---
title: 'In the News: November 17, 2026 (Extra)'
description: 'tokendoctor replays your own Claude Code sessions to show what a smaller context window would have saved, and logs the runs where it did not.'
canonical_url: 'https://darkfactory.dev/news/2026-11-17-extra'
markdown_url: 'https://darkfactory.dev/news/2026-11-17-extra.md'
collection: news
date_published: '2026-11-17T10:05:00-05:00'
date_modified: '2026-11-17T10:05:00-05:00'
---

# In the News: November 17, 2026 (Extra)


tokendoctor, a small open-source plugin for Claude Code, prices a smaller context window against your own past sessions instead of someone else's benchmark. Its author reports a 38% drop in cost per request, labels that figure as one person's before and after, and publishes the benchmark runs where the plugin's own advice cost more.

## 1. tokendoctor replays your Claude Code sessions to price a smaller context window

**[tokendoctor](https://github.com/localixai/tokendoctor)** · Localix Lab · GitHub, MIT license, repository created October 25, 2026

The first command, `/tokendoctor:sessions`, replays the last 30 days of Claude Code transcripts request by request at compaction windows of 200K, 300K, 400K and 600K tokens and reports what each would have cost. It makes no model calls and changes nothing. The README calls the result "an upper bound" because a replay cannot see work the model would redo after a summary. A second command proposes up to five settings changes, shows each as a diff, applies only the ones the user picks, and can undo them. An optional mode drops old tool output instead of summarizing when Claude Code would otherwise compact.

The README's argument is about where long sessions spend money. It says Claude Code on 1M-window models compacts only near 967K tokens, and that for the author "two thirds of main-thread cost went to requests with more than 400K in context." After switching to a 400K window and Sonnet for subagents, the author reports context per request falling from 412K to 219K tokens and cost per request falling 38% at API list prices, comparing about 23K requests in the month before with 19K in the week after. The README calls this "one person's before and after, not a controlled experiment," and names the trade-off: more summaries, each a pause of 79 seconds on average in the author's sessions.

The repository also ships a benchmark in which Claude Code, on Haiku 5.5, builds a TypeScript project against tests it cannot change. There the settings advice alone was more expensive than stock Claude Code in three of four runs; the README says these tasks rarely pass 400K, where the window starts to matter. Dropping tool output instead of summarizing cut cost 19% on a day of work and 35% on an eval harness, and raised it 6% on a short task. Every run, including the losing ones, is logged in `bench/HISTORY.md`, and many rows rest on a single run. The plugin had 28 points and 14 comments on [Hacker News](https://news.ycombinator.com/item?id=50526584) about 13 hours after posting.

**Why it matters:** The 38% is one developer's number. The method is the useful part: a team paying for long agent sessions can see what a smaller compaction window would have saved on its own transcripts before changing a setting, and the published benchmark already shows the change does not pay when sessions stay short.
