The AI chat is where you spend most of your time in AICODESIT, so we spent a week making it easier to read. Two things changed: the AI's written answers now look like GitHub, and every tool the AI runs shows up as a small, tidy card instead of a wall of text.
Answers that look like GitHub
Instead of a custom set of heading and list styles, the chat now renders Markdown the way GitHub does. If you have ever read a README or a pull request, nothing here will surprise you:
- Tables with aligned columns and a clear header row.
- Task lists with real checkboxes:
- [x] doneand- [ ] to do. - Code blocks with syntax highlighting for the language the AI names after the opening backticks.
- Alerts for notes, tips, important points, warnings and cautions, written as
> [!WARNING]. - Footnotes, strikethrough, links that turn into clickable addresses, and emoji shortcodes such as
:rocket:. - A safe subset of HTML for the cases where Markdown is not enough.
A reply like this now draws exactly as you would expect:
> [!WARNING]
> This will overwrite index.html.
| Step | Status |
|--------|--------|
| Read | done |
| Edit | next |
- [x] Read the file
- [ ] Update the headerIt follows your light or dark theme automatically, and long code blocks scroll inside their own box instead of stretching the chat.
Every tool gets its own card
When the AI works on your project it calls tools: it reads files, searches, runs commands, queries your database. Some of these already had a nice display, like file edits. Others, such as SSH, Memory or Search, showed up as plain text. Now every tool has the same compact card:
- Bash for commands the AI runs,
- SSH for commands on a remote machine,
- Read and Search for looking at your files,
- Database and Storage for your project's data and files,
- Memory for what the AI remembers about your project.
Each card has a one-line header: what ran, on what, and a status dot (green when it worked, red when it failed). Click the header to open the output. Failed cards say so in red, so you can see at a glance where a run went wrong. For the full story on why we show tool calls at all, read Pipeline Tool Call Transparency.
Nothing overflows, and the cards are much shorter
Two complaints came up again and again: a long tool output could push the end of the AI's answer out of view, and the cards took up a lot of vertical space. Both are fixed:
- Long paths, long lines and long outputs now wrap or scroll inside their card. The chat itself never gets wider than your screen, so you can always see the full reply.
- Cards use roughly 70% less height than before. An expanded output shows a few lines and scrolls for the rest, so a run with a dozen tool calls no longer fills the screen.
When the AI forgets a bracket
Models occasionally slip when they write a tool call. They leave off the opening < and write cmd>ls</cmd>, or they write an empty open-and-close pair for something that takes no content, like <read url="..."></read>. Before, those calls were silently ignored and the AI carried on as if the tool had run.
The platform now repairs these slips before the call is read, both when it runs the tool and when the chat draws it. The tool runs, the card appears, and the AI gets its result.
One thing we tried and took out
We also added LaTeX math rendering. It looked great, but it interfered with typing in the chat box, so we removed it. We would rather have a chat that is always comfortable to type in than formulas that look nicer; we will revisit math once we can do it without side effects.
Want to see it? Open a demo project, ask the AI to build something small, and watch the cards appear.