Drafts, attachments and message editing
Supply context deliberately and revise a request without losing provenance.
Prepare a request
State the intended result, relevant dimensions or files, constraints and the checks you expect. Select the correct project before sending. Enter sends while the composer has focus; Shift+Enter adds a line. A failed connection or missing model should leave your draft available.
Attach context
Use the paperclip, select files and inspect their chips before sending. Remove anything unrelated. The app stages its own attachment copies so a later change to the selected source does not silently change the request's evidence.
Default bounds are 10 attachments, 50 MiB per file and 100 MiB total. A rejected attachment must not erase the draft. Images require provider/model support; an attachment is supplied context, not permission to execute a program.
Only send material you are authorized to share with the remote provider. Local project storage does not mean AI inference is local. See privacy and security.
Edit a previous request
- Find the latest editable user message and choose Edit.
- Make the change, or choose Cancel to keep the original.
- If a turn is running, stop it before resending an edited request.
- Inspect the new run separately from the original revision.
Editing does not undo candidate files, roll back an applied model or inherit earlier Apply consent. Existing candidate work remains a separate artifact to inspect.
Copy output
Use the final-output Copy action for the assistant's combined response. Select text directly when you need an excerpt. Copying does not export candidate files or validate their contents.
If something goes wrong
Keep the draft and read the specific failure. For a size limit, remove or reduce only the rejected item. For missing credentials or models, restore the connection first and send deliberately—never assume the app retried automatically. For an interrupted request, follow interrupted-run recovery.