Agents

The same comments, from outside Obsidian. A model reviews a note and its remarks appear in the sidebar of the note you already have open.

Or the other direction, which is the one that will get more use: you leave comments asking for things, a model reads them, makes the edits, and resolves them.

Why there was little to build

The storage format is the interface. Comments are plain YAML under one frontmatter key, a suggested edit is a fenced block inside a markdown body, and unknown keys survive a round-trip. Any process that can edit a file can already leave a comment the plugin will render.

So the question was not how to give a model a channel. It was what a model gets wrong when handed the format directly, and the answer is narrow:

What an agent needs	Already true	What was missing
Somewhere to write	frontmatter	nothing
A body format	markdown, suggestion fences	nothing
To be told apart from a person	author	attribution that survives a batch
An anchor	anchor.quote plus a hint	a model cannot count lines
Not to corrupt the note	in-app writes only	a safe path from outside

The anchor contract

An agent never supplies a line or a column. It supplies the quote; the tool does the arithmetic.

Models cannot count lines, and will produce a plausible number under any prompt that asks for one. A wrong number is not a visible error. It writes a structurally valid anchor whose hint points somewhere else, which costs nothing until the quote turns out to be ambiguous and the comment silently attaches to the wrong twin.

The write path mirrors resolution exactly:

• found once → compute the range from the match, write the comment.
• found several → refuse, unless a near-line was given, in which case take the nearest occurrence.
• not found → refuse, and say the quote must be verbatim.

The third case fires constantly, because a model asked to quote a passage will paraphrase it, normalize its whitespace, or straighten its quotation marks. That refusal is the feature. Accepting a near miss writes a comment that is born detached, which reads to the reader as the plugin losing their comment.

Matching is strict. A looser search runs only after a failure: if straightening the punctuation and collapsing the whitespace finds exactly one match, the refusal quotes the note’s own wording back and says to copy it. Nothing is stored that did not come out of the note.

Attribution

An agent’s comment carries an origin with the model that wrote it and a run id shared by every comment in that pass.

The run id is the half that earns its place. A model does not leave a comment; it leaves twenty, in one pass, and the gesture a reader wants afterwards is all of these, gone. A shared id makes a pass an addressable thing: one filter chip, one dismissal, one undo.

A budget

At most 20 comments per run per note, enforced, against 8 stated in the tool description. A review is the few remarks worth reading, not every remark that could be made. A sidebar with forty cards in it does not get opened twice.

Writing while Obsidian is open

An external write re-reads the file and proceeds only if it is byte-for-byte what it was; otherwise it refuses and says to run it again.

The write itself splices a new frontmatter block in front of unchanged body text, so a note Obsidian has open keeps its editor state and every one of its anchors.