Asking an agent to tighten a section, or to answer a reviewer, is easy. Seeing what it did is not.
In a .tex file a paragraph is one long line, so git diff tells you the
whole paragraph changed when it was four words, and a paragraph it moved shows up as one deleted
and another added.
Mâché is suggesting mode, the Google Docs kind, for a LaTeX paper in VS Code. The agent edits the files however it likes. You see its edits as suggestions in place, talk about them in comment threads it can read and answer, and nothing becomes part of the paper until you accept it.
Thompson's lesson applies here as well to the checks themselves. A check performed by a tool you trust is only as trustworthy as that tool, and the tools are many.
moved from §4An endpoint comparison needs something to compare against. We call the accepted output of a revision the golden rendering, a term borrowed from golden-master testing.
Is "golden rendering" a standard term?
Not in typesetting. I've said where it comes from.
removed added commented moved
Install
Install Mâché from the VS Code Marketplace. That is all: open a .tex file of your paper and click Mâché: start reviewing in the status bar, and it offers to set up the rest, the mache command your agent runs and a note telling the agent about it. To install the command on its own, where there is no VS Code:
npm install -g --allow-git=all github:narunraman/mache
Use
Starting makes the paper's current text the accepted text (mache start does the same from a terminal). Comment on anything with Alt+M, then ask your agent, in whatever chat you already use,
to go through your comments. It finds its way by running mache, which lists what's
waiting for it. Accept, reject or reply to what comes back, and when the round is done:
mache commit # one commit of the text you accepted
What it keeps
All of it lives in a .review/ folder in the paper, out of git. Your repository only
ever sees one tidy commit per round, with a message drafted from what happened in it.