fred-capability-html-artifact

A Fred agent capability (html_artifact) that lets an agent produce a static HTML/CSS artifact rendered live in a sandboxed viewer beside the chat — the "artifact" experience, scoped to static markup (no JavaScript) for v1.

Design: docs/swift/rfc/HTML-ARTIFACT-CAPABILITY-RFC.md (issue #2478).

What it ships

• Tool render_html_artifact(title, html, css, artifact_id?) — HTML and CSS kept separate (for the viewer's tabs). Combined size is capped at 256 KB.
• Chat part HtmlArtifactPart (type="html_artifact") carrying the markup inline — no owned table, no router, no migration (v1 is read-only; chat ui_parts persist across reload).
• Prompt fragment steering the model to call the tool (static only) instead of pasting code into the chat.
• Side panel html_artifact_pane (frontend): read-only tabs Preview (sandboxed <iframe srcdoc>) / HTML / CSS + download.

execution_models=("react",): the prompt overlay is a wrap_model_call hook, so the tool is carried by the capability's middleware (mirrors writable_document).

Registration

Installing this package IS the registration — the fred.capabilities entry point in pyproject.toml points the fred-agents pod at HtmlArtifactCapability. It is wired into the pod as an editable path dependency of apps/fred-agents.

Security

The markup is untrusted LLM output; the backend carries only inert strings. Safe rendering is a frontend concern (RFC §4.7): the Preview iframe uses sandbox without allow-scripts/allow-same-origin and a restrictive CSP, so no script runs and no network egress is possible.

Dev

make code-quality   # ruff + format + type-check
make test           # offline unit tests