@juspay/neurolink/dist/localUsage/codexReader.d.ts

The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002

/**
 * Reads token usage out of Codex's own rollout transcripts.
 *
 * Layout: `~/.codex/sessions/<yyyy>/<mm>/<dd>/rollout-<iso>-<sessionId>.jsonl`.
 * Each line is one of `session_meta`, `turn_context`, `response_item` or
 * `event_msg`; usage lives on `event_msg` records whose `payload.type` is
 * `token_count`.
 *
 * The counter semantics are the whole story here, and they are the opposite of
 * the Claude Code reader's. Every `token_count` event carries BOTH a
 * cumulative `total_token_usage` for the session and a per-turn
 * `last_token_usage`. Summing the per-turn values is the obvious move and it is
 * wrong: measured across all 108 sessions on a real machine, summing yields
 * 9,191,613,238 tokens against a true 5,653,217,442 — an overstatement of
 * 62.6%, and 195% on the worst single session. The per-turn value repeats
 * across events within a turn, so it double-counts.
 *
 * The cumulative counter is authoritative and was monotonic in all 108
 * sessions — it never resets mid-file — so the last one in the file is the
 * session's true total. `Math.max` is still used rather than "last seen",
 * because a counter that is only monotonic in every case observed is not the
 * same as one that is guaranteed to be, and the max costs nothing.
 *
 * Cost is deliberately not computed. Codex is a ChatGPT subscription — the
 * rollouts carry `rate_limits.plan_type` — so a per-token dollar figure would
 * be an invention, not a measurement. The tokens are real; the cost is
 * `unavailable`.
 */
import type { LocalUsageReader } from "../types/index.js";
export declare function createCodexReader(): Promise<LocalUsageReader>;