AgentHub/server/lib/pricing-sync.js

/**
 * @file Automatic pricing for Claude, OpenAI (a Codex session's — see
 * codex-usage.js) and Gemini (an Antigravity session's — antigravity-usage.js)
 * models, from LiteLLM's public price list fetched at most once a day. A row
 * (`source = 'sync'`) is INSERTED for every family used in `token_usage` with
 * no `<family>%` row of its own (a `-YYYYMMDD` / `-YYYY-MM-DD` snapshot suffix
 * dropped; "its own row" is the test — Opus 5.5 was once priced as Opus 5 by
 * the broader `claude-opus-5%`) and for every current-generation chat model of
 * the three vendors before anyone uses it (`isCurrent`). The sync's own rows
 * follow a published rate change; a row from before the `source` column is
 * adopted only while it still equals the list; a `'user'` row (edited in
 * Settings) and a deleted pattern (`model_pricing_dismissed`) are never
 * touched. Off with AGENTHUB_PRICING_SYNC=0|false|off. An entry is taken only
 * from its own vendor (`anthropic` for claude-…, `openai` for gpt-…, o<n>…,
 * codex-…, Vertex AI or the Gemini API for gemini-…); a router's model id
 * (`clinepass/…`, `Highest`) is left to the user.
 *
 * A non-Claude model's PRIORITY rates (`*_cost_per_token_priority`) fill the
 * fast columns: Codex's fast/priority service tier is booked as `speed = fast`
 * (`codex-usage.js`), and its flex tier as `service_tier = 'flex'` (half price,
 * like batch, in `ratesForBucket`).
 * A Claude model's `*_above_100k_tokens` input/output rates (Haiku 5.5) fill
 * the long_* columns, priced against token_usage's long-context subsets.
 * ponytail: OpenAI's long-context (>272k) and Gemini's >200k-token rates are
 * not modelled — Codex's context window (258k) stays under 272k, and a
 * threshold other than LONG_CONTEXT_TOKENS (100k) needs its own subsets.
 * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
 */

const PRICE_LIST_URL =
  "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json";
const CHECK_MS = 15 * 60_000;
const CACHE_MS = 24 * 60 * 60_000;
/** A per-MTok rate above this is a parse error, not a price. */
const MAX_PER_MTOK = 1000;

let cached = null; // { at, list }
let failedAt = -Infinity; // last failed fetch

const OPENAI = /^(gpt-|o\d|codex-)[a-z0-9.-]*$/;
const GEMINI = /^gemini-[a-z0-9.-]+$/;
const PROVIDERS = {
  claude: ["anthropic"],
  gemini: ["vertex_ai-language-models", "gemini"],
  openai: ["openai"],
};
const vendorOf = (fam) =>
  fam.startsWith("claude-") ? "claude" : fam.startsWith("gemini-") ? "gemini" : "openai";

/**
 * `claude-opus-5-5-20260101` → `claude-opus-5-5`, `gpt-5.4-2026-03-05` →
 * `gpt-5.4`; null for anything else.
 */
function familyOf(model) {
  if (typeof model !== "string") return null;
  if (model.startsWith("claude-")) {
    const fam = model.replace(/-\d{8}$/, "");
    return /^claude-[a-z0-9-]+$/.test(fam) ? fam : null;
  }
  if (model.startsWith("gemini-")) return GEMINI.test(model) ? model : null;
  const fam = model.replace(/-\d{4}-\d{2}-\d{2}$/, "");
  return OPENAI.test(fam) ? fam : null;
}

const cap = (w) => w[0].toUpperCase() + w.slice(1);

/**
 * `claude-opus-5-5` → `Claude Opus 5.5`, `gpt-5.6-sol` → `GPT-5.6 Sol`,
 * `gemini-3.8-flash` → `Gemini 3.8 Flash`.
 */
function displayName(fam) {
  if (fam.startsWith("gemini-")) return fam.split("-").map(cap).join(" ");
  if (!fam.startsWith("claude-")) {
    const [head, ...rest] = fam.split("-");
    if (head === "gpt" && rest.length) {
      const [ver, ...words] = rest;
      return [`GPT-${ver}`, ...words.map(cap)].join(" ");
    }
    return [head === "codex" ? "Codex" : head, ...rest.map(cap)].join(" ");
  }
  const [, ...parts] = fam.split("-");
  const words = [];
  const nums = [];
  for (const p of parts) (/^\d+$/.test(p) ? nums : words).push(p);
  const name = words.map(cap).join(" ");
  return `Claude ${name}${nums.length ? " " + nums.join(".") : ""}`;
}

/** One LiteLLM entry → a model_pricing row, or null when it is not usable. */
function toRow(fam, e) {
  const claude = fam.startsWith("claude-");
  if (!e || !PROVIDERS[vendorOf(fam)].includes(e.litellm_provider)) return null;
  const per = (v) => (typeof v === "number" ? v * 1e6 : NaN);
  const input = per(e.input_cost_per_token);
  const output = per(e.output_cost_per_token);
  // Published multipliers stand in for a field the list omits: Anthropic
  // charges cache writes (1.25x / 2x), OpenAI does not.
  const read = per(e.cache_read_input_token_cost ?? e.input_cost_per_token * 0.1);
  const w5m = per(
    e.cache_creation_input_token_cost ?? (claude ? e.input_cost_per_token * 1.25 : 0)
  );
  const w1h = per(
    e.cache_creation_input_token_cost_above_1hr ??
      (claude ? e.input_cost_per_token * 2 : (e.cache_creation_input_token_cost ?? 0))
  );
  const rates = [input, output, read, w5m, w1h];
  if (!rates.every((r) => Number.isFinite(r) && r >= 0 && r <= MAX_PER_MTOK)) return null;
  if (input === 0 && output === 0) return null;
  const fast = claude && e.supports_fast_mode ? Number(e.provider_specific_entry?.fast) : 0;
  const fastOk = Number.isFinite(fast) && fast > 0 && fast <= 10;
  const prio = claude
    ? []
    : [per(e.input_cost_per_token_priority), per(e.output_cost_per_token_priority)];
  const prioOk =
    prio.length > 0 && prio.every((v) => Number.isFinite(v) && v > 0 && v <= MAX_PER_MTOK);
  const long = claude
    ? [
        per(e.input_cost_per_token_above_100k_tokens),
        per(e.output_cost_per_token_above_100k_tokens),
      ]
    : [];
  const longOk =
    long.length > 0 && long.every((v) => Number.isFinite(v) && v > 0 && v <= MAX_PER_MTOK);
  // Rounded: 0.000004 * 1e6 is 3.9999999999999996, which Settings would show.
  const r = (n) => Math.round(n * 1e6) / 1e6;
  return {
    pattern: `${fam}%`,
    name: displayName(fam),
    input: r(input),
    output: r(output),
    read: r(read),
    w5m: r(w5m),
    w1h: r(w1h),
    fastIn: fastOk ? r(input * fast) : prioOk ? r(prio[0]) : 0,
    fastOut: fastOk ? r(output * fast) : prioOk ? r(prio[1]) : 0,
    longIn: longOk ? r(long[0]) : 0,
    longOut: longOk ? r(long[1]) : 0,
  };
}

/** Families used in token_usage that have no `<family>%` row of their own. */
function missingFamilies(db) {
  const have = new Set(
    db
      .prepare("SELECT model_pattern FROM model_pricing")
      .all()
      .map((r) => r.model_pattern)
  );
  const out = new Set();
  for (const { model } of db.prepare("SELECT DISTINCT model FROM token_usage").all()) {
    const fam = familyOf(model);
    if (fam && !have.has(`${fam}%`)) out.add(fam);
  }
  return [...out];
}

// ponytail: "current" is a fixed generation floor per vendor, not a release
// date — raise MIN_GEN when a generation retires. Anything older still gets a
// row the moment it is used (missingFamilies).
const MIN_GEN = { openai: 5, gemini: 3 };
const NOT_A_CODING_MODEL =
  /(audio|search|tts|realtime|transcribe|image|computer-use|robotics|embedding|moderation|-exp-|latest$)/;

/**
 * A price-list key worth a row before anyone uses it: a family id itself (no
 * snapshot date), a chat model, not deprecated, of the vendor's current
 * generation.
 */
/**
 * Retired models (Claude 3.x, Opus 4/4.1, Sonnet 4): out of the defaults and
 * pruned once from existing DBs by server/db.js. The sync prices one only when
 * it is actually used — never as a "current" model added in advance.
 */
const RETIRED_PRICING_PATTERNS = [
  "claude-opus-4-1%",
  "claude-opus-4-2%",
  "claude-sonnet-4-2%",
  "claude-3-7-sonnet%",
  "claude-3-5-sonnet%",
  "claude-3-5-haiku%",
  "claude-3-haiku%",
  "claude-3-opus%",
];
// `claude-opus-4-2%` is Opus 4's dated ids (claude-opus-4-20250514): family claude-opus-4.
const RETIRED_FAMILIES = new Set(RETIRED_PRICING_PATTERNS.map((p) => p.replace(/(-2)?%$/, "")));

function isCurrent(key, e, today) {
  const fam = familyOf(key);
  if (fam !== key || !e || !["chat", "responses"].includes(e.mode)) return false;
  if (typeof e.deprecation_date === "string" && e.deprecation_date < today) return false;
  if (NOT_A_CODING_MODEL.test(fam) || RETIRED_FAMILIES.has(fam)) return false;
  const vendor = vendorOf(fam);
  if (vendor === "claude") return !/^claude-\d/.test(fam); // claude-3-5-sonnet naming = retired
  const m = fam.match(/^(?:gpt|gemini)-(\d+(?:\.\d+)?)/);
  return Boolean(m) && Number(m[1]) >= MIN_GEN[vendor];
}

const RATE_COLS = [
  ["input", "input_per_mtok"],
  ["output", "output_per_mtok"],
  ["read", "cache_read_per_mtok"],
  ["w5m", "cache_write_per_mtok"],
  ["w1h", "cache_write_1h_per_mtok"],
  ["fastIn", "fast_input_per_mtok"],
  ["fastOut", "fast_output_per_mtok"],
  ["longIn", "long_input_per_mtok"],
  ["longOut", "long_output_per_mtok"],
];
const sameRates = (row, db) => RATE_COLS.every(([k, col]) => Math.abs(row[k] - db[col]) < 1e-9);

async function defaultFetch() {
  const res = await fetch(PRICE_LIST_URL, { signal: AbortSignal.timeout(20_000) });
  if (!res.ok) throw new Error(`price list HTTP ${res.status}`);
  return res.json();
}

let appliedAt = null; // cached.at of the list last applied

/**
 * One pass: fetch the list at most once a day, then (1) add a row for every
 * used family with none and every current model of the three vendors, (2)
 * keep the sync's own rows (`source = 'sync'`) at the published rate, and (3)
 * adopt a row from before the `source` column while it still equals the list.
 * A user's row (`'user'`) and a dismissed pattern are never touched. Returns
 * the patterns inserted. Never throws: pricing is a display concern and a
 * failed fetch must not reach the server's startup path.
 */
async function syncPricing(db, { fetchList = defaultFetch, now = Date.now() } = {}) {
  try {
    const missing = missingFamilies(db);
    const stale = !cached || now - cached.at > CACHE_MS;
    // The Codex poll calls this every 15 s while a session runs: nothing new
    // to apply and nothing missing is the common answer, and it is free.
    if (!stale && appliedAt === cached.at && missing.length === 0) return [];
    if (stale) {
      // A failed fetch is retried at the sync's own cadence, not on every caller.
      if (now - failedAt < CHECK_MS) return [];
      try {
        cached = { at: now, list: await fetchList() };
      } catch (err) {
        failedAt = now;
        throw err;
      }
    }
    const list = cached.list || {};
    const today = new Date(now).toISOString().slice(0, 10);
    const rows = new Map(
      db
        .prepare("SELECT * FROM model_pricing")
        .all()
        .map((r) => [r.model_pattern, r])
    );
    const dismissed = new Set(
      db
        .prepare("SELECT model_pattern FROM model_pricing_dismissed")
        .all()
        .map((r) => r.model_pattern)
    );
    const families = new Set(missing);
    for (const [key, e] of Object.entries(list)) if (isCurrent(key, e, today)) families.add(key);
    for (const pattern of rows.keys()) {
      const fam = pattern.endsWith("%") ? pattern.slice(0, -1) : null;
      if (fam && familyOf(fam) === fam) families.add(fam);
    }

    const insert = db.prepare(
      `INSERT OR IGNORE INTO model_pricing (model_pattern, display_name, input_per_mtok,
         output_per_mtok, cache_read_per_mtok, cache_write_per_mtok, cache_write_1h_per_mtok,
         fast_input_per_mtok, fast_output_per_mtok, long_input_per_mtok, long_output_per_mtok, source)
         VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 'sync')`
    );
    const update = db.prepare(
      `UPDATE model_pricing SET display_name = ?, input_per_mtok = ?, output_per_mtok = ?,
         cache_read_per_mtok = ?, cache_write_per_mtok = ?, cache_write_1h_per_mtok = ?,
         fast_input_per_mtok = ?, fast_output_per_mtok = ?,
         long_input_per_mtok = ?, long_output_per_mtok = ?,
         updated_at = strftime('%Y-%m-%dT%H:%M:%fZ', 'now')
       WHERE model_pattern = ? AND source = 'sync'`
    );
    const adopt = db.prepare(
      "UPDATE model_pricing SET source = 'sync' WHERE model_pattern = ? AND source IS NULL"
    );
    const added = [];
    const changed = [];
    db.transaction(() => {
      for (const fam of families) {
        const pattern = `${fam}%`;
        if (dismissed.has(pattern)) continue;
        const row = toRow(fam, list[fam]);
        if (!row) continue;
        const rates = [
          row.input,
          row.output,
          row.read,
          row.w5m,
          row.w1h,
          row.fastIn,
          row.fastOut,
          row.longIn,
          row.longOut,
        ];
        const have = rows.get(pattern);
        if (!have) {
          if (insert.run(pattern, row.name, ...rates).changes) added.push(pattern);
        } else if (have.source === "sync") {
          if (!sameRates(row, have) && update.run(have.display_name, ...rates, pattern).changes) {
            changed.push(pattern);
          }
        } else if (have.source == null && !have.intro_until && sameRates(row, have)) {
          adopt.run(pattern);
        }
      }
    })();
    appliedAt = cached.at;
    if (added.length) console.log(`[pricing] added rates for ${added.join(", ")}`);
    if (changed.length) console.log(`[pricing] updated rates for ${changed.join(", ")}`);
    return added;
  } catch (err) {
    console.warn("[pricing] sync failed:", err.message);
    return [];
  }
}

function syncEnabled() {
  return !/^(0|false|off)$/i.test(process.env.AGENTHUB_PRICING_SYNC || "");
}

function startPricingSync(db) {
  if (!syncEnabled()) return;
  void syncPricing(db);
  setInterval(() => void syncPricing(db), CHECK_MS).unref();
}

function _resetCache() {
  cached = null;
  failedAt = -Infinity;
  appliedAt = null;
}

/** Make the next pass re-apply the cached list (after a reset emptied the table). */
function invalidate() {
  appliedAt = null;
}

/** Test seam: pretend the price list was just fetched. */
function _setCache(list, at = Date.now()) {
  cached = { at, list };
  appliedAt = null;
}

module.exports = {
  RETIRED_PRICING_PATTERNS,
  syncPricing,
  invalidate,
  syncEnabled,
  startPricingSync,
  familyOf,
  displayName,
  toRow,
  isCurrent,
  _resetCache,
  _setCache,
};