos Flux/SDXL/Juggernaut/Turbo sem moderation filter REMOVIDO Google Imagen (nunca disponível no plano free)0fb38e7
-
fix: respect THINKING_COT_TIMEOUT + retry tier 2-3 image providers
Isaac Quarenta1. thinking_engine.py:855 — clamp OpenRouter CoT timeout to THINKING_COT_TIMEOUT instead of hardcoded 45s, allowing config-driven cap. 2. cellcog_integration.py:1023 _try_provider — accepts retries param (default 1) for transient 429/timeout on tiers 2-3 (CellCog/HF). Pollinations fallback left unchanged (last tier, no retry). Driven by subagent analysis 2026-07-28.22eecd9
-
fix: image gen provider order - HF Inference first, Pollinations last
Isaac Quarenta0c8e0ad
-
fix: move Pollinations Flux to TIER 1 (only working image provider)
Isaac QuarentaAll paid providers (CellCog, HF, Google, Cloudflare, Stability) are depleted/down. Pollinations Flux is the only working provider. Moving it first saves 30-60s per image.9120847 -
fix: Pollinations back to base64 (BotCore was failing on URL download due to ESM require)
Isaac Quarentab2db1e6 -
fix: Pollinations returns URL instead of base64 for image delivery
Isaac Quarenta- Pollinations fallback now returns the direct URL instead of downloading and encoding to base64 (~1.4MB of text for 1024x1024 image) - BotCore already handles URLs in handleMediaResponse (line 2335) - Much faster delivery, less RAM usage, avoids JSON size issues - Updated skill_library comment to reflect URL/base64 flexibility6ad1c96
-
fix: move Stability AI before Pollinations in image gen chain, increase Pollinations timeout to 90s
Isaac Quarentaafeac99
-
fix(quality): Pollinations flux-realism before Stability, stable-image-ultra fallback, CLOUDFALRE_API_KEY typo fix
Isaac Quarentad3b8ccb -
feat(logo): overlay text with PIL instead of relying on AI rendering
Isaac QuarentaAI image models cannot reliably render specific text (SoftEdge, slogans). Now generate_brand_identity: 1. Generates the visual logo WITHOUT text (prompt says 'no text') 2. Overlays brand name + slogan using PIL with proper font rendering, shadow for readability, text wrapping 3. Works for all 3 fallback paths (CellCog, HF, Factory chain) This eliminates AI text artifacts completely.b3571b6 -
fix(image): reduce AI artifacts and reorder factory for best quality
Isaac Quarenta- Stability: added negative_prompt + cfg_scale=7 to reduce deformed/blurry/artifact-heavy outputs - Cloudflare: added guidance_scale=7.5 for better prompt adherence - Factory: reordered Cloudflare (FLUX.1-schnell) before Stability (SD3.5-large) since FLUX has fewer AI artifactsdbf5c17 -
fix(brand): add media factory fallback to generate_brand_identity
Isaac Quarentagenerate_brand_identity so tentava CellCog -> HF Image Gen, mas nao usava a factory chain (Stability/Cloudflare/Pollinations). Agora quando HF falha (402), cai na media factory como ultimo recurso antes de retornar erro.af523fc -
feat(image-gen): add Stability AI + Cloudflare providers with rotation, promote Google Imagen in chain
Isaac Quarenta- New stability_rotation.py: multi-key rotation (STABILITY_API_KEY, _2..._9) - New cloudflare_rotation.py: multi-account rotation (CLOUDFLARE_ACCOUNT_ID/API_TOKEN, _2..._9) - New stability_image_gen.py: SD3.5 with rotation, marks 402/429 as limited - New cloudflare_image_gen.py: SDXL+FLUX+sdxl-lightning via Workers AI - cellcog_integration.py: factory reordered (CellCog->HF->Google->Stability->Cloudflare->Pollinations), unified _try_provider - google_image_gen.py: removed internal Pollinations fallback (factory handles chain) - api.py: FastRouter CoT 402 now logged explicitly, falls to OpenRouter0fb5b45
-
fix: brand_identity uses image_data from HF + full traceback for SQL error diagnosis
Isaac Quarentad3a2b37 -
fix: brand_identity returns media_response format for agent loop + SQL error shows query params
Isaac Quarenta2ca844a -
fix: brand_identity HF fallback uses HuggingFaceImageGen directly instead of CellCogClient.generate
Isaac Quarentab1fecde -
fix: brand_identity fallback to HF image when CellCog fails + remove short-circuit
Isaac Quarenta07e9acb
-
fix: multi-provider image generation - HF Inference Tier 2, Google Imagen real Tier 4, HF_TOKEN rotation, standardized image_data format
Isaac Quarenta49fc39f
-
fix: timer skill, LSTM speakers DB, persona datetime, image gen fallback
Antigravity Assistant211e460 -
fix: lock timeout 30s + training worker + PostgreSQL setup
Antigravity Assistant51b711f