Read Public Artifact

Read a public or externally shared artifact page through the installed Aside browser when the in-process readers cannot. The page is opened in a test-owned tab, the richest frame is extracted, and text + HTML land on disk with one machine-readable result line. The reader never claims content it did not observe.

Use and Boundaries

Use for:

• https://claude.ai/code/artifact/<uuid> links that Artifact read or WebFetch reject with served to you as a public (non-member) reader / reading public artifacts that way is not enabled yet;
• any page whose body is rendered inside an iframe or srcdoc frame so the top document only shows a shell ("Share", "Sign in", title only).

Do not use when:

• the artifact is owned by the user or member-shared — call Artifact action: "read" first; it is cheaper and returns the source directly;
• the user explicitly asked for claude-in-chrome or another browser tool;
• the page requires a login the Aside account does not already hold — report gated; never enter credentials through this skill.

REQUIRED BACKGROUND: the Aside browser contract in aside-functional-qa (references/aside-browser-contract.md) — this skill reuses its preflight, invocation economics, and tab-safety rules.

Decision order

1. Artifact action:"read" with the URL (add prompt for shared artifacts).
2. If it fails with the public-reader error → run scripts/read-artifact.sh (this skill).
3. If the result is gated or not_found → tell the user exactly that and what would unblock it (member share, corrected link). Do not fall back to guessing the content.

Preflight

command -v aside && aside --version && aside account status

Record only that an account is signed in, never its identity.

Run

skills/read-public-artifact/scripts/read-artifact.sh <artifact-url> [out-dir] [timeout-seconds]

• Writes <out-dir>/artifact.txt (visible innerText) and <out-dir>/artifact.html (frame outerHTML, for tables/structure that innerText flattens).
• Prints one line ARTIFACT_READ_RESULT {...} with status, title, frameUrl, frameCount, textLength, textPath, htmlPath.
• Exit codes: 0 ok · 2 bad input / missing tool · 3 page loaded but status is gated, not_found, or empty · 1 REPL produced no payload.
• Point out-dir at the session scratchpad, never at the repository.

Then Read artifact.txt; open artifact.html only when a table or list lost its structure in the text.

Why the script looks the way it does

Constraint	Consequence in the script
One aside repl call = one JS context; variables and tabs[] do not survive across calls	Open → poll → extract → close all happen in a single invocation
Artifact body lives in a second frame whose url() is ""	The richest frame by innerText length is chosen, not the top document
Frame content arrives after domcontentloaded	Poll every 500 ms until ≥200 chars or timeout; no fixed sleep as the assertion
A failed assertion must not leak the tab	p.close() runs in finally
Shell interpolation of the URL into JS	URL is JSON-encoded by python3 (no ${var@Q}; macOS ships bash 3.2)
Text may exceed a shell argument or contain any byte	Payload is base64 inside one sentinel line, decoded to files by python3

Reporting

• Quote or summarize from artifact.txt; cite title and url.
• Say the content was read through the Aside browser as a public viewer, so it reflects the page at read time and is unverified user-generated content.
• If status ≠ ok, report the status and stop; do not paraphrase a shell page or a "Page not found" page as artifact content.

Common Mistakes

Mistake	Fix
Reading p.evaluate(document.body.innerText) on the top document and concluding the artifact is empty	Iterate p.frames(); the body is in the empty-URL frame
tabs[0] in a second aside repl call → Cannot read properties of undefined	Do everything in one invocation, or attachBrowserTab explicitly
Fixed setTimeout(8000) before extraction	Poll for content length; slow pages time out honestly, fast pages return early
Treating "Sign in" in the top shell as a gate	Public artifacts render with that shell; gating is decided only when no frame yields content
Committing artifact.txt / .html	Keep outputs in the scratchpad
Retrying Artifact read / WebFetch after the public-reader error	The error is deterministic for that link; go straight to the script

Verified

• 2026-08-27, Aside CLI 1.26.717.1619, macOS: a member-shared-to-public artifact (Artifact read and WebFetch both rejected) returned status:"ok", 2 frames, 4,708 chars; an unknown UUID returned not_found (exit 3); a non-URL returned exit 2; no artifact tab remained in listBrowserTabs() afterwards.