Every assistant reply leaves a trace on your device: the model calls and tools it used, how long each took and how many tokens it sent and received. Traces help you understand a slow or expensive reply, or share the details when you report a problem.
Settings
Section titled “Settings”Open Settings › Privacy › Turn traces:
| Setting | Default | What it does |
|---|---|---|
| Developer mode | Off | Shows a Trace button under replies. |
| Keep traces | On | Keeps a trace for each reply — sizes and timings only. |
| Include prompt text | Off | Also keeps what the last model call received and the reply, with secrets removed. |
| Clear traces | — | Deletes every trace on this device (asks you to confirm). |
Traces stay on this device for 7 days and are never uploaded.
The Trace drawer
Section titled “The Trace drawer”With Developer mode on, click Trace under a reply. The drawer shows:
- Turn stats — total time, model calls, tool calls (and how many failed), tokens in / out and from cache, and the error if the turn failed (or stopped if you stopped it).
- Timeline — every model call and tool call on a time line, with its duration.
- What the last model call received — the prompt split into its sections (instructions, memory, tools, conversation, attachments…) with their token counts, and the model’s context window.
- Prompt and reply text — only when Include prompt text is on. You can turn it on from the drawer too.
- Export JSON — downloads the trace, for a bug report.
A trace older than 7 days is gone: the drawer says so.
Privacy
Section titled “Privacy”- By default a trace holds sizes and timings, not your text.
- Turning on Include prompt text keeps the text of the last model call on this device only; secrets are removed.
- Anonymous assistant usage events (for example, a reply finished or failed, with counts and the error kind — never your text) are sent only while Share anonymous usage statistics is on, in the same Settings › Privacy page. Turn it off to stop them.