Reading the agent trace and the ghost preview

When you send a request, two things appear: a ghost preview on the model, and an agent trace in a panel. The ghost preview answers "what will change?". The trace answers "how did it decide?". Together they are how Solid Vibes keeps the promise that nothing happens behind your back.

The ghost preview

The proposed change is drawn translucent, in the accent color, on top of your real model. Your model has not changed yet.

Alongside the ghost you get a short change summary:

• a title ("Add snap-fit lid") and a size ("+3 steps to your recipe"),
• which Parameters changed, shown as before and after (wall_thickness 2 to 3),
• before and after numbers for things like weight and overall size,
• callouts drawn on the model, such as a "0.3 mm gap" label,
• the credits the work cost.

Your three choices

Accepting is not a one-way door. Everything remains undoable, and the Recipe shows exactly which steps came from the agent, so you can suppress or edit them later.

The agent trace

The trace is a live list of what the agent did, in order. Beginners can keep it collapsed, and pros usually leave it open. Each line is a step in the work:

• The route. What the fast router decided you meant (for example "edit geometry"), how confident it was, and whether it ran automatically or stopped to ask. High confidence runs on its own. Low confidence asks you.
• Tool calls. Each action the agent took, such as sketch, extrude, fillet, hole, or measure. Each shows a status (running, done, or error), a timing, and sometimes a confidence value.
• Verification. After building, the agent renders the result and checks it against your request. The note says what it checked and whether it passed.
• Credits. What was charged and your remaining balance.

Reading an error line

A tool call can fail. A fillet radius may be too big for the wall, or a boolean may produce nothing. The line says so in plain language. Often the agent adapts and retries on its own, and you see a successful attempt a few lines later. If it cannot recover, it tells you, and the proposal is a safe fallback or a question rather than a broken model.

When to open the trace

• Something surprised you. Check the route line first. If the agent classified your message as "ask a question" when you wanted an edit, that explains it.
• You want to learn. The trace shows how an expert would approach the problem.
• A change went further than you asked. Look at which tool calls ran.
• You are reporting a problem. The trace is the best evidence.

Teaching mode

Any agent action can be replayed slowly with the "why" attached to each step. Use it to learn techniques ("why did it shell the box before adding the lid lip?"). In the Recipe, steps written by the agent can carry a short note explaining the reasoning.

Try it

Ask for something small and read everything it produces:

Add a 2 mm fillet to all the outer edges on the top face.

Before accepting, check three things:

1. In the trace, was the route "edit geometry" with high confidence?
2. In the change summary, how many steps were added?
3. In the ghost, are the rounded edges the ones you meant?

Then Tweak with "make it 1.5 mm" and watch the ghost update before you commit.