pro-vi/cc-stretch

Claude writes a lot. Now you can read a little.

Replies arrive short. Click a linked phrase and it grows right where it stands. Click « and it shrinks back. Claude is told which parts you opened, so it doesn't assume you read the rest.

In a Claude Code session:

/plugin install cc-stretch --marketplace pro-vi/cc-stretch
/tui fullscreen

In the terminal, clicks need Claude Code's fullscreen renderer. /tui fullscreen switches to it, restarting and resuming your session. If stretches still show open, see Limits. In the Claude desktop app stretches are links: click one to open it in place, click « to close it. No setting needed.

You read about a sixth of what Claude writes. Across 69 real replies from four of my coding projects, the version you see before clicking was 17% of the words Claude wrote without the plugin. It carries about two of every three most important points; the rest is one click away. My replies run long, so if yours are short you will save less. How this was measured, and how to run it on your own transcripts: bench/.

In 1967 Ted Nelson described stretchtext: text that "stretches, becoming longer, with replacement phrases, new details and additional clauses popping into place", and shrinks back on demand, while "the reader remains oriented" (Hypertext Note 8). In 1970 he put it like this: "Gaps appear between phrases; new words and phrases pop into the gaps, an item at a time."

It never caught on. The hard part, as later writers put it, was "how to author the text in the first place" (Wadge and schraefel, 1999): someone had to write every level by hand.

Language models write every level for free, and they are good at it. So all this plugin does is plug a 59-year-old idea into a terminal. The idea is Nelson's. Most of the code is Claude's. I mostly clicked things.

Here is what that looks like. The reply opens with

A monad is a type that wraps values and lets you chain steps that each return a wrapped value.

and one click on that sentence turns it into this (new words in bold, the way Dream Machines appears to set the words Nelson's own example adds):

In programming, a monad is a generic type that wraps values, and it lets you chain steps that each take a plain value and return a wrapped value**, without unwrapping anything by hand between the steps.**

The words you already read never move. New ones open up around them, on both sides.

• skim: a reply with every stretch closed. What you read by default.
• full: the same reply with every stretch open. Everything Claude wrote.
• read: what you actually read, the skim plus the stretches you opened.

The point of the plugin is the gap between them. Nelson's word for this was levels; ours is progressive attention. Progressive disclosure is the interface deciding what to hide from you. Here you spend your attention one phrase at a time, and because nothing you have read moves, you never lose your place.

• Stretches nest. A long form can hold more stretches.
• Streaming stays clean. While a reply streams, stretches already show as their short forms.
• Quieter tool calls. A finished group of tool calls folds to Claude Code's own one-line summary (Ran 3 shell commands ▸ show). Running calls, and calls that errored, stay open.

• Clicks work in the terminal's fullscreen layout and in the Claude desktop app. Everywhere else every stretch shows open, so nothing is hidden: in other apps, and on the terminal's main screen, where a printed reply can't change. There the first such reply says why. As of Claude Code 2.1.300, some setups start on the main screen even with fullscreen set: tmux -CC, SSH from Windows, or a machine where an earlier fullscreen start failed. Starting Claude Code with CLAUDE_CODE_NO_FLICKER=1 forces fullscreen there.
• On Windows, clicks are reported not to work yet (#1).
• Claude writes every level up front, so a reply costs about as many tokens as before. You read less; you don't pay less.
• The {{short|long}} markup shows as raw text wherever the plugin isn't loaded.
• Built on Claude Code's function-hook plugin API, which is early access and changes between releases. Checked in the terminal on macOS with Claude Code 2.1.292 to 2.1.294 and 2.1.300, and in the Claude desktop app with 2.1.295.
• In the Claude desktop app the plugin's tool-call folding is not drawn; the app shows its own.
• A terminal session that also has the desktop app attached shows the markup while a reply streams, because the desktop builds its links from it.

npm test            # claude plugin validate . && claude plugin test .
npm run typecheck   # the benchmark scripts

The film in film/ is made from a real recorded session: film/take.mjs drives Claude Code in a WezTerm pane and records its screen, and film/compose.mjs renders the film. Archive pictures in the film: photographs of the Hypertext Editing System at Brown University, October 1969, by Gregory Lloyd (CC BY-SA 4.0, via Wikimedia Commons), and a photograph of Ted Nelson, 2010, by Beatrice Murch (CC BY 2.0, via Wikimedia Commons). They are not stored in this repo; film/CREDITS.md lists each source, its license and where to fetch it into film/out/assets/.

Not affiliated with Anthropic. Clawd belongs to Anthropic; stretchtext belongs to Ted Nelson.