/**
 * Schemastery config schema for the cortiq-router plugin.
 *
 * This schema is the settings form shown in the DSH web UI (Settings → Models →
 * Cortiq Router), so every field description is the UI's explanation of it.
 *
 * The form reads top-down as the story of one request: connect to the router,
 * decide what it classifies, then say which model serves each answer.
 */

import z from '@deepseek-ai/schemastery';
import { LOCALES, type Locale } from './i18n.js';

/** The provider route id used in DSH's LLM seam. */
export const PROVIDER = 'cortiq-router';

/** Display name shown in model pickers and settings tabs. */
export const DISPLAY_NAME = 'Cortiq Router';

/** Settings namespace — the plugin's config lives under `llm-cortiq-router` in settings.yaml. */
export const SETTINGS_NS = 'llm-cortiq-router';

/** Where to get a key, repeated in the one field that needs it. */
export const SIGNUP_URL = 'https://api.allaigate.com/';

/** The hosted classifier endpoint. */
export const DEFAULT_ROUTER_URL = 'https://router.allaigate.com';

/**
 * Entry points offered in the model picker. Each one routes; the three
 * profile-pinned ids override the configured policy profile for that call, so
 * "cheap for this chat, careful for that one" needs no settings trip.
 */
export const ENTRY_PROFILES: Record<string, string | undefined> = {
  auto: undefined,
  'auto-cost-saver': 'cost-saver',
  'auto-balanced': 'balanced',
  'auto-quality-first': 'quality-first',
};

// ── Public config types ────────────────────────────────────────────────
//
// Written out rather than inferred from the schema: the emitted declarations
// then name only this package's own types, and the settings section, the
// resolver, and DEFAULT_CONFIG all agree on one shape.

/** Per-tier model lists for one task label. */
export interface TaskRuleTiers {
  low: string[];
  medium: string[];
  high: string[];
  any: string[];
}

/** One task rule: a single model, a chain, or per-tier chains. */
export type TaskRuleConfig = string | string[] | TaskRuleTiers;

/** One entry offered in the model picker. */
export interface CatalogModelConfig {
  id: string;
  /** Overrides this plugin's own localized label; absent keeps the shipped one. */
  name?: string;
  contextWindow: number;
  inputModalities: ('text' | 'image')[];
}

/** Every setting this plugin owns, fully populated. */
export interface CortiqRouterSettings {
  language: 'auto' | Locale;
  apiKeyEnv: string;
  routerUrl: string;
  timeoutMs: number;
  taxonomyId: string;
  verifyTls: boolean;
  routerProfile: 'cost-saver' | 'balanced' | 'quality-first';
  textStrategy: 'last_user' | 'last_user_plus_system' | 'concat_all';
  maxChars: number;
  complexityBands: { low: number; medium: number };
  globalTiers: { low: string[]; medium: string[]; high: string[] };
  taskRules: Record<string, TaskRuleConfig>;
  defaultModel: string;
  echoRouting: boolean;
  models: CatalogModelConfig[];
}

/** What a composition entry or a settings snapshot may actually carry. */
export type ConfigType = Partial<CortiqRouterSettings>;

// ── Per-model entry in the catalog ─────────────────────────────────────
const CatalogModel = z.object({
  id: z.string().required().description('Id shown in the model picker (e.g. auto)'),
  name: z.string().description('Label shown in the picker; leave empty to use this plugin\'s own translated label'),
  contextWindow: z.number().step(1).min(1).default(128_000)
    .description('Context window advertised to the agent, in tokens. Keep it at or below the smallest routed model.'),
  inputModalities: z.array(z.union(['text', 'image'])).min(1).default(['text'])
    .description('Modalities every routed model must accept'),
});

// ── Task rule schema ───────────────────────────────────────────────────
/**
 * A task rule: either a simple array of models, or a per-tier map.
 *
 * In the UI this renders as a card per task with three tier model pickers
 * (low / medium / high) + an "any" fallback.
 *
 * Storage form:
 *   translation: ["model-a", "model-b"]            ← any-tier
 *   code:
 *     low:  ["fast-model"]
 *     high: ["powerful-model"]
 *     any:  ["fallback-model"]
 */
const TaskRuleValue = z.union([
  z.array(z.string()),
  z.string(),
  z.object({
    low: z.array(z.string()).default([]).description('Model(s) for simple requests'),
    medium: z.array(z.string()).default([]).description('Model(s) for medium complexity'),
    high: z.array(z.string()).default([]).description('Model(s) for hard requests'),
    any: z.array(z.string()).default([]).description('Fallback model(s) at any tier'),
  }),
]);

// ── Main config schema ─────────────────────────────────────────────────
const ConfigSchema = z.object({
  // ── Language ────────────────────────────────────────────
  language: z.union(['auto', ...LOCALES])
    .default('auto')
    .description('Language of this plugin\'s own copy — the model picker entries and the message a refused call shows. "auto" follows the harness UI language, which offers English and Chinese; name one of the other locales here to use it.'),

  // ── Connection ──────────────────────────────────────────
  apiKeyEnv: z.string()
    .role('credential-ref')
    .default('CORTIQ_ROUTER_KEY')
    .description(`Credential holding your allaigate API key (cortiq_…). Get one at ${SIGNUP_URL}. Without a key nothing breaks — every call goes straight to the fallback model below.`),

  routerUrl: z.string()
    .default(DEFAULT_ROUTER_URL)
    .description('Classifier endpoint. Point this at your own deployment to keep prompts in-house.'),

  timeoutMs: z.number().min(1000).max(60_000).default(15_000)
    .description('How long to wait for a routing decision, in ms. Hard prompts escalate to an oracle (~10s), so keep this ≥ 12000; on timeout the call still runs, on the fallback model.'),

  taxonomyId: z.string()
    .default('data-assistant')
    .description('Task taxonomy the router classifies against. Change only if your account has a custom one.'),

  verifyTls: z.boolean()
    .default(true)
    .description('Verify the classifier\'s TLS certificate. Node\'s fetch cannot skip verification, so turning this off only logs a warning; a self-hosted endpoint with a private CA is trusted by pointing NODE_EXTRA_CA_CERTS at that CA, and one with a self-signed certificate by giving it a real one (Let\'s Encrypt is free).'),

  // ── What gets classified ────────────────────────────────
  routerProfile: z.union(['cost-saver', 'balanced', 'quality-first'])
    .default('balanced')
    .description('Default policy profile. Picking the auto-cost-saver / auto-balanced / auto-quality-first model overrides it for that conversation.'),

  textStrategy: z.union(['last_user', 'last_user_plus_system', 'concat_all'])
    .default('last_user')
    .description('Which part of the conversation is sent to the classifier. "last_user" routes best and leaks least; add the system prompt when it carries the task.'),

  maxChars: z.number().min(100).max(32_000).default(4_000)
    .description('Hard cap on the characters sent to the router'),

  complexityBands: z.object({
    low: z.number().min(0.01).max(0.99).default(0.35)
      .description('Score ≤ this → "low" complexity'),
    medium: z.number().min(0.02).max(0.99).default(0.65)
      .description('Score ≤ this → "medium"; above it → "high"'),
  })
    .default({ low: 0.35, medium: 0.65 })
    .description('Where your low / medium / high tiers sit on the router\'s 0…1 complexity score. Raise them to send more work to cheap models.'),

  // ── Which model answers ─────────────────────────────────
  globalTiers: z.object({
    low: z.array(z.string()).default(['deepseek-v4-flash', 'deepseek-v3-chat'])
      .description('Simple requests — fast and cheap'),
    medium: z.array(z.string()).default(['deepseek-v4-pro', 'deepseek-v4-flash'])
      .description('Medium complexity'),
    high: z.array(z.string()).default(['deepseek-v4-pro'])
      .description('Hard requests — the capable, expensive models'),
  }).description('Model chain per complexity tier. Each list is tried in order, so put the preferred model first; ids come from your other provider plugins ("deepseek-v4-pro"), or name a route explicitly as "provider:model".'),

  taskRules: z.dict(TaskRuleValue)
    .default({
      code: {
        low: ['deepseek-v3-chat'],
        medium: [],
        high: ['deepseek-v4-pro'],
        any: ['deepseek-v4-flash'],
      },
      math: {
        low: ['deepseek-v3-chat'],
        medium: [],
        high: ['deepseek-v4-pro'],
        any: [],
      },
      translation: ['deepseek-v3-chat'],
      'creative-writing': {
        low: ['deepseek-v4-flash'],
        medium: [],
        high: ['deepseek-v4-pro'],
        any: [],
      },
      summarization: ['deepseek-v4-flash'],
      extraction: ['deepseek-v3-chat'],
      qa: ['deepseek-v4-flash', 'deepseek-v3-chat'],
      chitchat: ['deepseek-v3-chat'],
    })
    .description('Per-task overrides, applied before the tier chains above: which model serves each task label at each complexity.'),

  defaultModel: z.string()
    .default('deepseek-v4-flash')
    .description('Last resort: the model used when the router is unreachable, unpaid, or nothing else in the chain is registered.'),

  // ── Diagnostics ────────────────────────────────────────
  echoRouting: z.boolean()
    .default(true)
    .description('Log every routing decision (task, tier, score, chosen provider/model)'),

  // ── Model picker entries ────────────────────────────────
  // An empty name means "use this plugin's own translated label".
  models: z.array(CatalogModel).default([
    { id: 'auto', name: '', contextWindow: 128_000, inputModalities: ['text'] },
    { id: 'auto-cost-saver', name: '', contextWindow: 128_000, inputModalities: ['text'] },
    { id: 'auto-balanced', name: '', contextWindow: 128_000, inputModalities: ['text'] },
    { id: 'auto-quality-first', name: '', contextWindow: 128_000, inputModalities: ['text'] },
  ]).description('Entries this provider offers in the model picker. Every one of them routes — the three profile-pinned ones just override the policy profile for that conversation.'),
});

/** The settings form, as the web UI renders it. */
export const Config = ConfigSchema as unknown as Schemastery<ConfigType>;

/**
 * Default config as a plain object (for programmatic use without Schemastery).
 * Matches the defaults in the schema above.
 */
export const DEFAULT_CONFIG: CortiqRouterSettings = {
  language: 'auto',
  apiKeyEnv: 'CORTIQ_ROUTER_KEY',
  routerUrl: DEFAULT_ROUTER_URL,
  timeoutMs: 15_000,
  taxonomyId: 'data-assistant',
  verifyTls: true,
  routerProfile: 'balanced',
  textStrategy: 'last_user',
  maxChars: 4_000,
  complexityBands: { low: 0.35, medium: 0.65 },
  globalTiers: {
    low: ['deepseek-v4-flash', 'deepseek-v3-chat'],
    medium: ['deepseek-v4-pro', 'deepseek-v4-flash'],
    high: ['deepseek-v4-pro'],
  },
  taskRules: {
    code: { low: ['deepseek-v3-chat'], medium: [], high: ['deepseek-v4-pro'], any: ['deepseek-v4-flash'] },
    math: { low: ['deepseek-v3-chat'], medium: [], high: ['deepseek-v4-pro'], any: [] },
    translation: ['deepseek-v3-chat'],
    'creative-writing': { low: ['deepseek-v4-flash'], medium: [], high: ['deepseek-v4-pro'], any: [] },
    summarization: ['deepseek-v4-flash'],
    extraction: ['deepseek-v3-chat'],
    qa: ['deepseek-v4-flash', 'deepseek-v3-chat'],
    chitchat: ['deepseek-v3-chat'],
  },
  defaultModel: 'deepseek-v4-flash',
  echoRouting: true,
  models: [
    { id: 'auto', contextWindow: 128_000, inputModalities: ['text'] },
    { id: 'auto-cost-saver', contextWindow: 128_000, inputModalities: ['text'] },
    { id: 'auto-balanced', contextWindow: 128_000, inputModalities: ['text'] },
    { id: 'auto-quality-first', contextWindow: 128_000, inputModalities: ['text'] },
  ],
};