server/src/db/migrations.ts

import crypto from 'crypto';
import Database from 'better-sqlite3';
import { initEncryptionKey } from '../lib/crypto.js';
import { applyModelPricing } from './model-pricing.js';

export function migrateDbSchema(db: Database.Database) {
  createTables(db);
  initEncryptionKey(db);
  seedModels(db);
  migrateModels(db);
  migrateModelsV2(db);
  migrateModelsV3Ranks(db);
  migrateModelsV4(db);
  migrateModelsV5(db);
  migrateModelsV6(db);
  migrateModelsV7(db);
  migrateModelsV8(db);
  migrateModelsV9(db);
  migrateModelsV10(db);
  migrateModelsV11(db);
  migrateModelsV12(db);
  migrateModelsV13(db);
  migrateModelsV14(db);
  migrateModelsV15(db);
  migrateModelsV16Vision(db);
  migrateModelsV17IntelligenceTiers(db);
  migrateModelsV18OpenCodeZen(db);
  migrateModelsV19Gemma4(db);
  migrateModelsV20KiloFree(db);
  migrateModelsV21PruneDead(db);
  migrateModelsV22Tools(db);
  migrateModelsV23FreeTierAudit(db);
  migrateModelsV24ZenRefresh(db);
  // After all model migrations: add/refresh paid-equivalent pricing
  // (drives the realistic "Est. savings" analytics stat).
  applyModelPricing(db);
  migrateEmbeddingsV1(db);
  ensureUnifiedKey(db);
}

function createTables(db: Database.Database) {
  db.exec(`
    CREATE TABLE IF NOT EXISTS models (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      platform TEXT NOT NULL,
      model_id TEXT NOT NULL,
      display_name TEXT NOT NULL,
      intelligence_rank INTEGER NOT NULL,
      speed_rank INTEGER NOT NULL,
      size_label TEXT NOT NULL DEFAULT '',
      rpm_limit INTEGER,
      rpd_limit INTEGER,
      tpm_limit INTEGER,
      tpd_limit INTEGER,
      monthly_token_budget TEXT NOT NULL DEFAULT '',
      context_window INTEGER,
      enabled INTEGER NOT NULL DEFAULT 1,
      supports_vision INTEGER NOT NULL DEFAULT 0,
      UNIQUE(platform, model_id)
    );

    CREATE TABLE IF NOT EXISTS api_keys (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      platform TEXT NOT NULL,
      label TEXT NOT NULL DEFAULT '',
      encrypted_key TEXT NOT NULL,
      iv TEXT NOT NULL,
      auth_tag TEXT NOT NULL,
      status TEXT NOT NULL DEFAULT 'unknown',
      enabled INTEGER NOT NULL DEFAULT 1,
      created_at TEXT NOT NULL DEFAULT (datetime('now')),
      last_checked_at TEXT
    );

    CREATE TABLE IF NOT EXISTS requests (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      platform TEXT NOT NULL,
      model_id TEXT NOT NULL,
      key_id INTEGER,
      status TEXT NOT NULL,
      input_tokens INTEGER NOT NULL DEFAULT 0,
      output_tokens INTEGER NOT NULL DEFAULT 0,
      latency_ms INTEGER NOT NULL DEFAULT 0,
      error TEXT,
      created_at TEXT NOT NULL DEFAULT (datetime('now'))
    );

    CREATE TABLE IF NOT EXISTS rate_limit_usage (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      platform TEXT NOT NULL,
      model_id TEXT NOT NULL,
      key_id INTEGER NOT NULL,
      kind TEXT NOT NULL CHECK (kind IN ('request', 'tokens')),
      tokens INTEGER NOT NULL DEFAULT 0,
      created_at_ms INTEGER NOT NULL,
      created_at TEXT NOT NULL DEFAULT (datetime('now'))
    );

    CREATE TABLE IF NOT EXISTS rate_limit_cooldowns (
      platform TEXT NOT NULL,
      model_id TEXT NOT NULL,
      key_id INTEGER NOT NULL,
      expires_at_ms INTEGER NOT NULL,
      created_at TEXT NOT NULL DEFAULT (datetime('now')),
      PRIMARY KEY (platform, model_id, key_id)
    );

    CREATE TABLE IF NOT EXISTS fallback_config (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      model_db_id INTEGER NOT NULL REFERENCES models(id),
      priority INTEGER NOT NULL,
      enabled INTEGER NOT NULL DEFAULT 1,
      UNIQUE(model_db_id)
    );

    CREATE TABLE IF NOT EXISTS settings (
      key TEXT PRIMARY KEY,
      value TEXT NOT NULL
    );

    -- Dashboard accounts (email + password) gating the /api/* admin surface (#35).
    CREATE TABLE IF NOT EXISTS users (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      email TEXT NOT NULL UNIQUE,
      password_hash TEXT NOT NULL,
      created_at TEXT NOT NULL DEFAULT (datetime('now'))
    );

    CREATE TABLE IF NOT EXISTS sessions (
      token_hash TEXT PRIMARY KEY,
      user_id INTEGER NOT NULL REFERENCES users(id) ON DELETE CASCADE,
      expires_at_ms INTEGER NOT NULL,
      created_at TEXT NOT NULL DEFAULT (datetime('now'))
    );
    CREATE INDEX IF NOT EXISTS idx_sessions_user ON sessions(user_id);

    CREATE INDEX IF NOT EXISTS idx_requests_created_at ON requests(created_at);
    CREATE INDEX IF NOT EXISTS idx_requests_platform ON requests(platform);
    CREATE INDEX IF NOT EXISTS idx_rate_limit_usage_lookup ON rate_limit_usage(platform, model_id, key_id, kind, created_at_ms);
    CREATE INDEX IF NOT EXISTS idx_rate_limit_cooldowns_expires ON rate_limit_cooldowns(expires_at_ms);
    CREATE INDEX IF NOT EXISTS idx_api_keys_platform ON api_keys(platform);
  `);

  ensureRequestKeyIdColumn(db);
  ensureApiKeysBaseUrlColumn(db);
  ensureModelsKeyIdColumn(db);
  ensureRequestTtfbColumn(db);
  ensureRequestRequestedModelColumn(db);
}

// `requested_model` is the model id the CLIENT pinned in the request body.
// NULL when the request was auto-routed ('auto' or omitted model field).
// requested_model = model_id means the pin was honored; a different model_id
// means rate limits or failures forced a failover to another model.
function ensureRequestRequestedModelColumn(db: Database.Database) {
  const columns = db.prepare('PRAGMA table_info(requests)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'requested_model')) {
    db.prepare('ALTER TABLE requests ADD COLUMN requested_model TEXT').run();
  }
}

// `ttfb_ms` is the time-to-first-byte for streaming responses (ms from dispatch
// to the first chunk). NULL for non-streaming or pre-existing rows. Feeds the
// bandit router's latency axis (server/src/services/scoring.ts).
function ensureRequestTtfbColumn(db: Database.Database) {
  const columns = db.prepare('PRAGMA table_info(requests)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'ttfb_ms')) {
    db.prepare('ALTER TABLE requests ADD COLUMN ttfb_ms INTEGER').run();
  }
}

function ensureRequestKeyIdColumn(db: Database.Database) {
  const columns = db.prepare('PRAGMA table_info(requests)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'key_id')) {
    db.prepare('ALTER TABLE requests ADD COLUMN key_id INTEGER').run();
  }
  db.prepare('CREATE INDEX IF NOT EXISTS idx_requests_key_id ON requests(key_id)').run();
}

// `base_url` is the upstream endpoint for the user-configured 'custom' provider
// (#117). NULL for every built-in platform — they use their hardcoded base URL.
function ensureApiKeysBaseUrlColumn(db: Database.Database) {
  const columns = db.prepare('PRAGMA table_info(api_keys)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'base_url')) {
    db.prepare('ALTER TABLE api_keys ADD COLUMN base_url TEXT').run();
  }
}

// `key_id` binds a custom model to the api_keys row that carries ITS endpoint,
// so several custom providers can coexist (#212). NULL for built-in platforms
// (any key of the platform serves any of its models).
function ensureModelsKeyIdColumn(db: Database.Database) {
  const columns = db.prepare('PRAGMA table_info(models)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'key_id')) {
    db.prepare('ALTER TABLE models ADD COLUMN key_id INTEGER').run();
    // Backfill: bind pre-existing custom models to the (single) legacy custom
    // endpoint key so they keep routing to the URL they were created for.
    db.prepare(`
      UPDATE models
         SET key_id = (SELECT id FROM api_keys WHERE platform = 'custom' ORDER BY id LIMIT 1)
       WHERE platform = 'custom' AND key_id IS NULL
    `).run();
  }
}

function seedModels(db: Database.Database) {
  const count = db.prepare('SELECT COUNT(*) as cnt FROM models').get() as { cnt: number };
  if (count.cnt > 0) return;

  const insert = db.prepare(`
    INSERT INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);

  // NOTE: Limits current as of April 2026. See migrateModels() for in-place updates.
  const models = [
    // Google — gemini-2.5-flash free quotas were cut Dec 2025 (now ~20 RPD, budget much lower than before)
    ['google', 'gemini-2.5-pro', 'Gemini 2.5 Pro', 1, 8, 'Frontier', 5, 100, 250000, null, '~12M', 1048576],
    ['google', 'gemini-2.5-flash', 'Gemini 2.5 Flash', 4, 5, 'Large', 10, 20, 250000, null, '~3M', 1048576],
    ['google', 'gemini-2.5-flash-lite', 'Gemini 2.5 Flash-Lite', 8, 3, 'Medium', 15, 1000, 250000, null, '~120M', 1048576],
    // OpenRouter — upgraded DeepSeek R1 -> V3.1 (stronger reasoning); default RPD ~200
    ['openrouter', 'deepseek/deepseek-v3.1:free', 'DeepSeek V3.1 (free)', 2, 10, 'Frontier', 20, 200, null, null, '~6M', 131072],
    ['openrouter', 'moonshotai/kimi-k2:free', 'Kimi K2 (free)', 2, 9, 'Frontier', 20, 200, null, null, '~6M', 131072],
    ['openrouter', 'qwen/qwen3-coder:free', 'Qwen3 Coder (free)', 3, 9, 'Frontier', 20, 200, null, null, '~6M', 262144],
    ['openrouter', 'z-ai/glm-4.5-air:free', 'GLM-4.5 Air (free)', 4, 9, 'Large', 20, 200, null, null, '~6M', 131072],
    // Cerebras — same 30 RPM / 1M TPD free pool; adding frontier coder, Llama 4 Maverick, GPT-OSS
    ['cerebras', 'qwen-3-coder-480b', 'Qwen3-Coder 480B', 2, 1, 'Frontier', 30, null, 60000, 1000000, '~30M', 131072],
    ['cerebras', 'llama-4-maverick-17b-128e-instruct', 'Llama 4 Maverick', 3, 1, 'Frontier', 30, null, 60000, 1000000, '~30M', 131072],
    ['cerebras', 'qwen3-235b', 'Qwen3 235B', 3, 1, 'Large', 30, null, 60000, 1000000, '~30M', 8192],
    ['cerebras', 'gpt-oss-120b', 'GPT-OSS 120B', 3, 1, 'Large', 30, null, 60000, 1000000, '~30M', 131072],
    // GitHub Models — GPT-4o replaced with GPT-5 (same free tier key)
    ['github', 'openai/gpt-5', 'GPT-5 (GitHub)', 1, 7, 'Frontier', 10, 50, null, null, '~18M', 128000],
    // SambaNova — 70B RPM bumped to 20
    ['sambanova', 'Meta-Llama-3.3-70B-Instruct', 'Llama 3.3 70B', 6, 9, 'Large', 20, null, null, 200000, '~6M', 8192],
    // Mistral — Experiment pool ~1B tokens/mo shared across all models
    ['mistral', 'mistral-large-latest', 'Mistral Large 3', 7, 8, 'Large', 2, null, 500000, null, '~50-100M', 131072],
    ['mistral', 'magistral-medium-latest', 'Magistral Medium', 4, 8, 'Large', 2, null, 500000, null, '~50-100M', 40000],
    ['mistral', 'codestral-latest', 'Codestral', 6, 6, 'Medium', 2, null, 500000, null, '~50-100M', 32000],
    // Groq — scout TPM corrected to 6k (not 30k)
    ['groq', 'llama-3.3-70b-versatile', 'Llama 3.3 70B', 9, 2, 'Medium', 30, 1000, 6000, 500000, '~15M', 131072],
    ['groq', 'llama-4-scout-17b-16e-instruct', 'Llama 4 Scout', 10, 2, 'Medium', 30, 1000, 6000, 1000000, '~30M', 131072],
    // NVIDIA NIM — moved to credit-based model in 2025; no longer truly recurring monthly. Disabled by default.
    ['nvidia', 'meta/llama-3.1-70b-instruct', 'Llama 3.1 70B (NV)', 11, 6, 'Large', 40, null, null, null, 'credits-based', 131072],
    // Cohere — trial tier is 1000 calls/mo total → realistic budget 1-2M
    ['cohere', 'command-r-plus-08-2024', 'Command R+ (08-2024)', 12, 11, 'Large', 20, 33, null, null, '~1-2M', 131072],
    ['cloudflare', '@cf/meta/llama-3.1-70b-instruct', 'Llama 3.1 70B (CF)', 13, 11, 'Medium', null, null, null, null, '~18-45M', 131072],
    // Hugging Face — free Inference credits are ~$0.10/mo → budget closer to 1-3M on a 70B model
    ['huggingface', 'accounts/fireworks/models/llama-v3p3-70b-instruct', 'Llama 3.3 70B (HF)', 14, 11, 'Medium', null, null, null, null, '~1-3M', 131072],
    // New providers — recurring monthly free tiers, no card required
    ['zhipu', 'glm-4.5-flash', 'GLM-4.5 Flash', 5, 4, 'Large', null, null, null, 1000000, '~30M', 131072],
    ['moonshot', 'kimi-latest', 'Kimi Latest', 4, 8, 'Large', 60, null, null, 500000, '~15M', 200000],
    ['minimax', 'MiniMax-M1', 'MiniMax M1', 5, 8, 'Large', 20, null, 1000000, null, '~30M', 200000],
  ];

  const insertMany = db.transaction(() => {
    for (const m of models) {
      insert.run(...m);
    }
  });
  insertMany();

  // Seed default fallback config from models
  const allModels = db.prepare('SELECT id, intelligence_rank FROM models ORDER BY intelligence_rank ASC').all() as { id: number; intelligence_rank: number }[];
  const insertFallback = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
  const insertFallbacks = db.transaction(() => {
    for (let i = 0; i < allModels.length; i++) {
      insertFallback.run(allModels[i].id, i + 1);
    }
  });
  insertFallbacks();

  console.log(`Seeded ${models.length} models and fallback config`);
}

/**
 * Idempotent migration to bring existing DBs up to the April 2026 pool.
 * Covers: replaces outdated models (DeepSeek R1 → V3.1, GPT-4o → GPT-5),
 * corrects stale rate-limits / monthly budgets, adds new smarter models
 * and three new providers (Zhipu, Moonshot, MiniMax).
 */
function migrateModels(db: Database.Database) {
  // 1) Replace outdated models in-place (preserves fallback_config & any references)
  const renames: Array<[string, string, string, string, number, string, number | null, number | null, number]> = [
    // platform, oldModelId, newModelId, newDisplayName, intelligenceRank, monthlyBudget, rpdLimit, contextWindow, sizeLabelPriority(unused)
  ];
  const renameStmt = db.prepare(`
    UPDATE models
       SET model_id = ?, display_name = ?, intelligence_rank = ?,
           monthly_token_budget = ?, rpd_limit = COALESCE(?, rpd_limit),
           context_window = COALESCE(?, context_window),
           size_label = COALESCE(?, size_label)
     WHERE platform = ? AND model_id = ?
  `);
  // DeepSeek R1 (free) -> DeepSeek V3.1 (free)
  renameStmt.run('deepseek/deepseek-v3.1:free', 'DeepSeek V3.1 (free)', 2, '~6M', 200, 131072, 'Frontier', 'openrouter', 'deepseek/deepseek-r1:free');
  // GitHub GPT-4o -> GPT-5
  renameStmt.run('openai/gpt-5', 'GPT-5 (GitHub)', 1, '~18M', null, 128000, 'Frontier', 'github', 'gpt-4o');

  // 2) Correct stale limits / budgets on existing rows
  db.prepare(`UPDATE models SET rpd_limit = 20, monthly_token_budget = '~3M' WHERE platform = 'google' AND model_id = 'gemini-2.5-flash'`).run();
  db.prepare(`UPDATE models SET rpm_limit = 20 WHERE platform = 'sambanova' AND model_id = 'Meta-Llama-3.3-70B-Instruct'`).run();
  db.prepare(`UPDATE models SET tpm_limit = 6000 WHERE platform = 'groq' AND model_id = 'llama-4-scout-17b-16e-instruct'`).run();
  db.prepare(`UPDATE models SET monthly_token_budget = '~1-2M' WHERE platform = 'cohere' AND model_id = 'command-r-plus-08-2024'`).run();
  db.prepare(`UPDATE models SET monthly_token_budget = '~1-3M' WHERE platform = 'huggingface' AND model_id = 'accounts/fireworks/models/llama-v3p3-70b-instruct'`).run();
  // NVIDIA moved to credit model — disable and label accordingly
  db.prepare(`UPDATE models SET monthly_token_budget = 'credits-based', enabled = 0 WHERE platform = 'nvidia' AND model_id = 'meta/llama-3.1-70b-instruct'`).run();

  // 3) Insert new models (UNIQUE(platform, model_id) makes this idempotent)
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);

  const newModels: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // Cerebras — same free pool as qwen3-235b
    ['cerebras', 'qwen-3-coder-480b', 'Qwen3-Coder 480B', 2, 1, 'Frontier', 30, null, 60000, 1000000, '~30M', 131072],
    ['cerebras', 'llama-4-maverick-17b-128e-instruct', 'Llama 4 Maverick', 3, 1, 'Frontier', 30, null, 60000, 1000000, '~30M', 131072],
    ['cerebras', 'gpt-oss-120b', 'GPT-OSS 120B', 3, 1, 'Large', 30, null, 60000, 1000000, '~30M', 131072],
    // OpenRouter free tier
    ['openrouter', 'deepseek/deepseek-v3.1:free', 'DeepSeek V3.1 (free)', 2, 10, 'Frontier', 20, 200, null, null, '~6M', 131072],
    ['openrouter', 'moonshotai/kimi-k2:free', 'Kimi K2 (free)', 2, 9, 'Frontier', 20, 200, null, null, '~6M', 131072],
    ['openrouter', 'qwen/qwen3-coder:free', 'Qwen3 Coder (free)', 3, 9, 'Frontier', 20, 200, null, null, '~6M', 262144],
    ['openrouter', 'z-ai/glm-4.5-air:free', 'GLM-4.5 Air (free)', 4, 9, 'Large', 20, 200, null, null, '~6M', 131072],
    // Mistral Experiment pool — shared ~1B/mo across models
    ['mistral', 'magistral-medium-latest', 'Magistral Medium', 4, 8, 'Large', 2, null, 500000, null, '~50-100M', 40000],
    ['mistral', 'codestral-latest', 'Codestral', 6, 6, 'Medium', 2, null, 500000, null, '~50-100M', 32000],
    // New providers
    ['zhipu', 'glm-4.5-flash', 'GLM-4.5 Flash', 5, 4, 'Large', null, null, null, 1000000, '~30M', 131072],
    ['moonshot', 'kimi-latest', 'Kimi Latest', 4, 8, 'Large', 60, null, null, 500000, '~15M', 200000],
    ['minimax', 'MiniMax-M1', 'MiniMax M1', 5, 8, 'Large', 20, null, 1000000, null, '~30M', 200000],
  ];

  const apply = db.transaction(() => {
    for (const m of newModels) insert.run(...m);

    // Ensure every model has a fallback_config row (new inserts + any orphans)
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL
      ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFallback = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) {
        addFallback.run(missing[i].id, maxPriority + i + 1);
      }
    }
  });
  apply();
}

/**
 * Second-pass migration after live-testing every model against its provider.
 * Corrects model IDs verified wrong, removes models not actually available on
 * the current free tier, and adds real :free OpenRouter models found in the
 * live catalog (April 2026).
 */
function migrateModelsV2(db: Database.Database) {
  // Helper: delete a model and its fallback_config entry (FK is RESTRICT-by-default)
  const deleteModel = db.prepare(`DELETE FROM models WHERE platform = ? AND model_id = ?`);
  const deleteFallback = db.prepare(`
    DELETE FROM fallback_config WHERE model_db_id IN (
      SELECT id FROM models WHERE platform = ? AND model_id = ?
    )
  `);
  const removals: Array<[string, string]> = [
    // GitHub free tier does NOT include GPT-5 (only catalog-listed). Revert handled below.
    // Cerebras: qwen-3-coder-480b and llama-4-maverick not on free tier; gpt-oss-120b is listed
    // but requires special access — our key gets 404. Remove all three.
    ['cerebras', 'qwen-3-coder-480b'],
    ['cerebras', 'llama-4-maverick-17b-128e-instruct'],
    ['cerebras', 'gpt-oss-120b'],
    // These OpenRouter :free variants do not exist in the live catalog (April 2026)
    ['openrouter', 'deepseek/deepseek-v3.1:free'],
    ['openrouter', 'moonshotai/kimi-k2:free'],
  ];
  const applyRemovals = db.transaction(() => {
    for (const [p, m] of removals) {
      deleteFallback.run(p, m);
      deleteModel.run(p, m);
    }
  });
  applyRemovals();

  // GitHub: gpt-5 is in the model catalog but returns "unavailable_model" on free tier
  // inference. Revert to gpt-4o which works. This only runs if the gpt-5 row exists.
  db.prepare(`
    UPDATE models
       SET model_id = 'gpt-4o', display_name = 'GPT-4o', intelligence_rank = 5,
           size_label = 'Large', context_window = 8000, monthly_token_budget = '~18M'
     WHERE platform = 'github' AND model_id = 'openai/gpt-5'
  `).run();

  // Groq: scout requires the meta-llama/ publisher prefix
  db.prepare(`
    UPDATE models SET model_id = 'meta-llama/llama-4-scout-17b-16e-instruct'
     WHERE platform = 'groq' AND model_id = 'llama-4-scout-17b-16e-instruct'
  `).run();

  // Add real OpenRouter :free models that exist in the live catalog
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // Frontier-tier free models verified in OR catalog 2026-04
    ['openrouter', 'nvidia/nemotron-3-super-120b-a12b:free', 'Nemotron 3 Super 120B (free)', 2, 9, 'Frontier', 20, 200, null, null, '~6M', 262144],
    ['openrouter', 'qwen/qwen3-next-80b-a3b-instruct:free', 'Qwen3-Next 80B (free)', 3, 9, 'Large', 20, 200, null, null, '~6M', 262144],
    ['openrouter', 'minimax/minimax-m2.5:free', 'MiniMax M2.5 (free)', 3, 9, 'Large', 20, 200, null, null, '~6M', 196608],
    ['openrouter', 'google/gemma-4-31b-it:free', 'Gemma 4 31B (free)', 5, 9, 'Medium', 20, 200, null, null, '~6M', 262144],
  ];
  const applyAdditions = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    // Fallback entries for new models
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  applyAdditions();
}

/**
 * Re-rank intelligence based on April 2026 coding + agentic tool-use benchmarks:
 * SWE-bench Verified, Terminal-Bench 2, TAU-Bench, Aider Polyglot.
 * Higher rank = weaker. Ties are allowed (same weights across providers).
 */
function migrateModelsV3Ranks(db: Database.Database) {
  const setRank = db.prepare(`UPDATE models SET intelligence_rank = ? WHERE platform = ? AND model_id = ?`);
  const ranks: Array<[number, string, string]> = [
    // #1-10 frontier coders / agents
    [1,  'openrouter',  'minimax/minimax-m2.5:free'],                     // SWE-V ~80%, TB2 ~57%
    [2,  'openrouter',  'qwen/qwen3-coder:free'],                         // SWE-V ~70%
    [3,  'openrouter',  'qwen/qwen3-next-80b-a3b-instruct:free'],         // SWE-V ~70.6%
    [4,  'moonshot',    'kimi-latest'],                                   // K2: SWE-V ~71%
    [5,  'cerebras',    'qwen-3-235b-a22b-instruct-2507'],                // SWE-V ~65-72%
    [6,  'google',      'gemini-2.5-pro'],                                // SWE-V 63.8%, Aider 83%
    [7,  'openrouter',  'z-ai/glm-4.5-air:free'],                         // ~58% SWE-V (distill of 4.5)
    [8,  'openrouter',  'openai/gpt-oss-120b:free'],                      // SWE-V 62.4%
    [9,  'openrouter',  'nvidia/nemotron-3-super-120b-a12b:free'],        // SWE-V 53.7%
    [10, 'minimax',     'MiniMax-M1'],                                    // M1 predecessor, ~45-55%
    // #11-15 mid-tier specialists
    [11, 'mistral',     'codestral-latest'],                              // HumanEval 86.6%
    [12, 'mistral',     'mistral-large-latest'],
    [13, 'mistral',     'magistral-medium-latest'],                       // reasoning, not code-tuned
    [14, 'google',      'gemini-2.5-flash'],
    [15, 'zhipu',       'glm-4.5-flash'],
    // #16 Llama 3.3 70B — identical weights across providers (tie)
    [16, 'groq',        'llama-3.3-70b-versatile'],
    [16, 'sambanova',   'Meta-Llama-3.3-70B-Instruct'],
    [16, 'openrouter',  'meta-llama/llama-3.3-70b-instruct:free'],
    [16, 'huggingface', 'accounts/fireworks/models/llama-v3p3-70b-instruct'],
    // #17-23 weaker
    [17, 'openrouter',  'nousresearch/hermes-3-llama-3.1-405b:free'],     // L3.1 base with tool-use tune
    [18, 'groq',        'meta-llama/llama-4-scout-17b-16e-instruct'],     // multimodal focus
    [19, 'openrouter',  'google/gemma-4-31b-it:free'],
    [20, 'google',      'gemini-2.5-flash-lite'],
    [21, 'github',      'gpt-4o'],                                        // Aug 2024, SWE-V ~33%
    [22, 'nvidia',      'meta/llama-3.1-70b-instruct'],                   // older Llama 3.1 tune
    [22, 'cloudflare',  '@cf/meta/llama-3.1-70b-instruct'],               // same base weights
    [23, 'cohere',      'command-r-plus-08-2024'],                        // RAG-focused, weakest on code
  ];
  const apply = db.transaction(() => {
    for (const [rank, platform, modelId] of ranks) {
      setRank.run(rank, platform, modelId);
    }
  });
  apply();
}

/**
 * V4: Agentic-tool-use focus. Live-probed every candidate against real free-tier
 * keys (April 2026) with a weather-tool function-calling test. Keeps only models
 * that return a structured tool_calls response and are reachable on the free tier.
 *
 * Adds SambaNova DeepSeek/Llama-4/gpt-oss, Groq gpt-oss & qwen3-32b, OpenRouter
 * ling-2.6-flash + nemotron-nano + gpt-oss + trinity, Mistral devstral/medium,
 * GitHub gpt-4.1, Cohere command-a, Cloudflare llama-4/gpt-oss/glm-4.7. Removes
 * moonshot/kimi (paid-only now), minimax/M1 (superseded), HF/Fireworks route
 * (no structured tools), OR/gemma-4 (weak at tools). Renames CF llama-3.1 → 3.3
 * fp8-fast. Corrects stale limits.
 */
function migrateModelsV4(db: Database.Database) {
  // 1) Remove entries that are unavailable or fail agentic tool use
  const deleteModel = db.prepare(`DELETE FROM models WHERE platform = ? AND model_id = ?`);
  const deleteFallback = db.prepare(`
    DELETE FROM fallback_config WHERE model_db_id IN (
      SELECT id FROM models WHERE platform = ? AND model_id = ?
    )
  `);
  const removals: Array<[string, string]> = [
    ['moonshot', 'kimi-latest'],                                            // paid-only now ($1 min deposit)
    ['minimax', 'MiniMax-M1'],                                              // superseded; use OR minimax-m2.5:free
    ['openrouter', 'google/gemma-4-31b-it:free'],                           // weak at tool use
    ['huggingface', 'accounts/fireworks/models/llama-v3p3-70b-instruct'],  // emits tool call as text content, not structured
  ];
  const applyRemovals = db.transaction(() => {
    for (const [p, m] of removals) {
      deleteFallback.run(p, m);
      deleteModel.run(p, m);
    }
  });
  applyRemovals();

  // 2) Cloudflare: replace Llama 3.1 70B with the current-gen 3.3 70B fp8-fast
  db.prepare(`
    UPDATE models
       SET model_id = '@cf/meta/llama-3.3-70b-instruct-fp8-fast',
           display_name = 'Llama 3.3 70B fp8-fast (CF)',
           context_window = 131072
     WHERE platform = 'cloudflare' AND model_id = '@cf/meta/llama-3.1-70b-instruct'
  `).run();

  // 3) Field corrections verified via primary sources + live probe
  db.prepare(`UPDATE models SET tpm_limit = 12000 WHERE platform = 'groq' AND model_id = 'llama-3.3-70b-versatile'`).run();
  db.prepare(`UPDATE models SET rpd_limit = 20 WHERE platform = 'sambanova' AND model_id = 'Meta-Llama-3.3-70B-Instruct'`).run();
  db.prepare(`UPDATE models SET rpd_limit = 14400 WHERE platform = 'cerebras' AND model_id = 'qwen-3-235b-a22b-instruct-2507'`).run();
  db.prepare(`UPDATE models SET rpd_limit = 250, monthly_token_budget = '~25M' WHERE platform = 'google' AND model_id = 'gemini-2.5-flash'`).run();
  // gemini-2.5-pro is at-risk: April 2026 Google moved Pro-class off free tier in practice.
  // Our live probe hit "quota exceeded" immediately. Cut rpd in half to reduce 429 blast radius.
  db.prepare(`UPDATE models SET rpd_limit = 50, monthly_token_budget = '~6M' WHERE platform = 'google' AND model_id = 'gemini-2.5-pro'`).run();

  // 4) Add live-probed, tool-capable models
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // OpenRouter :free — shared 20 RPM / 200 RPD / ~6M tokens across :free pool
    ['openrouter', 'inclusionai/ling-2.6-flash:free',        'Ling 2.6 Flash (free)',         7,  9,  'Large',    20, 200, null, null, '~6M', 262144],
    ['openrouter', 'arcee-ai/trinity-large-preview:free',    'Trinity Large Preview (free)',  13, 9,  'Frontier', 20, 200, null, null, '~6M', 131072],
    ['openrouter', 'nvidia/nemotron-3-nano-30b-a3b:free',    'Nemotron 3 Nano 30B (free)',    22, 9,  'Medium',   20, 200, null, null, '~6M', 262144],
    ['openrouter', 'openai/gpt-oss-120b:free',               'GPT-OSS 120B (free)',           6,  9,  'Large',    20, 200, null, null, '~6M', 131072],
    ['openrouter', 'openai/gpt-oss-20b:free',                'GPT-OSS 20B (free)',            18, 9,  'Medium',   20, 200, null, null, '~6M', 131072],
    ['openrouter', 'meta-llama/llama-3.3-70b-instruct:free', 'Llama 3.3 70B (free)',          17, 9,  'Medium',   20, 200, null, null, '~6M', 131072],

    // SambaNova — 20 RPM / 20 RPD / 200K TPD shared free Developer tier
    ['sambanova',  'DeepSeek-V3.1',                          'DeepSeek V3.1',                 5,  9,  'Frontier', 20, 20,  null, 200000, '~3M', 131072],
    ['sambanova',  'DeepSeek-V3.2',                          'DeepSeek V3.2',                 4,  9,  'Frontier', 20, 20,  null, 200000, '~3M', 131072],
    ['sambanova',  'Llama-4-Maverick-17B-128E-Instruct',     'Llama 4 Maverick',              11, 9,  'Large',    20, 20,  null, 200000, '~3M', 8192],
    ['sambanova',  'gpt-oss-120b',                           'GPT-OSS 120B (SambaNova)',      6,  9,  'Large',    20, 20,  null, 200000, '~3M', 131072],

    // Groq — very fast; 30 RPM per model, 1000 RPD on most, 14.4k on the 8B
    ['groq',       'openai/gpt-oss-120b',                    'GPT-OSS 120B (Groq)',           6,  2,  'Large',    30, 1000, 8000, 200000,  '~6M',  131072],
    ['groq',       'openai/gpt-oss-20b',                     'GPT-OSS 20B (Groq)',            18, 2,  'Medium',   30, 1000, 8000, 200000,  '~6M',  131072],
    ['groq',       'qwen/qwen3-32b',                         'Qwen3 32B (Groq)',              19, 2,  'Medium',   60, 1000, 6000, 500000,  '~15M', 131072],
    ['groq',       'llama-3.1-8b-instant',                   'Llama 3.1 8B Instant',          28, 2,  'Small',    30, 14400, 6000, 500000, '~15M', 131072],

    // Mistral Experiment tier — shared 2 RPM / 500k TPM / 1B tokens/mo across all models
    ['mistral',    'devstral-latest',                        'Devstral',                      16, 8,  'Medium',   2, null, 500000, null, '~50-100M', 131072],
    ['mistral',    'mistral-medium-latest',                  'Mistral Medium 3.5',            14, 8,  'Large',    2, null, 500000, null, '~50-100M', 131072],

    // GitHub Models — Low-tier category (15 RPM / 150 RPD, 8K in / 4K out per call)
    ['github',     'openai/gpt-4.1',                         'GPT-4.1 (GitHub)',              20, 7,  'Large',    10, 50,  null, null, '~9M', 128000],

    // Cohere — shared 1000 calls/mo trial pool, 20 RPM Chat
    ['cohere',     'command-a-03-2025',                      'Command-A (03-2025)',           27, 11, 'Large',    20, 33,  null, null, '~1-2M', 131072],

    // Cloudflare Workers AI — shared 10K Neurons/day free pool across all @cf/* models
    ['cloudflare', '@cf/openai/gpt-oss-120b',                'GPT-OSS 120B (CF)',             6,  11, 'Large',    null, null, null, null, '~18-45M', 131072],
    ['cloudflare', '@cf/zai-org/glm-4.7-flash',              'GLM-4.7 Flash (CF)',            10, 11, 'Large',    null, null, null, null, '~18-45M', 131072],
    ['cloudflare', '@cf/meta/llama-4-scout-17b-16e-instruct', 'Llama 4 Scout (CF)',            12, 11, 'Large',    null, null, null, null, '~18-45M', 131072],
  ];

  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();

  // 5) Re-rank the live catalog by agentic tool-use capability (lower = smarter).
  //    Grounded in April 2026 SWE-Bench Verified + BFCL v3 + Tau-Bench numbers.
  const setRank = db.prepare(`UPDATE models SET intelligence_rank = ? WHERE platform = ? AND model_id = ?`);
  const ranks: Array<[number, string, string]> = [
    [1,  'openrouter',  'minimax/minimax-m2.5:free'],
    [2,  'openrouter',  'qwen/qwen3-coder:free'],
    [3,  'openrouter',  'qwen/qwen3-next-80b-a3b-instruct:free'],
    [4,  'sambanova',   'DeepSeek-V3.2'],
    [5,  'sambanova',   'DeepSeek-V3.1'],
    [6,  'cerebras',    'qwen-3-235b-a22b-instruct-2507'],
    [6,  'openrouter',  'openai/gpt-oss-120b:free'],
    [6,  'groq',        'openai/gpt-oss-120b'],
    [6,  'sambanova',   'gpt-oss-120b'],
    [6,  'cloudflare',  '@cf/openai/gpt-oss-120b'],
    [7,  'openrouter',  'inclusionai/ling-2.6-flash:free'],
    [8,  'openrouter',  'z-ai/glm-4.5-air:free'],
    [10, 'cloudflare',  '@cf/zai-org/glm-4.7-flash'],
    [11, 'sambanova',   'Llama-4-Maverick-17B-128E-Instruct'],
    [12, 'groq',        'meta-llama/llama-4-scout-17b-16e-instruct'],
    [12, 'cloudflare',  '@cf/meta/llama-4-scout-17b-16e-instruct'],
    [13, 'openrouter',  'arcee-ai/trinity-large-preview:free'],
    [14, 'google',      'gemini-2.5-pro'],
    [14, 'mistral',     'mistral-large-latest'],
    [14, 'mistral',     'mistral-medium-latest'],
    [16, 'mistral',     'devstral-latest'],
    [16, 'mistral',     'codestral-latest'],
    [17, 'groq',        'llama-3.3-70b-versatile'],
    [17, 'sambanova',   'Meta-Llama-3.3-70B-Instruct'],
    [17, 'cloudflare',  '@cf/meta/llama-3.3-70b-instruct-fp8-fast'],
    [17, 'openrouter',  'meta-llama/llama-3.3-70b-instruct:free'],
    [17, 'nvidia',      'meta/llama-3.1-70b-instruct'],
    [18, 'openrouter',  'openai/gpt-oss-20b:free'],
    [18, 'groq',        'openai/gpt-oss-20b'],
    [19, 'groq',        'qwen/qwen3-32b'],
    [20, 'google',      'gemini-2.5-flash'],
    [20, 'github',      'openai/gpt-4.1'],
    [21, 'mistral',     'magistral-medium-latest'],
    [22, 'openrouter',  'nvidia/nemotron-3-super-120b-a12b:free'],
    [23, 'openrouter',  'nvidia/nemotron-3-nano-30b-a3b:free'],
    [24, 'zhipu',       'glm-4.5-flash'],
    [25, 'github',      'gpt-4o'],
    [26, 'google',      'gemini-2.5-flash-lite'],
    [27, 'cohere',      'command-a-03-2025'],
    [27, 'cohere',      'command-r-plus-08-2024'],
    [28, 'groq',        'llama-3.1-8b-instant'],
  ];
  const applyRanks = db.transaction(() => {
    for (const [r, p, m] of ranks) setRank.run(r, p, m);
  });
  applyRanks();
}

/**
 * V5: Google moved all Pro-tier Gemini off the free tier on 2026-04-01 — disable
 * gemini-2.5-pro. Add Cerebras `zai-glm-4.7` (355B z.ai GLM preview, newly on
 * free tier but throttled to 10 RPM / 100 RPD due to high demand; context capped
 * at 8192 on free tier).
 */
function migrateModelsV5(db: Database.Database) {
  db.prepare(`UPDATE models SET enabled = 0 WHERE platform = 'google' AND model_id = 'gemini-2.5-pro'`).run();

  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const apply = db.transaction(() => {
    insert.run('cerebras', 'zai-glm-4.7', 'GLM-4.7 (Cerebras)', 7, 1, 'Frontier', 10, 100, null, null, '~3M', 8192);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V6: Live-probed against real free-tier keys on 2026-04-25.
 *
 * Corrections (Google free-tier RPD): the documented "250" / "1000" RPD numbers
 * for gemini-2.5-flash and gemini-2.5-flash-lite are stale — both share a 20
 * RPD per-model-per-project free pool now. Confirmed by the
 * `generate_content_free_tier_requests` quota error, limit 20.
 *
 * Removals: arcee-ai/trinity-large-preview:free returns 404 "No endpoints found"
 * — pulled from OpenRouter's free pool. (Other previously-suspected dead OR :free
 * IDs are still live in /api/v1/models, so they stay.)
 *
 * Additions (all probe-verified to return 200 with content on the user's keys):
 *   - 3 Cloudflare Workers AI reasoning models
 *   - 3 Google preview models, including Pro (which returned a free-tier 429
 *     against the same 20 RPD pool, confirming free-tier eligibility)
 *   - 2 OpenRouter :free models with no expiration_date
 */
function migrateModelsV6(db: Database.Database) {
  // 1) Remove confirmed-dead OR route
  const deleteModel = db.prepare(`DELETE FROM models WHERE platform = ? AND model_id = ?`);
  const deleteFallback = db.prepare(`
    DELETE FROM fallback_config WHERE model_db_id IN (
      SELECT id FROM models WHERE platform = ? AND model_id = ?
    )
  `);
  const removals: Array<[string, string]> = [
    ['openrouter', 'arcee-ai/trinity-large-preview:free'],
  ];
  const applyRemovals = db.transaction(() => {
    for (const [p, m] of removals) {
      deleteFallback.run(p, m);
      deleteModel.run(p, m);
    }
  });
  applyRemovals();

  // 2) Correct stale Google free-tier RPD numbers
  db.prepare(`
    UPDATE models SET rpd_limit = 20, monthly_token_budget = '~3M'
     WHERE platform = 'google' AND model_id = 'gemini-2.5-flash'
  `).run();
  db.prepare(`
    UPDATE models SET rpd_limit = 20, monthly_token_budget = '~3M'
     WHERE platform = 'google' AND model_id = 'gemini-2.5-flash-lite'
  `).run();

  // 3) Add live-probed models
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // Cloudflare Workers AI — 10K Neurons/day shared free pool. Reasoning traces
    // burn output tokens fast, so per-call effective budget is small. Estimates
    // assume 1K-in/500-out typical: kimi-k2.5 ≈ 50/day, qwen3-30b ≈ 200/day,
    // r1-distill ≈ 5/day on the reasoning-heavy path.
    ['cloudflare', '@cf/moonshotai/kimi-k2.5',                    'Kimi K2.5 (CF)',                  3,  11, 'Frontier', null, null, null, null, '~10-20M', 262144],
    ['cloudflare', '@cf/qwen/qwen3-30b-a3b-fp8',                  'Qwen3 30B-A3B fp8 (CF)',          7,  11, 'Large',    null, null, null, null, '~18-45M', 131072],
    ['cloudflare', '@cf/deepseek-ai/deepseek-r1-distill-qwen-32b', 'DeepSeek R1 Distill Qwen 32B (CF)', 9, 11, 'Large',  null, null, null, null, '~3-5M',   131072],

    // Google preview tier — shares the 20 RPD per-model free pool. Pro confirmed
    // free-tier-eligible by the `free_tier_requests` quota metric in 429 errors.
    ['google',     'gemini-3.1-flash-lite-preview',               'Gemini 3.1 Flash-Lite Preview',   18, 3,  'Medium',   15, 20,  250000, null, '~3M',  1048576],
    ['google',     'gemini-3-flash-preview',                       'Gemini 3 Flash Preview',          11, 5,  'Large',    10, 20,  250000, null, '~3M',  1048576],
    ['google',     'gemini-3.1-pro-preview',                       'Gemini 3.1 Pro Preview',          1,  8,  'Frontier',  5, 20,  250000, null, '~3M',  1048576],

    // OpenRouter :free pool — 20 RPM / 50 RPD (1000 once $10 credits bought).
    ['openrouter', 'google/gemma-4-31b-it:free',                   'Gemma 4 31B (free)',             19, 9,  'Medium',   20, 200, null, null, '~6M', 262144],
    ['openrouter', 'liquid/lfm-2.5-1.2b-instruct:free',            'Liquid LFM 2.5 1.2B (free)',     30, 10, 'Small',    20, 200, null, null, '~6M', 32768],
  ];
  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V7 (April 2026): live-probed delta against OpenRouter's free pool + Z.ai.
 * - Removes inclusionai/ling-2.6-flash:free (transitioned to paid, 404 on chat).
 * - Adds 8 new :free routes confirmed via /v1/models + chat-completion probe.
 * - Adds zhipu/glm-4.7-flash (probe: 429 "overloaded" — free-pool throttle, not
 *   "insufficient balance" which paid models return). Same baseUrl works for both
 *   api.z.ai and open.bigmodel.cn keys.
 * HF and NVIDIA left as-is: HF still serves chat with current key; NVIDIA already disabled.
 */
function migrateModelsV7(db: Database.Database) {
  const deleteModel = db.prepare(`DELETE FROM models WHERE platform = ? AND model_id = ?`);
  const deleteFallback = db.prepare(`
    DELETE FROM fallback_config WHERE model_db_id IN (
      SELECT id FROM models WHERE platform = ? AND model_id = ?
    )
  `);
  const removals: Array<[string, string]> = [
    ['openrouter', 'inclusionai/ling-2.6-flash:free'],
  ];
  const applyRemovals = db.transaction(() => {
    for (const [p, m] of removals) {
      deleteFallback.run(p, m);
      deleteModel.run(p, m);
    }
  });
  applyRemovals();

  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  // OpenRouter :free quotas: 20 RPM / 50 RPD without credits, 1000 RPD with $10 lifetime topup.
  // Catalog convention is rpd=200 (matches existing rows).
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    ['openrouter', 'inclusionai/ling-2.6-1t:free',                           'Ling 2.6 1T (free)',                       4,  9,  'Frontier', 20, 200, null, null, '~6M', 262144],
    ['openrouter', 'tencent/hy3-preview:free',                               'Tencent HY3 Preview (free)',               7,  9,  'Frontier', 20, 200, null, null, '~6M', 262144],
    ['openrouter', 'poolside/laguna-m.1:free',                               'Poolside Laguna M.1 (free)',               13, 9,  'Large',    20, 200, null, null, '~6M', 131072],
    ['openrouter', 'google/gemma-4-26b-a4b-it:free',                         'Gemma 4 26B-A4B (free)',                   22, 9,  'Medium',   20, 200, null, null, '~6M', 262144],
    ['openrouter', 'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free',     'Nemotron 3 Nano 30B Reasoning (free)',     23, 9,  'Medium',   20, 200, null, null, '~6M', 262144],
    ['openrouter', 'poolside/laguna-xs.2:free',                              'Poolside Laguna XS.2 (free)',              26, 10, 'Medium',   20, 200, null, null, '~6M', 131072],
    ['openrouter', 'nvidia/nemotron-nano-9b-v2:free',                        'Nemotron Nano 9B v2 (free)',               28, 10, 'Medium',   20, 200, null, null, '~6M', 128000],
    ['openrouter', 'liquid/lfm-2.5-1.2b-thinking:free',                      'Liquid LFM 2.5 1.2B Thinking (free)',      30, 10, 'Small',    20, 200, null, null, '~6M', 32768],
    // Zhipu (Z.ai) — free pool. glm-4.7-flash quotas unpublished; mirror glm-4.5-flash row shape.
    ['zhipu',      'glm-4.7-flash',                                          'GLM-4.7 Flash',                            18, 4,  'Large',    null, null, null, 1000000, '~30M', 131072],
  ];
  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V8 (May 2026): 3-day delta. SambaNova's /v1/models added two free-tier models;
 * Cloudflare's @cf catalog added two new text models. All four probe-verified 200
 * with the user's keys. SambaNova's paid-only MiniMax-M2.5 explicitly returns 422
 * "Couldn't find valid service tier", so the 200s on these rows confirm free-tier
 * access. Cloudflare's @cf/* models share the 10K Neurons/day free pool.
 */
function migrateModelsV8(db: Database.Database) {
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // SambaNova free pool: 20 RPM / 20 RPD / 200K TPD shared across all free models.
    ['sambanova',  'DeepSeek-V3.1-cb',                          'DeepSeek V3.1 (CB)',             5,  9,  'Frontier', 20, 20, null, 200000, '~3M',     131072],
    ['sambanova',  'gemma-3-12b-it',                            'Gemma 3 12B (SambaNova)',        22, 9,  'Medium',   20, 20, null, 200000, '~3M',     131072],
    // Cloudflare @cf — 10K Neurons/day shared pool.
    ['cloudflare', '@cf/moonshotai/kimi-k2.6',                  'Kimi K2.6 (CF)',                 2,  11, 'Frontier', null, null, null, null, '~10-20M', 262144],
    ['cloudflare', '@cf/ibm-granite/granite-4.0-h-micro',       'Granite 4.0 H Micro (CF)',       29, 11, 'Small',    null, null, null, null, '~5-10M',  131072],
  ];
  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V9 (May 2026): disable cerebras/zai-glm-4.7. The model still appears in
 * Cerebras's /v1/models listing but the chat-completions endpoint returns
 * 404 "Model does not exist or you do not have access" for free-tier keys —
 * matches their docs note about temporarily reducing free-tier access on
 * zai-glm-4.7 due to high demand. Row kept (not deleted) so it can be
 * re-enabled later without losing fallback_config history.
 */
function migrateModelsV9(db: Database.Database) {
  db.prepare(
    "UPDATE models SET enabled = 0 WHERE platform = 'cerebras' AND model_id = 'zai-glm-4.7'"
  ).run();
}

/**
 * V10 (May 2026): Ollama Cloud — first new platform since Z.ai/Zhipu in V7.
 * Free plan: GPU-time-based quota (not per-token), 1 concurrent model,
 * 5h session caps, no card required. /v1/models lists 39 SKUs but only 28
 * respond on the Free tier — paid models return 403 with an explicit
 * "this model requires a subscription" message.
 *
 * Curated to ~10 representative free models that either (a) aren't reachable
 * elsewhere in the catalog or (b) provide a useful alternate route through
 * Ollama's independent rate-limit pool. Probe-verified May 2 2026.
 *
 * Quota shape: GPU-time, not tokens. monthly_token_budget reflects rough
 * Free-tier "session" capacity rather than a hard token cap.
 */
function migrateModelsV10(db: Database.Database) {
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // Budget strings are estimates: Ollama publishes no token cap (quota is GPU-time +
    // 7-day rolling). Frontier ~5-10M, Large ~10-20M, Medium ~20-30M reflect that
    // heavier models burn quota faster. Numeric limits stay null — real provider
    // throttling is the source of truth, not these display strings.
    ['ollama', 'qwen3-coder:480b',     'Qwen3-Coder 480B (Ollama)',    2,  9, 'Frontier', null, null, null, null, '~5-10M',  262144],
    ['ollama', 'mistral-large-3:675b', 'Mistral Large 3 675B (Ollama)', 3,  9, 'Frontier', null, null, null, null, '~5-10M',  131072],
    ['ollama', 'deepseek-v3.2',        'DeepSeek V3.2 (Ollama)',        4,  9, 'Frontier', null, null, null, null, '~5-10M',  131072],
    ['ollama', 'cogito-2.1:671b',      'Cogito 2.1 671B (Ollama)',      4,  9, 'Frontier', null, null, null, null, '~5-10M',  131072],
    ['ollama', 'kimi-k2-thinking',     'Kimi K2 Thinking (Ollama)',     5,  9, 'Frontier', null, null, null, null, '~5-10M',  131072],
    ['ollama', 'glm-4.7',              'GLM-4.7 (Ollama)',              6,  9, 'Frontier', null, null, null, null, '~5-10M',  131072],
    ['ollama', 'gpt-oss:120b',         'GPT-OSS 120B (Ollama)',         6,  9, 'Large',    null, null, null, null, '~10-20M', 131072],
    ['ollama', 'devstral-2:123b',      'Devstral 2 123B (Ollama)',      8, 10, 'Large',    null, null, null, null, '~10-20M', 131072],
    ['ollama', 'gpt-oss:20b',          'GPT-OSS 20B (Ollama)',         18, 10, 'Medium',   null, null, null, null, '~20-30M', 131072],
    ['ollama', 'gemma4:31b',           'Gemma 4 31B (Ollama)',         22, 10, 'Medium',   null, null, null, null, '~20-30M', 131072],
  ];
  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V11 (May 2026):
 * 1. Fix long-standing bug: Cerebras `qwen3-235b` was inserted with the
 *    wrong model_id in the original seed (real id is
 *    `qwen-3-235b-a22b-instruct-2507`). Subsequent rank/limit updates that
 *    target the correct id have been silent no-ops since V0 on fresh deploys.
 * 2. Re-enable NVIDIA NIM — `meta/llama-3.1-70b-instruct` was disabled in V2
 *    when NIM moved to credits. Per May 2026 audit it's free again (~1,000
 *    starter credits never expire, 40 RPM/model).
 * 3. Add three new aggregator/anon-friendly platforms confirmed live May 2026:
 *    Kilo Gateway, Pollinations, LLM7.io — all three accept anonymous
 *    requests on at least one model.
 *    - The user still needs a placeholder key entry (any non-empty string
 *      works) because the router filters on `keys.length === 0` to decide
 *      whether a platform is routable.
 *    Chutes was evaluated and dropped: probe with a free-tier key returned
 *    402 on every model — "Quota exceeded and account balance is $0.0,
 *    please pay with fiat or send tao". The "free" tier requires a paid
 *    balance, which conflicts with the no-card criterion.
 */
function migrateModelsV11(db: Database.Database) {
  // 1) Rename cerebras qwen3-235b → qwen-3-235b-a22b-instruct-2507 if the
  //    old id still exists on this DB. Safe to re-run because of the WHERE.
  db.prepare(`
    UPDATE models SET model_id = 'qwen-3-235b-a22b-instruct-2507'
     WHERE platform = 'cerebras' AND model_id = 'qwen3-235b'
  `).run();

  // 2) Re-enable NVIDIA NIM (still has 1,000+ starter credits free-tier).
  db.prepare(`
    UPDATE models SET enabled = 1, monthly_token_budget = '~3M (1k credits)'
     WHERE platform = 'nvidia' AND model_id = 'meta/llama-3.1-70b-instruct'
  `).run();

  // 3) Add catalog rows for the four new platforms. Numeric limits are
  //    conservative — provider docs publish best-effort bounds that fluctuate.
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // NVIDIA NIM — live-probed May 2026 with a free-tier key. All 8 returned
    // 200 + content. Limits are per-model: 40 RPM, shared 1k starter credits
    // (never-expire) used for the rough budget estimate. The existing
    // meta/llama-3.1-70b-instruct row stays (re-enabled above).
    ['nvidia',       'meta/llama-3.3-70b-instruct',                       'Llama 3.3 70B (NV)',                17, 6, 'Large',    40, null, null, null, '~3M (credits)', 131072],
    ['nvidia',       'meta/llama-4-maverick-17b-128e-instruct',           'Llama 4 Maverick (NV)',             11, 6, 'Large',    40, null, null, null, '~3M (credits)', 131072],
    ['nvidia',       'deepseek-ai/deepseek-v4-pro',                       'DeepSeek V4 Pro (NV)',               3, 9, 'Frontier', 40, null, null, null, '~2M (credits)', 131072],
    ['nvidia',       'mistralai/mistral-large-3-675b-instruct-2512',      'Mistral Large 3 675B (NV)',          3, 9, 'Frontier', 40, null, null, null, '~2M (credits)', 131072],
    ['nvidia',       'minimaxai/minimax-m2.7',                            'MiniMax M2.7 (NV)',                  3, 9, 'Frontier', 40, null, null, null, '~2M (credits)', 196608],
    ['nvidia',       'nvidia/nemotron-3-super-120b-a12b',                 'Nemotron 3 Super 120B (NV)',        22, 9, 'Frontier', 40, null, null, null, '~2M (credits)', 262144],
    ['nvidia',       'nvidia/nemotron-3-nano-30b-a3b',                    'Nemotron 3 Nano 30B (NV)',          22, 9, 'Medium',   40, null, null, null, '~3M (credits)', 262144],
    ['nvidia',       'google/gemma-4-31b-it',                             'Gemma 4 31B (NV)',                  19, 9, 'Medium',   40, null, null, null, '~3M (credits)', 262144],
    ['nvidia',       'moonshotai/kimi-k2.6',                              'Kimi K2.6 (NV)',                     3, 9, 'Frontier', 40, null, null, null, '~2M (credits)', 131072],

    // Cerebras — live-probed May 2026 with a free-tier key. Both 200 + content.
    // gpt-oss-120b was removed in V2 ("requires special access, 404 on our
    // key") but is reachable on the current free tier — re-add. llama3.1-8b
    // is the fast small-model alternative (no hyphen, distinct from Groq's
    // llama-3.1-8b-instant id). Free-pool limits match qwen-3-235b row.
    ['cerebras',     'gpt-oss-120b',                              'GPT-OSS 120B (Cerebras)',        6,  1, 'Large',    30, 1000, 60000, 1000000, '~30M', 131072],
    ['cerebras',     'llama3.1-8b',                               'Llama 3.1 8B (Cerebras)',       28,  1, 'Small',    30, 1000, 60000, 1000000, '~30M', 131072],

    // Groq compound — agent system that internally routes through gpt-oss
    // models and exposes the trace in usage metadata. Standard chat-completions
    // shape works (200 + content). Same free-tier limits as other Groq rows.
    ['groq',         'groq/compound',                             'Compound (Groq)',                6,  2, 'Large',    30, 1000, 8000, 200000, '~6M', 131072],
    ['groq',         'groq/compound-mini',                        'Compound Mini (Groq)',          18,  2, 'Medium',   30, 1000, 8000, 200000, '~6M', 131072],

    // Kilo Gateway — 200 req/hr per IP anon. Most named :free routes have
    // transitioned to paid ("free period ended"); probe-confirmed live:
    ['kilo',         'nvidia/nemotron-3-super-120b-a12b:free',  'Nemotron 3 Super 120B (Kilo)',  22, 9,  'Frontier', null, null, null, null, '~2-3M (200/hr)', 262144],

    // Pollinations — anonymous /openai endpoint. Public model list returns
    // just one anonymous-tier entry. Tool calls supported per their metadata.
    ['pollinations', 'openai-fast',                              'GPT-OSS 20B (Pollinations)',    18, 10, 'Medium',   null, null, null, null, '~? (anon)',      131072],

    // LLM7.io — 100 req/hr free (anonymous works). Probe-confirmed list:
    ['llm7',         'gpt-oss-20b',                              'GPT-OSS 20B (LLM7)',            18, 10, 'Medium',   100, null, null, null, '~2-3M (100/hr)', 131072],
    ['llm7',         'meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo', 'Llama 3.1 8B Turbo (LLM7)', 28, 10, 'Small',    100, null, null, null, '~2-3M (100/hr)', 131072],
    ['llm7',         'codestral-latest',                          'Codestral (LLM7)',              16, 8,  'Medium',   100, null, null, null, '~2-3M (100/hr)',  32000],
    ['llm7',         'ministral-8b-2512',                         'Ministral 8B (LLM7)',           28, 10, 'Small',    100, null, null, null, '~2-3M (100/hr)', 131072],
    ['llm7',         'GLM-4.6V-Flash',                            'GLM-4.6V Flash (LLM7)',         15, 9,  'Large',    100, null, null, null, '~2-3M (100/hr)', 131072],
  ];

  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V12 (May 2026): live-probed delta against OpenRouter's free pool.
 *
 * Removals (both confirmed 404 "no longer available as a free model" — moved
 * to paid SKUs at the same id without the :free suffix):
 *   - inclusionai/ling-2.6-1t:free
 *   - tencent/hy3-preview:free
 *
 * Additions (probe-verified 200 + tool_calls with tool_choice=auto on the
 * user's OR key; tool_choice=required is rejected by all three providers,
 * but the router's tool requests use the OpenAI default of auto):
 *   - arcee-ai/trinity-large-thinking:free  — Arcee's Trinity *Thinking*
 *     successor to the trinity-large-preview:free row pulled in V6 (404 then).
 *   - baidu/cobuddy:free                    — Baidu Qianfan coding/agent
 *     model with native tool use + reasoning.
 *   - openrouter/owl-alpha                  — OR-house agentic foundation
 *     model. Zero-priced but no :free suffix, so it sits on a different rate
 *     pool than the shared :free 20 RPM / 200 RPD bucket. 1M ctx.
 *   - nousresearch/hermes-3-llama-3.1-405b:free — 405B route ranked in V3
 *     but never inserted. Currently 429 on probe (upstream throttle), not
 *     gone. No tools support listed; router falls past it on tool requests.
 *
 * Dropped from the add list:
 *   - deepseek/deepseek-v4-flash:free — listed at $0 in /v1/models but the
 *     Crucible provider returns 402 "Out of credits" — not actually free.
 *
 * Context-window corrections — OR raised these to match upstream provider
 * caps; our seeded values were stale:
 *   - nvidia/nemotron-3-super-120b-a12b:free  262144 → 1000000
 *   - qwen/qwen3-coder:free                   262144 → 1048576
 */
function migrateModelsV12(db: Database.Database) {
  const deleteModel = db.prepare(`DELETE FROM models WHERE platform = ? AND model_id = ?`);
  const deleteFallback = db.prepare(`
    DELETE FROM fallback_config WHERE model_db_id IN (
      SELECT id FROM models WHERE platform = ? AND model_id = ?
    )
  `);
  const removals: Array<[string, string]> = [
    ['openrouter', 'inclusionai/ling-2.6-1t:free'],
    ['openrouter', 'tencent/hy3-preview:free'],
  ];
  const applyRemovals = db.transaction(() => {
    for (const [p, m] of removals) {
      deleteFallback.run(p, m);
      deleteModel.run(p, m);
    }
  });
  applyRemovals();

  // Context-window upgrades for existing rows.
  db.prepare(`
    UPDATE models SET context_window = 1000000
     WHERE platform = 'openrouter' AND model_id = 'nvidia/nemotron-3-super-120b-a12b:free'
  `).run();
  db.prepare(`
    UPDATE models SET context_window = 1048576
     WHERE platform = 'openrouter' AND model_id = 'qwen/qwen3-coder:free'
  `).run();

  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  // :free pool quotas as elsewhere in catalog: 20 RPM / 200 RPD / ~6M tokens.
  // openrouter/owl-alpha sits on the non-:free zero-priced pool — quotas
  // unpublished; mirror :free numbers conservatively so it cools down on 429.
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    ['openrouter', 'arcee-ai/trinity-large-thinking:free',         'Trinity Large Thinking (free)',  5,  9, 'Frontier', 20, 200, null, null, '~6M', 262144],
    ['openrouter', 'baidu/cobuddy:free',                           'CoBuddy (free)',                 6,  9, 'Large',    20, 200, null, null, '~6M', 131072],
    ['openrouter', 'openrouter/owl-alpha',                         'Owl Alpha (OR-house)',           5,  9, 'Frontier', 20, 200, null, null, '~6M', 1048576],
    ['openrouter', 'nousresearch/hermes-3-llama-3.1-405b:free',    'Hermes 3 405B (free)',          17,  9, 'Large',    20, 200, null, null, '~6M', 131072],
  ];
  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V13 (May 2026): cross-provider catalog refresh — live-probed against the
 * user's keys for Cerebras, Groq, SambaNova, Cloudflare, Google, NVIDIA,
 * Mistral, Cohere, Ollama, HuggingFace, Z.ai, Chutes. Z.ai/Chutes confirmed
 * no-change (non-flash Z.ai SKUs require recharge; Chutes still gated on
 * $0 balance).
 *
 * DISABLES (row kept for re-enable history; pattern follows V5/V9):
 *   - google/gemini-3.1-pro-preview — 429 with quotaMetric `free_tier_requests`
 *     limit=0; moved off free tier. (V6 inferred Pro was free; lapsed since.)
 *   - ollama/kimi-k2-thinking, ollama/mistral-large-3:675b, ollama/deepseek-v3.2
 *     — all return 403 "this model requires a subscription" on the Free plan
 *     now. (V10 catalogued them as Free; Ollama paywalled them since.)
 *
 * HARD REMOVALS (model is gone, not paywalled):
 *   - sambanova/DeepSeek-V3.1-cb — absent from live /v1/models; V8 was over-eager.
 *   - cloudflare/@cf/moonshotai/kimi-k2.5 — CF changelog: aliases to k2.6 on
 *     2026-05-30. Two rows for one backend wastes a fallback slot.
 *
 * UPDATES (numbers verified from live response headers or /v1/models):
 *   - cerebras: all 3 enabled rows → 5 RPM / 30K TPM / 2400 RPD / 1M TPD
 *     (V4's 30/14400/60K/1M was stale; current free pool is 5 RPM shared).
 *   - groq/llama-3.3-70b-versatile: TPD 500K → 100K (docs).
 *   - groq/llama-4-scout: TPM 6K → 30K (live header).
 *   - groq/compound + groq/compound-mini: RPD 1000→250, TPM 8K→70K, TPD→null
 *     (compound systems run on a separate budget per docs + live header).
 *   - sambanova/DeepSeek-V3.2: ctx 131K → 32K (/v1/models reports 32768).
 *   - cloudflare/@cf/meta/llama-3.3-70b-instruct-fp8-fast: ctx 131K → 24K
 *     (CF model page: fp8-fast variant capped at 24K).
 *   - mistral context windows (all stale vs /v1/models):
 *       codestral-latest        32K → 256K
 *       devstral-latest        131K → 262K
 *       magistral-medium       40K → 131K
 *       mistral-large-latest  131K → 262K
 *
 * ADDITIONS (all chat-probed; tools verified with tool_choice=auto):
 *   - groq: openai/gpt-oss-safeguard-20b (tool-tuned 20B; same pool as gpt-oss-20b)
 *   - cloudflare: @cf/nvidia/nemotron-3-120b-a12b, @cf/google/gemma-4-26b-a4b-it
 *     (both 256K ctx, function_calling=true; 429'd on probe due to daily
 *     neuron exhaustion, not model-level — same Free pool as other CF rows)
 *   - google: gemini-3.5-flash (2026-05 release; 1M ctx; tool calls verified)
 *   - nvidia: deepseek-ai/deepseek-v4-flash, z-ai/glm-5.1 (~2min cold start),
 *     qwen/qwen3-coder-480b-a35b-instruct (full id — discovered via /v1/models)
 *   - mistral: mistral-small-latest (Small 4), ministral-8b-latest (edge 8B)
 *   - cohere: command-a-reasoning-08-2025, command-r-08-2024. README ToS
 *     table marks Cohere "❌ Avoid for personal use" — these rows expand
 *     fallback coverage but inherit the same caveat.
 *   - ollama: qwen3-coder-next (~80B-A3B coder)
 *
 * NEW PLATFORM (huggingface):
 *   - V4 removed HF for "tool-call format issues" on the legacy serverless
 *     endpoint that emitted tool calls as text. The new router.huggingface.co
 *     meta-router uses each backend's native protocol and normalizes the
 *     response — chat-probed clean tool_calls on DeepSeek-V4-Flash, Kimi-K2.6,
 *     Qwen3-Coder-Next, GLM-4.7, Qwen3-235B. Recurring $0.10/mo router credit
 *     on the free tier (no card, no expiry). Budget ~1-3M tokens/mo depending
 *     on backend. Seeded with 3 frontier rows.
 *   - Provider registration lives in server/src/providers/index.ts.
 *
 * NO-CHANGE PROVIDERS (probed, nothing to update):
 *   - openrouter (V12 just shipped)
 *   - z.ai/zhipu — non-flash SKUs all 1113 "insufficient balance"; flash-only
 *     pool is correctly catalogued.
 *   - chutes — every probe 402 on $0 balance; V11 drop decision stands.
 *
 * DEFERRED:
 *   - cerebras/qwen-3-235b-a22b-instruct-2507 + llama3.1-8b: docs flag
 *     hard-deprecation 2026-05-27 (4d from migration), but both still 200
 *     today. Disable in V14 on/after that date.
 *   - sambanova: every chat probe returned 402 PAYMENT_METHOD_REQUIRED on the
 *     user's account. /v1/models still lists rows so REMOVE/UPDATE above is
 *     valid, but the SambaNova free tier may have lapsed account-wide.
 *     Investigate before adding new SambaNova rows.
 *   - github: gpt-5 family + xai/grok-3 both 400 "unavailable_model" on free
 *     tier — keep gpt-4o (V2 verdict still holds).
 */
function migrateModelsV13(db: Database.Database) {
  // 1) Disables (row kept; can be re-enabled without losing fallback history).
  const disable = db.prepare(`UPDATE models SET enabled = 0 WHERE platform = ? AND model_id = ?`);
  const disables: Array<[string, string]> = [
    ['google', 'gemini-3.1-pro-preview'],
    ['ollama', 'kimi-k2-thinking'],
    ['ollama', 'mistral-large-3:675b'],
    ['ollama', 'deepseek-v3.2'],
  ];
  for (const [p, m] of disables) disable.run(p, m);

  // 2) Hard removals.
  const deleteModel = db.prepare(`DELETE FROM models WHERE platform = ? AND model_id = ?`);
  const deleteFallback = db.prepare(`
    DELETE FROM fallback_config WHERE model_db_id IN (
      SELECT id FROM models WHERE platform = ? AND model_id = ?
    )
  `);
  const removals: Array<[string, string]> = [
    ['sambanova', 'DeepSeek-V3.1-cb'],
    ['cloudflare', '@cf/moonshotai/kimi-k2.5'],
  ];
  const applyRemovals = db.transaction(() => {
    for (const [p, m] of removals) {
      deleteFallback.run(p, m);
      deleteModel.run(p, m);
    }
  });
  applyRemovals();

  // 3) Cerebras free-pool limit correction (3 enabled rows; zai-glm-4.7 stays
  //    on its V5-set per-model 10/100 cap since it's gated separately).
  db.prepare(`
    UPDATE models
       SET rpm_limit = 5, rpd_limit = 2400, tpm_limit = 30000, tpd_limit = 1000000
     WHERE platform = 'cerebras'
       AND model_id IN ('qwen-3-235b-a22b-instruct-2507', 'gpt-oss-120b', 'llama3.1-8b')
  `).run();

  // 4) Groq limit corrections.
  db.prepare(`UPDATE models SET tpd_limit = 100000 WHERE platform = 'groq' AND model_id = 'llama-3.3-70b-versatile'`).run();
  db.prepare(`UPDATE models SET tpm_limit = 30000 WHERE platform = 'groq' AND model_id = 'meta-llama/llama-4-scout-17b-16e-instruct'`).run();
  db.prepare(`
    UPDATE models SET rpd_limit = 250, tpm_limit = 70000, tpd_limit = NULL
     WHERE platform = 'groq' AND model_id IN ('groq/compound', 'groq/compound-mini')
  `).run();

  // 5) Single-row context-window corrections.
  db.prepare(`UPDATE models SET context_window = 32768 WHERE platform = 'sambanova' AND model_id = 'DeepSeek-V3.2'`).run();
  db.prepare(`UPDATE models SET context_window = 24000 WHERE platform = 'cloudflare' AND model_id = '@cf/meta/llama-3.3-70b-instruct-fp8-fast'`).run();

  // 6) Mistral context-window corrections.
  db.prepare(`UPDATE models SET context_window = 256000 WHERE platform = 'mistral' AND model_id = 'codestral-latest'`).run();
  db.prepare(`UPDATE models SET context_window = 262144 WHERE platform = 'mistral' AND model_id = 'devstral-latest'`).run();
  db.prepare(`UPDATE models SET context_window = 131072 WHERE platform = 'mistral' AND model_id = 'magistral-medium-latest'`).run();
  db.prepare(`UPDATE models SET context_window = 262144 WHERE platform = 'mistral' AND model_id = 'mistral-large-latest'`).run();

  // 7) Additions across providers (chat-probed; tools verified where claimed).
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    // Groq — shared 30 RPM / 1K RPD / 8K TPM / 200K TPD per-model pool.
    ['groq',       'openai/gpt-oss-safeguard-20b',          'GPT-OSS Safeguard 20B (Groq)',   18, 2, 'Medium',   30, 1000, 8000, 200000, '~6M', 131072],

    // Cloudflare — 10K Neurons/day shared free pool.
    ['cloudflare', '@cf/nvidia/nemotron-3-120b-a12b',       'Nemotron 3 120B (CF)',            9, 11, 'Frontier', null, null, null, null, '~5-10M',  262144],
    ['cloudflare', '@cf/google/gemma-4-26b-a4b-it',         'Gemma 4 26B-A4B it (CF)',        22, 11, 'Medium',   null, null, null, null, '~10-20M', 262144],

    // Google — same 20 RPD per-model free pool. 3.5 Flash is the current Flash flagship.
    ['google',     'gemini-3.5-flash',                      'Gemini 3.5 Flash',                3, 5, 'Large',    10, 20,  250000, null, '~3M', 1048576],

    // NVIDIA NIM — credits-based; per-model 40 RPM.
    ['nvidia',     'deepseek-ai/deepseek-v4-flash',         'DeepSeek V4 Flash (NV)',          4, 9, 'Frontier', 40, null, null, null, '~3M (credits)', 131072],
    ['nvidia',     'z-ai/glm-5.1',                          'GLM-5.1 (NV, slow cold-start)',   5, 9, 'Frontier', 40, null, null, null, '~3M (credits)', 200000],
    ['nvidia',     'qwen/qwen3-coder-480b-a35b-instruct',   'Qwen3-Coder 480B (NV)',           2, 9, 'Frontier', 40, null, null, null, '~3M (credits)', 262144],

    // Mistral — Experiment plan 2 RPM / 500K TPM / shared ~1B/mo.
    ['mistral',    'mistral-small-latest',                  'Mistral Small 4',                14, 8, 'Medium',   2, null, 500000, null, '~50-100M', 262144],
    ['mistral',    'ministral-8b-latest',                   'Ministral 3 8B',                 28, 8, 'Small',    2, null, 500000, null, '~50-100M', 262144],

    // Cohere — trial 20 RPM / 1000 RPM total. ToS table marks ❌ Avoid for personal use.
    ['cohere',     'command-a-reasoning-08-2025',           'Command A Reasoning (08-2025)',  13, 11, 'Large',   20, 33, null, null, '~1-2M', 256000],
    ['cohere',     'command-r-08-2024',                     'Command R (08-2024)',            25, 11, 'Medium',  20, 33, null, null, '~1-2M', 131072],

    // Ollama Cloud — GPU-time quota.
    ['ollama',     'qwen3-coder-next',                      'Qwen3-Coder Next (Ollama)',       3, 9, 'Large',   null, null, null, null, '~10-20M', 262144],

    // HuggingFace router (new platform) — recurring $0.10/mo credit, no card.
    ['huggingface', 'deepseek-ai/DeepSeek-V4-Flash',        'DeepSeek V4 Flash (HF)',          4, 9, 'Frontier', null, null, null, null, '~1-3M', 131072],
    ['huggingface', 'moonshotai/Kimi-K2.6',                 'Kimi K2.6 (HF)',                  3, 9, 'Frontier', null, null, null, null, '~1-3M', 262144],
    ['huggingface', 'Qwen/Qwen3-Coder-Next',                'Qwen3-Coder Next (HF)',           3, 9, 'Large',    null, null, null, null, '~1-3M', 262144],
  ];

  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V14 (May 2026): Cerebras hard-deprecation 2026-05-27.
 *
 * Per inference-docs.cerebras.ai/models/overview, both
 * `qwen-3-235b-a22b-instruct-2507` and `llama3.1-8b` hit a hard deprecation
 * on 2026-05-27 with no announced free-tier replacement at the same
 * parameter class. Disable both ahead of the cutover so the router stops
 * sending traffic to them. Row kept (not deleted) so it can be re-enabled
 * if Cerebras restores or renames either model — same pattern as V9's
 * disable of `zai-glm-4.7`.
 *
 * Cerebras `gpt-oss-120b` is NOT in the deprecation list and stays enabled
 * as the sole free-tier Cerebras route.
 */
function migrateModelsV14(db: Database.Database) {
  db.prepare(`
    UPDATE models SET enabled = 0
     WHERE platform = 'cerebras'
       AND model_id IN ('qwen-3-235b-a22b-instruct-2507', 'llama3.1-8b')
  `).run();
}

/**
 * V15 (May 2026): purge SiliconFlow.
 *
 * SiliconFlow was briefly added (#131) on the belief Qwen/Qwen3-8B was a $0
 * "free model". Re-verification showed it is PAID ($0.06/M in+out): the
 * account balance dropped 0.9999 -> 0.9998 across ~2.7K tokens, and the
 * official pricing page lists it at $0.06/M. The earlier "zero-cost" read was
 * a 4-decimal rounding artifact on a tiny call. SiliconFlow's .com endpoint is
 * a one-time $1 trial credit, not a recurring free tier — same disqualifier
 * as Chutes — so the provider was reverted. This removes any orphaned row from
 * a DB that already ran the original V15. No-op on DBs that never had it.
 */
function migrateModelsV15(db: Database.Database) {
  db.prepare(`
    DELETE FROM fallback_config WHERE model_db_id IN (
      SELECT id FROM models WHERE platform = 'siliconflow'
    )
  `).run();
  db.prepare(`DELETE FROM models WHERE platform = 'siliconflow'`).run();
}

// Adds the supports_vision column to existing DBs and (re)applies the vision
// flags by rule. Rule-based rather than a hardcoded id list because the catalog
// churns through migrations (model ids get renamed/replaced) — rules survive
// that. Two constraints decide a model can accept images:
//   1. The model is genuinely multimodal (every Gemini is; Llama 4 Scout/
//      Maverick are; GitHub's GPT-4o/4.1/5 are).
//   2. Its provider adapter forwards image content. OpenAICompat and Google do;
//      Cohere and Cloudflare flatten content to text, so models on those
//      platforms are excluded even when the underlying model can see.
// Conservative on purpose: with hard-fail routing a false negative is just a
// clear "no vision model" error, while a false positive routes an image to a
// model that chokes. Idempotent — safe on fresh seeds and upgrades alike.
function migrateModelsV16Vision(db: Database.Database) {
  const columns = db.prepare('PRAGMA table_info(models)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'supports_vision')) {
    db.prepare('ALTER TABLE models ADD COLUMN supports_vision INTEGER NOT NULL DEFAULT 0').run();
  }
  const apply = db.transaction(() => {
    // Reset first so de-flagged models (e.g. an id that moved to Cloudflare)
    // don't keep a stale flag across re-runs.
    db.prepare('UPDATE models SET supports_vision = 0').run();
    // Every Gemini is multimodal (the 'google' platform is all Gemini).
    db.prepare("UPDATE models SET supports_vision = 1 WHERE platform = 'google'").run();
    // Llama 4 (Scout/Maverick) is natively multimodal — but only where the
    // adapter forwards images (exclude the text-flattening providers).
    db.prepare(`
      UPDATE models SET supports_vision = 1
      WHERE LOWER(model_id) LIKE '%llama-4%'
        AND platform NOT IN ('cloudflare', 'cohere')
    `).run();
    // GitHub's OpenAI vision models.
    db.prepare(`
      UPDATE models SET supports_vision = 1
      WHERE platform = 'github'
        AND (model_id LIKE '%gpt-4o%' OR model_id LIKE '%gpt-4.1%' OR model_id LIKE '%gpt-5%')
    `).run();
    // V23 vision additions, both live-probed 2026-06-07 (answered a color
    // question about an inline data-URL PNG): Zhipu's free GLM-4.6V Flash and
    // NVIDIA's Nemotron Nano 12B v2 VL (via OpenRouter).
    db.prepare(`
      UPDATE models SET supports_vision = 1
      WHERE LOWER(model_id) LIKE '%glm-4.6v%'
         OR LOWER(model_id) LIKE '%nemotron-nano-12b-v2-vl%'
    `).run();
  });
  apply();
}

// ── V17: intelligence tier audit (2026-06) ──
// `size_label` is the cross-provider capability tier that DOMINATES the router's
// intelligence axis (intelligenceComposite = tier*1000 - intelligence_rank in
// services/router.ts), so it must track real benchmarks rather than the
// release-day guesses several rows were seeded with. This normalizes every model
// family that has a published Artificial Analysis Intelligence Index v4.0 score
// (served/default mode, June 2026) to the correct tier — and as a side effect
// fixes cross-provider inconsistencies where the same model family had landed in
// different tiers per provider (e.g. Llama 4 Scout, Llama 3.3 70B).
//
// Tier bands by AA Index v4.0:  Frontier ≥45 · Large 26–44 · Medium 13–25 · Small ≤12.
//
// Rules are keyed by model_id LIKE patterns so one rule covers a family across
// all providers. Models with NO published AA Index are intentionally left at
// their seeded tier (Cogito 2.1, Owl Alpha, Poolside Laguna, Hermes 3 405B,
// Baidu CoBuddy, Groq Compound, GLM-4.6V Flash, GLM-4.5 Flash, Mistral Small 4,
// Devstral). intelligence_rank (the within-tier tiebreak, low impact) is left
// untouched. Idempotent — every statement is an absolute SET, safe to re-run.
function migrateModelsV17IntelligenceTiers(db: Database.Database) {
  const apply = db.transaction(() => {
    // Frontier (AA ≥ 45): genuine frontier-class. Promotes Gemini 3.5 Flash (55)
    // and Gemini 3 Flash Preview (46) up from Large.
    db.prepare(`
      UPDATE models SET size_label = 'Frontier' WHERE
           LOWER(model_id) LIKE '%gemini-3.1-pro%'
        OR LOWER(model_id) LIKE '%gemini-3.5-flash%'
        OR LOWER(model_id) LIKE '%gemini-3-flash%'
        OR LOWER(model_id) LIKE '%kimi-k2.6%'
        OR LOWER(model_id) LIKE '%kimi-k2-thinking%'
        OR LOWER(model_id) LIKE '%deepseek-v4-pro%'
        OR LOWER(model_id) LIKE '%deepseek-v4-flash%'
        OR LOWER(model_id) LIKE '%glm-5.1%'
        OR LOWER(model_id) LIKE '%minimax-m2.7%'
    `).run();

    // Large (AA 26–44). Demotes Gemini 2.5 Pro (35), Nemotron 3 Super/120B (36),
    // GLM-4.7 (42), DeepSeek V3.1/V3.2 (28/32), Trinity (32) down from Frontier;
    // promotes Gemma 4 31B (39) / 26B (31) and Gemini 3.1 Flash-Lite (34) up.
    db.prepare(`
      UPDATE models SET size_label = 'Large' WHERE
           LOWER(model_id) LIKE '%minimax-m2.5%'
        OR LOWER(model_id) LIKE '%qwen3-next%'
        OR LOWER(model_id) LIKE '%qwen3-coder-next%'
        OR LOWER(model_id) LIKE '%gpt-oss-120b%' OR LOWER(model_id) LIKE '%gpt-oss:120b%'
        OR LOWER(model_id) LIKE '%glm-4.7%'
        OR LOWER(model_id) LIKE '%nemotron-3-super%' OR LOWER(model_id) LIKE '%nemotron-3-120b%'
        OR LOWER(model_id) LIKE '%gemini-2.5-pro%'
        OR LOWER(model_id) LIKE '%deepseek-v3.2%'
        OR LOWER(model_id) LIKE '%deepseek-v3.1%'
        OR LOWER(model_id) LIKE '%trinity-large%'
        OR LOWER(model_id) LIKE '%mistral-medium%'
        OR LOWER(model_id) LIKE '%magistral-medium%'
        OR LOWER(model_id) LIKE '%gpt-4.1%'
        OR LOWER(model_id) LIKE '%gemma-4-31b%' OR LOWER(model_id) LIKE '%gemma4:31b%'
        OR LOWER(model_id) LIKE '%gemma-4-26b%'
        OR LOWER(model_id) LIKE '%gemini-3.1-flash-lite%'
    `).run();

    // Medium (AA 13–25). Demotes Qwen3-Coder 480B (25) and Mistral Large 3 (23)
    // down from Frontier; Llama 4 Maverick (18), GPT-4o (17), Gemini 2.5 Flash
    // (21), GLM-4.5 Air (23), DeepSeek R1 Distill (17), Command A/R+ down from
    // Large; unifies Llama 4 Scout (14) and Llama 3.3 70B (14) across providers.
    db.prepare(`
      UPDATE models SET size_label = 'Medium' WHERE
           (LOWER(model_id) LIKE '%qwen3-coder%' AND LOWER(model_id) NOT LIKE '%qwen3-coder-next%')
        OR LOWER(model_id) LIKE '%qwen-3-235b%' OR LOWER(model_id) LIKE '%qwen3-235b%'
        OR LOWER(model_id) LIKE '%mistral-large%'
        OR LOWER(model_id) LIKE '%gpt-oss-20b%' OR LOWER(model_id) LIKE '%gpt-oss:20b%'
        OR LOWER(model_id) LIKE '%gpt-oss-safeguard-20b%' OR model_id = 'openai-fast'
        OR LOWER(model_id) LIKE '%glm-4.5-air%'
        OR LOWER(model_id) LIKE '%devstral-2%'
        OR LOWER(model_id) LIKE '%deepseek-r1-distill%'
        OR LOWER(model_id) LIKE '%qwen3-30b%'
        OR LOWER(model_id) LIKE '%qwen3-32b%'
        OR LOWER(model_id) LIKE '%llama-4-maverick%'
        OR LOWER(model_id) LIKE '%llama-4-scout%'
        OR LOWER(model_id) LIKE '%llama-3.3-70b%'
        OR LOWER(model_id) LIKE '%llama-3.1-70b%'
        OR (LOWER(model_id) LIKE '%gemini-2.5-flash%' AND LOWER(model_id) NOT LIKE '%flash-lite%')
        OR LOWER(model_id) LIKE '%gemini-2.5-flash-lite%'
        OR LOWER(model_id) LIKE '%gpt-4o%'
        OR LOWER(model_id) LIKE '%command-a-03-2025%'
        OR LOWER(model_id) LIKE '%command-r-plus%'
        OR LOWER(model_id) LIKE '%nemotron-3-nano%'
        OR LOWER(model_id) LIKE '%nemotron-nano-9b%'
    `).run();

    // Small (AA ≤ 12). Demotes Gemma 3 12B (9), Command R 08-2024 (legacy ~7),
    // and Codestral (8) down from Medium.
    db.prepare(`
      UPDATE models SET size_label = 'Small' WHERE
           LOWER(model_id) LIKE '%gemma-3-12b%'
        OR LOWER(model_id) LIKE '%command-r-08-2024%'
        OR LOWER(model_id) LIKE '%codestral%'
        OR LOWER(model_id) LIKE '%llama-3.1-8b%' OR LOWER(model_id) LIKE '%llama3.1-8b%'
        OR LOWER(model_id) LIKE '%meta-llama-3.1-8b%'
        OR LOWER(model_id) LIKE '%ministral-8b%'
        OR LOWER(model_id) LIKE '%granite-4.0-h-micro%'
        OR LOWER(model_id) LIKE '%lfm-2.5-1.2b%'
    `).run();
  });
  apply();
}

// ── V18: OpenCode Zen provider (2026-06) ──
// Adds the OpenCode Zen gateway (#128, originally contributed by @Aldo-f). Zen is
// an OpenAI-compatible service whose paid models bill pay-as-you-go but which
// also exposes a rotating set of *promotional* free models. Access is via a FREE
// account key from https://opencode.ai/auth — no credit card; billing only
// applies if you call paid models. We require that key like any other provider
// (we do NOT use Zen's unauthenticated path).
//
// We seed only the four models the Zen docs explicitly label free — big-pickle,
// deepseek-v4-flash-free, mimo-v2.5-free, nemotron-3-super-free. The live
// /v1/models list also surfaces qwen3.6-plus-free and minimax-m3-free, but those
// are NOT documented as free, so they're intentionally omitted (same "verified
// only" bar as V12-V14). Tiers follow V17's bands (deepseek-v4-flash → Frontier,
// nemotron-3-super → Large). Caveats per docs: promotional/limited-time, "trial
// use only — not for production", and prompts/outputs may be used to improve the
// models (NVIDIA logs Nemotron traffic). Conservative shared 20 RPM / 200 RPD,
// matching the OpenRouter :free pool pattern. Idempotent (INSERT OR IGNORE +
// fallback_config backfill), safe to re-run.
function migrateModelsV18OpenCodeZen(db: Database.Database) {
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    ['opencode', 'big-pickle',             'Big Pickle (OpenCode Zen, stealth)',     10, 4, 'Large',    20, 200, null, null, 'promo (trial)', 131072],
    ['opencode', 'deepseek-v4-flash-free', 'DeepSeek V4 Flash Free (OpenCode Zen)',   4, 4, 'Frontier', 20, 200, null, null, 'promo (trial)', 131072],
    ['opencode', 'mimo-v2.5-free',         'MiMo-V2.5 Free (OpenCode Zen)',          14, 4, 'Medium',   20, 200, null, null, 'promo (trial)', 131072],
    ['opencode', 'nemotron-3-super-free',  'Nemotron 3 Super Free (OpenCode Zen)',   12, 4, 'Large',    20, 200, null, null, 'promo (trial)', 131072],
  ];

  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    const missing = db.prepare(`
      SELECT m.id FROM models m
      LEFT JOIN fallback_config f ON m.id = f.model_db_id
      WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
    `).all() as { id: number }[];
    if (missing.length > 0) {
      const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
      const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
      for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
    }
  });
  apply();
}

/**
 * V19 (June 2026): Google Gemma 4 (released Mar 2026) on the AI Studio free tier.
 * Both IDs are reachable with the same key as Gemini (live-probed 200). The 26B
 * is an MoE model (~3.8B active) — Google's real id is `gemma-4-26b-a4b-it`, not
 * a plain "26B". Tiers match V17's bands (both Large; V17 re-asserts on every
 * boot). Free limits are now AI-Studio-dashboard-driven and were cut 50-80% in
 * Dec 2025, so rpd/tpm are conservative. Idempotent (INSERT OR IGNORE + fallback
 * backfill), safe to re-run.
 */
function migrateModelsV19Gemma4(db: Database.Database) {
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    ['google', 'gemma-4-31b-it',     'Gemma 4 31B IT', 19, 4, 'Large', 15, 1000, 250000, null, '~30M', 32768],
    ['google', 'gemma-4-26b-a4b-it', 'Gemma 4 26B IT', 20, 4, 'Large', 15, 1000, 250000, null, '~30M', 32768],
  ];

  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    backfillFallback(db);
  });
  apply();
}

/**
 * V20 (June 2026): Kilo Gateway anonymous free models. Live-probed keyless
 * (no API key, cost:0); Kilo documents anonymous access for `:free` routes,
 * rate-limited 200 req/hr per IP shared across ALL free models. Per-model rate
 * limits are left null on purpose — the 200/hr budget is per-IP, not per-model,
 * so we rely on Kilo's own 429s + gateway failover rather than guessing a split.
 * Prompts/outputs are logged for training (don't send sensitive data); the 120B
 * Nemotron is additionally flagged trial-use by Kilo. Tiers follow V17 bands
 * (V17 re-asserts nemotron-3-super → Large on every boot). Routing also needs a
 * 'kilo' api_keys sentinel row, added via the keyless Keys-page flow. Idempotent
 * (INSERT OR IGNORE + fallback backfill), safe to re-run.
 */
function migrateModelsV20KiloFree(db: Database.Database) {
  const insert = db.prepare(`
    INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null]> = [
    ['kilo', 'poolside/laguna-m.1:free',               'Poolside Laguna M.1 (Kilo)',    13, 8, 'Large',  null, null, null, null, 'free · 200/hr per IP',         262144],
    ['kilo', 'poolside/laguna-xs.2:free',              'Poolside Laguna XS.2 (Kilo)',   16, 4, 'Medium', null, null, null, null, 'free · 200/hr per IP',         262144],
    ['kilo', 'nvidia/nemotron-3-super-120b-a12b:free', 'Nemotron 3 Super 120B (Kilo)',  12, 5, 'Large',  null, null, null, null, 'free · 200/hr per IP (trial)', 1000000],
    ['kilo', 'stepfun/step-3.7-flash:free',            'StepFun Step 3.7 Flash (Kilo)', 14, 3, 'Medium', null, null, null, null, 'free · 200/hr per IP',         262144],
  ];

  const apply = db.transaction(() => {
    for (const a of additions) insert.run(...a);
    backfillFallback(db);
  });
  apply();
}

/**
 * V21 (June 2026): Remove models confirmed DEAD by live probing (2026-06-03 with
 * production keys), and re-enable Cerebras zai-glm-4.7 which serves free again.
 *
 * Deleted (fallback_config row removed first — foreign_keys=ON forbids deleting a
 * referenced models row):
 *   - llm7/gpt-oss-20b, llm7/meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo,
 *     llm7/ministral-8b-2512 — LLM7 silently serves `mistral-small-3.2` for these
 *     (200 OK but wrong model — a routing footgun).
 *   - llm7/GLM-4.6V-Flash — now pro-gated (402 "upgrade required").
 *   - openrouter/arcee-ai/trinity-large-thinking:free,
 *     openrouter/minimax/minimax-m2.5:free, openrouter/baidu/cobuddy:free —
 *     404 "no endpoints found" (delisted / moved to paid).
 *
 * Re-enabled: cerebras/zai-glm-4.7 (V9 disabled it; live-probed 200 free again).
 * These ids are re-inserted by their original migrations on each boot, so this
 * later DELETE is what keeps them out. Idempotent, safe to re-run.
 */
function migrateModelsV21PruneDead(db: Database.Database) {
  const dead: Array<[string, string]> = [
    ['llm7', 'gpt-oss-20b'],
    ['llm7', 'meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo'],
    ['llm7', 'ministral-8b-2512'],
    ['llm7', 'GLM-4.6V-Flash'],
    ['openrouter', 'arcee-ai/trinity-large-thinking:free'],
    ['openrouter', 'minimax/minimax-m2.5:free'],
    ['openrouter', 'baidu/cobuddy:free'],
  ];
  const apply = db.transaction(() => {
    const getId = db.prepare('SELECT id FROM models WHERE platform = ? AND model_id = ?');
    const delFb = db.prepare('DELETE FROM fallback_config WHERE model_db_id = ?');
    const delModel = db.prepare('DELETE FROM models WHERE id = ?');
    for (const [platform, modelId] of dead) {
      const row = getId.get(platform, modelId) as { id: number } | undefined;
      if (!row) continue;
      delFb.run(row.id);   // remove the fallback chain entry first (FK)
      delModel.run(row.id);
    }
    // Re-enable Cerebras zai-glm-4.7 (model + its fallback chain entry).
    db.prepare("UPDATE models SET enabled = 1 WHERE platform = 'cerebras' AND model_id = 'zai-glm-4.7'").run();
    db.prepare(`
      UPDATE fallback_config SET enabled = 1
       WHERE model_db_id = (SELECT id FROM models WHERE platform = 'cerebras' AND model_id = 'zai-glm-4.7')
    `).run();
  });
  apply();
}

// ── V22: tools-aware routing (2026-06-04) ──
// Adds the supports_tools column and flags the models that reliably emit
// STRUCTURED tool_calls. Motivation: an OpenAI-compatible agent client
// (Paperclip/Codex via /v1/responses) sent tool-bearing requests that the
// chain cascaded down to models with no real function-calling support —
// nemotron-3-nano answered 72 requests averaging ~38 output tokens, and the
// tool call leaked into chat text as literal `</tool_call>` XML, so the agent
// harness never saw a status update and every issue dead-ended in manual
// recovery. The router now keeps tool-bearing requests on this flagged subset
// (see routeRequest's requireTools).
//
// Rule-based by model family rather than a hardcoded id list, same reasoning
// as V16Vision: the catalog churns through migrations and LIKE rules survive
// renames. A family is flagged only when (a) it has native, documented
// function calling AND (b) we've seen structured tool_calls from it through
// this gateway (the 2026-06-01 live tool-calling benchmark, V4's probe pass,
// or production traffic). Conservative on purpose: a false negative just
// narrows the pool, while a false positive reproduces the silent-garbage
// failure above. Deliberately NOT flagged: gemma (weak at tools — V4),
// nemotron nano/9b (the incident model), poolside laguna (returns ~2 tokens),
// hermes-3 (emits tool calls as text — V4), groq compound (built-in tools
// only, rejects user functions), r1-distills, and the small/stealth/unknown
// tail (granite, lfm, stepfun, big-pickle, mimo, owl-alpha, cogito,
// pollinations). Idempotent — reset-then-set, safe on fresh seeds and
// upgrades alike.
function migrateModelsV22Tools(db: Database.Database) {
  const columns = db.prepare('PRAGMA table_info(models)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'supports_tools')) {
    db.prepare('ALTER TABLE models ADD COLUMN supports_tools INTEGER NOT NULL DEFAULT 0').run();
  }
  const apply = db.transaction(() => {
    // Reset first so a de-flagged model doesn't keep a stale flag across re-runs.
    db.prepare('UPDATE models SET supports_tools = 0').run();
    db.prepare(`
      UPDATE models SET supports_tools = 1
      WHERE (
           LOWER(model_id) LIKE '%gpt-oss%'        -- groq/OR/cerebras/CF/sambanova/ollama; incl. safeguard (tool-tuned)
        OR ((LOWER(model_id) LIKE '%llama-3%' OR LOWER(model_id) LIKE '%llama-4%')
            AND LOWER(model_id) NOT LIKE '%hermes%'   -- hermes-3-llama emits text tool calls
            AND LOWER(model_id) NOT LIKE '%llama-3.2%') -- 3B route declares no tool support (V23)
        OR LOWER(model_id) LIKE '%gemini-%'        -- every Gemini; gemma intentionally NOT matched
        OR LOWER(model_id) LIKE '%glm-%'           -- GLM 4.5+/5.x are agentic-tuned (zai/zhipu/CF/ollama)
        OR LOWER(model_id) LIKE '%qwen3%'
        OR LOWER(model_id) LIKE '%qwen-3%'         -- Qwen3 incl. coder/next variants
        OR LOWER(model_id) LIKE '%deepseek-v%'     -- V3.x/V4 function calling; excludes r1-distill
        OR LOWER(model_id) LIKE '%kimi-k2%'        -- K2 family is tool-native
        OR LOWER(model_id) LIKE '%minimax-m2%'     -- M2.x is agent-focused
        OR LOWER(model_id) LIKE '%mistral-large%'  -- Mistral API function calling (whole family)
        OR LOWER(model_id) LIKE '%mistral-medium%'
        OR LOWER(model_id) LIKE '%mistral-small%'
        OR LOWER(model_id) LIKE '%magistral%'
        OR LOWER(model_id) LIKE '%codestral%'
        OR LOWER(model_id) LIKE '%devstral%'
        OR LOWER(model_id) LIKE '%ministral%'
        OR LOWER(model_id) LIKE '%command-a%'      -- Cohere native tool use (benchmarked top-20)
        OR LOWER(model_id) LIKE '%command-r%'
        OR LOWER(model_id) LIKE '%gpt-4o%'         -- GitHub's OpenAI models
        OR LOWER(model_id) LIKE '%gpt-4.1%'
        OR LOWER(model_id) LIKE '%gpt-5%'
        OR LOWER(model_id) LIKE '%nemotron-3-super%' -- benchmarked #8 with real tool calls; nano stays excluded
        OR LOWER(model_id) LIKE '%nemotron-nano-12b-v2-vl%' -- unlike the 30B nano: live-probed structured tool_calls (V23, 2026-06-07)
        OR LOWER(model_id) LIKE '%nemotron-3-ultra%' -- structured tool_calls live-verified via Zen's dedicated endpoint (V24, 2026-06-07); covers the disabled OR row too
        OR LOWER(model_id) LIKE '%minimax-m3%'       -- finish_reason:tool_calls live-verified on Zen (V24)
      )
    `).run();
  });
  apply();
}

/**
 * V23 (June 2026): recurring-free audit — drop SambaNova + Chutes, add
 * live-verified free models (all probed 2026-06-07 with production keys).
 *
 * Removed platforms (models, fallback rows, AND key rows; earlier migrations
 * re-insert the model rows on every boot, so this later DELETE is what keeps
 * them out — same pattern as V21):
 *   - sambanova: free tier is gone for good. The always-free tier was retired
 *     in early 2025 ("We do not have any current plans to maintain the free
 *     tier in any state" — staff, community.sambanova.ai/t/847) in favor of a
 *     one-time $5 trial credit that expires in 3 months. Once it lapses with
 *     no card on file, every chat call 402s "payment method required" — live
 *     probe confirmed on all 6 models, while /v1/models still 200s (billing
 *     gate, not a key problem). No recurring no-card path remains.
 *   - chutes: never a registered provider — rows were hand-added during the
 *     V11 evaluation. The Early Access free plan (200 req/day) fully retired
 *     2026-03-15 (chutes.ai/news/community-announcement-february); PAYG and
 *     paid subs only now, so V11's drop verdict is permanent.
 *
 * Added:
 *   - openrouter kimi-k2.6:free — chat probed 200; tools flagged via the V22
 *     %kimi-k2% family rule (the tool probe itself only hit upstream 429s —
 *     it's a popular new route — but K2 is tool-native everywhere we run it).
 *   - openrouter nemotron-nano-12b-v2-vl:free — structured tool_calls AND
 *     vision both live-verified; V16/V22 rules added.
 *   - zhipu glm-4.6v-flash — listed "Free" on Z.AI pricing; 200 with our
 *     existing bigmodel.cn key; structured tool_calls + vision verified.
 *   - openrouter nemotron-3-ultra-550b-a55b:free — 1M ctx, currently the
 *     biggest free model anywhere, but generation takes 180s+ even on trivial
 *     prompts (heavily congested), so it's seeded enabled=0. Flip it on when
 *     it serves sanely. (Tools were unverifiable on the OR route while it
 *     hangs; V24 verified the model family's structured tool_calls via Zen's
 *     dedicated endpoint and added the V22 rule, so this seed is now 1.)
 *   - openrouter llama-3.2-3b-instruct:free and
 *     dolphin-mistral-24b-venice-edition:free — both heavily contended
 *     (persistent upstream 429s at probe time) but real routes; no tools.
 *   - NOT added: nvidia/nemotron-3.5-content-safety:free — it's a safety
 *     classifier, not a chat model: probe returned content:null with a "User
 *     Safety: safe/unsafe" verdict in reasoning, which would surface as empty
 *     responses whenever the fallback chain landed on it.
 *
 * Idempotent: DELETEs re-run harmlessly; INSERT OR IGNORE means the ultra
 * row's enabled=0 is seed-only, so a dashboard re-enable sticks across boots.
 * vision/tools seeds match the V16/V22 rules (which re-assert next boot);
 * tiers match V17's bands.
 */
function migrateModelsV23FreeTierAudit(db: Database.Database) {
  const apply = db.transaction(() => {
    for (const platform of ['sambanova', 'chutes']) {
      db.prepare(`
        DELETE FROM fallback_config WHERE model_db_id IN (
          SELECT id FROM models WHERE platform = ?
        )
      `).run(platform);
      db.prepare('DELETE FROM models WHERE platform = ?').run(platform);
      db.prepare('DELETE FROM api_keys WHERE platform = ?').run(platform);
    }

    const insert = db.prepare(`
      INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window, enabled, supports_vision, supports_tools)
      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
    `);
    const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null, number, number, number]> = [
      ['openrouter', 'moonshotai/kimi-k2.6:free',                                     'Kimi K2.6 (OR free)',                 3, 9,  'Frontier', 20, 200, null, null, '~6M',  262144,  1, 0, 1],
      ['openrouter', 'nvidia/nemotron-3-ultra-550b-a55b:free',                        'Nemotron 3 Ultra 550B (free, slow)',  7, 11, 'Frontier', 20, 200, null, null, '~6M',  1000000, 0, 0, 1],
      ['openrouter', 'nvidia/nemotron-nano-12b-v2-vl:free',                           'Nemotron Nano 12B VL (free)',        26, 9,  'Medium',   20, 200, null, null, '~6M',  128000,  1, 1, 1],
      ['openrouter', 'meta-llama/llama-3.2-3b-instruct:free',                         'Llama 3.2 3B (free)',                30, 9,  'Small',    20, 200, null, null, '~6M',  131072,  1, 0, 0],
      ['openrouter', 'cognitivecomputations/dolphin-mistral-24b-venice-edition:free', 'Dolphin Mistral 24B Venice (free)',  25, 9,  'Medium',   20, 200, null, null, '~6M',  32768,   1, 0, 0],
      ['zhipu',      'glm-4.6v-flash',                                                'GLM-4.6V Flash',                     21, 4,  'Large',    null, null, null, null, '~30M', 131072,  1, 1, 1],
    ];
    for (const a of additions) insert.run(...a);
    backfillFallback(db);
  });
  apply();
}

/**
 * V24 (June 2026): OpenCode Zen roster refresh + NIM gemma pause. All claims
 * live-probed 2026-06-07 with a real Zen account key through the gateway.
 *
 * Zen's promo roster rotated since V18:
 *   - ADDED nemotron-3-ultra-free — newly docs-confirmed free, and unlike the
 *     OpenRouter ultra route (which hangs: 0 tokens in 140s even streaming),
 *     Zen serves it from a dedicated vLLM endpoint in ~2s WITH structured
 *     tool_calls. Currently the biggest usable free model anywhere.
 *   - ADDED minimax-m3-free — requested in #242 (thanks @Naster17). Live
 *     probe: 2s chat, finish_reason:tool_calls. CAVEAT: not in Zen's docs
 *     free table (same undocumented status qwen3.6-plus-free had before its
 *     promo ended), so it may transition without notice — watchlist.
 *   - qwen3.6-plus-free now errors "Free promotion has ended" — confirms the
 *     docs-confirmed-only bar; never added, nothing to remove.
 *   - nemotron-3-super-free DROPPED OUT of Zen's docs free table but still
 *     serves (routes to the same congested NVIDIA :free upstream as
 *     OpenRouter, ~15s). Kept enabled; prune if it starts billing or 402s.
 *   - big-pickle unmasked: responses report model "deepseek-v4-flash".
 *
 * Also pauses nvidia/google/gemma-4-31b-it: still listed on NIM's /v1/models
 * (only gemma-4 there) but every chat probe hangs 90s+ even on tiny prompts —
 * NIM free-tier serverless capacity starvation (forum 504 reports, May–Jun
 * 2026) compounded by a known gemma-4-31b dense Flash-Attention prefill
 * deadlock upstream. Disabled like the V13 disables (re-asserted each boot);
 * re-enable via a future migration when NIM serves it again. Gemma-4 coverage
 * remains via google/openrouter/cloudflare rows.
 *
 * Idempotent: INSERT OR IGNORE + always-run UPDATEs, safe to re-run.
 */
function migrateModelsV24ZenRefresh(db: Database.Database) {
  const apply = db.transaction(() => {
    const insert = db.prepare(`
      INSERT OR IGNORE INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, size_label, rpm_limit, rpd_limit, tpm_limit, tpd_limit, monthly_token_budget, context_window, enabled, supports_vision, supports_tools)
      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
    `);
    const additions: Array<[string, string, string, number, number, string, number | null, number | null, number | null, number | null, string, number | null, number, number, number]> = [
      ['opencode', 'nemotron-3-ultra-free', 'Nemotron 3 Ultra Free (OpenCode Zen)',  7, 4, 'Frontier', 20, 200, null, null, 'promo (trial)', 131072, 1, 0, 1],
      ['opencode', 'minimax-m3-free',       'MiniMax M3 Free (OpenCode Zen)',        4, 4, 'Frontier', 20, 200, null, null, 'promo (trial)', 131072, 1, 0, 1],
    ];
    for (const a of additions) insert.run(...a);
    backfillFallback(db);

    db.prepare(`
      UPDATE models SET enabled = 0
       WHERE platform = 'nvidia' AND model_id = 'google/gemma-4-31b-it'
    `).run();
  });
  apply();
}

// Embeddings V1 (2026-06): per-family embedding catalog. A "family" is one
// model identity + dimension — vectors from different families live in
// incompatible spaces, so /v1/embeddings only ever fails over WITHIN a family
// (same model served by another provider), never across families.
// Every entry was live-verified against the provider on 2026-06-04.
function migrateEmbeddingsV1(db: Database.Database) {
  db.exec(`
    CREATE TABLE IF NOT EXISTS embedding_models (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      family TEXT NOT NULL,
      platform TEXT NOT NULL,
      model_id TEXT NOT NULL,
      display_name TEXT NOT NULL,
      dimensions INTEGER NOT NULL,
      max_input_tokens INTEGER,
      priority INTEGER NOT NULL DEFAULT 0,
      enabled INTEGER NOT NULL DEFAULT 1,
      quota_label TEXT NOT NULL DEFAULT '',
      UNIQUE(platform, model_id)
    );
  `);

  // Tag request rows so embeddings traffic doesn't pollute the chat token
  // budget / headroom math. Existing rows backfill to 'chat' via the default.
  const columns = db.prepare('PRAGMA table_info(requests)').all() as { name: string }[];
  if (!columns.some(col => col.name === 'request_type')) {
    db.prepare("ALTER TABLE requests ADD COLUMN request_type TEXT NOT NULL DEFAULT 'chat'").run();
  }

  const seed = db.prepare(`
    INSERT OR IGNORE INTO embedding_models
      (family, platform, model_id, display_name, dimensions, max_input_tokens, priority, enabled, quota_label)
    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
  `);
  const rows: Array<[string, string, string, string, number, number | null, number, number, string]> = [
    // family, platform, provider model id, display name, dims, max input, priority, enabled, quota label
    ['gemini-embedding-001', 'google', 'gemini-embedding-001', 'Gemini Embedding', 3072, 2048, 1, 1, '100 rpm · 1K req/day'],
    ['llama-nemotron-embed-vl-1b-v2', 'nvidia', 'nvidia/llama-nemotron-embed-vl-1b-v2', 'Nemotron Embed VL 1B', 2048, 8192, 1, 1, '~40 rpm'],
    ['llama-nemotron-embed-vl-1b-v2', 'openrouter', 'nvidia/llama-nemotron-embed-vl-1b-v2', 'Nemotron Embed VL 1B (OR)', 2048, 8192, 2, 1, '$0/M tok'],
    ['llama-nemotron-embed-1b-v2', 'nvidia', 'nvidia/llama-nemotron-embed-1b-v2', 'Nemotron Embed 1B', 2048, 8192, 1, 1, '~40 rpm'],
    ['nv-embedqa-e5-v5', 'nvidia', 'nvidia/nv-embedqa-e5-v5', 'NV-EmbedQA E5 v5', 1024, 512, 1, 1, '~40 rpm'],
    ['text-embedding-3-small', 'github', 'openai/text-embedding-3-small', 'Text Embedding 3 Small', 1536, 8191, 1, 1, 'rate-limited free'],
    ['text-embedding-3-large', 'github', 'openai/text-embedding-3-large', 'Text Embedding 3 Large', 3072, 8191, 1, 1, 'rate-limited free'],
    ['bge-m3', 'cloudflare', '@cf/baai/bge-m3', 'BGE-M3', 1024, 8192, 1, 1, '10K neurons/day (shared)'],
    ['bge-m3', 'huggingface', 'BAAI/bge-m3', 'BGE-M3 (HF)', 1024, 8192, 2, 1, '$0.10/mo credits'],
    ['embeddinggemma-300m', 'cloudflare', '@cf/google/embeddinggemma-300m', 'EmbeddingGemma 300M', 768, 2048, 1, 1, '10K neurons/day (shared)'],
    ['qwen3-embedding-0.6b', 'cloudflare', '@cf/qwen/qwen3-embedding-0.6b', 'Qwen3 Embedding 0.6B', 1024, 4096, 1, 1, '10K neurons/day (shared)'],
    // Cohere trial keys allow 1,000 calls/month TOTAL shared with chat —
    // disabled by default so embedding traffic can't silently eat chat quota.
    ['embed-v4.0', 'cohere', 'embed-v4.0', 'Cohere Embed v4', 1536, 128000, 1, 0, '1K calls/mo (shared w/ chat)'],
  ];
  const apply = db.transaction(() => { for (const r of rows) seed.run(...r); });
  apply();

  const def = db.prepare("SELECT value FROM settings WHERE key = 'embeddings_default_family'").get();
  if (!def) {
    db.prepare("INSERT INTO settings (key, value) VALUES ('embeddings_default_family', 'gemini-embedding-001')").run();
  }
}

/** Append any models not yet in the fallback chain, lowest priority, ordered by
 * intelligence_rank. Shared by the recent model migrations (V18–V20). */
function backfillFallback(db: Database.Database) {
  const missing = db.prepare(`
    SELECT m.id FROM models m
    LEFT JOIN fallback_config f ON m.id = f.model_db_id
    WHERE f.id IS NULL ORDER BY m.intelligence_rank ASC
  `).all() as { id: number }[];
  if (missing.length > 0) {
    const maxPriority = (db.prepare('SELECT COALESCE(MAX(priority), 0) AS mx FROM fallback_config').get() as { mx: number }).mx;
    const addFb = db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, ?, 1)');
    for (let i = 0; i < missing.length; i++) addFb.run(missing[i].id, maxPriority + i + 1);
  }
}

function ensureUnifiedKey(db: Database.Database) {
  const existing = db.prepare("SELECT value FROM settings WHERE key = 'unified_api_key'").get() as { value: string } | undefined;
  if (!existing) {
    const key = `freellmapi-${crypto.randomBytes(24).toString('hex')}`;
    db.prepare("INSERT INTO settings (key, value) VALUES ('unified_api_key', ?)").run(key);
    console.log(`\n  Your unified API key: ${key}\n`);
  }
}