@latten/adapter-codex

See what your Codex sessions cost. Locally, in seconds.

This adapter parses Codex CLI rollouts (~/.codex/sessions/**/*.jsonl) into a Latten session model, so the @latten/cost detectors can point at recoverable spend — redundant round trips, stuck loops, uncached context — on your machine, with nothing sent anywhere.

Install

npm install @latten/adapter-codex @latten/cost

Usage

Three lines to your last 24 hours:

import { readCodexSessions } from '@latten/adapter-codex'
import { detectStuckLoop } from '@latten/cost'

const sessions = await readCodexSessions({ sinceMs: Date.now() - 86_400_000 })
sessions.forEach((s) => console.log(s.sessionId, detectStuckLoop(s)))

Need control? findCodexRollouts() and parseCodexRollout() are the pieces readCodexSessions() is made of — bring your own file list, inject a price function per model, or parse a single transcript string.

Prefer zero code? npx @latten/glow reads the same rollouts and shows your usage across every coding agent in one view.

What it extracts

• Per-call token counts from Codex token_count events — input, output, and the cached portion split out, so cache misses are visible.
• Model name per call (from the turn context), for accurate pricing.
• Tool calls (shell, exec, web_search, …) with stable hashes of their arguments, and tool outputs hashed for equality checks.
• Read-only discovery classification for common tools and shell commands, so detectors never mistake git status for real work.
• Assistant text hashes for repeated-output loop detection.

Privacy model

The session files this adapter reads contain your prompts, code, and command output. None of that survives parsing:

• prompts and developer messages are dropped entirely, not even hashed;
• tool arguments, outputs, and assistant text are reduced to on-device SHA-256 hashes, used only for local equality checks;
• token counts and tool types are the only facts that leave the parser;
• pricing is injected by the caller — the parser never touches the network;
• malformed or partial JSONL lines are ignored.

This is the same values-never contract as every Latten package: types and counts, never values. The test suite asserts it — serialize a parsed session and the raw content is provably absent.

Part of the Latten session model

@latten/adapter-claude-code does the same for Claude Code transcripts, and @latten/adapter-gemini covers Gemini CLI (honestly — see its README). One session model, every agent your team runs.

You moved fast with AI. Latten helps you stay in control of it. See what your AI costs — and where your data goes: Latten turns the AI calls across your company into one live graph — cost, reach, and PII exposure, attributed and observed from real traffic. One prompt to your coding agent, a PR you review, live in minutes. latten.io