ClaudeMods
☰
EN
● — online · Views — times
SponsorsSubmit a project
GitHub repositories · by korro_ai

I made a Claude Code mod that cuts ~30 % of my token cost, plus a pixel-art theme to go with it

Seuil is an open-source Claude Code mod that monitors conversation size after each turn and auto-compacts past a configurable threshold (default 300k tokens) to reduce token cost, with a live statusline and /seuil commands; a pixel-art Campfire theme is included. The author reports ~30% savings across 144 replayed sessions, plus a read-only local simulator script.

Translated

About this mod

The post describes Seuil: after each turn it measures the real conversation size and, past 300k tokens (adjustable via /seuil 250, disabled via /seuil off), compacts between turns, keeping decisions, work state and file paths. A statusline shows progress (Seuil · 182k / 300k). A benchmark script replays local sessions read-only to estimate savings, which the author reports as ~30% overall (27% with a pessimistic 50k-token loss assumption), concentrated in long sessions. An optional Windows Terminal pixel-art 'Campfire' theme is bundled. Install: /plugin install seuil --marketplace KorroAi/seuil; repo at github.com/KorroAi/seuil.

Installation

Check the author's README for the marketplace and plugin name first. Commands may change as the repository evolves.

claude plugin marketplace add KorroAi/seuil
claude plugin install seuil@seuil
Original text / README
<div align="center">

Seuil

Automatic compaction for Claude Code: between two turns, once the conversation passes a threshold.

Every model call re-reads the whole conversation, so a long session pays mostly for re-reading.<br> Seuil keeps the conversation short without you thinking about it. Seuil is French for threshold.

License: MIT Claude Code plugin Tests Paper

Seuil in 15 seconds: every call re-reads the conversation; past 300 k tokens Seuil compacts it between two turns

<sub>Seuil in 15 seconds. The compaction shown (366 k → 47 k) comes from a real session, the 30 % from the paper. <a href="https://github.com/KorroAi/seuil/raw/main/docs/media/seuil-demo.mp4">Watch it in HD (1080p)</a>.</sub>

</div>

Why

On 144 real sessions (30 days, 33,634 model calls), replayed with automatic compaction at 300 k tokens:

  • about 30 % of the total cost saved, 27 % if each summary forgets 50 k tokens that must be read again;
  • about 20 % for a user who already restarts long sessions by hand;
  • nothing for sessions that stay short: all the gain is in long sessions.

Trimming tool outputs instead, the obvious idea, would save 0.1 % without loss and 5 % at most. The method, the tables and the limits are in the paper. bench/simulate.mjs runs the same replay on your own sessions.

Saving by threshold

Install

At the prompt of a Claude Code terminal session:

/plugin install seuil --marketplace KorroAi/seuil

Answer y to add the marketplace, then pick a scope (user scope = every session). From a shell:

claude plugin install seuil --marketplace KorroAi/seuil

In the desktop app: click + next to the prompt box, then Plugins, Add marketplace with KorroAi/seuil, then Add plugin.

Seuil uses the Claude Code mod API (function hooks), which is in early access. Checked on Claude Code 2.1.325.

What it does

  • At the end of each turn it reads the real size of the conversation from Claude Code.

  • Above the threshold (300 k tokens by default) it starts Claude Code's own compaction, with instructions that ask the summary to keep the goal, every decision and correction you gave (quoted), the exact state of the work, the files with their paths, what remains to do, and the pitfalls met.

  • It never compacts in the middle of a turn or inside a subagent.

  • It does not start again right after a compaction that left the conversation close to the threshold: then it waits until the conversation has grown back by half.

  • On a model whose context window is smaller than the setting, the threshold drops to 75 % of the window.

  • You always see where you are, live, even during a long turn. In the terminal, at the end of the hint line under the prompt:

    ? for shortcuts · Seuil · 182 k / 300 k · last compaction 341 k → 64 k
    

    In the desktop app and remote sessions, as the plugin's status line. Each compaction also leaves a line in the transcript.

  • In a headless session (claude -p, an SDK host) it goes through /compact, since direct compaction is not available there yet.

| Command | Effect | |---|---| | /seuil | status, last compaction, total taken out since install | | /seuil 250 | sets the threshold to 250 k (minimum 100 k) | | /seuil now | compacts as soon as the turn ends | | /seuil off, /seuil on | switches automatic compaction off or on |

The model also gets a diagnostic tool that reports the real state read from the engine.

Measure it on your own sessions

node bench/simulate.mjs        # the last 30 days
node bench/simulate.mjs 60     # the last 60 days

It reads ~/.claude/projects without changing anything and sends nothing anywhere. Projects are numbered, never named, so the output can be shared. Node 18 or newer, no dependency.

The Campfire theme (Windows Terminal, optional)

A night landscape in pixel art, drawn by a shader behind the text: starry sky, shooting stars, mountains, forest, and a camp with a fire and a cat that sits, blinks and falls asleep. It stays dark and in the back, fades out under text, and never passes under the line you type in. It is the background of the demo above.

The camp, enlarged

node theme/install.mjs               # adds a "Seuil · Campfire" profile to the new-tab menu
node theme/install.mjs --everywhere  # applies it to every Windows Terminal profile
node theme/install.mjs --remove      # removes everything it added

A copy of Windows Terminal's settings is kept before the first change (settings.json.before-seuil-theme). The theme uses Windows Terminal's experimental pixel shader settings; it was checked on Windows Terminal 1.24. Other terminals (VS Code, macOS, Linux) have no pixel shaders: the theme is Windows Terminal only, the mod works everywhere Claude Code runs. Everything is code: paint.mjs paints the camp, campfire.hlsl draws the rest and animates it.

Limits

  • A compaction loses detail. The instructions limit the damage, they do not remove it.
  • It happens only between turns: a long agentic turn can grow well past the threshold before it ends.
  • It takes time: 45 to 58 s in real sessions.
  • The savings come from a simulation on one heavy user's sessions; see the paper for what it does not cover.

Develop

claude --plugin-dir .          # run Claude Code with this folder as a mod
claude plugin test .           # 15 tests
claude plugin validate .

What changed between versions: CHANGELOG.md.

License

MIT

Get a badge

Is this your project? Add this badge to your README or website. It links back to this page.

Featured on ClaudeMods
Size
Theme
Markdown
HTML

Similar projects