@latten/adapter-gemini

An adapter that refuses to guess.

Gemini CLI keeps local records of your sessions in two generations. Newer CLIs write chats/session-*.jsonl whose turns carry real token counts — input, output, cached, thoughts, tool — and this adapter parses them (fidelity: 'tokens-only'), splitting the cached portion out of input so cache misses are visible. Older CLIs wrote logs.json and chats/session-*.json with messages only — for those it reports exactly that: sessions, message counts, timestamps, labeled fidelity: 'messages-only'.

What it never does is estimate. Every result tells you which of the two it is; a number you can't observe is a number you shouldn't report.

Install

npm install @latten/adapter-gemini

Usage

Three lines to your last 24 hours:

import { readGeminiChats } from '@latten/adapter-gemini'

const chats = await readGeminiChats({ sinceMs: Date.now() - 86_400_000 })
chats.forEach((c) => console.log(c.sessionId, c.fidelity, c.tokens ?? `${c.messages} messages`))

Every chat comes back at the fidelity its file supports: tokens-only (real per-turn counts, per model) or messages-only (honest counts, nothing invented). readGeminiLogs() does the same for the older logs.json records.

Need control? findGeminiChats() / parseGeminiChat() are the pieces the one-liner is made of — bring your own file list or parse a single string.

npx @latten/glow uses the same honest read to show Gemini alongside Claude Code and Codex in one usage view.

What it reads — and what it deliberately doesn't

Reads: session ids, message ids, message types, timestamps, models, and token counts. Never reads: message text, thought text, or tool arguments. The message, content, displayContent, thoughts, and toolCalls payloads are not accessed, not hashed, not counted by size — they simply never enter the model. The test suite asserts it: serialize any parsed result and the raw text is provably absent.

Types and counts, never values — the same contract as every Latten package.

Part of the Latten session model

@latten/adapter-claude-code and @latten/adapter-codex parse full token-level sessions for their agents. One session model, every agent your team runs — each reported at the fidelity its files actually support.

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