5. Using the Chat
The Chat panel is where you talk to the AI and put it to work on your project.
Agent mode vs. Chat mode
| Mode | What it does | When to use |
|---|---|---|
| Agent (default) | The AI can act on the project: read, edit code, build, run, drive the Form Designer, git… via MCP tools | When you want the AI to do something in the project |
| Chat | Conversation only — no acting on the project | Questions, brainstorming, explanations |
Agent mode is the default. The two modes are kept distinct — choose according to your intent.
Sending a message
Type your request and press Enter. The CLI output is streamed live and rendered as Markdown (with syntax highlighting) when WebView2 is installed; without it, it appears as plain text.
Skills with /agent
You can invoke skills (predefined flows) by typing /agent followed by the skill.
Skills live in a canonical project folder (.aefos/skills/) and are automatically
replicated to the format the active CLI expects.
- Free-form text (no skill) goes straight to the CLI as a normal prompt.
- A skill builds a Delphi-aware prompt and dispatches it to the CLI.
Project context
In Agent mode, Aefos packages your Delphi project context (via OTA) so the AI answers with knowledge of what is open — instead of generic answers.
Reviewing changes: see the before and the after
Whenever the agent changes code, Aefos shows the change inside the editor as a stacked before/after diff: the original text appears struck out in red on top, and the new text in green right below — so you see exactly what will change before deciding (it also works for wide and multi-line edits).
In the gutter (on the left) each change has three buttons:
- ✓ accept — keep the new text.
- ✗ reject — undo it and restore the original.
- ✎ annotate — write a note that is delivered to the agent (tagged accepted or rejected), so it learns your feedback on that edit.
There's also an Accept all / Reject all pill when several changes are pending — they accumulate without blocking the agent. Saving the file (or running) accepts the pending changes and clears the review.
The same review applies to edits made from the Chat and the Terminal — it lives in a shared IDE layer, so the accept/reject/annotate flow is identical in both.
The review is shown for code edits (e.g. editing a unit, replacing a snippet). See What the agent does in your project.
Safety: consent and audit
Actions that modify the project go through a consent prompt before they run, and every tool call is recorded in an audit log. You have traceability of what the agent did.
➡️ Next: What the agent does in your project