All documentation
Replay, review & export a run
Understand what happened at each step and export the right evidence.
Every run is stored on the iPhone. Replay is useful for checking what the model saw, where it tried to tap, and whether the visible result matched your goal.
Review a run
Finish or stop a run, then open it in Runs. Step through the screenshots and actions. Where supported by the review screen, add Correct / Wrong verdicts, corrected tap locations, and notes.
Those annotations are stored separately from the original trace. A final done action is the model’s completion claim; your review is separate evidence about whether it actually succeeded.
Export one raw trace
- Open a run in Runs.
- Save any review annotations.
- Choose ⋯ → Export trace.
- Use the share sheet to save the ZIP to Files or transfer it to your computer.
This exports the run folder, including original screenshots and typed text. You do not need to enable Keep run data for sharing for the raw export.
Export a batch
Use Settings → Diagnostics → Export traces to export trace folders, including older runs beyond those shown in the list. Settings → About → Logs → Export all logs is broader and may include other Documents data; use it only when needed.
Finder file sharing is also enabled: connect a paired iPhone to your Mac, open Finder → iPhone → Files → ESP32 Agent, and copy the traces folder.
What is in the archive?
| File | Contents |
|---|---|
steps.jsonl | Task, model-call, and action events, one JSON object per line |
| Root JPG images | Screenshots referenced by the events |
review.json | Saved verdicts, corrections, and notes, when present |
video/*.jpg | Low-frame-rate screen history, when recorded |
metrics.json | A diagnostic snapshot, when written |
ZIP wrapper folders vary. Find the run by locating steps.jsonl. Review edits can insert or remove steps; do not join edited and original records by array position alone.
Raw export or sharing export?
The separate Settings → Privacy → Keep run data for sharing → Prepare Export flow processes the content to reduce exposure of selected text and image areas. It is not a guarantee that all personal information is removed.
Choose a raw trace when analysis requires original images and text. Choose a processed sharing export when that better suits your purpose, then inspect the result before sending it anywhere.
Keep expectations precise
Saved exports help investigate behavior but do not reproduce every original request. Some request configuration and failed attempts are not fully captured. Diagnostic metrics can include a shared recent-step buffer, so they are not necessarily an isolated summary of one run.
Read the upstream export reference for event fields and analysis limits, and privacy and data before sharing archives.