Skip to main content
Chat displays PNG, JPEG and WebP files from the assistant reply’s run output folder (runs/<run-id>/) directly below the reply. A chart exported by the agent as PNG uses the same viewer. Each image has its filename as alt text and caption, plus a SHA-256 byte-version identifier. The layout fits the conversation on phones. These are current files, not an immutable historical archive. Overwriting an output changes its displayed version on refresh. A write detected between listing and download is refused until a subsequent refresh. The viewer checks file signatures and never renders SVG or HTML as an image, including files renamed with a PNG extension. SVG outputs remain outside inline image discovery. The viewer first verifies the chat execution profile. Pending and restricted chats never request generic agent files. Only files returned by the existing authorized agent-files API are considered; image URLs or filenames in model prose cannot create an inline file reference. Chat uses the same output filters as Artifacts, excludes internal/configuration folders and attachments, and displays up to eight images per reply run. Only replies near the viewport load images. Scrolling away releases their file bytes and stops polling, while retaining the measured space to keep scrolling stable. Browsers without viewport observation offer explicit Load run images and Hide run images controls instead of automatically loading history. File events, completed runs, reconnects and window focus refresh the visible viewer. A one-minute foreground recheck covers missed events and access changes. When a read fails, the displayed images and filenames disappear and Refresh images retries the authorized read. A revoked viewer receives no image bytes from the file endpoint. Existing bytes may remain visible until the next recheck.