Skip to content

inline-visualizations.md

Latest commit

 

History

History
41 lines (34 loc) · 2.09 KB

inline-visualizations.md

File metadata and controls

41 lines (34 loc) · 2.09 KB
Raw

Inline visualizations

Codex's bundled Visualize skill can return an interactive HTML fragment in a Task or Project Conversation. XpressClaw recognizes the exact visualization reference in a final Agent reply and shows the result in place. Normal and wide layouts are supported, and any visual can be expanded for a larger or mobile-friendly view.

The generated file must be an HTML fragment of at most 1 MiB inside that Agent's writable workspace or an explicitly configured writable volume. When the final reply arrives, XpressClaw copies the fragment into its own durable message storage. Reloading the page or replacing the Agent container therefore does not depend on the original file still existing. Missing, invalid, oversized, or out-of-workspace references remain visible as an unavailable card with a useful reason.

Security boundary

Visualization HTML is untrusted. It is never inserted into the XpressClaw DOM. The viewer loads a capability-protected copy in a sandboxed, opaque-origin iframe nested inside a trusted navigation-guard frame. The guard permits only the artifact's initial blob document, so script-driven document navigation cannot become an outbound request even in browsers that do not implement the experimental navigate-to CSP directive. A restrictive content security policy additionally ensures that:

  • scripts may run inside the iframe, but it receives no same-origin access;
  • host cookies, storage, XpressClaw APIs, forms, popups, objects, workers, top-level navigation, and network connections are unavailable;
  • static scripts, styles, fonts, and images are limited to the CDN origins in the Codex Visualize contract; and
  • referrers are disabled.

The only host bridge is window.openai.sendFollowUpMessage({ prompt, title }). XpressClaw validates the request and shows the exact prompt in a confirmation dialog. Nothing is sent to the originating Task or Conversation until the user confirms it.

Visualization references in user or system messages are always inert text. Malformed, escaped, or lookalike references are also rendered through the normal raw-HTML-safe Markdown path.