Watching it work

Most of a turn is the agent doing things rather than talking. Semantix spends that time telling you what's happening — in the transcript, in the editor, and in the 3D view of your project.

The indicator#

While the agent works, a line above the message box shows a spinning mark and a word — Warming up, Cogitating, Skimming, Etching, Sleuthing, Churning, Composing and a couple of dozen others. Often a detail trails it: the file being read, the pattern being searched for, the command being run.

The word tells you the phase, not the specific action. Each phase — booting, reasoning, reading, writing, searching, listing, running, responding — has its own small pool, and one is drawn at random and re-rolled as it shimmers. Two different files being read look the same apart from the trailing filename. The phase itself is real; it's driven by the tool the agent actually called.

If the model goes quiet composing a long tool call, the detail becomes preparing write file (6s) and the seconds tick up. A frozen number there means the run has died.

On a Claude Code model a second line sits above it: Thinking · 1.34k tokens, counting output tokens for the run.

On every provider except Claude Code, the transcript also gets a collapsible reasoning row — click its header to read the model's thinking. The header word is chosen once per turn and stays put. Claude Code has no such panel; its thinking is the transient line above.

Following the agent#

The eye icon above the message box opens a three-way picker:

• Visual follow (3D camera) — the default. The camera flies to whatever the agent is touching.
• Editor follow (open agent's file) — a tab opens on the file the agent is working on.
• Follow off.

In editor follow, one tab is reused as the follow channel: a second file closes the first and retargets. If you had already opened that file yourself, your tab wins and is simply focused.

In the 3D view, the camera moves to whatever the agent touches, and a small glass window floats beside the node showing the lines going past — the last few being written, or the first few being read. A file that doesn't exist in the scene yet appears as a ghost node; a file already on screen simply lights up and its wires glow.

Up to three of those windows at once. A write window fades a second or two after it lands; a read window lingers around five seconds.

Two more things about the camera. Touching it takes it back — orbit or drag during a turn and the agent stops steering it for that step. And in visual mode the camera returns to an overview of the project at each end of a turn, which will undo a folder you had drilled into.

Notices in the transcript#

Some things are reported as a centered line rather than as chat: Stopped, Response truncated when the model hit its output limit, Response filtered when a safety filter cut in, Agent error with the message, and on Claude Code a family of run notices — API retry, Model fallback, Context compacted, Rate limit — that tell you when the provider did something to your run.

They're informational and they persist with the conversation.

Stopping#

The send button becomes a pause button while the agent works. Press it, or press Escape in the message box. You get a Stopped line, the indicator clears, and anything you had queued is dropped.

Two honest caveats.

Denying a permission also ends the turn, without a Stopped line — the denial row says it. Closing a chat tab stops its run silently, since there's no longer anywhere to show the notice.