Agent reply formatting and showing images

What it is

A reply from an agent is drawn as formatted text, not raw characters. Beyond ordinary markdown, an agent can use a fixed set of GitHub-style shapes that GitHub users already reach for, and it can put a picture in front of you — including one saved anywhere on your computer.

Where to find it

Turning it off

The whole thing has a switch (Settings → Features, "Agent Markdown Formatting"), on by default. With it off, every message is drawn exactly as it was before this existed — the shapes above appear as the plain characters the agent wrote.

How it behaves

The formatting an agent can use

Written as	Drawn as
> [!NOTE] / [!TIP] / [!IMPORTANT] / [!WARNING] / [!CAUTION]	A colour-coded block with its type name and an icon
<div align="center">…</div>	Centred text
<div align="right">…</div>	Right-aligned text
<img src="…" width="400">	The image at that width instead of full size
<details><summary>…</summary>…</details>	A section you can open and close
<kbd>Ctrl+K</kbd>	A keyboard key cap, drawn the same way the rest of the app draws shortcuts
<mark>…</mark>	Highlighted text
<u>…</u>	Underlined text

Table columns also follow the alignment you specify in markdown — | :---: | centres a column and | ---: | right-aligns it.

Only those shapes are honoured. Anything else an agent writes as HTML is shown as plain text, exactly as characters — a deliberate limit, so a reply can never inject markup into the app. Unrecognised callout types behave the same way: > [!SOMETHING] stays an ordinary quote rather than being guessed at.

Showing you a picture

An agent shows you an image by handing the app a file path and the conversation to show it in. The app copies the file in and gives the agent back a markdown line; when the agent writes that line into its reply, the picture appears.

• Any image file on the machine works, from any folder — not just the project it is working in. PNG, JPEG, GIF, WebP, SVG, BMP and TIFF, up to 50 MB.
• The picture is visible only in the conversation it was shown in.

What happens automatically

Separately from that, an agent that merely mentions an image path in its reply has the file picked up automatically — but only when it sits inside the project the agent is working in and was written during that same turn. That narrow rule exists on purpose: without it, narration like `icon.png` would silently attach whatever unrelated file happened to match. The explicit route above is the way to show anything outside that.

When a picture cannot be shown

You are told why, in place of the image — never left with a blank gap. You may see:

• that the message pointed at a file on the computer, which a message cannot load
• that the picture belongs to another conversation
• that the message gave no address for it

A refusal from the explicit route says what was wrong and what to do about it — not an image format, no readable file at that path, a file over the size limit, or a conversation id that does not exist.

Worth knowing

• The app's own bracketed markers never appear on screen. Tokens like [[OMNISCIO_FINAL]] are plumbing that tells the app where a message ends; they are removed before the message is drawn. That holds for one an agent invents, too — a made-up token is cleaned up rather than shown. To genuinely display a token like that, put it inside a code block or backticks, and it is left exactly as written.
• Copying a message gives you the agent's original text, not the formatted result — the drawing happens only on screen, so what you paste elsewhere is what the agent wrote.
• Naming a file is not the same as showing it. A path on its own tells you nothing you can open, which is why agents are taught to use the route above rather than print a path.

Related

How an agent's prose and its tool activity are laid out in the feed is Agent Message Display. The interactive multiple-choice pills an agent can put inside its prose are Question widget, and a file you attach to your own message is Chat attachments.