src/config/runtime.ts

/**
 * src/config/runtime.ts — RuntimeConfig
 * Source of truth unica per tutti i default runtime, limiti e feature flags.
 *
 * NON duplicare questi valori in altri file.
 * Importa sempre da qui, mai hardcodarli altrove.
 */

import { ENV } from "./env";

// ─── Endpoints ──────────────────────────────────────────────────────────────
export const ENDPOINTS = {
  HF_ROUTER:        "https://router.huggingface.co/v1/chat/completions",
  HF_INFERENCE:     "https://api-inference.huggingface.co/models",
  GROQ:             "https://api.groq.com/openai/v1/chat/completions",
  OPENROUTER:       "https://openrouter.ai/api/v1/chat/completions",
  OPENAI:           "https://api.openai.com/v1/chat/completions",
  TOGETHER:         "https://api.together.xyz/v1/chat/completions",
  GEMINI_BASE:      "https://generativelanguage.googleapis.com",
  LOCAL_BACKEND:    ENV.BACKEND_URL || "http://localhost:8000",
  // S325: Cerebras — 2000+ tok/s su Llama 3.3 70B, 1M tok/giorno gratis, free tier
  CEREBRAS:         "https://api.cerebras.ai/v1/chat/completions",
  // S408: SambaNova — 2200+ tok/s, OpenAI-compatible, free tier (cloud.sambanova.ai)
  // S-LOOP9: modelli verificati live 2026-06-12 -- vedi SAMBANOVA_MODELS
  SAMBANOVA:        "https://api.sambanova.ai/v1/chat/completions",
  // S408: Cloudflare Workers AI — gratis 10k req/giorno, non serve account separato
  // URL dinamico: dipende da CF_ACCOUNT_ID — costruito in cfAiEndpoint()
  CF_AI_BASE:       "https://api.cloudflare.com/client/v4/accounts",
} as const;

// ─── CF Workers AI endpoint builder ──────────────────────────────────────────
// Costruisce l'URL per un modello Cloudflare Workers AI dato l'account ID.
export function cfAiEndpoint(accountId: string, model: string): string {
  return `${ENDPOINTS.CF_AI_BASE}/${accountId}/ai/run/${model}`;
}

// ─── Model defaults ──────────────────────────────────────────────────────────
export const DEFAULT_MODELS = {
  HF_ROUTER:    "deepseek-ai/DeepSeek-V3-0324",
  GROQ:         "llama-3.3-70b-versatile",
  OPENROUTER:   "openai/gpt-oss-20b:free",
  OPENAI:       "gpt-4o-mini",
  GEMINI:       "gemini-2.5-flash",
  // S325: Cerebras — gpt-oss-120b (llama-3.3-70b rimosso da Cerebras a giugno 2026)
  CEREBRAS:     "gpt-oss-120b",
  // S408: SambaNova — modelli free tier (giugno 2026)
  SAMBANOVA:    "Meta-Llama-3.3-70B-Instruct",
  // S408: CF Workers AI — fp8-fast = versione ottimizzata per latenza
  CF_AI:        "@cf/meta/llama-3.3-70b-instruct-fp8-fast",
} as const;

// ─── SambaNova model list (free tier, giugno 2026) ───────────────────────────
// S-LOOP9: verificato live 2026-06-12 -- specchio di SAMBANOVA_MODELS_FAST in providerConstants.ts
// Rimossi (non piu in /v1/models): Llama-4-Scout-17B-16E, Meta-Llama-3.1-8B-Instruct, DeepSeek-R1-671B
export const SAMBANOVA_MODELS = [
  "DeepSeek-V3.1",               // 131k ctx, best quality
  "Meta-Llama-3.3-70B-Instruct", // 131k ctx, 2200+ tok/s -- default
  "MiniMax-M2.7",                // 196k ctx -- contesti lunghi
  "gemma-4-31B-it",              // 131k ctx, Gemma 4 31B
  "gpt-oss-120b",                // 131k ctx, GPT-class quality
  "DeepSeek-V3.2",               // 32k ctx, versione alternativa
] as const;

// ─── CF Workers AI model list (free tier) ────────────────────────────────────
export const CF_AI_MODELS = [
  "@cf/meta/llama-3.3-70b-instruct-fp8-fast",         // 70B fast -- massima qualita CF
  "@cf/deepseek-ai/deepseek-r1-distill-llama-70b",     // reasoning 70B con <think>
  "@cf/deepseek-ai/deepseek-r1-distill-qwen-32b",      // reasoning 32B alternativo
  "@cf/meta/llama-3.1-8b-instruct",                    // 8B fallback (ancora disponibile)
] as const;

// ─── Token / request limits ──────────────────────────────────────────────────
export const LIMITS = {
  MAX_TOKENS_DEFAULT:     4096,
  MAX_TOKENS_AGENT:       8192,
  MAX_TOKENS_SERVERLESS:  2048,   // HF Inference API — modelli piccoli
  TEMPERATURE_DEFAULT:    0.7,
  TEMPERATURE_AGENT:      0.5,
  TIMEOUT_FAST_MS:        10_000, // Groq, Cerebras, SambaNova — provider veloci
  TIMEOUT_DEFAULT_MS:     22_000, // HF Router, OpenRouter
  TIMEOUT_SLOW_MS:        45_000, // Gemini, backend locale
  TIMEOUT_AGENT_MS:       120_000,
  LOCAL_BACKEND_CHECK_TTL_MS: 30_000,
  PROVIDER_RETRY_AFTER_MS:    20_000,
} as const;

// ─── Provider chain order (usata dal fallback automatico) ────────────────────
// Ordine: più affidabile → meno affidabile per uso generale.
export const PROVIDER_CHAIN_ORDER = [
  "gemini",
  "deepseek-hf",
  "deepseek-openrouter",
  "llama-groq",
  "llama-sambanova",
  "llama-openrouter",
  "hf-router",
  "hf-serverless",
  "openai",
  "together",
  "browser-llm",
] as const;

// ─── Feature flags ───────────────────────────────────────────────────────────
export const FEATURES = {
  // WebGPU/WebLLM: carica modelli nel browser (disabilitato di default — crash su iOS)
  BROWSER_LLM_AUTO_PRELOAD: false,
  // Virtualizzazione lista messaggi (riduce DOM su Safari)
  VIRTUALIZE_MESSAGES: true,
  // Backend locale (Ollama) — check automatico all'avvio
  LOCAL_BACKEND_CHECK: !!ENV.BACKEND_URL,
  // Diagnostics panel in dev
  DIAGNOSTICS_PANEL: import.meta.env.DEV,
} as const;

// ─── OpenRouter branding headers ─────────────────────────────────────────────
export const OPENROUTER_HEADERS = {
  "HTTP-Referer": "https://baida98.github.io",
  "X-Title":      "Agente AI",
} as const;