Images
Image input supports PNG, JPEG, and WebP: at most four images per prompt, with
a 10 MB source-file limit and decoded Base64 limited to 5 MiB. Use startup
--image (repeatable) or TUI image paste:
one run "describe this diagram" --image diagram.pngone --mode text "describe this diagram" --image diagram.pngImages are transient for the current turn and are not persisted in session JSONL.
Errors and events use only source basenames. Vision-incapable models reject
images before an HTTP request. ChatGPT/Codex vision-capable models send native
Responses API input_image parts; invalid or missing blobs fail explicitly.
TUI, text/json, and one run accept startup --image; RPC accepts startup
images and prompt attachments. read_image is a runtime tool the agent can
call during its turn to inspect a file; it is not a startup image-input channel
for --mode text or one run. Steering and follow-up messages are text-only.